Life System Modeling and Simulation, LSMS 2007

Lecture Notes in Bioinformatics

4689

Edited by S. Istrail, P. Pevzner, and M. Waterman Editorial Board: A. Apostolico S. Brunak M. Gelfand T. Lengauer S. Miyano G. Myers M.-F. Sagot D. Sankoff R. Shamir T. Speed M. Vingron W. Wong

Subseries of Lecture Notes in Computer Science

Kang Li Xin Li George William Irwin Guosen He (Eds.)

Life System Modeling and Simulation International Conference, LSMS 2007 Shanghai, China, September 14-17, 2007 Proceedings

13

Series Editors Sorin Istrail, Brown University, Providence, RI, USA Pavel Pevzner, University of California, San Diego, CA, USA Michael Waterman, University of Southern California, Los Angeles, CA, USA Volume Editors Kang Li George William Irwin Queen’s University Belfast School of Electronics, Electrical Engineering and Computer Science Ashby Building, Stranmillis Road, BT9 5AH Belfast, UK E-mail: {K.Li, g.irwin}@ee.qub.ac.uk Xin Li Guosen He Shanghai University, School of Mechatronics and Automation, China E-mail: {su_xinli, guosenhe}@yahoo.com.cn

Library of Congress Control Number: 2007933845

CR Subject Classification (1998): F.2.2, F.2, E.1, G.1, I.2, J.3 LNCS Sublibrary: SL 8 – Bioinformatics ISSN ISBN-10 ISBN-13

1865-0929 3-540-74770-2 Springer Berlin Heidelberg New York 978-3-540-74770-3 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. Springer is a part of Springer Science+Business Media springer.com © Springer-Verlag Berlin Heidelberg 2007 Printed in Germany Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper SPIN: 12118502 06/3180 543210

Preface

The International Conference on Life System Modeling and Simulation (LSMS) was formed to bring together international researchers and practitioners in the field of life system modeling and simulation as well as life system-inspired theory and methodology. The concept of a life system is quite broad. It covers both micro and macro components ranging from cells, tissues and organs across to organisms and ecologic niches. These interact and evolve to produce an overall complex system whose behavior is difficult to comprehend and predict. The arrival of the 21st century has been marked by a resurgence of research interest both in arriving at a systems-level understanding of biology and in applying such knowledge in complex real-world applications. Consequently, computational methods and intelligence in systems, biology, as well as bio-inspired computational intelligence, have emerged as key drivers for new computational methods. For this reason papers dealing with theory, techniques and real-world applications relating to these two themes were especially solicited. Building on the success of a previous workshop in 2004, the 2007 International Conference on Life System Modeling and Simulation (LSMS 2007) was held in Shanghai, China, September 14–17, 2007. The conference was jointly organized by The Shanghai University, Queen's University Belfast together with The Life System Modeling and Simulation Special Interest Committee of the Chinese Association for System Simulation. The conference program offered the delegates keynote addresses, panel discussions, special sessions and poster presentations, in addition to a series of social functions to enable networking and future research collaboration. LSMS 2007 received a total of 1,383 full paper submissions from 21 countries. All these papers went through a rigorous peer-review procedure, including both prereview and formal referring. Based on the referee reports, the Program Committee finally selected 333 good-quality papers for presentation at the conference, from which 147 were subsequently selected and recommended for publication by Springer in one volume of Lecture Notes in Computer Science (LNCS) and one volume of Lecture Notes in Bioinformatics (LNBI). This particular volume of Lecture Notes in Computer Science (LNCS) includes 84 papers covering 6 relevant topics. The organizers of LSMS 2007 would like to acknowledge the enormous contributions made by the following: the Advisory Committee and Steering Committee for their guidance and advice, the Program Committee and the numerous referees worldwide for their efforts in reviewing and soliciting the papers, and the Publication Committee for their editorial work. We would also like to thank Alfred Hofmann, from Springer, for his support and guidance. Particular thanks are of course due to all the authors, as without their high-quality submissions and presentations, the LSMS 2007 conference would not have been possible. Finally, we would like to express our gratitude to our sponsor – The Chinese Association for System Simulation, – and a number of technical co-sponsors: the IEEE United Kingdom and Republic of Ireland Section, the IEEE CASS Life Science Systems and Applications Technical Committee, the IEEE CIS Singapore Chapter, the

VI

Preface

IEEE Shanghai Section for their technical co-sponsorship and the Systems and Synthetic Biology (Springer) for their financial sponsorship. The support of the Intelligent Systems and Control group at Queen’s University Belfast, Fudan University, the Shanghai Institute for Biological Sciences, the Chinese Academy of Sciences, the Shanghai Association for System Simulation, the Shanghai Association for Automation, Shanghai Association for Instrument and Control, the Shanghai Rising-star Association, the Shanghai International Culture Association, Shanghai Medical Instruments Trade Association is also acknowledged.

June 2007

Bohu Li Guosen He Mitsuo Umezu Min Wang Minrui Fei George W. Irwin Kang Li Luonan Chen Shiwei Ma

LSMS 2007 Organization

Advisory Committee Panos J. Antsaklis, USA Aike Guo, China Huosheng Hu, UK Iven Mareels, Australia Shuzhi Sam Ge, Singapore Yishan Wang, China Zhenghai Xu, China Xiangsun Zhang, China Mengchu Zhou, USA

John L. Casti, Austria Roland Hetzer, Germany Okyay Kaynak, Turkey Kwang-Hyun Park, Korea Eduardo Sontag, USA Paul Werbos, USA Hao Ying, USA Guoping Zhao, China

Joseph Sylvester Chang, Singapore Tom Heskes, Netherlands

Kwang-Hyun Cho, Korea

Seung Kee Han, Korea

Yan Hong, HK China

Fengju Kang, China Yixue Li, China Sean McLoone, Ireland Dhar Pawan, Singapore

Young J Kim, Korea Zaozhen Liu, China David McMillen, Canada Chen Kay Tan, Singapore

Stephen Thompson, UK Tianyuan Xiao, China Tianshou Zhou, China

Svetha Venkatesh, Australia Jianxin Xu, Singapore Quanmin Zhu, UK

Jenn-Kang Hwang, Taiwan China Gang Li, UK Zengrong Liu, China Yi Pan, USA Kok Kiong Tan, Singapore YuguoWeng, Germany Wu Zhang, China

Kazuyuki Aihara, Japan Zongji Chen, China Alfred Hofmann, Germany Frank L. Lewis, USA Xiaoyuan Peng, China Steve Thompson, UK Stephen Wong, USA Minlian Zhang, China Yufan Zheng, Australia

Steering Committee

Honorary Chairs Bohu Li, China Guosen He, China Mitsuo Umezu, Japan

General Chairs Min Wang, China Minrui Fei, China George W. Irwin, UK

VIII

Organization

International Program Committee IPC Chairs Kang Li, UK Luonan Chen, Japan IPC Local Chairs Luis Antonio Aguirre, Brazil Xingsheng Gu, China WanQuan Liu, Australia T.C. Yang, UK

Yongsheng Ding, China

Orazio Giustolisi, Italy

Pheng-Ann Heng, HK, China Zhijian Song, China Jun Zhang, USA

Nicolas Langlois, France Shu Wang, Singapore

IPC Members Akira Amano, Japan Ming Chen, China Xiaochun Cheng, UK Patrick Connally, UK Huijun Gao, China Ning Gu, China Liqun Han, China Guangbin Huang, Singapore Ping Jiang, UK

Vitoantonio Bevilacqua, Italy Zengqiang Chen, China Minsen Chiu, Singapore Rogers Eric, UK Xiaozhi Gao, Finland Weihua Gui, China Jiehuan He, China Sunan Huang, Singapore

Weidong Cai, Australia Wushan Cheng, China Sally Clift, UK Haiping Fang, China Zhinian Gao, China Lingzhong Guo, UK Liangjian Hu, China Peter Hung, Ireland

Prashant Joshi, Austria

Abderrafiaa Koukam, France Keun-Woo Lee, Korea Jun Li, Singapore Xiaoou Li, Mexico Guoqiang Liu, China Junfeng Liu, USA Zuhong Lu, China Kezhi Mao, Singapore Carlo Meloni, Italy Manamanni Noureddine, France Girijesh Prasad, UK

Xuecheng Lai, Singapore

Tetsuya J Kobayashi, Japan Ziqiang Lang, UK

Raymond Lee, UK Shaoyuan Li, China Yunfeng Li, China Han Liu, China Mandan Liu, China Guido Maione, Italy Marco Mastrovito, Italy Zbigniew Mrozek, Poland Philip Ogunbona, Australia

Donghai Li, China Wanqing Li, Australia Paolo Lino, Italy Julian Liu, UK Wei Lou, China Fenglou Mao, USA Marion McAfee, UK Antonio Neme, Mexico Jianxun Peng, UK

Yixian Qin, USA

Wei Ren, China

Organization

Qiguo Rong, China Ziqiang Sun, China Nigel G Ternan, UK Bing Wang, UK Ruiqi Wang, Japan Xiuying Wang, Australia Guihua Wen, China Lingyun Wu, China Qingguo Xie, China Jun Yang, Singapore Ansheng Yu, China Jingqi Yuan, China Jun Zhang, USA Cishen Zhang, Singapore Yisheng Zhu, China

Da Ruan, Belgium Sanjay Swarup, Singapore Shanbao Tong, China Jihong Wang, UK Ruisheng Wang, Japan Yong Wang, Japan

Chenxi Shao, China Shin-ya Takane, Japan Gabriel Vasilescu, France Ning Wang, China Xingcheng Wang, China Zhuping Wang, Singapore

Peter A. Wieringa, Netherlands Xiaofeng Wu, China Meihua Xu, China Tao Yang, USA Weichuan Yu, HK China Dong Yue, China Yi Zhang, China Xingming Zhao, Japan

Guangqiang Wu, China Hong Xia, UK Zhenyuan Xu, China Maurice Yolles, UK Wen Yu, Mexico Zhoumo Zeng, China Zuren Zhang, China Huiyu Zhou, UK

Secretary General Shiwei Ma, China Ping Zhang, China

Co-Secretary-General Li Jia, China Qun Niu, China Banghua Yang, China

Lixiong Li, China Yang Song, China

Publication Chairs Xin Li, China Sanjay Swarup, Singapore

Special Session Chair Hai Lin, Singapore

IX

Xin Li, China Ling Wang, China

X

Organization

Organizing Committee OC Chairs Jian Wang, China Yunjie Wu, China Zengrong Liu, China Yuemei Tan, China OC Co-chairs Tingzhang Liu, China Shiwei Ma, China Weiyi Wang, China Xiaojin Zhu, China OC Members Jian Fan, China Zhihua Li, China Zhongjie Wang, China

Weiyan Hou, China Hai Lin, Singapore Lisheng Wei, China

Aimei Huang, China Xin Sun, China Xiaolei Xia, UK

Reviewers Jean-Francois Arnold Xiaojuan Ban Leonora Bianchi Mauro Birattari Ruifeng Bo Jiajun Bu Dongsheng Che Fei Chen Feng Chen Guochu chen Hang Chen Mingdeng Chen Lijuan Chen Zengqiang Chen Cheng Cheng Guojian Cheng Jin Cheng Maurizio Cirrincione Patrick Connally Marco Cortellino

Jean-Charles Creput Shigang Cui Dan Diaper Chaoyang Dong Guangbo Dong Shuhai Fan Lingshen Fang Dongqing Feng Hailin Feng Zhanshen Feng Cheng Heng Fua Jie Gao Padhraig Gormley Jinhong Gu Lan Guo Qinglin Guo Yecai Guo Yu Guo Dong-Han Ham Zhang Hong

Aimin Hou Yuexian Hou Jiangting Hu Qingxi Hu Wenbin Hu Xianfeng Huang Christian Huyck George W. Irwin Yubin Ji Li Jian Shaohua Jiang Guangxu Jin Hailong Jin Xinsheng Ke Mohammad Khalil Yohei Koyama Salah Laghrouche Usik Lee Chi-Sing Leung Gun Li

Organization

Honglei Li Kan Li Kang Li Ning Li Xie Li Yanbo Li Yanyan Li Zhonghua Li Xiao Liang Xiaomei Lin Binghan Liu Chunan Liu Hongwei Liu Junfang Liu Lifang Liu Renren Liu Wanquan Liu Weidong Liu Xiaobing Liu Xiaojie Liu Xuxun Liu Yumin Liu Zhen Liu Zhiping Liu Xuyang Lou Tao Lu Dajie Luo Fei Luo Suhuai Luo Baoshan Ma Meng Ma Xiaoqi Ma Quentin Mair Xiong Men Zhongchun Mi Claude Moog Jin Nan Jose Negrete Xiangfei Nie Xuemei Ning Dongxiao Niu Jingchang Pan Paolo Pannarale Konstantinos Pataridis Jianxun Peng Son Lam Phung Xiaogang Qi

Chaoyong Qin peng qin Zhaohui Qin Lipeng Qiu Yuqing Qiu Yi Qu Qingan Ren Didier Ridienger Giuseppe Romanazzi R Sanchez Jesus Savage Ssang-Hee Seo Tao Shang Zichang Shangguan Chenxi Shao JeongYon Shim Chiyu Shu Yunxing Shu Vincent Sircoulomb Anping Song Chunxia Song Guanhua Song Yuantao Song Yan Su Yuheng Su Suixiulin Shibao Sun Wei Sun Da Tang Pey Yuen Tao Shen Tao Keng Peng Tee Jingwen Tian Han Thanh Trung Callaghan Vic Ping Wan Hongjie Wang Kundong Wang Lei Wang Lin Wang Qing Wang Qingjiang Wang Ruisheng Wang Shuda Wang Tong Wang Xiaolei Wang Xuesong Wang

Ying Wang Zhelong Wang Zhongjie Wang Hualiang Wei Liang Wei Guihua Wen Qianyong Weng Xiangtao Wo Minghui Wu Shihong Wu Ting Wu Xiaoqin Wu Xintao Wu Yunna Wu Zikai Wu Chengyi Xia Linying Xiang Xiaolei Xia Yougang Xiao Jiang Xie Jun Xie Xiaohui Xie Lining Xing Guangning Xu Jing Xu Xiangmin Xu Xuesong Xu Yufa Xu Zhiwen Xu Qinghai Yang Jin Yang Xin Yang Yinhua Yang Zhengquan Yang Xiaoling Ye Changming Yin Fengqin Yu Xiaoyi Yu Xuelian Yu Guili Yuan Lulai Yuan Zhuzhi Yuan Peng Zan Yanjun Zeng Chengy Zhang Kai Zhang Kui Zhang

XI

XII

Organization

Hongjuan Zhang Hua Zhang Jianxiong Zhang Limin Zhang Lin Zhang Ran Zhang Xiaoguang Zhang Xing Zhang

Haibin Zhao Shuguang Zhao Yi Zhao Yifan Zhao Yong Zhao Xiao Zheng Yu Zheng Hongfang Zhou

Huiyu Zhou Qihai Zhou Yuren Zhou Qingsheng Zhu Xinglong Zhu Zhengye Zhu Xiaojie Zong

Table of Contents

The First Section: Modeling and Simulation of Societies and Collective Behavior Phase Synchronization of Circadian Oscillators Induced by a Light-Dark Cycle and Extracellular Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ying Li, Jianbao Zhang, and Zengrong Liu

1

Detecting RNA Sequences Using Two-Stage SVM Classifier . . . . . . . . . . . Xiaoou Li and Kang Li

8

Frequency Synchronization of a Set of Cells Coupled by Quorum Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jianbao Zhang, Zengrong Liu, Ying Li, and Luonan Chen

21

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Min Xu, Yongsheng Ding, and Liangjian Hu

28

Simulation of Artificial Life of Bee’s Behaviors . . . . . . . . . . . . . . . . . . . . . . . Bin Wu, Hongying Zhang, and Xia Ni

38

Hybrid Processing and Time-Frequency Analysis of ECG Signal . . . . . . . Ping Zhang, Chengyuan Tu, Xiaoyang Li, and Yanjun Zeng

46

Robust Stability of Human Balance Keeping . . . . . . . . . . . . . . . . . . . . . . . . Minrui Fei, Lisheng Wei, and Taicheng Yang

58

Modelling Pervasive Environments Using Bespoke and Commercial Game-Based Simulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Marc Davies, Vic Callaghan, and Liping Shen

67

The Research of Artificial Animal’s Behavior Memory Based on Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiaojuan Ban, Shurong Ning, Jing Shi, and Dongmei Ai

78

The Second Section: Computational Methods and Intelligence in Biomechanical Systems, Tissue Engineering and Clinical Bioengineering How to Ensure Safety Factors in the Development of Artificial Heart: Verified by the Usage of “Modeling and Simulation” Technology . . . . . . . Mitsuo Umezu

88

XIV

Table of Contents

Parametric-Expression-Based Construction of Interior Features for Tissue Engineering Scaffold with Defect Bone . . . . . . . . . . . . . . . . . . . . . . . Chunxiang Dai, Qingxi Hu, and Minglun Fang Computation of Uniaxial Modulus of the Normal and Degenerated Articular Cartilage Using Inhomogeneous Triphasic Model . . . . . . . . . . . . Haijun Niu, Qing Wang, Yongping Zheng, Fang Pu, Yubo Fan, and Deyu Li Effect of the Plantar Ligaments Injury on the Longitudinal Arch Height of the Human Foot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yunfeng Yang, Guangrong Yu, Wenxin Niu, Jiaqian Zhou, Yanxi Chen, Feng Yuan, and Zuquan Ding Internet Living Broadcast of Medical Video Stream . . . . . . . . . . . . . . . . . . Shejiao Li, Bo Li, and Fan Zhang Predicting Syndrome by NEI Specifications: A Comparison of Five Data Mining Algorithms in Coronary Heart Disease . . . . . . . . . . . . . . . . . . Jianxin Chen, Guangcheng Xi, Yanwei Xing, Jing Chen, and Jie Wang Application of Image Processing and Finite Element Analysis in Bionic Scaffolds’ Design Optimizing and Fabrication . . . . . . . . . . . . . . . . . . . . . . . . Liulan Lin, Huicun Zhang, Yuan Yao, Aili Tong, Qingxi Hu, and Minglun Fang The Mechanical Properties of Bone Tissue Engineering Scaffold Fabricating Via Selective Laser Sintering . . . . . . . . . . . . . . . . . . . . . . . . . . . . Liulan Lin, Aili Tong, Huicun Zhang, Qingxi Hu, and Minglun Fang

97

104

111

120

129

136

146

The Third Section: Computational Intelligence in Bioinformatics and Biometrics Informational Structure of Agrobacterium Tumefaciens C58 Genome . . . Zhihua Liu and Xiao Sun

153

Feature Extraction for Cancer Classification Using Kernel-Based Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shutao Li and Chen Liao

162

A New Hybrid Approach to Predict Subcellular Localization by Incorporating Protein Evolutionary Conservation Information . . . . . . . . . ShaoWu Zhang, YunLong Zhang, JunHui Li, HuiFeng Yang, YongMei Cheng, and GuoPing Zhou

172

Table of Contents

Support Vector Machine for Prediction of DNA-Binding Domains in Protein-DNA Complexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jiansheng Wu, Hongtao Wu, Hongde Liu, Haoyan Zhou, and Xiao Sun

XV

180

Feature Extraction for Mass Spectrometry Data . . . . . . . . . . . . . . . . . . . . . Yihui Liu

188

An Improved Algorithm on Detecting Transcription and Translation Motif in Archaeal Genomic Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Minghui Wu, Xian Chen, Fanwei Zhu, and Jing Ying

197

Constructing Structural Alignment of RNA Sequences by Detecting and Assessing Conserved Stems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiaoyong Fang, Zhigang Luo, Bo Yuan, Zhenghua Wang, and Fan Ding

208

Iris Verification Using Wavelet Moments and Neural Network . . . . . . . . . . Zhiqiang Ma, Miao Qi, Haifeng Kang, Shuhua Wang, and Jun Kong

218

Comprehensive Fuzzy Evaluation Model for Body Physical Exercise Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yizhi Wu, Yongsheng Ding, and Hongan Xu

227

The Fourth Section: Brain Stimulation, Neural Dynamics and Neural Interfacing The Effect of Map Information on Brain Activation During a Driving Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tao Shang, Shuoyu Wang, and Shengnan Zhang

236

Worm 5: Pseudo-organics Computer and Natural Live System . . . . . . . . . Yick Kuen Lee and Ying Ying Lee

246

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics of Coupled Neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jiang Wang, Xiumin Li, and Dong Feng

254

Distinguish Different Acupuncture Manipulations by Using Idea of ISI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jiang Wang, Wenjie Si, Limei Zhong, and Feng Dong

264

The Study on Internet-Based Face Recognition System Using PCA and MMD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jong-Min Kim

274

Simulation of Virtual Human’s Mental State in Behavior Animation . . . . Zhen Liu

284

XVI

Table of Contents

Hemodynamic Analysis of Cerebral Aneurysm and Stenosed Carotid Bifurcation Using Computational Fluid Dynamics Technique . . . . . . . . . . Yi Qian, Tetsuji Harada, Koichi Fukui, Mitsuo Umezu, Hiroyuki Takao, and Yuichi Murayama

292

Active/Inactive Emotional Switching for Thinking Chain Extraction by Type Matching from RAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . JeongYon Shim

300

Pattern Recognition for Brain-Computer Interfaces by Combining Support Vector Machine with Adaptive Genetic Algorithm . . . . . . . . . . . . Banghua Yang, Shiwei Ma, and Zhihua Li

307

The Fifth Section: Biological and Biomedical Data Integration, Mining and Visualization Improved Locally Linear Embedding by Cognitive Geometry . . . . . . . . . . Guihua Wen, Lijun Jiang, and Jun Wen

317

Predicting the Free Calcium Oxide Content on the Basis of Rough Sets, Neural Networks and Data Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yunxing Shu, Shiwei Yun, and Bo Ge

326

Classification of Single Trial EEG Based on Cloud Model for Brain-Computer Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shaobin Li and Chenxi Shao

335

The Modified Self-organizing Fuzzy Neural Network Model for Adaptability Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zuohua Miao, Hong Xu, and Xianhua Wang

344

Predicting Functional Protein-Protein Interactions Based on Computational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Luwen Zhang and Wu Zhang

354

The Chaos Model Analysis Based on Time-Varying Fractal Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jianrong Hou, Dan Huang, and Hui Zhao

364

Bi-hierarchy Medical Image Registration Based on Steerable Pyramid Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiuying Wang and David Feng

370

Table of Contents

XVII

The Sixth Section: Computational Methods and Intelligence in Organism Modeling and Biochemical Networks and Regulation A Multiagent Quantum Evolutionary Algorithm for Global Numerical Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chaoyong Qin, Jianguo Zheng, and Jiyu Lai

380

Developing and Optimizing a Finite Element Model of Phalange Using CT Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qingxi Hu, Quan Zhang, and Yuan Yao

390

Reverse Engineering Methodology in Broken Skull Surface Model Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Luyue Ju, Gaojian Zhong, and Xia Liu

399

Identification and Application of Nonlinear Rheological Characteristics of Oilseed Based on Artificial Neural Networks . . . . . . . . . . . . . . . . . . . . . . . Xiao Zheng, Guoxiang Lin, Dongping He, Jingzhou Wang, and Yan You

406

Prediction of Death Rate of Breast Cancer Induced from Average Microelement Absorption with Neural Network . . . . . . . . . . . . . . . . . . . . . . Shouju Li, Jizhe Wang, Yingxi Liu, and Xiuzhen Sun

414

An Adaptive Classifier Based on Artificial Immune Network . . . . . . . . . . . Zhiguo Li, Jiang Zhong, Yong Feng, and ZhongFu Wu

422

Investigation of a Hydrodynamic Performance of a Ventricular Assist Device After Its Long-Term Use in Clinical Application . . . . . . . . . . . . . . . Yuma Kokuzawa, Tomohiro Shima, Masateru Furusato, Kazuhiko Ito, Takashi Tanaka, Toshihiro Igarashi, Tomohiro Nishinaka, Kiyotaka Iwasaki, and Mitsuo Umezu

429

The Seventh Section: Computational Methods and Intelligence in Modeling of Molecular, Cellular, Multi-cellular Behavior and Design of Synthetic Biological Systems QSAR and Molecular Docking Study of a Series of Combretastatin Analogues Tubulin Inhibitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yubin Ji, Ran Tian, and Wenhan Lin

436

A Software Method to Model and Fabricate the Defective Bone Repair Bioscaffold Using in Tissue Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qingxi Hu, Hongfei Yang, and Yuan Yao

445

XVIII

Table of Contents

Using Qualitative Technology for Modeling the Process of Virus Infection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hailin Feng and Chenxi Shao

453

AOC-by-Self-discovery Modeling and Simulation for HIV . . . . . . . . . . . . . . Chunxiao Zhao, Ning Zhong, and Ying Hao

462

A Simulation Study on the Encoding Mechanism of Retinal Ganglion Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chao-Feng Cai, Pei-Ji Liang, and Pu-Ming Zhang

470

Modelling the MAPK Signalling Pathway Using a Two-Stage Identification Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Padhraig Gormley, Kang Li, and George W. Irwin

480

The Eighth Section: Others Design and Path Planning for a Remote-Brained Service Robot . . . . . . . . Shigang Cui, Xuelian Xu, Zhengguang Lian, Li Zhao, and Zhigang Bing

492

Adaptive Fuzzy Sliding Mode Control of the Model of Aneurysms of the Circle of Willis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peijun Ju, Guocai Liu, Li Tian, and Wei Zhang

501

Particle Swarm Optimization Applied to Image Vector Quantization . . . . Xubing Zhang, Zequn Guan, and Tianhong Gan

507

Face Detection Based on BPNN and Wavelet Invariant Moment in Video Surveillance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hongji Lin and Zhengchun Ye

516

Efficient Topological Reconstruction for Medical Model Based on Mesh Simplification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chunxiang Dai, Ying Jiang, Qingxi Hu, Yuan Yao, and Hongfei Yang

526

Repetitive Motion Planning of Redundant Robots Based on LVI-Based Primal-Dual Neural Network and PUMA560 Example . . . . . . . . . . . . . . . . Yunong Zhang, Xuanjiao Lv, Zhonghua Li, and Zhi Yang

536

Tensile Test to Ensure a Safety of Cannula Connection in Clinical Ventricular Assist Device (VAD) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Takashi Tanaka, Tomohiro Shima, Masateru Furusato, Yuma Kokuzawa, Kazuhiko Ito, Kiyotaka Iwasaki, Yi Qian, and Mitsuo Umezu

546

Table of Contents

A Reproduction of Inflow Restriction in the Mock Circulatory System to Evaluate a Hydrodynamic Performance of a Ventricular Assist Device in Practical Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Masateru Furusato, Tomohiro Shima, Yuma Kokuzawa, Kazuhiko Ito, Takashi Tanaka, Kiyotaka Iwasaki, Yi Qian, Mitsuo Umezu, ZhiKun Yan, and Ling Zhu Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

XIX

553

559

Phase Synchronization of Circadian Oscillators Induced by a Light-Dark Cycle and Extracellular Noise Ying Li1 , Jianbao Zhang1 , and Zengrong Liu2 1

2

Department of Mathematics, Shanghai University, 200444 Shanghai, China [email protected] Institute of Systems Biology, Shanghai University, 200444 Shanghai, China [email protected]

Abstract. In mammals, the master circadian pacemaker is considered the suprachiasmatic nucleus (SCN) of the hypothalamus. Individual cellular clocks in SCN, the circadian center, are integrated into a stable and robust pacemaker with a period length of about 24 hours, which are remarkably accurate at timing biological events despite the randomness of their biochemical reactions. In this paper, we study the effect of the Light-Dark cycle and environment noise on the daily rhythms of mammals and give some numerical analysis. The results show that the environment noise makes for phase synchronization but it can not make the oscillators get phase synchronization with period of 24-h. On the contrary, the threshold of the strength of light that makes the oscillators to get the phase synchronization with period of 24-h with environment noise is larger than that in the case without environment noise.

1

Introduction

Circadian rhythms are observed in the physiology of mammals and other higher organisms. In mammals, physiological and behavioral circadian rhythms are controlled by a pacemaker located in the suprachiasmatic nucleus(SCN) of the hypothalamus[1,2]. SCN consists of 16000 neurons arranged in a symmetric bilateral structure, and it is generally believed that each isolated SCN neuron behaves as an oscillator by itself. It has been shown that isolated individual neurons are able to produce circadian oscillations, with periods ranging from 20 to 28 hours[3,4]. Daily rhythms in behavior, physiology and metabolism are controlled by endogenous circadian clocks. At the heart of these clocks is a circadian oscillator that keeps circadian time. There are many factors that entrain circadian oscillators such as the intercellular coupling, a 24-h LD cycle and intercellular and extracellular noise and the structure of SCN. In this article, we mainly analyze the effect of a 24-h LD cycle and the environment noise. Firstly, a mathematical model to describe the behavior of a population of SCN neurons is presented. The single cell oscillator is described by a K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 1–7, 2007. c Springer-Verlag Berlin Heidelberg 2007 

2

Y. Li, J. Zhang, and Z. Liu

three-variable model similar to the widely used Goodwin model that was able to simulate physiological oscillations on the basis of a negative feedback. This model, based on the negative feedback loop, accounts for the core molecular mechanism leading to self-sustained oscillations of clock genes. Then under the condition to have periodic solution, we show that, the models of cells with different parameter values, their individual periods are different. Our attention is mainly focused on the analysis and comparison of their effects on the circadian oscillator from numerical viewpoint. The result of this paper is that the environment noise can make for phase synchronization but it increases the threshold of the strength of light to make the oscillators synchronize to 24-h compared to the case without noise and the 24-h LD cycle plays the crucial role on entraining the 24-h clocks.

2

Model of Self-sustained Oscillation in a SCN Neuron

To simulate circadian oscillations in single mammalian cells, we resort to a threevariable model, based on the Goodwin oscillator[5]. In this model, a clock gene mRNA (X ) produces a clock protein (Y ) which, in turn, actives a transcriptional inhibitor (Z ). The latter inhibits the transcription of the clock gene, closing a negative feedback loop. In circadian clocks, protein degradation is controlled by phosphorylation, ubiquitination and proteasomal degradation and thus it is reasonable to assume Michaelian kinetics. Here,we advise the model of an individual cell as following: ⎧ K1n X ⎪ ⎨ X˙ = v1 K1n +Z n − v2 K2 +X , (1) Y˙ = k3 X − v4 K4Y+Y , ⎪ ⎩ ˙ Z = k5 Y − v6 K6Z+Z . where v1 , v2 , v4 , v6 , K1 , K2 , K4 , K6 , k3 , k5 are parameters. In this version, self-sustained oscillation can be obtained for a Hill coefficient of n = 4. The variable X represents mRNA concentration of a clock gene, per or cry; Y is the resulting protein, PER or CRY; and Z is the active protein or the nuclear form of the protein(inhibitor). This model is closely related to those proposed by Ruoff and Rensing[7], Leloup and co-workers[6], or Ruoff and co-workers[8] for the circadian clock in Neurospora. In [10], we analyzed the dynamics of model (1) for which we gave the sufficient conditions to be a self-sustained oscillator. Now, we consider the multi-cell system under the effect of 24-h LD cycle and environment noise. The evolution equations for the population composed of N oscillators(denoted by i = 1, 2, ..., N ) are then written as ⎧ K1n Xi ⎪ ⎨ X˙ i = v1 K1n +Zin − v2 K2 +Xi + L + Dξi , i (2) Y˙i = k3 Xi − v4 K4Y+Y , i ⎪ ⎩ ˙ Zi Zi = k5 Yi − v6 K6 +Zi . L is a square-wave function which reflects the effect of a LD cycle. The term L switches from L = 0 in dark phase to L = L0 in light phase. That is to say

Phase Synchronization of Circadian Oscillators

 L(t) =

L0 t ∈ [24k, 24k + 12), 0 t ∈ [24k + 12, 24(k + 1)),

3

(3)

where k is a natural number. The parameter D describes the strength of noise. ξi , called as extracellular noises representing external noises originated outside the cells due to environment perturbations, are assumed as independent Gaussian white noises with zero mean < ξi (t) >= 0 and covariances < ξi (t), ξj (t ) >= δ(t − t ).

3

The Main Results

In paper [10], we proved theoretically that the oscillators can get phase synchronization with period of 24-h under the effect of the 24-h LD cycle as if the strength of light was large enough. Here we invest the effect of the environment noises. When L0 = D = 0, system Eq.(2) is exactly the self-sustained oscillators of individual cells. Here we consider 20 cells that is to say N = 20. The values of parameters consult Ref.[9]. From Fig.1, we can see that the twenty self-sustained oscillators have different periods ranging from 20-h to 30-h. Now we consider the effect of environment noise that is to say D = 0. The simulation shows that the extracellular noise makes for phase synchronization(See Fig.3). But at the same time, the noise increases the threshold of L0 that makes the oscillators get phase synchronization with period of 24-h compared to the case without noise, which is verified by Fig.3-6. Fig.3 shows that the noise makes the oscillators to get phase synchronization but their periods are not 24-h. At the same time the light is added, when L0 = 0.01 the oscillators can not get the phase synchronization with period of 24-h while they can without 0.25

Variable X for 20 oscillators

0.2

0.15

0.1

0.05

1

1.002

1.004

1.006 Time(h)

1.008

1.01

1.012 4

x 10

Fig. 1. The time evolution of variable X for 20 oscillators, when L0 = D = 0

4

Y. Li, J. Zhang, and Z. Liu 0.4

0.35

L and Variable X for 20 oscillators

0.3

0.25

0.2

0.15

0.1

0.05

0

1

1.002

1.004

1.006 Time(h)

1.008

1.01

1.012 4

x 10

Fig. 2. The time evolution of L and variable X for 20 oscillators, when L0 = 0.02 and D=0 1

0.9

0.8

Variable X for 20 oscillators

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

2

2.002

2.004

2.006 Time(h)

2.008

2.01

2.012 4

x 10

Fig. 3. The time evolution of variable X for 20 oscillators, when L0 = 0 and D = 0.1

noise. With the increase of L0 , the periods of these synchronized oscillators approach 24-h gradually. Until L0 ≥ 0.2 with noise, these oscillators get the phase synchronization with period of 24-h. That is to say the threshold is increased greatly. When only the 24-h LD cycle is added, from paper [10], we know that as long as L0 is large enough, these oscillators can get phase synchronization and their periods are all 24-h. Our simulation results show that when L0 ≥ 0.015, the phase synchronization with period of 24-h is got(See Fig.2).

Phase Synchronization of Circadian Oscillators

5

0.8

0.7

L and variable X for 20 oscillators

0.6

0.5

0.4

0.3

0.2

0.1

0

1

1.002

1.004

1.006 Time(h)

1.008

1.01

1.012 4

x 10

Fig. 4. The time evolution of L and variable X for 20 oscillators, when L0 = 0.01 and D = 0.1 1

0.9

L and variable X for 20 oscillators

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

1

1.002

1.004

1.006 Time(h)

1.008

1.01

1.012 4

x 10

Fig. 5. The time evolution of L and variable X for 20 oscillators, when L0 = 0.1 and D = 0.1

From the Fig.1-6, we get a summary result that the extracellular noise and the 24-h LD cycle can accelerate phase synchrony of oscillators with different respective periods. But only the 24-h LD cycle can entrain the 24-h circadian clocks of the population of cells and the extracellular noise increases the threshold of the strength of light that makes the oscillators to get phase synchronization with period of 24-h.

6

Y. Li, J. Zhang, and Z. Liu 2

1.8

L and variable X for 20 oscillators

1.6

1.4

1.2

1

0.8

0.6

0.4

0.2

0

1

1.002

1.004

1.006 Time(h)

1.008

1.01

1.012 4

x 10

Fig. 6. The time evolution of L and variable X for 20 oscillators, when L0 = 0.2 and D = 0.1

4

Discussion and Conclusions

In this letter, we have introduced a molecular model for the regulatory network underlying the circadian oscillations in the SCN. We analyze the effect of the 24h LD cycle and extracellular noise numerically. From these numerical simulation we can see that the 24-h LD plays a crucial role in entraining the circadian clocks 24-h which accords with the biologic experiments. The results of this paper establish quantitative basis for understanding the essential cooperative dynamics. The effect of intercellular coupling and the SCN structure will be discussed later. Acknowledgments. This research is supported by the NNSF of China (Grants: 70431002) and Innovation Foundation of Shanghai University for Postgraduates. We express special thanks.

References 1. Reppert, S.M., Weaver, D.R.: Coordination of circadian timing in mammals. Nature 418, 935–941 (2002) 2. Moore, R.Y., Speh, J.C., Leak, R.K.: Suprachiasmatic nucleus organizarion. Cell Tissue Res. 309, 89–98 (2002) 3. Welsh, D.K., Logothetis, D.E., Meister, M., Reppert, S.M.: Individual neurons dissociated from rat suprachiasmatic nucleus express independently phased circadian firing rhythms. Neuron 14, 697–706 (1995) 4. Honma, S., Nakamura, W., Shirakawa, T., Honma, K.: Diversity in the circadian periods of single neurons of the rat suprachiasmatic nucleus on nuclear structure and intrinsic period. Neurosci. Lett. 358, 173–176 (2004)

Phase Synchronization of Circadian Oscillators

7

5. Goodwin, B.C.: Oscillatory behavior in enzymatic control processes. Adv. Enzyme Regul. 3, 425–438 (1965) 6. Leloup, J., Gonze, C.D., Goldbeter, A.: Limit cycle models for circadian rhythms based on thanscriptional regulation in Drosophila and Neurospora. J. Bilo. Rhy. 14, 433–448 (1999) 7. Rouff, P., Rensing, L.: The temperature-conpensated Goodwin model simulates many circadian clock properties. J. Theor. Biol. 179, 275–285 (1996) 8. Rouff, P., Vinsjevik, M., Monnerjahn, C., Rensing, L.: The Goodwin model: simulating the effect of light pulses on the circadian sporulation rhythm of Neurospora crassa. J. Theor. Biol. 209, 29–42 (2001) 9. Didier, G., Bernard, S., Waltermann, C., Kramer, A., Herzel, H.: Spontaneous synchronization of coupled circadian oscillators. Biophysical J. 89, 120–129 (2005) 10. Li, Y., Zhang, J.B., Liu, Z.R.: Circadian oscillators and phase synchronization under a light-dark cycle. Int. J. Nonlinear Science 1(3), 131–138 (2006)

Detecting RNA Sequences Using Two-Stage SVM Classifier Xiaoou Li1 and Kang Li2 1

2

Departamento de Computaci´ on CINVESTAV-IPN A.P. 14-740, Av.IPN 2508, M´exico D.F., 07360, M´exico School of Electronics, Electrical Engineering and Computer Science Queen’s University Belfast Ashby Building, Stranmillis Road, Belfast, BT9 5AH, UK [email protected]

Abstract. RNA sequences detection is time-consuming because of its huge data set size. Although SVM has been proved to be useful, normal SVM is not suitable for classification of large data sets because of its high training complexity. A two-stage SVM classification approach is introduced for fast classifying large data sets. Experimental results on several RNA sequences detection demonstrate that the proposed approach is promising for such applications.

1

Introduction

RNA plays many important biological roles other than as a transient carrier of amino acid sequence information [14]. It catalyzes peptide bond formation, participates in protein localization, serves in immunity, catalyzes intron splicing and RNA degradation, serves in dosage compensation. It is also an essential subunit in telomeres, guides RNA modification, controls development, and has an abundance of other regulatory functions [29]. Non-coding RNAs (ncRNAs) are transcripts that have function without being translated to protein [12]. The number of known ncRNAs is growing quickly, and their significance had been severely underestimated in classic models of cellular processes. It is desirable to develop high-throughput methods for discovery of novel ncRNAs for greater biological understanding and for discovering candidate drug targets. However, novel ncRNAs are difficult to detect in conventional biochemical screens. They are frequently short, often not polyadenylic, and might only be expressed under specific cellular conditions. Experimental screens have found many ncRNAs, but have demonstrated that no single screen is capable of discovering all known ncRNAs for an organism. A more effective approach, demonstrated in previous studies [2,28], may be to first detect ncRNA candidates computationally, then verify them biochemically. Considering the number of available whole genome sequences, SVM can be applied to a large and diverse data set, and has massive potential for novel ncRNA discovery[21,27]. However, long training K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 8–20, 2007. c Springer-Verlag Berlin Heidelberg 2007 

Detecting RNA Sequences Using Two-Stage SVM Classifier

9

time is needed. Therefore, it is impossible to repeat the SVM classification on the updated data set in an acceptable time when new data are included into the data set frequently or continuously. Many researchers have tried to find possible methods to apply SVM classification for large data sets. Generally, these methods can be divided into two types: 1) modify SVM algorithm so that it could be applied to large data sets, and 2) select representative training data from a large data set so that a conventional SVM could handle. For the first type, a standard projected conjugate gradient (PCG) chunking algorithm can scale somewhere between linear and cubic in the training set size [8,16]. Sequential Minimal Optimization (SMO) is a fast method to train SVM [23,7]. Training SVM requires the solution of QP optimization problem. SMO breaks this large QP problem into a series of smallest possible QP problems, it is faster than PCG chunking. [10] introduced a parallel optimization step where block diagonal matrices are used to approximate the original kernel matrix so that SVM classification can be split into hundreds of subproblems. A recursive and computational superior mechanism referred as adaptive recursive partitioning was proposed in [17], where the data is recursively subdivided into smaller subsets. Genetic programming is able to deal with large data sets that do not fit in main memory [11]. Neural networks technique can also be applied for SVM to simplify the training process [15]. For the second type, clustering has been proved to be an effective method to collaborate with SVM on classifying large data sets. For examples, hierarchical clustering [31,1], k-means cluster [4] and parallel clustering [7]. Clustering based methods can reduce the computations burden of SVM, however, the clustering algorithms themselves are still complicated for large data set. Rocchio bundling is a statistics-based data reduction method [25]. The Bayesian committee machine is also reported to be used to train SVM on large data sets, where the large data set is divided into m subsets of the same size, and m models are derived from the individual sets [26]. But, it has higher error rate than normal SVM and the sparse property does not hold. Falling into the second type of SVM classification methods for large data sets, a two stages SVM classification approach has been proposed in our previous work[4,5,18]. At first, we select representative training data from the original data set using the results of clustering, and these selected data are used to train the first stage SVM. Note that the first stage SVM is not precise enough because of the great reduction on original data set. So we use a second stage SVM to refine the classification. The obtained support vectors of the first stage SVM are used to select data for the second stage SVM by recovering their clustermates (we call the process de-clustering). At last, the second stage SVM is applied on those de-clustered data. Our experimental results show that the accuracy obtained by our approach is very close to the classic SVM methods, while the training time is significantly shorter. Furthermore, the proposed approach can be applied on huge data sets regardless of their dimensionality. In this paper, we apply our approach on several RNA sequences data sets. The rest of the paper is organized as follows: Section II introduces our two

10

X. Li and K. Li

stages SVM classifier. Section III show the experiment results on RNA sequence detection with comparisons with other well known classifiers. Conclusion is given in Section IV.

2

Two-Stage SVM Classifier

By the sparse property of SVM, data samples which are not support vectors will not contribute the optimal hyperplane. The input data sets which are far away from the decision hyperplane should be eliminated, meanwhile the data sets which are possibly support vectors should be used. In this paper, we select the cluster centers and data of mix-labeled clusters as training data for the first stage SVM. We believe these data are the most useful and representative in a large data set for finding support vectors. Nota that, the training data set in the first stage SVM classification is only a small percentage of the original data. Data of the clusters near the hyperplane are not used totally for training SVM, since we only select the cluster centers. This may affect the classification precision, i.e., the obtained decision hyperplane may not be precise enough. However, at least it gives us a reference on data distribution. According to above analysis, we make the following modification on the training data set of the first stage SVM. 1). Remove the data far from the hyperplane from the training data set because they will not contribute to find the support vectors, 2). Retain the data of the mix-labeled clusters since they are more likely support vectors. 3). Additionally, we add the data of the clusters whose centers are support vectors of the first stage SVM. In general, our approach consists of the four steps which are shown in Figure 1: 1) data selection, 2) the first stage SVM classification, 3) de-clustering, 4) the second stage SVM classification. The following subsections will give a detailed explanation on each step. 2.1

Selecting Training Data

The goal of clustering is to separate a finite number of unlabeled items into a finite and discrete set of “natural” hidden data structures, such that items in the same cluster are more similar to each other, and those in different clusters tend to be dissimilar according to certain measure of similarity or proximity. A large number of clustering methods have been developed, e.g., squared error-based kmeans [3], fuzzy C-means [22], kernel-base clustering [13]. By our experience, fuzzy C-means clustering, Minimum enclosing ball(MEB) clustering and random selection have been proved very effective for selecting training data for the first stage SVM.[4,5,18] Let l be the cluster number, then the process of clustering is to find l partitions (or clusters) Ωi from input data set X, i = 1, . . . , l, l < n, Ωi = ∅, ∪li=1 Ωi = X.

Detecting RNA Sequences Using Two-Stage SVM Classifier

11

Fig. 1. Two-stage SVM classification

Note that the data in a cluster may have same label (positive or negative) or different labels (both positive and negative). The obtained clusters can be classified into three types: 1) clusters with only positive labeled data, denoted by Ω + , i.e., Ω + = {∪Ωi | y = +1}; 2) clusters with only negative labeled data, denoted by Ω − , i.e., Ω − = {∪Ωi | y = −1}; 3) clusters with both positive and negative labeled data (or mix-labeled), denoted by Ωm , i.e., Ωm = {∪Ωi | y = ±1}. Figure 2 (a) illustrates the clusters after clustering, where the clusters with only red points are positive labeled (Ω + ), the clusters with green points are negative labeled (Ω − ) , and clusters A and B are mix-labeled (Ωm ).

Fig. 2. Data selection: (a) Clusters (b) The first stage SVM

12

X. Li and K. Li

We select not only the centers of the clusters but also all the data of mixlabeled clusters as training data in the first SVM classification stage. If we denote the set of the centers of the clusters in Ω + and Ω − by C + and C − respectively, i.e., C + = {∪Ci | y = +1} positive labeled centers C − = {∪Ci | y = −1} negative labeled centers Then the selected data which will be used in the first stage SVM classification is the union of C + , C − and Ωm , i.e., C + ∪ C − ∪ Ωm . In Figure 2 (b), the red points belongs to C + , and the green points belong to C − . It is clear that the data in Figure 2 (b) are all cluster centers except the data in mix-labeled clusters A and B. 2.2

The First Stage SVM Classification

We consider binary classification. Let (X, Y ) be the training patterns set, X = {x1 , · · · , xn }, Y = {y1 , · · · , yn } yi = ±1, xi = (xi1, . . . , xip )T ∈ Rp

(1)

The training task of SVM classification is to find the optimal hyperplane from the input X and the output Y , which maximize the margin between the classes. That is, training SVM yields to find an optimal hyperplane or to solve the following quadratic programming problem (primal problem), 1 T 2w w

n 

+ c ξk minw,b J (w) = k=1  T  subject : yk w ϕ (xk ) + b ≥ 1 − ξk

(2)

where ξk is slack variables to tolerate mis-classifications ξk > 0, k = 1 · · · n,  c > 0, wk is the distance from xk to the hyperplane wT ϕ (xk ) + b = 0, ϕ (xk ) is a nonlinear function. The kernel which satisfies the Mercer condition [9] is T K (xk , xi ) = ϕ (xk ) ϕ (xi ) . (2) is equivalent to the following quadratic programming problem which is a dual problem with the Lagrangian multipliers αk ≥ 0, n n   maxα J (α) = − 21 yk yj K (xk , xj ) αk αj + αk subject :

n 

k,j=1

αk yk = 0,

k=1

(3)

0 ≤ αk ≤ c

k=1

Many solutions of (3) are zero, i.e., αk = 0, so the solution vector is sparse, the sum is taken only over the non-zero αk . The xi which corresponds to nonzero αi is called a support vector (SV). Let V be the index set of SV, then the optimal hyperplane is  αk yk K (xk , xj ) + b = 0 (4) k∈V

Detecting RNA Sequences Using Two-Stage SVM Classifier

The resulting classifier is



y(x) = sign



13

 αk yk K (xk , x) + b

k∈V

where b is determined by Kuhn-Tucker conditions. Sequential minimal optimization (SMO) breaks the large QP problem into a series of smallest possible QP problems [23]. These small QP problems can be solved analytically, which avoids using a time-consuming numerical QP optimization as an inner loop. The memory required by SMO is linear in the training set size, which allows SMO to handle very large training sets [16]. A requirement l  αi yi = 0, it is enforced throughout the iterations and implies that in (3) is i=1

the smallest number of multipliers can be optimized at each step is two. At each step SMO chooses two elements αi and αj to jointly optimize, it finds the optimal values for these two parameters while all others are fixed. The choice of the two points is determined by a heuristic algorithm, the optimization of the two multipliers is performed analytically. Experimentally the performance of SMO is very good, despite needing more iterations to converge. Each iteration uses few operations such that the algorithm exhibits an overall speedup. Besides convergence time, SMO has other important features, such as, it does not need to store the kernel matrix in memory, and it is fairly easy to implement [23]. In the first stage SVM classification, we use SVM classification with SMO algorithm to get the decision hyperplane. Here, the training data set is C + ∪ C − ∪ Ωm , which has been obtained in the last subsection. Figure 2 (b) shows the results of the first stage SVM classification. 2.3

De-clustering

We propose to recover the data into the training data set by including the data in the clusters whose centers are support vectors of the first stage SVM, we call this process de-clustering. Then, more original data near the hyperplane can be found through the de-clustering. The de-clustering results of the support vectors in Figure 2 (b) are shown in Figure 3 (a). The de-clustering process not only overcomes the drawback that only small part of the original data near the support vectors are trained, but also enlarge the training data set size of the second stage SVM which is good for improving the accuracy. 2.4

The Second Stage SVM Classification

Taking the recovered data as new training data set, we use again SVM classification with SMO algorithm to get the final decision hyperplane  yk α∗2,k K (xk , x) + b∗2 = 0 (5) k∈V2

where V2 is the index set of the support vectors in the second stage. Generally, the hyperplane (4) is close to the hyperplane (5).

14

X. Li and K. Li

Fig. 3. (a) De-clustering (b) The second stage SVM

In the second stage SVM, we use the following two types of data as training data: 1). The data of the clusters whose centers are support vectors, i.e., ∪Ci ∈V {Ωi }, where V is a support vectors set of the first stage SVM; 2). The data of mix-labeled clusters, i.e, Ωm . Therefore, the training data set is ∪Ci ∈V {Ωi } ∪ Ωm . Figure 3 (b) illustrates the second stage SVM classification results. One can observe that the two hyperplanes in Figure 2 (b) and Figure 3 (b) are different but similar.

3

RNA Sequence Detection

We use three case studies to show the two-stage SVM classification approach introduced in the last section. The first example is done to show the necessities of the second stage SVM by comparing the accuracy of both stages. The second example is not a large data set, but it shows that training time, accuracy can be improved through adjusting the cluster number. The third example is a real large data set, we made a complete comparison with several well known algorithms as well as our two stage SVM with different clustering methods. Example 1. The training data is at www.ghastlyfop.com/blog/tag index svm.html/ .To train the SVM classifier, a training set containing every possible sequence pairing. This resulted in 47, 5865 rRNA and 114, 481 tRNA sequence pairs. The input data were computed for every sequence pair in the resulting training set of 486, 201 data points. Each record has 8 attributes with continuous values between 0 to 1. In [27], a SVM-based method was proposed to predict the common structure of two RNA sequences on the basis of minimizing folding free energy change. RNA, the total free energy change of an input sequence pair can either be compared with the total free energy changes of a set of control sequence pairs, or be used in combination with sequence length and nucleotide frequencies as input to a classification support vector machine.

Detecting RNA Sequences Using Two-Stage SVM Classifier

15

Fig. 4. The first stage SVM classification on the RNA sequences data set used in [27] with 103 data

In our experiments, we obtained 12 clusters from 1000 original data using FCM clustering. Then, in the first stage SVM, 113 training data including the cluster centers and data of mix-labeled clusters are obtained using the data selection process introduced in section II, and we got 23 support vectors. Figure 4 shows the result of the first stage SVM. Following the de-clustering technique, 210 data were recovered as training data for the second stage SVM. In the second stage SVM, we got 61 support vectors, see Figure 5. Table 1 shows the comparisons on training time and accuracy between our two SVM stages. The training time of our two-stage SVM and LIBSVM is first compared. For training 103 data, our classifier needs 67 seconds while LIBSVM needs about 100 seconds. For training 104 data, our classifier needs 76 seconds while LIBSVM needs about 1, 000 seconds. And, for 486, 201 data, our classifier needs only 279 seconds while the LIBSVM should use a very long time, it is not reported in [27], (we guess it maybe around 105 seconds). On the other hand, there is almost no difference between their accuracies. This implies that our approach has great advantage on gaining training time. Then, the accuracy between the first stage SVM and two-stage SVM is compared. From the figures and Table 1, it is obvious that the accuracy of two-stage SVM is much better than the first stage SVM. This shows that the two stages are necessary. Example 2. 3mer Dataset. The original work on string kernels – kernel functions defined on the set of sequences from an alphabet S rather than on a vector space [9] – came from the field of computational biology and was motivated by algorithms for aligning DNA and protein sequences. The recently presented k-spectrum (gap-free k-gram) kernel and the (k,m) mismatch kernel provide an alternative model for string kernels for biological sequences, and were designed, in particular, for the application of SVM

16

X. Li and K. Li

Fig. 5. The two stages SVM classification on the RNA sequences data set used in [27] with 103 data

protein classification. These kernels use counts of common occurrences of short k-length subsequences, called k-mers, rather than notions of pairwise sequence alignment, as the basis for sequence comparison. The k-mer idea still captures a biologically-motivated model of sequence similarity, in that sequences that diverge through evolution are still likely to contain short subsequences that match or almost match. We use SVM to classify proteins based on sequence data into homologous groups (evolutionary similar) to understand the structure and functions of proteins. The 3mer data set has 2000 data points, and each record has 84 attributes with continuous values between 0 to 1. The data set contains 1000 positive sequences and 1000 negative sequences. The data set is available at noble.gs.washington.edu/proj/hs/ In [21], spectrum kernel was used as a feature set representing the distribution of every possible k-mer in a RNA sequence. The value for each feature is the number of times that particular feature appears in the sequence divided by the number of times any feature of the same length appears in the sequence. In our experiments, we used MEB clustering to select data for the first stage SVM, and k=3 (i.e., 3mers, 2mers and 1mers are our features) to train our two stages SVM classifier. Table 2 shows the accuracy and training time of our classifier and LIBSVM, where the accuracies are almost the same. Also there is no much deference on training time, this is because that the data set contains only 2,000 data. However, we did experiments with cluster number (l) 400 and 100. We can see that, when we use less cluster number, the training time is less too, since the training data size is smaller, but, we get a worse accuracy. Example 3. This RNA data set is available at http://www.pubmedcentral. nih.gov /articlerender.fcgi?artid=1570369#top from Supplementary Material (additional file 7). The data set consists of 23605 data points, each record has 8 attributes with

Detecting RNA Sequences Using Two-Stage SVM Classifier

17

Table 1. Accuracy and training time on the RNA sequences data set in [27]

Data set size 103 104 486, 201

First stage SVM T (s) Acc (%) 31 67.2 70 76.9 124 81.12

Two-stage SVM T (s) Acc (%) 76 88.9 159 92.7 279 98.4

LIBSVM T (s) Acc (%) 102 87.4 103 92.6 98.3 105 ?

Table 2. Accuracy and training time on the RNA sequences data set in [21] Two-stage SVM # t(s) Acc(%) l 2000 17.18 75.9 400 2000 7.81 71.7 100

LIBSVM # t(s) 2000 8.71 — —

Acc(%) 73.15 —

continuous values between 0 to 1. The data set contains 3919 ncRNAs and 19686 negative sequences. We used sizes 500, 1, 000, 2, 500, 5, 000, 10, 000 and 23, 605 in our experiments. Experiments were done using MEB two-stage, RS two-stage, SMO, LIBSVM and simple SVM. Table 3 shows our experiment results on different data size with MEB two-stage and RS two-stage. Table 4 shows the comparisons between our approach and other algorithms. In Table 3 and 4, the notations are as explained as follows. “#” is the data size; “t” is the training time of the whole classification which includes the time of clustering, the first stage SVM training, de-clustering and the second stage SVM training; “Acc” is the accuracy; “l” is the number of clusters used in the experiment; “TrD2” is the number of training data for the second stage SVM training; “SV1” is the number of support vectors obtained in the first stage SVM; “SV2” is the number of support vectors obtained in the second stage SVM. Table 3 shows our experiment results on different data size with MEB twostage and RS two-stage. For example, in the experiment on 10, 000 data points, we sectioned it into 650 clusters using MEB clustering and random selection. In the first stage classification of MEB two-stage, we got 199 support vectors. Following the de-clustering technique, 862 data were recovered as training data for the second stage SVM, which is much less than the original data size 10, 000. In the second stage SVM, 282 support vectors were obtained. From Table 2, we can also see that MEB two-stage has a little better accuracy than random selection (RS) two-stage, while its training time is longer than that of RS two-stage. Table 4 shows the comparison results on training time and accuracy between our two-stage classification and some other SVM algorithms including SMO, simple SVM and LIBSVM. For example, to classify 5000 data, LIBSVM is the fastest, and SMO has the best accuracy, our two approaches are not better than them,

18

X. Li and K. Li

although the time and accuracy are similar to them. However, to classify 23, 605 data, Simple SVM and SMO have no better accuracy than the others, but their training time is tremendous longer. Comparing to our two approaches, LIBSVM takes almost double training time of MEB two-stage, and almost 7 times of the time of RS two-stage, although it has the same accuracy as ours. This experiment implies that our approach has great advantage on large data sets since it can reach the same accuracy as the other algorithm can in a very short training time. Table 3. Two-stage SVM classification results on RNA sequence data set MEB two-stage # t Acc 500 4.71 85.3 1000 5.90 86.2 2500 15.56 86.3 5000 26.56 86.7 10000 69.26 86.9 23605 174.5 88.5

l 350 400 450 500 650 1500

SV1 87 108 124 149 199 278

TrD2 397 463 529 656 862 1307

RS two-stage # t Acc 500 4.07 85.3 1000 4.37 85.7 2500 11.2 86.5 5000 15.8 86.1 10000 30.2 86.5 23605 65.7 88.3

SV2 168 162 209 227 282 416

l 350 400 450 500 650 1500

SV1 88 97 132 146 187 257

TrD2 421 453 581 637 875 1275

SV2 172 153 221 211 278 381

Table 4. Training time and accuracy on RNA sequence data set

# 500 1000 2500 5000 10000 23605

4

MEB two-stage RS two-stage t Acc t Acc 4.71 85.3 4.07 85.3 5.90 86.2 4.37 85.7 15.56 86.3 11.21 86.5 26.56 86.7 15.79 86.1 69.26 87.9 30.22 86.5 174.5 88.2 65.7 88.3

LIBSVM t Acc 0.37 86 0.72 87.2 3.06 87.4 12.53 87.6 48.38 88.2 298.3 88.6

SMO t 1.56 3.54 4.20 212.43 1122.5 —-

Acc 87.7 88.3 87.7 88.8 89.6 —-

Simple SVM t Acc 2.78 86.7 8.18 87.1 561.3 88.1 ——————-

Conclusions and Discussions

Our two-stage SVM classification approach is much faster than other SVM classifiers without loss of accuracy when data set is large enough. From the results of the experiments we made on biological data sets in this work, our approach has been showed suitable for classifying large and huge biological data sets. Additionally, another promising application on genomics machine learning is under study.

References 1. Awad, M.L., Khan, F., Bastani, I., Yen, L.: An Effective support vector machine(SVMs) Performance Using Hierarchical Clustering. In: Proceedings of the 16th IEEE International Conference on Tools with Artificial Intelligence, pp. 663– 667. IEEE Computer Society Press, Los Alamitos (2004) 2. Axmann, I.M., Kensche, P., Vogel, J., Kohl, S., Herzel, H., Hess, W.R.: Identification of cyanobacterial non-coding RNAs by comparative genome analysis. Genome Biol R73 6 (2005)

Detecting RNA Sequences Using Two-Stage SVM Classifier

19

3. Babu, G., Murty, M.: A near-optimal initial seed value selection in K-means algorithm using a genetic algorithm. Pattern Recognit. Lett. 14, 763–769 (1993) 4. Cervantes, J., Li, X., Yu, W.: Support Vector Machine Classification Based on Fuzzy Clustering for Large Data Sets. In: Gelbukh, A., Reyes-Garcia, C.A. (eds.) MICAI 2006. LNCS (LNAI), vol. 4293, pp. 572–582. Springer, Heidelberg (2006) 5. Cervantes, J., Li, X., Yu, W., Li, K.: Support vector machine classification for large data sets via minimum enclosing ball clustering. Neurocomputing (accepted for publication) 6. Chang, C.C., Lin, C.J.: LIBSVM: A library for support vector machines (2001), http://www.csie.ntu.edu.tw/∼ cjlin/libsvm 7. Chen, P.H., Fan, R.E., Lin, C.J.: A Study on SMO-Type Decomposition Methods for Support Vector Machines. IEEE Trans. Neural Networks 17, 893–908 (2006) 8. Collobert, R., Bengio, S.: SVMTorch: Support vector machines for large regression problems. Journal of Machine Learning Research 1, 143–160 (2001) 9. Cristianini, N., Shawe-Taylor, J.: An Introduction to Support Vector Machines and Other Kernel-based Learning Methods. Cambridge University Press, Cambridge (2000) 10. Dong, J.X., Krzyzak, A., Suen, C.Y.: Fast SVM Training Algorithm with Decomposition on Very Large Data Sets. IEEE Trans. Pattern Analysis and Machine Intelligence 27, 603–618 (2005) 11. Folino, G., Pizzuti, C., Spezzano, G.: GP Ensembles for Large-Scale Data Classification. IEEE Trans. Evol. Comput. 10, 604–616 (2006) 12. Griffiths-Jones, S., Moxon, S., Marshall, M., Khanna, A., Eddy, S.R., Bateman, A.: RFAM: annotating non-coding RNAs in complete genomes. Nucleic Acids Res. 33, 121–124 (2005) 13. Girolami, M.: Mercer kernel based clustering in feature space. IEEE Trans. Neural Networks 13, 780–784 (2002) 14. Hansen, J.L., Schmeing, T.M., Moore, P.B., Steitz, T.A.: Structural insights into peptide bond formation. Proc. Natl. Acad. Sci. 99, 11670–11675 (2002) 15. Huang, G.B., Mao, K.Z., Siew, C.K., Huang, D.S.: Fast Modular Network Implementation for Support Vector Machines. IEEE Trans. on Neural Networks (2006) 16. Joachims, T.: Making large-scale support vector machine learning practice. Advances in Kernel Methods: Support Vector Machine. MIT Press, Cambridge (1998) 17. Kim, S.W., Oommen, B.J.: Enhancing Prototype Reduction Schemes with Recursion: A Method Applicable for Large Data Sets. IEEE Trans. Syst. Man, Cybern. B. 34, 1184–1397 (2004) 18. Li, X., Cervantes, J., Yu, W.: Two Stages SVM Classification for Large Data Sets via Randomly Reducing and Recovering Training Data. In: IEEE International Conference on Systems, Man, and Cybernetics, Montreal Canada (2007) 19. Lin, C.T., Yeh, L.C.M., S, F., Chung, J.F., Kumar, N.: Support-Vector-Based Fuzzy Neural Network for Pattern Classification. IEEE Trans. Fuzzy Syst. 14, 31–41 (2006) 20. Mavroforakis, M.E., Theodoridis, S.: A Geometric Approach to Support Vector Machine(SVM) Classification. IEEE Trans. Neural Networks 17, 671–682 (2006) 21. Noble, W.S., Kuehn, S., Thurman, R., Yu, M., Stamatoyannopoulos, J.: Predicting the in vivo signature of human gene regulatory sequences. Bioinformatics 21, 338– 343 (2005) 22. Pal, N., Bezdek, J.: On cluster validity for the fuzzy c-means model. IEEE Trans. Fuzzy Syst. 3, 370–379 (1995)

20

X. Li and K. Li

23. Platt, J.: Fast Training of support vector machine using sequential minimal optimization. Advances in Kernel Methods: support vector machine. MIT Press, Cambridge, MA (1998) 24. Prokhorov, D.: IJCNN 2001 neural network competition. Ford Research Laboratory (2001), http://www.geocities.com/ijcnn/nnc ijcnn01.pdf 25. Shih, L., Rennie, D.M., Chang, Y., Karger, D.R.: Text Bundling: Statistics-based Data Reduction. In: Proc. of the Twentieth Int. Conf. on Machine Learning, Washington DC (2003) 26. Tresp, V.: A Bayesian Committee Machine. Neural Computation 12, 2719–2741 (2000) 27. Uzilov, A.V., Keegan, J.M., Mathews, D.H.: Detection of non-coding RNAs on the basis of predicted secondary structure formation free energy change. BMC Bioinformatics 173 (2006) 28. Washietl, S., Hofacker, I.L., Lukasser, M., Huttenhofer, A., Stadler, P.F.: Mapping of conserved RNA secondary structures predicts thousands of functional noncoding RNAs in the human genome. Nat. Biotechnol. 23, 1383–1390 (2005) 29. Weilbacher, T., Suzuki, K., Dubey, A.K., Wang, X., Gudapaty, S., Morozov, I., Baker, C.S., Georgellis, D., Babitzke, P., Romeo, T.: A novel sRNA component of the carbon storage regulatory system of Escherichia coli. Mol. Microbiol. 48, 657– 670 (2003) 30. Xu, R., WunschII, D.: Survey of Clustering Algorithms. IEEE Trans. Neural Networks 16, 645–678 (2005) 31. Yu, H., Yang, J., Han, J.: Classifying Large Data Sets Using SVMs with Hierarchical Clusters. In: Proc. of the 9th ACM SIGKDD (2003)

Frequency Synchronization of a Set of Cells Coupled by Quorum Sensing Jianbao Zhang2 , Zengrong Liu1 , Ying Li2 , and Luonan Chen1 1

Institute of Systems Biology, Shanghai University, Shanghai. 200444, China [email protected], [email protected] 2 College of Sciences, Shanghai University, Shanghai, 200444, China

Abstract. Collective behavior of a set of cells coupled by quorum sensing is a hot topic of biology. Noticing the potential applications of frequency synchronization, the paper studies frequency synchronization of a set of cells with different frequencies coupled by quorum sensing. By phase reduced method, the multicell system is transformed to a phase equation, which can be studied by master stability function method. The sufficient conditions for frequency synchronization of the multicell system is obtained under two general hypotheses. Numerical simulations confirm the validity of the results.

1

Introduction

Recently, collective dynamics in a population of cells communicated with each other through intercellular signaling have attracted much attention from many fields of biology and many researches have been carried out[1,2,3] . In the sight of biology, this type of collective dynamics is caused by intercellular signaling from identical and unreliable components, whereas bacteria display various social behaviors and cellular differentiations, such as quorum sensing in grampositive and gram-negative strains because of intercellular communications[4,5,6] . In order to study the theoretical mechanism of such phenomena, many interesting studies were carried out such as Ref.[2]. But most of studies are based on theory of synchronization and two hypotheses are adopted for the convenience of mathematical analysis[2,7] : (1). The network consists of identical cells, in other words, all the individual cells are with identical parameters; (2). The hypothesis of quasi-steady-state approximation of the network holds. Under the two hypotheses above, the authors give the conditions for complete synchronization. It is very difficult to study theoretically without the two hypotheses. Luckily, there are many different synchronization states such as frequency synchronization, which is more important than complete synchronization in the fields of biology. Plenty of phenomena such as the circadian rhythms of mammals[3] verify the importance of frequency synchronization. Therefore, we try to deduce 

Supported by National Natural Science Foundation of China (70431002,10672093).

K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 21–27, 2007. c Springer-Verlag Berlin Heidelberg 2007 

22

J. Zhang et al.

the theoretical mechanism of frequency synchronization caused by intercellular signaling without the two hypotheses mentioned above. In this direction, we give an introduction to the advances of phase synchronization. In order to study phase synchronization of a set of limit-cycle oscillators, some researchers proposed the phase reduced method[8,9,10] . Initially, the method transformed an oscillator under a small perturbation to a compact dynamical equation of phase. Then Kuramoto considered a network consisting of n identical subsystems with intrinsic frequency ωi and obtain the famous Kuramoto model[12] . Later[13] , phase synchronization in random sparse complex networks of small-world and scale-free topologies with long-range decayed couplings was studied. Recently, population of identical limit-cycle oscillators driven by common additive noise[14] and uncoupled limit-cycle oscillators under random impulses[15] was also studied. The results mentioned above show that the phase reduced method is a valid method to study phase synchronization. In our point, the more important collective dynamics in multicell systems or multicellular structures is frequency synchronization. The paper studies frequency synchronization of cells communicated with each other by intercellular signaling through the method mentioned above.

2

A Synthetic Multicellular Network

Now, we introduce a synthetic gene network in Escherichia coli[7] and all cells communicate with each other by signaling molecules based on the quorumsensing mechanism. The repressilator is a network of three genes, a, b, c, the products of which inhibit the transcription of each other in a cyclic way. Specifically (see Fig. 1), the gene c expresses protein C, which inhibits transcription of the gene a. The product of gene a inhibits transcription of the gene b, the protein product B of which in turn inhibits expression of gene b, completing the cycle. These bacteria exhibit cell-to-cell communication through a mechanism that makes use of two proteins, the first one of which (LuxI) synthesizes a small molecule known as an autoinducer (AI), which can diffuse freely through the cell membrane. When a second protein (LuxR) binds to this molecule, the resulting complex activates transcription of various genes, including some coding for light-producing enzymes. The scheme of the network is shown in Figure 1, and the detailed description is provided by Garcia-Ojalvo and others (2004). The dynamics of genes a, b, c and proteins A, B, C are given respectively as dai (t) αC = −d1i ai (t) + μC +C , m dt i (t) dbi (t) αA , dt = −d2i bi (t) + μA +Am i (t) dci (t) αS Si (t) αB dt = −d3i ci (t) + μB +Bim (t) + μS +Si (t) , dAi (t) = −d4i Ai (t) + βa ai (t), dt dBi (t) = −d5i Bi (t) + βb bi (t), dt dCi (t) = −d6i Ci (t) + βc ci (t), dt

(1)

Frequency Synchronization of a Set of Cells

Bi

23

Ci

cI (

bi

)

lac I (

ci

) AI

te tR ( a i )

Ai

Ci la cI (

ci

lu x I

)

AI

Ai

LuxR

AI AI

the ith cell CELL

CELL

CELL

AI

CELL

C ELL CELL

C ELL

Fig. 1. Scheme of the repressilators communicated with each other through autoinducer (AI), which can diffuse freely through the cell membrane

dSi (t) = −ds Si (t) + βs Ai (t), (2) dt where ai , bi , and ci are the concentrations of mRNA transcribed from genes a, b, and c in cell i, respectively; concentrations of the corresponding proteins are represented by Ai , Bi , and Ci , respectively. Concentration of AI inside each cell is denoted by Si . αA , αB , and αC are the dimensionless transcription rates in the absence of repressor. αs is the maximal contribution to the gene c transcription in the presence of saturating amounts of AI. Parameters βa , βb and βc are the translation rates of the proteins from the mRNAs. βs is the synthesis rate of AI. m is the Hill coefficient, and dji are the respective dimensionless degradation rates of mRNA or proteins for genes a, b, and c and the corresponding proteins in cell i. Consider that AI can diffuse freely through the cell membrane and denote ηs and ηe as the diffusion rate of AI inward and outward the cell, the dynamics of AI can be rewritten as follows, dSi (t) dt dSe (t) dt

= −ds Si (t) + βs Ai (t) − ηs (Si (t) − Se (t)), N  Sj (t)−Se (t) = −de Se (t) + ηe . n j=1

(3)

24

J. Zhang et al.

Then the individual cells (1)-(2) communicated with each other through the signaling molecules and form a biological network (1)-(3). One can regard the signaling molecules in the extracellular environment as the (n+1)th node, which implies that the n cells communicated with the (n + 1)th node and construct a star-type network. Then we can apply the method to study frequency synchronization of oscillator networks to study frequency synchronization of multicell systems (1)-(3). The dynamics of most of biologic systems are periodic. For example, it can be concluded that system (1)-(2) exhibits limit-cycle oscillations in a wide region of parameter ds space from Fig. 2 of Ref[2]. Then denote xi = (ai , bi , ci , Ai , Bi , Ci , Si ) , it is reasonable to assume that each individual cells in network exhibits limitcycle oscillations. Let x0i (t) and Oi denote the stable periodic solution of individual cells and the orbit corresponding to x0i (t). In the light of phase reduced method, we can produce a definition of phase φi on the orbit Oi such that φ˙i (x0i ) = ωi . Extending the definition of φi to the absorbing domain Di of the orbit Oi , which is essentially the same as asymptotic phase[9] , one obtains φ˙i (xi ) = ωi , xi ∈ Di .

(4)

Then consider the network (1)-(3). The exact expression of Se is easily to obtain as follows, n  ηe  t −(de +ηe )t Se (t) = Se0 e + Sj (s)e(d+ηe )(s−t) ds. n j=1 0 Based on equation (3) and φ˙i (xi ) = gradxi φi · x˙ i , one gets ηe φ˙i = ωi +ηgradSi φi {Se0 1e−(de +ηe )t + n

n   j=1

t 0

(de +ηe )(s−t) xm ds − Si (φi (t))}. j (φj (s))e

(5)

It is reasonable to assume the following hypothesis holds, especially for a periodic function Sj (φj (t)). φi as the gradient of φi along xm (H1 ). Denote gradxm i , there holds i  t  t (d+ηy )(s−t) φi xm ds = Gij (φi , φj ) e(d+ηy )(s−t) ds. (6) gradxm j (φj (s))e i 0

0

Motivated by Ref.[16], we can further assume that hypothesis (H2 ) holds and a numerical example has been given to illustrate the rationality of such hypothesis. (H2 )[16] There exists d + 1 constants φ01 , . . . , φ0d , ω such that the following conditions. 1. Denoting ϕi = ωt + φ0i , Δωi = ωi − ω, there holds, n ηq  Δωi + Gij (ϕi , ϕj ) − ηHi (ϕi ) = 0. n j=1

(7)

2. G(φi , φj ) is differentiable and Gyij (φi , φj )|φ=ϕ = −Gxij (φi , φj )|φ=ϕ , where ∂ ∂ Gxij (φi , φj ) = ∂φ Gij (φi , φj ), Gyij (φi , φj ) = ∂φ Gij (φi , φj ). i j Then one obtain the following results.

Frequency Synchronization of a Set of Cells

25

Theorem 1. Under hypotheses (H1 ) and (H2 ), system (1)-(3) realizes collective rhythms if q is large enough. The detailed derivation and proof are similar to our recent work[16]. Though the conditions of the results relate with the initial values of system (2)-(4), but once some initial values are verified to satisfy Theorem.1, Theorem.1 still holds when the initial values deviates slightly. Therefore, Theorem.1 has many practical applications in the fields of biology.

3

Numerical Simulations

As far as the system mentioned in section 2 is concerned, we can not give the explicit expression of Gij (ϕi , ϕj ), Hi (φi ), but we can presume that Gij (ϕi , ϕj ), Hi (φi ) satisfy hypotheses (H1 ), (H2 ) because of the generality of the two hypotheses. Then we consider the multicell systems (2)-(4) consists of 6 cells with the initial values and parameters in Tab.1 and Tab.2.It has been shown thatparameters βa , βb , , βc and d4i , d5i , , d6i affects most markedly the oscillation frequency. Then all the repressilators oscillate at different frequencies because of the different values of β and d. Figure 2 shows that frequency synchronization Table 1. The initial values of the 6-cell system (2)-(4) n

1

2

3

4

5

6

ai0 8.2803 9.1756 1.1308 8.1213 9.0826 1.5638 bi0

1.2212 7.6267 7.2180 6.5164 7.5402 6.6316

ci0

8.8349 2.7216 4.1943 2.1299 0.3560 0.8116

Ai0 8.5057 3.4020 4.6615 9.1376 2.2858 8.6204 Bi0 6.5662 8.9118 4.8814 9.9265 3.7333 5.3138 Ci0 1.8132 5.0194 4.2219 6.6043 6.7365 9.5733 Si0 1.9187 1.1122 5.6505 9.6917 0.2374 8.7022 Se0 0.2688

Table 2. The parameters of the 6-cell system (2)-(4) n

1

2

3

4

5

6

αAi

2.9823 3.7146 3.8071 2.6168 2.8173 2.8836

βi

0.1302 0.1301 0.1303 0.1301 0.1304 0.1301

di

0.5008 0.5004 0.5012 0.5004 0.5015 0.5004

others αB = αC = 1.96, αs = 1.βs = 0.018, ds = 0.016, μA = μB = μC = 0.2, m = 4, ηs = 0.4.

26

J. Zhang et al.

can be realized, which confirms the validity of the results. The numerical simulations are consistent with the previous researches on system (2)-(4) which show that a network of coupled repressilators with identical parameters can realize e is large enough. complete synchronization when Q = ηeη+d e

ai:tetR in the ith cell

20

15

10

5

0 0

50

t→

100

150

1750

1800

ai:tetR in the ith cell

20

15

10

5

0 1650

1700

t→

Fig. 2. Time evolution of ai (tetR in the ith cell) of multicell systems (2)-(4) consists of 6 cells with different frequencies. Fig.a implies that the 6 cells oscillate at different frequencies when t ∈ [0, 150], but the 6 cells oscillate at identical frequency, different amplitudes when t ∈ [1650, 1800] (see Fig.b).

References 1. Wang, R., Jing, Z., Chen, L.: Modelling periodic oscillation in gene regulatory networks by cyclic feedback networks. Bull. Math. Biol. 67, 339–367 (2004) 2. Wang, R., Chen, L.: Synchronizing Genetic Oscillators by Signaling Molecules. Journal of biological Rhythms 20, 257–269 (2005) 3. Li, Y., Zhang, J., Liu, Z.: Circadian Oscillators and Phase Synchronization under a Light-Dark Cycle. International Journal of Nonlinear Science 1, 131–138 (2006) 4. Taga, M.E., Bassler, B.L.: Chemical communication among bacteria. PNAS 100, 14549–14554 (2003) 5. Weiss, R., Knight, T.F.: Engineering communications for microbial robotics. DNA 6, 13–17 (2000)

Frequency Synchronization of a Set of Cells

27

6. Chen, L., Wang, R., Zhou, T., Aihara, K.: Noise-induced cooperative behavior in a multi-cell system. Bioinformatics 21, 51–62 (2005) 7. Garcia-Ojalvo, J., Elowitz, M., Strogatz, S.H.: Modeling a synthetic multicellular clock: repressilators coupled by quorum sensing. PNAS 101, 10955–10960 (2004) 8. Guckenheimer, J.: Isochrons and phaseless sets. J. Math. Biol. 1, 259C273 (1975) 9. Kuramoto, Y.: Chemical Oscillations, Waves and Turbulence. Springer, New York (1984) 10. Winfree, A.T.: Biological Rhythms and the Behavior of Populations of Coupled Oscillators. J. Theoret. Biol. 16, 15–42 (1967) 11. Izhikevich, E.M.: Phase equations for relaxation oscillators. SIAM Journal on Applied Mathematics 60, 1789–1804 (2000) 12. Strogatz, S.H.: From Kuramoto to Crawford: Exploring the onset of synchronization in populations of coupled oscillators. Physica D 143, 1–20 (2000) 13. Li, X.: Phase synchronization in complex networks with decayed long-range interactions. Physica D 223, 242C247 (2006) 14. Teramae, J., Tanaka, D.: Robustness of the noise-induced phase synchronization in a general class of limit cycle oscillators. Physical Review Letters 93, 20 (2004) 15. Nakao, H., Arai, K., Nagai, K., Tsubo, Y., Kuramoto, Y.: Synchrony of limit-cycle oscillators induced by random external impulses. Phys. Rev. E. 72, 026220 (2005) 16. Zhang, J., Liu, Z., Li, Y.: An approach to analyze phase synchronization in oscillator networks with weak coupling. Chinese Physics Review Letters 24(6) (2007) 17. Pecora, L.M., Carroll, T.L.: Master Stability Functions for Synchronized Coupled Systems. Phycical Review Letters 80, 2109 (1998) 18. Jordi, G.O., Elowitz, M.B., Steven, H.S.: Modeling a synthetic multicellular clock: Repressilators coupled by quorum sensing. PNAS 101, 10955–10960 (2004)

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics Min Xu1, Yongsheng Ding1,2, and Liangjian Hu3 1

Glorious Sun School of Business and Management, Donghua University, Shanghai 200051, China 2 College of Information Sciences and Technology, Donghua University, Shanghai 201620, China [email protected] 3 Department of Applied Mathematics, Donghua University, Shanghai 201620, China

Abstract. In this paper, we first present a stochastic model of the proportion of the population infected with HIV against total population, and prove the existence and uniquess of its solution. Through computer simulation, we forecast the proportion of the population infected with HIV against the total population in the transmission course of AIDS in China in next 20 years. Especially, we study the control index of the transmission rate β to obtain its effect on the epidemic trend of AIDS when it fluctuates. As such, we present a strategy to adjust β to reach a certain control aim based on the analysis of the mean value and variance of the proportion.

1 Introduction The AIDS epidemic is spreading very fast according to the report from the United Union. The number of people infected with HIV increased from 35 million in 2001 to 38 million in 2003. According to a report from the Ministry of Health of the People’s Republic of China, the number of people infected with HIV was 135630 at the end of September of 2005. More and more researchers focus on prevention and control of AIDS. And most of them work on the spread of HIV based on ordinary differential equations. Haynatzka [1] studied the spread of AIDS among interactive transmission. Castillo [2] formulated a group of ordinary differential equations and studied the incubation in the dynamics of AIDS. Blythe and Anderson [3-5] studied various models, including heterogeneity population among whom AIDS is spreading. Jacquez et al. [6-7] presented a compartmental model of homosexual population for AIDS spreading. Greenhalgh et al. [8] discussed a two group model for the susceptible and infected populations. However, the above models have not considered the stochastic attributes during the transmission course of AIDS. Roberts [9] formulated a stochastic differential equation model for a fatal epidemic’s. In this paper, based on the above model, we extend it for AIDS and prove the existence and uniquess of the solution. K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 28–37, 2007. © Springer-Verlag Berlin Heidelberg 2007

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics

29

In our model, the control index of the transmission rate β has been changed from a constant in the deterministic model to a stochastic one in the stochastic process because of environment factor. And what’s more, the proportion of population infected with HIV against the total population in the future will also be a stochastic process correspondingly. In the deterministic model, a corresponding control aim can be obtained only if β is a certain constant. But it can not be obtained with 100% probability when the environment disturbance with a certain intension exists. In reality, the stochastic model can describe the transmission course exactly. Based on the formulation of the stochastic model, we analyses the proportion of the population infected with HIV against the total population in next few years under the environmental disturbance with a certain intensity. We present a method to adjust β to reach a certain control aim.

2 Stochastic Differential Equation Model of AIDS Transmission 2.1 The Stochastic Differential Equation Model When the infection rate of the disease is a constant, the proportion of the population infected with HIV against the total population satisfies the ordinary differential equation, dZ = ( p − 1) BZ + ( β C − α )(1 − Z ) Z dt

(1)

where Z is the proportion of the population infected with HIV against the total population; B is the birth rate independent on the total population; C is the contact rate between individuals; α is the increase of the death rate suffering from AIDS; p is the vertical distribution probability, which is more than 0 and less than 1. β is the constant transmission rate. Because of the environmental effect, β should be a stochastic process. We regard it as a Gaussian white noise β0 + ρη(t ) , and replace it in the Eq. (1). Then we get the stochastic differential equation (SDE) model for AIDS transmission, dZ = F ( Z )dt + G ( Z ) dW

(2)

F ( Z ) = ( p − 1) BZ + ( β 0 C − α )(1 − Z ) Z

(3)

G ( Z ) = ρC (1 − Z ) Z

(4)

Where

E (η (t )) = 0, D(η (t )) = 1,

β0 is the average transmission rate and ρ is the intensity of the

environmental disturbance. Z is a stochastic process reflecting the trend of fluctuation of the proportion under the environmental disturbance. 2.2 Existence and Uniquess of Solution Following, we prove the solution of the SDE model (2) is existent and unique. Theorem 1. [10] (Existence and uniquess of the solution of SDE model (2)) If X t satisfies n dimensional SDE

30

M. Xu, Y. Ding, and L. Hu

b( x, t ) = (b1 ( x, t ), b2 ( x, t ),..., bn ( x, t ))T

(5)

σ ( x, t ) = (σ ij ( x, t )) 1 ≤ i, j ≤ n

(6)

X 0 = x0 , x0 ∈ R n

(7)

The initial condition is

Assume that

b( x, t ) and σ ( x, t ) are continuous vector value function and matrix

value function about conditions below,

( x, t ) ∈ R n × [0, T ] , respectively, and satisfy the − − ⎧ ⎪ b( x, t ) − b( x, t ) ≤ c* x − x ⎪ ⎪ ο ( x, t ) − σ ( x− , t ) ≤ c x − x− * ⎨ ⎪ b( x, t ) ≤ c(1 + x ) ⎪ ⎪ σ ( x, t ) ≤ c(1 + x ) ⎩

Lipschitz

(8)

where c* and c are constants. Then the unique strong solution of Eqs. (1) and (2) exists in

(Ω, F , Ft , P) as far as any

x ∈ R n is concerned. Here, we always assume that Wt is Brown motion in the probability space (Ω, F , P ) , Ft is its natural σ algebra.

Theorem 2. Assume that p , B , β , α and C are positive real numbers. Then for any initial condition Z 0 (0 < Z 0 < 1) , there is a unique solution to Eq. (2). Proof. We just need prove the coefficients satisfy Lipschitz condition. Here, b( z, t ) = ( p − 1) Bz + ( β0C − α )(1 − z ) z

(9)

σ ( z, t ) = ρC (1 − z ) z

(10)

Then we verify that they satisfy the conditions in Theorem 1, −

−

−

−

−

_

b( z, t ) − b( z , t ) = ( p − 1) B( z − z ) + ( β 0C − α )[( z − z )(1 − ( z + z ))]

(11)

By means of three triangle inequality, we get −

−

b( z, t ) − b( z , t ) ≤ ( p − 1) B ( z − z ) + ( β 0C − α )[( z − z )(1 − ( z + z ))]

(12)

Noticing that z represents the proportion of the population infected with HIV against the total population, so −

z ≤ 1 . Then we get −

b( z, t ) − b( z , t ) ≤ 2 max( ( p − 1) B , 3( β 0 C − α ) ) z − z

(13)

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics

31

In addition, −

−

σ ( z , t ) − σ ( z , t ) ≤ 3 ρC z − z

(14)

c* = max(2 max(( p − 1)B , 3(β0C − α ) ),3 ρC )

(15)

So if we let

then the first and second conditions are proved. Next we prove that the third and fourth conditions are satisfied. b( z , t ) ≤ ( ( p − 1) B + β 0 C − α )(1 + z )

(16)

σ ( z , t ) = ρC (1 − z ) z ≤ ρC (1 + z )

(17)

c = max(( p − 1) B + β0C − α , ρC )

(18)

Then we put

So the third and fourth conditions are proved. As a result, the solution of Eq. (1) is existent and unique according to Theorem 1.

3 Strategy for AIDS Prevention and Control We can predict the proportion of the population infected with HIV against the total population in next few years with the application of SDE (2) under a certain intensity of the environmental disturbance. Compared with the deterministic one, it shows the trend of fluctuation. Since only a few of SDEs can have explicit solution [11-15], we can not obtain the explicit solution of the Eq. (2). So we simulate the proportion of the population infected with HIV against the total population by means of numerical solution. Furthermore, we analyses how to adjust β to reach a certain control under the condition of a certain intensity of the environmental disturbance. 3.1 Numerical Solution of the SDE The numerical solution methods include Euler method, Milsteins method, RungeKutta method and so on [16, 17]. We can get the approximation of the sample trajectory on the node one by one by a definite partition on the time interval with consideration of stochastic increment. An iterated expression can be obtained when we use Euler method. Z ( j , k ) = Z ( j − 1, k ) + F ( Z ( j − 1, K ))( t j − t j −1 ) + G ( Z ( j − 1, k ))[W ( j ) − W ( j − 1)]

(19)

j = 1,2,..., M ; k = 1,2,..., K

(20)

F ( Z ( j − 1, k )) = ( p − 1) BZ ( j − 1, k ) + ( β 0 C − α )[1 − Z ( j − 1, k )]Z ( j − 1, k )

(21)

Where

32

M. Xu, Y. Ding, and L. Hu

G ( Z ( j − 1, k )) = ρC (1 − Z ( j − 1, k )) Z ( j − 1, k )

(22)

j and k represent the time node and trajectory, respectively. M and K are the number of time node and sample trajectory, respectively. According to the expression above and the known condition, we can get Z (1, k ), Z (2, k ),..., Z ( M , k ) , the values of the proportion of the population infected with HIV against the total population on different time node as far as every given trajectory is concerned. Furthermore, the mean value of this process is, −

Z ( j) =

1 K

K

∑ Z ( j, k ),

1≤ j ≤ M

k =1

(23)

and the variance is, S 2 ( j) =

− − 1 K 1 K Z ( j, k ) 2 − [Z ( j )]2 , 1 ≤ j ≤ M ∑ (Z ( j, k ) − Z ( j))2 = K − 1 ∑ K − 1 k =1 k =1

(24)

3.2 Adjusting for the Control Index of the Infection Rate β We put t1 as the beginning year and t2 as the end year in the deterministic model. The proportion of the population infected with HIV against the total population Z (t ) on [t1 , t 2 ] is a curve with constant β. So the control can be obtained in the deterministic model if we put γ as the given control aim, where γ is a little more than Z (t 2 ) . But in fact, β is changed to a stochastic process because of the environmental disturbance. Correspondingly, Z (t ) is also changed to a stochastic process. We write it as Zt in order to avoid confusion in the deterministic model. Then we can get a series of curves, Z ( j ,1), Z ( j ,2),..., Z ( j , K ) , representing the proportion of the population infected with HIV against the total population on [t1, t2 ] . These curves show the character of probability of the process. In other word, Z t , the proportion of the population in2

fected with HIV against the total population on time t 2 , will have M results, Z ( M ,1), Z ( M ,2),..., Z (M , K ) corresponding to the end points of k trajectories. So we can compute the number of trajectories, H (β 0 ) which can reach the aim on time t2 (we put it as H (β 0 ) , for it only has something to do with β0 when ρ , the intensity of the disturbance is a certain value.), H (β 0 ) =

∑l

(25)

1≤l ≤ K

where l ∈ L = {l : Z ( M , l ) ≤ γ } . Then we can compute the probability of reaching the aim P , P=

H (β 0 ) K

(26)

Obviously, P equals to 100% in the deterministic model, while it may reach 50% or less than 50% under the condition of stochastic disturbance. In order to ensure P to

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics

33

reach a bigger value, we must decrease the average transmission rate β 0 , when the intensity of the disturbance is given.

4 Strategy for AIDS Prevention and Control Plenty of investigations demonstrate that the year of 1995 is a very important year in the history of AIDS epidemic and control. The number of people infected with HIV reached more than 15000 according to the statistics from health ministry of People’s Republic of China this year. We estimate the value of β based on the statistics of year from 1995 to 2015. And then we put the year of 1995 as the beginning point and use the model to forecast the trend of change of the proportion of the population infected with HIV against the total population in next 20 years. According to the statistics offered by Chinese center for disease control and prevention, the number of people infected with HIV was about 15000, and the number of total population was 1207.78 million in 1995 in China. So the proportion of the population infected with HIV against the total population was about 1.25 × 10 −5 in 1995. The vertical transmission probability from mother to child is 35%. The birth rate B is about 18.3556 × 10-3 according to the statistics from the same resource. Generally speaking, people will die one year after developing symptoms of AIDS. Therefore, α equals to 1, and C equals to 0.5256 [9,18]. Furthermore, we postulate that the parameters are also the same during this time period. Then we forecast the proportion of the population infected with HIV against the total population in next 10 years with Eq. (2) and we put M and K as 20 and 100 in Eq. (3), respectively. Firstly, we find the value of β to make the deterministic model satisfy the fact according to the number of people infected with HIV from year 1995 to 2005. By computing, we determine that β is 10. 6

x 10 -4

5

4

3

2

1

0

1995 1997

1999

2001

2003

2005

2007

2009

2011

2013

2015

Fig. 1. The proportion of the population infected with HIV against the total population under deterministic situation from 1995 to 2015 (Line 1 represents the simulation result, while line 2 represents the actual data)

34

M. Xu, Y. Ding, and L. Hu

In Fig. 1, line 1 represents the simulation result from the deterministic model, while line 2 represents the actual data from year 1995 to 2005. From it, we can see that the proportion of the population infected with HIV against the total population will reach about 6 ×10 −4 in 2015. While, we must use the stochastic model (2) for the transmission rate of the AIDS is disturbed by the environment noise. Fig. 2 demonstrates that the fluctuation of the proportion of the population infected with HIV against the total population from 1995 to 2015 when ρ is 0.5. We conclude that the proportion exceeds 6 × 10 −4 under some circumstances because of environmental disturbance. 8

x 10-4

7 6 5 4 3 2 1 0 1995

1997

1999

2001 2003

2005

2007

2009

2011

2013

2015

Fig. 2. The proportion of the population infected with HIV from 1995 to 2005, where ρ the number of sample trajectory is 10

= 0.5 and

We postulate that the control reaches below 6 × 10 −4 in the next 20 years. Then we can obtain the aim when β equals to 10 in the deterministic model. But the probability of obtaining such an aim is not as big as expected when the control index β equals 0

to 10. Table 1. The value of the end point of every trajectory when

Value(1× 10–3)

β0 β0

＝9 ＝10

k=1

k=2

k=3

k=4

k=5

k=6

β0 k=7

0.34 0.31 0.37 0.33 0.34 0.35 0.42 0.71 0.61 0.54 0.63 0.56 0.62 0.42

is 9 and 10, respectively

k=8 k=9

k=10

0.51 0.38 0.49 0.70 0.63 0.39

We compute that the probability equals to about 40% with the application of numerical method of the stochastic model. Consequently, we must decrease β 0 in order to ensure the proportion not to exceed 6×10－4 with the probability of 95%. We find that β 0 should be put as 9 when we use numerical simulation and skip from 9.9 to 9.8 to 9.7, etc.

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics

35

Fig. 3 demonstrates the fluctuation of the proportion of the population infected with HIV from 1995 to 2005 when ρ is the same, while the average transmission rate β equals to 9. Fig. 4 demonstrates the mean value of the proportion of the population 0

infected with HIV from 1995 to 2005 when

ρ equals to 0.5 and β 0 equals to 9 and

10, respectively.

6

x 10-4

5

4

3

2

1

0 1995

1997 1999

2001

2003

2005

2007

2009

2011

2013

2015

Fig. 3. The proportion of the population infected with HIV from 1995 to 2005, where and, β 0 is 9. The number of sample trajectory is 10.

6

ρ is to 0.5

x 10-4

β =10 β =9 0 0

5

4

3

2

1

0 1995

1997 1999

2001 2003

2005 2007 2009

2011

2013

2015

Fig. 4. The mean value of the proportion of the population infected with HIV from 1995 to 2005, where ρ is 0.5 and β 0 is 9, respectively −4 We find that the mean value will reach about 6 × 10 in 1995 when

it will fall below 4×10

－4.

β0

is 10, while

36

M. Xu, Y. Ding, and L. Hu

5 Conclusions Based on the formulation of a stochastic model of the proportion of the population infected with HIV against the total population, we first prove its solution is existent and unique. we analyses the effect of the control index of the transmission rate on the proportion of the population based on the estimation of the distribution of the proportion from 1995 to 2015 in China with the application of stochastic model. Acknowledgements. This work was supported in part by Program for New Century Excellent Talents in University from Ministry of Education of China (No. NCET-04415), the Cultivation Fund of the Key Scientific and Technical Innovation Project from Ministry of Education of China (No. 706024), International Science Cooperation Foundation of Shanghai (No. 061307041), and Specialized Research Fund for the Doctoral Program of Higher Education from Ministry of Education of China (No. 20060255006).

References 1. Haynatzka, V.R., Gani, J., Rachevn, S.T.: The spread of AIDS among interactive transmission. Biosystems 73(3), 157–161 (2004) 2. Castillo, C.C., et al.: The role of long incubation periodic in the dynamics of acquired immunodeficiency syndrome—Single population models. Math. Biol. 7, 373–398 (1989) 3. Blythe, S.P., Anderson, R.M.: Distributed incubation and infections periods in models of transmission dynamics of human immunodeficiency virus (HIV). IMA J. Math. Appl. Med. Biol. 1–19 (1988) 4. Blythe, S.P., Anderson, R.M.: Variable infectiousness in HIV transmission models. IMA. Math. Appl. Med. Biol. 5, 181–200 (1988) 5. Blythe, S.P., Anderson, R.M.: Heterogeneous sexual activity models of HIV transmission in male homosexual populations. IMA J. Math. Appl. Med. Biol. 5, 237–260 (1988) 6. Jacquez, J.A., Simon, C.P., Koopman, J.S.: Structured mixing: Heterogeneous mixing by the definition of activity groups. In: Castillo-Chavez, C. (ed.) Mathematical and Statistical Approaches to AIDS Epidemiology. Lecture Notes in Biomath., pp. 301–315. Springer, Heidelberg (1989) 7. Jacquez, J.A., Simon, C.P., Koopman, J.S.: The reproduction number in deterministic models of contagious disease. Comments Theor. Biol. 2, 159–209 (1988) 8. Greenhalgh, D., Doyle, M., Lewis, F.: A mathematical model of AIDS and condom use. IMA J. Math. Appl. Med. Biol. 18, 225–262 (2001) 9. Roberts, M.G., Saha, A.K.: The asymptotic behavior of a logistic epidemic model with stochastic disease transmission. Applied Mathematics Letters 12, 37–41 (1999) 10. Friedman, A.: Stochastic Differential Equations and Their Applications. Academic Press, New York (1976) 11. Okesendel, B.: Stochastic Differential Equations. Springer, Heidelberg (1985) 12. Kloeden, P.E., Platen, E.: Numerical solution of stochastic differential equations. Springer, Heidelberg (1992) 13. Saito, Y., Higham, T.: Stability analysis of numerical scheme for stochastic differential equations. SIAM, Numer. Anal. 33, 2254–2267 (1996)

A Stochastic Model for Prevention and Control of HIV/AIDS Transmission Dynamics

37

14. Higham, D.J.: Mean-square and asymptotic stability of the stochastic theta method. SIAM, Numer. Anal. 38(3), 753–769 (2000) 15. Ryden, T., Wiktorsson, M.: On the simulation of iterated Ito integrals. Stochastic Processes and their Applications 91(1), 116–151 (2001) 16. Burrage, K., Burrage, P., Mitsui, T.: Numerical solutions of stochastic differential equations-implementation and stability issues. Computational and Applied Mathematics 125, 171–182 (2000) 17. Slominski, L.: Euler’s approximations of solutions of SDEs with reflecting boundary. Stochastic Processes and their Applications 94(2), 317–337 (2001) 18. Liu, M.X., Zhou, Y.C.: An age-structured dynamic model of HIV. Journal of North China Institute of Technology 25(2), 25–30 (2004)

Simulation of Artificial Life of Bee’s Behaviors Bin Wu, Hongying Zhang, and Xia Ni School of Information Engineering, Southwest University of Science and Technology, Sichuan, Mianyang 621002, China {wbxn,zhanghongying,nixia}@swust.edu.cn

Abstract. Based on the ‘bottom-up’ programming approach in Artificial Life, Finite-State Machine is adopted to describe the behavior of the individual bee, and the behavior in different period was realized in this paper. As a result, the interaction of the individual bee each other, individual bee and virtual environment produce the Emergence of swarm’s collective behaviors. Finally, we apply the graphical interfaces to realize the simulation.

1 Introduction Artificial life is a charming topic in researching of complexity. Bionic system of life phenomenon is an important research topic in Artificial Life. All kinds of complicated life phenomenon are reappeared through setting up bionic system of life phenomenon, which can help us to understanding the essence of life. Biologic behaviors in the nature are rich and colorful, especially, biologists are more and more interested in the complicated colony behaviors of insects. Recently, many of the biologists have thought that the communications of all these insects were based on the theory of cooperation without communication. The basic idea of this theory is that each of the insects adjusts its behavior according to the changes of its surroundings without any leaders and obvious communications of each other. Based on this, this entire colony can complete very complicated tasks. In general, individual behavior is sample, blind and random, while colony behaviors are coherent, fluent and accordant [1]. According to the above theory, the behaviors of the swarm are simulated by artificial life in this paper. We can see that although the behavior of the individual bee is signal, it emerges complicated structures of the swarm under the interaction of the bees and their surroundings.

2 Behavior of Bees Although the behavior of the individual bee is sample and disorderly, the behavior of the swarm is complicated and orderly. A queen bee, many worker bees and a few drones make up of a swarm. Their conformations and functions are different. They divide the work and cooperate with each other. Queen bee is the only bee which can lay eggs in a swarm. Generally speaking, she can live from three to five year after successful amphimixis. In her life, queen bee K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 38–45, 2007. © Springer-Verlag Berlin Heidelberg 2007

Simulation of Artificial Life of Bee’s Behaviors

39

goes through three phases. In the first phase, she becomes eclosion imago from eggs. During this period, the main task of the queen bee is eating and flying around the honeycomb. After twenty-one days, the queen bee step into her second phase. She begins to fly out and copulate with drones. When the queen bee is entire impregnation, she steps into her third phase.In this period, the queen bee does nothing but give birth to children. Drone is a male individual in the swarm which is developed by unoosperm in the male comb. His only task is copulated with the queen bee to multiply their offspring. The number of the drone is about two percent of the total number of the swarm. The age of the drone is about two months. The drone goes through two phases in his life. In the first phase, it speeds the drone about twenty-four day to become eclosion imago from eggs. After twelve days, the drone is mature. During this period, the main task of the queen bee is eating and flying around the honeycomb. After that, the drone steps into his second phase. He copulated with the queen bee. Once the amphimixis is successful, the drone is died at once. Worker bee is a female individual which developed by the germ cell. With the age be increasing, the worker bee takes on all works in the comb. The number of the worker bee is very large, it occupies the ninety-eight percents of the total number of the swarm. The age of the worker bee is about four or six weeks in the summer and three or six months in the winter. It speeds the worker bee about twenty-one days to become eclosion imago from eggs. In this period, the main tasks of the worker bee are keeping the comb clean, feeding the baby bees, excreting the queen bee milk, building the comb, and guarding. After twenty-one days, the worker bee flies out their comb and pick the farina, which is the last work in her life. Although the bees live in a colony life, the swarms do not communicate with each other. The bees have the abilities of guarding their comb in order to resisting the outside swarms and insects. If the outside swarms thrust themselves in the comb, the bees which guard the comb must be fighted with them until the outside bees are running away or died. On the contrary, when all the bees are in the outside combs, such as in the flowers or in the watering place, different swarms are not hostile and interference to each other. If the queen bee flies in the outside swarm, she must be killed by the worker bees. On the contrary, if the drone flies in the outside swarm, the worker bees do not hurt him. It is may be show that the swarm wants to avoid propagating with close relative to make their offspring better[2].

3 The Basic Ideas 3.1 Adaptability Ashby, who is the one of the outrunner of the control theory, has understood in the 1950 that it is must be look on the controlled objects and their surroundings as a whole. When researching on a control system, we not only establish the model of the controlled objects, but also the model of objects and their surroundings. The viewpoint of Ashby is in fact that we study the behaviors of the organism from the point of adaptability. To research the adaptability in artificial intelligence is more and more important[1].

40

B. Wu, H. Zhang, and X. Ni

The individual, which make up of the artificial life system, is a self-determination body which has the abilities of self-study. It has the characteristics of selfadaptability, self-organization and evolution. All these characteristics represent action and counteractive between the self-determination body and surroundings. Through the interaction of the elf-determination body each other and the self-determination body and surroundings, the behaviors which are applied to different environments will be selected by training and studying local information. All these embody macroscopically the intelligent performance of emergence[3]. 3.2 Emergence Unlike designing car or robot, artificial life is not beforehand designed. The most interesting example of artificial life exhibits the behavior of ‘emergence’. The meaning of ‘emergence’ is that the interactions of many simple units produce remarkably total performance in the complicated environment. In artificial life, the exhibition of the system is not deduced by the genotype, where the genotype represents sample rules of the system running, the exhibition represents total emergence of the system. Computer speaking, ‘bottom-up’ programming approach allows emerging newly unpredictable phenomenon on the up level. This phenomenon is the key to life system[4]. In this paper, we apply ‘bottom-up’ programming approach to simulate the behaviors of the individual bee in order to emerge the total behaviors of the swarm. 3.3 Finite-State Machine Finite-state machine is a mechanism which made up of finite state. A state in it is called current state. Finite-state machine can receive an input which will result in occurring state transformation, that is, the state transforms from current state to output state. This state transformation is based on a state transformation function. After finishing the state transformation, the output state is become current state. In the state transformation figure of finite-state machine, vertex is represented as state, arrowhead is represented as transfer because it describes FSM transforming from a state to another state. Label text of the transfer is divided into two parts by a bias, which the left part is the name of touching off transfer event and the right one is the name of behavior after touched off transfer. In this paper, we apply finite-state machine to describe the behavior of bee individual in the simulation.

4 Simulations In this section, we will apply the graphical interfaces to realize the simulation, which can visually show the behaviors of virtual swarm in virtual environment at different phases. 4.1 Finite-State Machine The object of our simulation is to simulate the behavior of bees with the method of artificial life. So we not only establish the model of bees, but also place these bee individuals into a dynamic virtual environment. When the environment changes, the

Simulation of Artificial Life of Bee’s Behaviors

41

individuals produce different behaviors vary from different environments. This phenomenon shows that the individuals are adapted to the environment. The environment and the bees can be looked as two objects which oppose and contact with each other. Based on these, we can establish a simulation system through a layered structure, this can be shown in Fig.1. Main program Module Management

Resource Module

Comb Module

Bee Module

flower

Comb

Bee

Fig. 1. Layered structure of the program

4.2 Realization of the Program The best method of implementing the synthesis of artificial life is the ‘bottom-up’ programming approach. The basic idea of this method is that we define a lot of small units in the bottom and several simple rules which relative to their inner and local interaction, then coherent collectivity behavior is produced from the interaction. This behavior does not beforehand organize based on the special rules. The aim of the simulation in our paper is not aim at implementing the behaviors of the swarm but implementing the individual behavior of bee. Although the individual behavior of the artificial bee is single, it produces complicated behaviors of the swarm under the interactions of many individual bees. (1)Establishing the Virtual Environment As we know, the behaviors of the bees are relative to the changes of the seasons or the temperature. So the first task of our simulation is to realize the changes of the seasons in the virtual environment. In general, the life of the swarm in one year can be divided into some periods[2]. According to this, we can divide the seasons as follows: Early spring: February. The queen bee begins to lay eggs in this period. If the queen bee lays eggs, the temperature must be beyond 10 . Based on this, we set the temperature of February 10 ~15 .

℃ ℃

℃

Spring: from March to May.

℃ ℃

Summer: from June to August. We set the temperature of March to July 15 ~35 and August 35 ~25 .

℃ ℃

Autumn: from September to October. We set the temperature of this period 25 ~15 .

℃ ℃

℃ ℃

Late autumn: November. We set the temperature of this period 15 ~5 Winter: from December to January in the following year.

42

B. Wu, H. Zhang, and X. Ni

Fig. 2. The relation between season and temperature

Fig.2 describes the temperature distribution in our simulation. Based on this figure, we can establish a simple function of the temperature change. In the virtual environment, we also simulate the food of the virtual bees, that is, establish the flower. In our simulation, we set some parameters to describe these flowers, such as the location of the flower (X, Y), florescence(BloomingP), the number of farina (UnitFPN), color and size. (2)The Behavior of Bees In our simulation, we describe the behaviors of the queen bee, the worker bee and the drone through finite-state machine. 4.3 The Interface of the Simulation The main interface of the simulation program is shown as Fig.3. The simulation program is divided into five parts: the control of original parameter, data statistic and

Fig. 3. The main interface of the simulation program

Simulation of Artificial Life of Bee’s Behaviors

43

analysis, establishing new comb, motion comb and motion monitor. The user can set original parameters through parameter control panel, such as season, month, the performance of the queen bee, the number of swarm in the system, and the total number of the bees. We set different original parameters in order to observe the affects of these parameters. During the program running, if you put down the button of “establish the new swarm”, you can add the swarm arbitrarily. The aim of this button is that you can observe the affects when the new swarm adds into the virtual environment. 4.4 The Results of the Simulation In this section, we give an experiment of moving bee colony. Firstly, we set the initial values which are as follows: Season: Spring; Month: April; Queen bee is impregnation; The number of the comb: one; The number of the bee: eighty The number of the farina in comb: ten thousands The number of the flower: one hundred; The content of the farina in one flower: one thoudred The position of the comb is (5496, 4577) which is established randomly by the system. Fig.4 (a) gives the emluator running statement after eleven days. From this picture, we can see that the number of flower is very small around the comb. This can affect the number of the farina in the comb. After that, we move the comb to a new position (1000, 5000). The picture after moving the comb is as shown in Fig.4(b). Just on this, the swarming is happened. Fig.4(c) is the statement picture of the original swarming. After producing the new queen bee, the old queen bee and some worker bees leave the comb to build a new comb which new position is at (6310,4769). Fig.5 is the increasing picture of the farina in the comb after twenty days. In this picture,

(a) Fig. 4. The progress picture of emluator running

44

B. Wu, H. Zhang, and X. Ni

(b)

(c) Fig. 4. (continued)

Fig. 5. The schedule graph of pollen in hive when emluator run for 20 days

there are two curves, C1 and C2. C1 is represent as the increasing curve of the farina of the old comb, while C2 is represent as the increasing curve of the farina of the new comb after swarming.

Simulation of Artificial Life of Bee’s Behaviors

45

5 Conclusions In this paper, we apply the graphical interfaces to realize the behaviors of the swarm in the virtual environment. The proposed program emerged the collectivity behaviors of the whole swarm through implementing the individual behaviors of the bees. That is not simple simulation but is virtual swarm which has artificial life. In the visual interface, we can see that new swarms produce and old swarms die. All of these are the results of the interactions of individual bees each other. If the program runs forever, we must be think that it is an ecosystem of the bees.

References 1. Zhou, D.Y., Dai, R.W.: Adaptability Behavior and Simulation. Journal of System Simulation 6, 578–583 (2000) 2. Sun, Y.X.: The Bees. Chinese curatorial Press, Beijing (2001) 3. Wu, J.B.: Emergent Behaviors of Autonomous Agents. Micromachine Development 6, 6–8 (1997) 4. Li, J.H.: Artificial Life: Explore New Form of Life. Research on Nature Dialectic 7, 1–5 (2001)

Hybrid Processing and Time-Frequency Analysis of ECG Signal Ping Zhang, Chengyuan Tu, Xiaoyang Li, and Yanjun Zeng* Beijing University of Technology, P.R. China 100022 [email protected]

Abstract. A new simple approach basing on the histogram and genetic algorithm(GA) to efficiently detect QRS-T complexes of the ECG curve is described, so as to easily get the P-wave (when AF does not happen) or the f-wave (when AF happens). By means of signal processing techniques such as the power spectrum function the auto-correlation function and cross-correlation function, two kinds of ECG signal when AF does or does not happen were successively analyzed, showing the evident differences between them.

，

1 Introduction The ECG (electrocardiogram) is a standard clinical examination method and as such a valuable tool in the diagnosis of cardiac disease. A key element in the automatic analysis of the ECG signal is the detection of QRS complexes [1]. Once the QRS complexex is detected, one can analyse the patient’s HRV (heart rate variability) along with a number of other parameters which are of diagnostic significance. ECG processing strategies which have been applied to detect QRS complexes are based on various methods such as band-pass filtering, adaptive filtering, nonlinear adaptive filtering, etc. [1-2]. Owing to the complexity and considerable computational expense associated with these filtering methods, their applicability (particularly, in clinical practise) is restricted. Recently, along with the development of analysis techniques by means of wavelets, a number of methods based on wavelet transform for detecting QRS complexes have been put forward[1,3-9]. These methods demand, however, the availability of a suitable mother-wave reflecting the properties of the signal to be analysed so that the important local features of this signal in the time domain as well as in the frequency domain are all preserved after a wavelet transform. Moreover, these methods rely on the definition of a suitable scale so that the signal abruptness may be preserved. If these conditions are not fulfilled, it is quite possible that there appears an omission (a QRS complex is not detected) or a mistake (something detected is not a QRS complex). In order to find a better solution to these problems, a simple and effective method based on histogram and the improved GA to search and detect QRS complexes has been developed by the authors. And as an example of its novel application, the P-wave can be extracted easily and efficiently from ECG curve whenever AF is absent or the f-wave can be analogously extracted when AF occurs. *

Corresponding author.

K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 46–57, 2007. © Springer-Verlag Berlin Heidelberg 2007

Hybrid Processing and Time-Frequency Analysis of ECG Signal

47

2 Detection of QRS Complexes from the ECG Curve 2.1 Fast Search for QRS Complexes The ECG plot appeared on a computer screen may be considered as a two-dimensional picture consisting of elements, organised in rows and columns, with value 0 (corresponds to white) and 1 (corresponds to black). The parameter QOP (quantity of pixels) is defined as the total number of picture elements per column, from which a histogram over the entire plot is obtained. The information embodied in the histogram can be used for a rapid detection of QRS complexes The ECG signal appears as a curve in the t-(time-) domain is shown in Fig. 1. For the further analys, we define several technical terms: X ECG picture matrix (binary: 0 corresponds to white, 1 corresponds to black)

Fig. 1. Typical ECG waveform

x(i, j) the related element of the matrix X d(j) QOP(the total number of picture elements of the related column) i, j the row coodinate and column coordinate of the element x(i, j) thereby

d ( j ) = ∑ x(i, j ) i

(1)

D={d(j)} where the set D = { d(j)} appears as a vector. The range of i and that of j are determined by the size and resolution of the ECG plot considered. For a typical ECG frame on a computer screen, we generally let: i = 1, 2, . . . , 600; j = 1, 2, . . . , 1200 Figure 2 shows the picture corresponding to the set D, namely the histogram, in which the abscissa is time, and the ordinate is QOP. Figure 3 shows the histogram after thresholding, i.e., the picture corresponding to a new set Dth obtained by means of selecting a suitable threshold value d0 (see Eq. (2)).

48

P. Zhang et al.

Fig. 2. The histogram corresponding to Fig. 1

Fig. 3. The threshold-form of Fig. 2

Dth={dth(j)} dth(j)= d(j), when d(j)≥d0 dth(j)=0 , when d(j)fm ), we reject this unfavorable operation, and let the population state return to its state previous to that operation. reconstruction Rc: while a crossover or a mutation fails many times, let the population be reconstructed. recording the better Rb: whenever a crossover or a mutation succeeds( as fm(t) bmiHeader); pmt.SetSubtype(& subtype); //set subtype pmt.SetFormat((BYTE∗)inFormat, inLength); //set video format type (2) After adding in the multicast group, a double-buffer queue is constructed. The queue constitutes by two single chained list, writing data queue and reading data queue:

PData

PData

PData

PData

Head

PData

PData Tail

PData

PData

PData

PData

After finished a frame data in write chain, this data block is delete from head of writing queue and send it to the tail of reading queue. After run out of this data block in reading chain, it is deleted from reading queue and send it to the tail of writing queue. 6.3

Video Stream Render

When the Filters are connected successfully, FilterGraph display the received Video stream under the control of Filter Graph Manager. IMediaControl ∗ pMC = NULL; IMediaEventEx ∗ pME = NULL; // media control interface pGraph − > Queryinterface ( IID-ImediaControl, (void ∗ ∗ ) & pMC); //media event interface pGraph − > Queryinterface ( IID-ImediaEventEx, (void ∗ ∗ ) & pME); pMC − > Run (); //run Filter chain

7

Experimental Result

Experiment is implemented in Huaihe Hospital of Henan University. The specification of digital subtraction angiography equipment of Huaihe Hospital is GE LCV+. Digital subtraction angiography video stream are captured by an advanced image capture board. The specification of image capture board is OKM30A. MPEG-4 and DirectShow framework technique are used for data’s compression and transmission. Data of DSA video stream are transported easily from server sender Filter to client receiver Filter according to TCP and UDP protocols. Table 1 shows the experimental environment and equipments of this scheme.

Internet Living Broadcast of Medical Video Stream

127

Experimental results show that the image is clear in 100M LAN, and the delay is less than one second. High performances of definition and real-time of the DSA video are both achieved in this scheme. Memory usage in server is 10M. Memory usage is 6M and 21M in Client(render model VMR) and Client(render model Normal) respectively. The frame frequency is always 29 frames/second. Experimental results under above environment are shown as Table 2. Table 1. Experimental environment Equipments Specification DSA equipment GE LCV+ Image capture board OK-M30A,1024 × 1024, 10bit MPEG-4 Codec/Decodecer Xvid Network 100M Lan Server CPU PV 2.0G, 512M memory Client CPU XP 1700+, 512M memory, 32M video card Table 2. Experimental results

Server Client(render model VMR) Client(render model Normal)

Memory usage Frames/second 10M 29 6M 29 21M 29

VMR (Video Mixing Render) is used firstly to render in client according to the data type. In addition, if the CPU occupancy rate is high in client, the packet dropping rate of data received from Internet will rise, and will result the Mosaic when the video render.If close some other programs, quality of video is all right.

8

Conclusion

In this scheme, Digital Subtraction Angiography (DSA) video stream are captured by an advanced image capture board. MPEG-4, and DirectShow framework technique are used for data’s compression and transmission. Data of DSA video stream can be transported easily from server sender Filter to client receiver Filter according to TCP and UDP protocols. Experimental results show that the image is clear in 100M LAN, and the delay is less than one second. High performances of definition and real-time of the DSA video are both achieved in this scheme. And by the image processing function, we can also process images received from Internet. Experimental results show that the living broadcast of interventional operation/surgery via Internet is a feasible scheme.

128

S. Li, B. Li, and F. Zhang

References 1. Perednia, D.A., Allen, A.: Telemedicine technology and clinical applications. Journal of the American Medical Association 273(6), 483–488 (1995) 2. Gandsas, A., Altrudi, R., Pleatman, M., Silva, Y.: Live interactive broadcast of laparoscopic surgery via the Internet. Current Surgery 60(2), 126–129 (2003) 3. Tao, Y., Miao, J.: Workstation scheme and implementation for a medical imaging information system. Chinese Medical Journal 116(5), 654–657 (2003) 4. Huang, Z., Zhuang, T.: Evolution of DICOM Standard and its Latest Changes. Chinese Journal of Medical Instrumentation 28(3), 203–207 (2004) 5. Jose, R., Pablo, G., Miguel, S.: A Compression and Transmission Scheme of Computer Tomography Images for Telemedicine Based on JPEG2000. Telemedicine Journal and e-Health 10, 40–44 (2004) 6. Ramakrishnan, B., Sriraam, N.: Internet transmission of DICOM images with effective low bandwidth utilization. Digital Signal Processing 16(6), 825–831 (2006) 7. Brody, W.: Digital Subtraction Angiography. IEEE Transactions on Nuclear Science 29(3), 1176–1180 (1982) 8. Prasad, R., Ramkishor, K.: Implementation of MPEG-4 Video Encoder on RISC Core. IEEE Transactions on Consumer Electronics 49(2), 1–6 (2003) 9. Dasu, A., Panchanathan, S.: A Survey of Media Processing Approaches. IEEE Transactions on Circuit and System for Video Technology 12(8), 1–13 (2002) 10. Lu, Q.: DirectShow Development Guidebook. Tsinghua University Press, Beijing (2003)

Predicting Syndrome by NEI Specifications: A Comparison of Five Data Mining Algorithms in Coronary Heart Disease Jianxin Chen1 , Guangcheng Xi1 , Yanwei Xing2 , Jing Chen1 , and Jie Wang2 1

2

Key Laboratory of Complex Systems and Intelligence Science Institute of Automation, Chinese Academy of Sciences 100080, Beijing, China {jianxin.chen,guangcheng.xi}@ia.ac.cn GuangAnMen Hospital, Chinese Academy of Chinese Medical Science 100053, Beijing, China

Abstract. Nowadays, most Chinese take a way of integration of TCM and western medicine to heal CHD. However, the relation between them is rarely studied. In this paper, we carry out a clinical epidemiology to collect 102 cases, each of which is a CHD instance confirmed by Coronary Artery Angiography. Moreover, each case is diagnosed by TCM experts as what syndrome and the corresponding nine NEI specifications are measured.We want to explore whether there exist relation between syndrome and NEI specifications. Therefore, we employ five distinct kinds of data mining algorithms: Bayesian model; Neural Network; Support vector machine ,Decision trees and logistic regression to perform prediction task and compare their performances. The results indicated that SVM is the best identifier with 90.5% accuracy on the holdout samples. The next is neural network with 88.9% accuracy, higher than Bayesian model with 82.2% counterpart. The decision tree is less worst,77.9%, logistic regression models performs the worst, only 73.9%. We concluded that there do exist relation between syndrome and western medicine and SVM is the best model for predicting syndrome by NEI specifications.

1

Introduction

Coronary heart disease (CHD)is a serious disease causing more than 1 million Chinese to death each year. [1].In China, most people take a way of integration of TCM and western medicine to heal CHD.The following is a brief introduction to TCM. 1.1

TCM

TCM has been always regarded as a key component in five thousand years of Chinese civilization history. It has a history of more than 3000 years, while 1000 years are spending on healing CHD, so it piles up extensive experience. TCM, K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 129–135, 2007. c Springer-Verlag Berlin Heidelberg 2007 

130

J. Chen et al.

whose core is syndrome, or ‘Zheng’ in Chinese, is on her way to modernization, aiming to be accepted, like western medicine, as sciences [2],[3]. The kernel of TCM is syndrome. Every herbal is prescribed in accord with syndromes. However, till now, the relation between syndrome and physical and chemical specifications of western medicine is rarely explored. Furthermore, during animal experiments in laboratories, disliking human, animals can not be felt pulse, so determination of syndrome in animals is significantly difficult. However, the bloods of human and animals are more easily to obtained, therefore, we can explore the relation between syndrome and some blood specifications. Here, we choose Neuro-Endocrine-Immune(NEI) specifications. The following is the backgroud of NEI. 1.2

NEI System

In modern Western medicine (WM), NEI system acts as a pivot in modulating host homeostasis and naturally optimising health through complex communications among chemical messengers (CMs), including hormones, cytokines and neurotransmitters [4]. If we consider CMs as the biochemical ingredients of the NEI system, then those genes that (directly or indirectly) encode these CMs can be considered as the genic ingredients of the NEI system. Here, our goal is using the information of NEI specifications to predicting whether a patient is a specific syndrome. We employ five kinds of classical data mining methods to perfom the task and we compare the methods to search the best one. Therefore, the problem is classification problem. 1.3

Data Mining Algorithms

Under the background of supervised classification problem, data mining algorithms mainly comprise of five broadly used kinds: Bayesian method, neural network, support vector machine, decision trees and logistic regression. Each kind is developed quickly and usually combines with each other to solve some hard problems [5]. Bayesian network (BN) is chosen from Bayesian method to perform classification here. Radial Basis Function Network is selected from neural networks for its higher performance in doing classification than other algorithm, such as recurrent neural network. We used a well-known support vector machine algorithm, Platt’s SMO Algorithm [6], as a representative of SVM classification since it can process both categorical and numerical variables. For decision tree kind, Quinlan’s C4.5 algorithm [7] is employed for performing tasks here. Logistic regression is a classical method and it is used here to perform the task in this paper.

2

Biological Variables and Clinical Data Collection

The 102 patients included in the survey are incoming patients of AnZhen Hosptial in Beijing from June 2005 to April 2006. Each patient is diagnosed by western

Predicting Syndrome by NEI Specifications

131

medicine experts as CHD by Coronary Artery Angiography, meanwhile, each patient is diagonosed by experts as what syndrome.Blood samples of all patients were collected after a 12-hour overnight fast before cardiovascular procedures. Totally, 9 NEI specifications are measured from the blood samples. There are Mpo1, Wbc1, Tni2, Tnf, Et, Il6, Il8, No and Hicrp. Additionally, the basic information of each patients,such as names,ages and so on are also recorded, but this part of data is not included in the process of data mining.

3

Prediction Models

New data always needs already existed algorithms to test its performance. We employed four types of classification algorithms: Bayesian model, neural networks, SVM and decision trees. These models were jigged for inclusion in this research due to their popularity in the recently published documents. The following is a brief introduction to the four classification algorithms and the parameter setting of each model. 3.1

Bayesian Network

A Bayesian network (BN) is a graphical model that encodes probabilistic relationships among attributes of interest. Several advances have been made to ameliorate Bayesian network to fit all kinds of realistic problems [8]. We select stimulated annealing as method for searching network structures. Estimator is BayesNetEstimator, which is the base class for estimating the conditional probability tables of a Bayesian network once the structure has been learned. 3.2

Radial Basis Function Network

As shown in Fig. 1, RBF network has two layers, not counting the input layer, and differs from a multilayer perceptron in the way that the hidden units perform computations.Each hidden unit essentially represents a particular point in input space, and its output, or activation, for a given instance depends on the distance between its point and the instance-which is just another point. Intuitively, the closer these two points, the stronger the activation. This is achieved by using a nonlinear transformation function to convert the distance into a similarity measure. A bell-shaped Gaussian activation function, whose width may be different for each hidden unit, is commonly used for this purpose. The hidden units are called RBFs because the points in instance space for which a given hidden unit produces the same activation form a hypersphere or hyperellipsoid. (In a multilayer perceptron, this is a hyperplane.) 3.3

Support Vector Machine

The SVM is a state-of-the-art maximum margin classification algorithm rooted in statistical learning theory [11],[12]. SVM performs classification tasks by maximizing the margin separating both classes while minimizing the classification errors. We used sequential minimal optimization algorithm to train the SVM here, as shown in As shown in Fig. 2

132

J. Chen et al.

Fig. 1. The topology of MLP network

Fig. 2. A illustration of SVM

Predicting Syndrome by NEI Specifications

3.4

133

Decision Trees C4.5

As the name implies, this algorithm recursively separates observations in branches to construct a tree for the purpose of improving the prediction accuracy. In doing so, they use mathematical algorithms information gain to identify a variable and corresponding threshold for the variable that splits the input observation into two or more subgroups. This step is repeated at each leaf node until the complete tree is constructed. This step is repeated at each leaf node until the complete tree is constructed Confidence factor is set as 0.01. The minimum number of instances per leaf is 2. 3.5

Logistic Regression

Logistic regression is a generalization of linear regression [13]. It is used primarily for predicting binary or multi-class dependent variables. Because the response variable is discrete, it cannot be modeled directly by linear regression. Therefore, rather than predicting point estimate of the event itself, it builds the model to predict the odds of its occurrence. In a two-class problem, odds greater than50% would mean that the case is assigned to the class designated as 1 and 0 otherwise. While logistic regression is a very powerful modeling tool, it assumes that the response variable (the log odds, not the event itself) is linear in the coefficients of the predictor variables.

4 4.1

Performance Evaluation and Results Performance Measures

We employed three hackneyed performance measures: accuracy, sensitivity and specificity. A distinguished confusion matrix is obtained to calculate the three measures. Confusion matrix is a matrix representation of the classification results. the upper left cell denotes the number of samples classifies as true while they were true (i.e., TP), and lower right cell denotes the number of samples classified as false while they were actually false (i.e., TF). The other two cells (lower left cell and upper right cell) denote the number of samples misclassified. Specifically, the lower left cell denoting the number of samples classified as false while they actually were true (i.e., FN), and the upper right cell denoting the number of samples classified as true while they actually were false (i.e., FP).Once the confusion matrixes were constructed, the accuracy, sensitivity and specificity are easily calculated as: sensitivity = TP/(TP + FN); specificity = TN/(TN + FP). Accuracy = (TP + TN)/(TP + FP + TN + FN);10-fold cross validation is used here to minimize the bias produced by random sampling of the training and test data samples. Extensive tests on numerous data sets, with different learning strategies, have shown that 10 is about the right number of folds to get the best estimate of error, and there is also some theoretical evidence that backs this up [9],[10].

134

4.2

J. Chen et al.

Results

Every model was evaluated based on the three measures discussed above (classification accuracy, sensitivity and specificity). The results were achieved using average value of 10 fold cross-validation for each algorithm. As shown in Figure 3, we found that the Bayesian model (BN) achieved classification accuracy of 82.2% with a sensitivity of 81.1% and a specificity of 85.5%. The neural network achieved classification accuracy of 88.9% with a sensitivity of 88.9% and a specificity of 88.8%. The decision trees (C4.5) achieved a classification accuracy of 77.9% with a sensitivity of 76.5% and a specificity of 81.9%. The logistic regression achieved a classification accuracy of 73.9% with a sensitivity of 84.6% and a specificity of 70.3%. However, SVM preformed the best of the five models evaluated. It achieved a classification accuracy of 90.5% with a sensitivity of 92% and a specificity of 90%.

Fig. 3. The performance of five data mining algorithms as classification problem

5

Conclusion

In this paper, we employ five kinds of popular data mining models to perform classification task in identifying syndrome by NEI specifications in CHD. The data was recruited from clinics with whole 102 cases. We used 10-fold cross validation to compute confusion matrix of each model and then calculate the three performance measures-sensitivity, specificity and accuracy to evaluate five kinds

Predicting Syndrome by NEI Specifications

135

of models. We found that the Bayesian model (BN) achieved classification accuracy of82.2% with a sensitivity of 81.1% and a specificity of 85.5%. The neural network achieved classification accuracy of 88.9% with a sensitivity of 88.9% and a specificity of 88.8%. The decision trees (C4.5) achieved a classification accuracy of 77.9% with a sensitivity of 76.5% and a specificity of 81.9%. The logistic regression achieved a classification accuracy of 70.3% with a sensitivity of 84.6% and a specificity of 73.9%. However, SVM preformed the best of the five models evaluated. It achieved a classification accuracy of 90.5% with a sensitivity of 92% and a specificity of 90%. We concluded that syndrome does exist strong relation with NEI specifications and our results shown that SVM will provide a better insight to predicting syndrome by NEI specifications. Acknowledgments. The work has been supported by 973 Program under grant No. (2003CB517106 and 2003CB517103) and NSFC Projects under Grant No. 60621001, China.

References 1. World Health Organization.: World Health statistics Annual. Geneva, Switzerland, World Health Organization (2006) 2. Normile, D.: The new face of Traditional Chinese Medicine. Science 299, 188–190 (2003) 3. Xue, T.H., Roy, R.: Studying Traditional Chinese Medicine. Science 300, 740–741 (2003) 4. Roth, J., LeRoith, D., et al.: The evolutionary origins of hormones, neurotransmitters, and other extracellular chemical messengers: implications for mammalian biology. The New England Journal of Medicine 306, 523–527 (2006) 5. Brudzewski, K., Osowski, S., Markiewicz, T.: Classification of milk by means of an electronic nose and SVM neural network. Sensors and Actuators B 98, 291–298 (2004) 6. Keerth, S., Shevade, K., et al.: Improvements to Platt’s SMO Algorithm for SVM classifier design. Neural Computation 13, 637–649 (2001) 7. Quinlan, R.: C4.5: Programs for Machine Learning. Morgan Kaufmann, San Mateo, CA (1993) 8. Huang, C.L., Shih, H.C., Chao, C.Y.: Semantic Analysis of Soccer Video Using Dynamic Bayesian Network. IEEE Transactions on Multimedia 8, 749–760 (2006) 9. Witten, I.H., FrankMichalewicz, E.Z.: Data Mining: Practical machine learning tools and techniques, 2nd edn. Morgan Kaufmann, San Francisco (2005) 10. Delen, D., Walker, G., Kadam, A.: Predicting breast cancer survivability. Artif. Intell. Med. 34, 113–127 (2005) 11. Vapnik, K.: Statistical learning theory. Wiley, New York (1998) 12. Graf, A., Wichmann, F., Bulthoff, H., et al.: Classification of faces in man and machine. Neural Computation 18, 143–165 (2006) 13. Hastie, T., Tibshirani, R., Friedman, J.: The elements of statistical learning. Springer, New York (2001)

Application of Image Processing and Finite Element Analysis in Bionic Scaffolds’ Design Optimizing and Fabrication Liulan Lin, Huicun Zhang, Yuan Yao, Aili Tong, Qingxi Hu, and Minglun Fang Rapid Manufacturing Engineering Center, Shanghai University, P.O. Box 113, 99 Shangda Road, Shanghai, 200444, China [email protected]

Abstract. Design optimizing is the key step in obtaining bionic scaffolds with proper shape and inner microstructure, which are two critical parameters for bionic scaffolds in Tissue Engineering. In this paper, the application of image processing and finite element analysis in the design optimizing of bionic scaffold’s shape and inner microstructure were studied respectively. The bionic scaffold’s shape was obtained through Mimics’ image processing and 3D reconstruction technologies. Finite element analysis (FEA) was used in evaluating the mechanical properties of scaffold’s structure models with different macropores shape and distribution to obtain the optimized parameters. Three groups of bioceramic scaffolds samples were fabricated through an indirect method combining stereolithography (SLA) and slurry casting, and then mechanical experiments were tested. The change trendy of the compressive strength obtained through mechanical experiments was consistent with the FEA results basically so the significance of FEA in bionic scaffolds’ design optimizing was proved.

1 Introduction In Tissue Engineering, temporary 3D bionic scaffolds are essential to guide cell proliferation and to maintain native phenotypes in regenerating biologic tissues or organs [1]. The shape and inner microstructure are the two critical properties of bionic scaffolds for repairing defective bone. Bionic scaffolds should have the same shape as restoration for repairing the defective bone, so the scaffolds could be placed well in body and guide the neonatal bone’s growth correctly. To satisfy tissue engineering’s requirement, bionic bone scaffolds must have exact shape with the defects, polygradient porous configuration with characteristics and properties such as porosity, surface area to volume ratio, pore size, pore interconnectivity, shape (or overall geometry), structural strength and biocompatibility. These characteristics and properties are often considered to be critical factors in their designing and fabrication [2]. Design optimizing is the key step in obtaining bionic scaffolds with proper shape and inner microstructure. Traditional methods of scaffold fabrication include fiber bonding, solvent casting and particulate leaching [3], membrane lamination, melt molding, gas foaming, cryogenic induced phase separation [4],[5] and so on. However, all of these techniques are mainly based on manual work and lack of K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 136–145, 2007. © Springer-Verlag Berlin Heidelberg 2007

Application of Image Processing and Finite Element Analysis

137

corresponding designing process, so extra procedure was needed to obtain suitable shape and the microstructure wasn’t able to be controlled well. These traditional also have many disadvantages such as long fabrication periods, poor repeatability and insufficient connectivity of pores [6]. To overcome the limitations of these conventional techniques, automated computer controlled fabrication techniques, such as rapid prototyping (RP), are being explored. Based on layer by layer manufacturing process, parts with complex shape or structure could be produced through RP technologies easily and rapidly. Several kinds of RP technologies such as stereolithography (SLA) [7],[8], selective laser sintering (SLS) [9],[10], fused deposition modeling (FDM) [11],[12], three-dimensional printing (TDP or 3DP) [13],[14] and so on, have been applied widely in fabricating bionic scaffolds for tissue engineering and achieved some progress. Using RP technologies in bionic scaffolds preparation could fully performs the significance of designing and improves the bionic scaffolds’ properties. In this paper, the application of image processing and finite element analysis in the design optimizing of bionic scaffold’s shape and inner microstructure was studied respectively. The bionic scaffold’s shape was obtained through Mimics’ image processing and 3D reconstruction technologies. The inner microstructure of bionic scaffolds should be polygradient porous configuration with macro-pores and micro-pores. The macro-pores’ size and distribution was designed by CAD software and could be manufactured through RP technologies, while the micro-pores were caused by burning off of the pore-forming agent forming the spacing. So the design optimizing means finding the optimized parameters for size and distribution of macro-pores. Finite element analysis (FEA) was used in evaluating the mechanical properties of the optimized bionic scaffold’s microstructure models and models without design optimizing. Several groups of bioceramic scaffolds samples were fabricated through an indirect method combining stereolithography (SLA) and slurry casting and mechanical experiments were done to validate the FEA results.

2 Designing and Analyzing Process 2.1 Image Processing and 3D Reconstruction The patient’s CT date of defective skull were imported into Mimics 9.11 (The Materialise Group, Leuven, Belgium) and the 3D STL model of the skull was obtained through image processing and 3D reconstruction technologies. Then the restoration’s 3D STL model of defect was constructed by the symmetrically repairing operation in Mimics. The shape of the restoration was just the shape of the bionic bone scaffold to prepare. 2.2 Scaffold’s Structure Model Design and Analysis The macro-pores of bionic scaffold’s inner microstructure should be 3D connective gridding structure with proper size to assure the scaffold’s connectivity and be suitable for cell’s growth and proliferation. There were lots of scaffold models with different pore shape, size and distribution. Considering the preparing technology, four kinds of structure model with different pore shape and distribution were created in UG NX3.0 (UGS PLM Solutions, Plano, TX, USA) to make the analysis and contrast.

138

L. Lin et al.

As shown in Fig.1, model A has cylindrical macro-pores with diameter of 0.80 mm and distance between adjacent pores of 2.0 mm. Model C has square macro-pores with edge length of 0.71 mm and distance between adjacent pores of 2.0 mm to have the equal porosity with model A. The macro-pore’s distribution of model B and model D are different of model A and model C. The three coordinate axis pores were intersected in one point in model A and model C, while in model B and model D the X-axis pores and Y-axis pores were not in the same horizontal plane.

Fig. 1. The structure models of bionic scaffolds

The compress stimulation of the microstructure models were solved in finite element analysis software Ansys. The change of mechanical strength between the models before and after design optimizing was contrast and analyzed. According formulas of the Mechanics of Materials, the ratio between loads and strain is a constant to the same material. The four models were set with same element type and material attributes and the deformation of each model under the same load (compressing or bending) was contrasted. Based on the max total strain density value of each model, the compressive strength or bending strength was calculated. 2.3 Mechanical Experiment Three groups of bioceramic scaffolds of these four microstructure model were fabricated through an indirect method combining stereolithography (SLA) and slurry casting technologies. Compression tests were conducted with an INSTRON 5542 material testing system with Merlin™ software using a 50N load-cell (Norwood, MA, USA). Compression tests with real materials were done to validate the effect of design optimizing and the finite element analysis’ accuracy.

Application of Image Processing and Finite Element Analysis

139

3 Results 3.1 Result of Defective Bone Repairing As shown in Fig.2, the CT scanning images of patient’s skull were imported into Mimics 9.11. After selecting suitable threshold, the 3D model of the skull was reconstructed exported to STL model (Fig.3 A). Threshold selecting was very important for the reconstructed 3D model’s quality. In this paper, the bone model was defined through masks with threshold between 226 and 1738. The defective skull was repaired by the symmetrically repairing operation in Mimics. The shape of the restoration was just the shape of the bionic bone scaffold to prepare. Fig.3 B showed the bionic scaffold model with macro-pores, which was created through Boolean operation in Magics X (The Materialise Group, Leuven, Belgium), which is a software especially for processing STL files.

Fig. 2. The structure models of bionic scaffolds

Fig. 3. Result of repairing of the defective skull (A) and the model of restoration and bionic scaffold with macropores (B)

140

L. Lin et al.

3.2 Result of Finite Element Analysis Bionic scaffolds are full with porous microstructure, so it could be assumed that this porous microstructure is a homogeneous, isotropic material. Bio-ceramic is a brittle material and performs flexibility property before break (the max strain value reaches 0.03). The Young’s Module of scaffolds material was worked out based on previous mechanical experiments (E=36.96MPa). The most common loads of scaffolds suffered were compression and bending, especially the compressive strength is the most important mechanical property of bionic scaffolds. Selected the same element type and material attributes, the compression and bending analysis of these four models were analyzed in Ansys. The max strain value was obtained and then the compressive strength and bending strength of every model was calculated respectively according to mechanical formula. Based on the compressive strength and bending strength of every model, the influences of macro-pores shape and distribution on the mechanical properties were analyzed and discussed. Fig.4 shows the compressive strength of each model under unidirectional compressive load. The compressive strength of model C is more than twice of model A’s and the compressive strength of model D is 24.7% higher than model B’s. Compared with model A, the compressive strength of model B increases 63.6%, while the compressive strength of model D decreases 6.2% compared with model C. From Fig.4, following conclusions could be obtained. Firstly, scaffolds with square macro-pores exhibit much better compressive property than scaffolds with cylindrical macro-pores having the same porosity. Secondly, altering the macro-pores distribution could improve the compressive strength of model with cylindrical macro-pores apparently but result a little reduce of the compressive strength of model with square macro-pores. ) 0.3 a P M ( 0.25 h t g n 0.2 e r t s 0.15 e v i 0.1 s s e r p 0.05 m o c

0.2536

0.2389

0.1916

0.1171

0

A

B

C

D

model

Fig. 4. Compressive strength of the four models under unidirectional compressive load

Application of Image Processing and Finite Element Analysis

141

Fig. 5. The total strain intensity of structure models under unidirectional compressive load

Fig.5 showed the total strain intensity of four structure models under unidirectional compressive load analyzed in Ansys. All of these were under the same load situation with the bottom surface fixed and pressure of 10000N/m2 was on the top surface. The total strain intensity reached the maximum value in the stress concentration area, which was the area that was broken firstly. The stress concentration area was usually in the smallest sections. As shown in Fig.7, both the stress concentration degree of model A and B were higher than the model C and D respectively and the smallest section of model with cylinder macro-pores was smaller than the model with square macro-pores under the same pores distribution. So the scaffolds with square macropores exhibit much better compressive property than scaffolds with cylindrical macropores having the same porosity. Altering the macro-pores distribution could increase the smallest section area. To the models with cylinder macro-pores, altering the macro-pores distribution could reduce the concentration degree of stress apparently, so the compressive strength of model B increased clearly from model A. To the models with square macro-pores, altering the macro-pores distribution aggravated the concentration degree of stress, so the compressive strength of model D reduced a little from model C. 3.3 Result of Finite Element Analysis As shown in Fig.6, three groups of bioceramic scaffold samples of the four kinds of structure model were fabricated through an indirect method combining with stereolithography (SLA) and slurry casting technologies. Compression tests were conducted with an INSTRON 5542 material testing system using a 50N load-cell. The compressive tests results were shown as Fig.7. From Fig.7, it was seen that the compressive strength of scaffold samples C is 59.2% higher than scaffold samples A and the compressive strength of scaffold samples D is 5.4% more than scaffold samples B. These distinctions were not as evident

142

L. Lin et al.

Fig. 6. Bioceramic scaffold samples of the model structure )a 0.7 P 0.6 M ( h t 0.5 g n e r 0.4 t s e 0.3 v i s s 0.2 e r p m 0.1 o c

0

0.5434

0.5698

0.5724

C

D

0.3579

A

B

scaffold sample

Fig. 7. Compressive strength of the scaffold samples under unidirectional compressive load

as the FEA results, but also showed that scaffolds with square macro-pores exhibit better compressive strength than scaffolds with cylinder macro-pores. Compared with scaffold samples A, the compressive strength of scaffold samples B increases 51.8%, while the compressive strength of scaffold samples D nearly equal with scaffold samples C. This trendy also exhibited that altering the macro-pores distribution could improve the compressive strength of scaffold samples with cylindrical macro-pores apparently while made less influence on scaffold samples with square macro-pores.

4 Discussion 4.1 The Advantage of Using Mimics in Scaffolds’ Shape Designing The shape of human bone is very complex and irregular, while the bone defects caused by accidents or disease are even more. Before the appearance of the medical image processing system, the traditional method to repair the defective bone needs a surgery first, and in this operation, the shape and the size of the defective part of the bone is estimated by the naked eyes of doctors and some measure tools. And then, the model of the defective part is made by a manual operation followed by a second operation to plant the model to the defective part of the bone. This traditional method to repair the defective bone is of low efficiency, long period, low precision, and has a high cost. And the second operation usually causes hurt to the patient.

Application of Image Processing and Finite Element Analysis

143

As the appearance and the development of medical image processing system, this kind of system is used more and more in the repair of the defective bone. First, the CT images are obtained, and then the 3D CAD model of the defective bone is reconstructed by the medical image processing system. Then using the function of the system or other CAD software, the defective bone is repair and the CAD model of the repair is obtained. At last, the CAD model of the repair is fabricated. This method could have a high precision, shorten the period, and avoid the hurt to the patient in the second operation [15-16]. Mimics is a kind of such digital 3D medical image processing system. It provides the function to read the medical images, to reconstruct the images, and to repair the defective bone. The unique mask operation makes the segmentation of the images and the reconstruction of the tissue easy and convenient. Using the reconstructing function of Mimics, the related area can be calculated and the 3D model can be obtained. Usually, these operations require a high speed. However, Mimics can successfully finish on normal PC. Mimics provides a great convenience to the doctors, shortens the treating time, reduces the hurt to the patient and brings a great improvement to the clinic effect. 4.2 Mechanical Experiment The macro-pores design optimizing contains the design of the size, shape, distribution and so on. These parameters are very important to improve bionic scaffold’s properties. As the biomaterials used to prepare scaffolds were always very expensive and the fabricating process is complicated and needs long time, using finite element analysis to evaluate the mechanical properties of scaffolds as the design changed could save the cost and preparation cycle. The change trendy of the compressive strength obtained through mechanical experiments was consistent with the FEA results basically in this study validated the significance of FEA in bionic scaffolds design optimizing. Although the change trendy of the compressive strength obtained through FEA and mechanical experiments was consistent basically, the value of compressive strength of each scaffold sample was different from the FEA results of structure models. The compressive strength of all the four scaffold samples obtained through compressive tests was higher obviously than the FEA results of the four structure models. This was because the factual strain value when the scaffolds were broken was much bigger than 0.03, which was set in Ansys. The bioceramic materials properties were not exactly consistent with the material attributes set in Ansys. The fabrication methods and processing technologies could also influence the mechanical properties to a large extent.

5 Conclusions In this paper, the application of image processing and finite element analysis in the design optimizing of bionic scaffolds’ shape and inner microstructure was studied respectively. The bionic scaffold’s shape was obtained through Mimics’ image processing and 3D reconstruction technologies. Finite element analysis (FEA) was used in evaluating the mechanical properties of scaffolds’ structure models with different macro-pores shape and distribution to obtain the optimized parameters. Three groups

144

L. Lin et al.

of bioceramic scaffolds samples were fabricated through an indirect method combining stereolithography (SLA) and slurry casting and mechanical experiments were tested. The change trendy of the compressive strength obtained through mechanical experiments was consistent with the FEA results basically, so the significance of FEA in bionic scaffolds’ design optimizing was validated. The compressive strength of all the four scaffold samples obtained through compressive tests was lower obviously than the FEA results of the four structure models and the possible reasons were discussed. Acknowledgments. The authors would like to acknowledge the support of Shanghai Education Fund (No. 5A281) and Shanghai Academic Excellent Youth Instructor Special Foundation.

References 1. Tan, K.H., Chua, C.K., Leong, K.F., Cheah, C.M.: Scaffold development using selective laser sintering polyetheretherketone-hydroxyapatite biocomposite blends. Biomaterials 26, 4281–4289 (2005) 2. Yang, S.F., Leong, K.F., Du, Z.H., Chua, C.K.: The design of scaffolds for use in tissue engineering: Part 1-Traditional factors. Tissue Eng. 7(6), 679–690 (2001) 3. Linbo, W., Jiandong, D.: Compression Molding of PCL Porous Scaffolds with complicated shape for Tissue Engineering. Polymer Material Science and Engineering 25(1), 296–299 (2005) 4. Deville, S., Saiz, E., Tomsia, A.P.: Freeze casting of hydroxyapatite scaffolds for bone tissue engineering. Biomaterials 27, 5480–5489 (2006) 5. Madihally, S.V., Howard, W.T.: Matthew Porous chitosan scaffolds for tissue engineering. Biomaterials 20, 1133–1142 (1999) 6. Junmin, Q., Kai, C., Hao, A., Zhihao, J.: Progress in research of preparation technologies of porous ceramics. Ordnance Material Science and Engineering 28(5), 60–64 (2005) 7. Woesz, A., Rumpler, M., Stampfl, J., Varga, F.: Towards bone replacement materials from calcium phosphates via rapid prototyping and ceramic gelcasting. Materials Science and Engineering C 25, 181–186 (2005) 8. Chen, Z., Li, D., Lu, B.: Fabrication of osteo-structure analogous scaffolds via fused deposition modeling. Scripta Materialia 52, 157–161 (2005) 9. Williams, J.M., Adewunmi, A., Schek, R.M., Flanagan, C.L.: Bone tissue engineering using polycaprolactone scaffolds fabricated via selective laser sintering. Biomaterials 26, 4817–4827 (2005) 10. Chen, V.J., Smith, L.A., Ma, P.X.: Bone regeneration on computer-designed nano-fibrous scaffolds. Biomaterials 27, 3973–3979 (2006) 11. Kalitaa, S.J., Bosea, S., Hosickb, H.L.: Amit Bandyopadhyay, Development of controlled porosity polymer-ceramic composite scaffolds via fused deposition modeling. Materials Science and Engineering C 23, 611–620 (2003) 12. Zein, I., Hutmacher, D.W., Tan, K.C.: Fused deposition modeling of novel scaffold architectures for tissue engineering applications. Biomaterials 23, 1169–1185 (2002) 13. Leea, M., Dunna, J.C.Y., Wu, B.M.: Scaffold fabrication by indirect three-dimensional printing. Biomaterials 26, 4281–4289 (2005)

Application of Image Processing and Finite Element Analysis

145

14. Leukers, B., Gulkan, H., Irsen, S.H.: Hydroxyapatite scaffolds for bone tissue engineering made by 3D printing. Journal of Materials Science: Materials in Medicine 16, 1121–1124 (2005) 15. Huanwen, D., Yingjun, W., Qingshui, Y.: Recent development of Computer-aided tissue Engineering. Chinese Journal of Reparative and Reconstruction of Surgery 5, 574–577 (2006) 16. Xi, H., Longbiao, Z., Zhisong, Z., Jianwei, Z.: CT-image Based Reverse of Custom Made Stem. Journal of Nantong University (Natural Science) 5, 52–56 (2006)

The Mechanical Properties of Bone Tissue Engineering Scaffold Fabricating Via Selective Laser Sintering Liulan Lin, Aili Tong, Huicun Zhang, Qingxi Hu, and Minglun Fang Rapid Manufacturing Engineering Center, Shanghai University, Shanghai, China 99 Shangda Road Shanghai 200444, China [email protected], tony_li@ shu.edu.cn

Abstract. Performance of bone tissue depends on porous scaffold microstructures with specific porosity characteristics that influence the behavior of the ingrown cells. The mechanical properties of porous tissue scaffolds are important for their biomechanical tissue engineering application. In this study, the composite materials powder was developed for the selective laser sintering process, and the parameters of selective laser sintering were optimized. With the

aim of evaluating the influence of porosity on mechanical properties, we have studied the load limits for three specimens of scaffolds which have different porosity. Young’s modulus was computed by determining the slope of the stress - strain curve along the elastic portion of the deformation. In addition, the final element analysis (FEA) module of UG NX4 was used to analyze these scaffolds. The results showed that the bone tissue engineering scaffolds were fabricated by SLS technology have good mechanical properties, which have good potential for tissue engineering applications.

1 Introduction In bone tissue engineering (BTE), 3D scaffolds are essential to guide cell proliferation and to maintain native phenotype in regenerating biologic tissues or organs [1], [2], [3]. These scaffolds give shape to regenerate tissue and temporarily fulfill the structural function of native tissue. In addition to fitting into the anatomical defect, they have possessed sufficient strength and stiffness that will bear in vivo loads so that the scaffolds can function before the growing tissue replaces the gradually degrading scaffolds matrix [4], [5], [6]. Conventional methods for making scaffolds include solvent casting, fiber meshes, phase separation, melt molding and gas foaming [7], [8]. These techniques lack precise control of pore shape, pore geometry and spatial distribution. Some methods also require the use of organic solvents that leave undesirable residues in the finished products, and thus create host reactions due to inflammation or toxicity [9]. The use of rapid prototyping (RP) allows the production of scaffolds with controlled hierarchical structures directly from computer data. Furthermore, the shape of the scaffold can be designed by taking anatomical information of the patient’s target defect (e.g. CT, MRI) to obtain a custom-tailored implant [10].One rapid prototyping method, selective laser sintering, may be advantageous for creating bone tissue engineering K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 146–152, 2007. © Springer-Verlag Berlin Heidelberg 2007

The Mechanical Properties of Bone Tissue Engineering Scaffold Fabricating

147

scaffolds because it provides a cost-effective, efficient method by which to construct scaffolds to match the complex anatomical geometry of bone defect structure. More important, this method which was different from other RP method can directly sinter biocompatible materials. Here, we report the example of β-TCP scaffolds with self-supporting feature fabricated by selective laser sintering. The feasibility of sintering such powder blends and the influence of SLS processing parameters on the sintering quality and resulting microstructure of the sintered specimens were studied. With the aim of evaluating the influence of porosity on mechanical properties, we have studied the load limits for three specimens of scaffolds which have different porosity. Young’s modulus was computed by determining the slope of the stress - strain curve along the elastic portion of the deformation. In addition, the final element analysis (FEA) module of UG NX4 was used to analyze these scaffolds.

2 Materials and Methods 2.1 Preparation of Scaffolds Cylindrical porous scaffolds (20mm diameter, 10mm height), with three-dimensional orthogonal periodic porous architectures, were designed using Unigraphics NX4 3D solid modeling software (UGS PLM Solution, Plano, TX). The design was exported to a Sinterstation machine (HPRS-IIIA, BinHu, Wuhan, China) in STL file format, then was used to construct scaffolds of β-tricalcium phosphate and binding material mixture powder by SLS processing. The sieve sizes (74μm) of mixture material were grade by using the vibratile-sizer. SLS processing of the mixture powder was conducted by 11W laser power and 2400mm/s scanning speed. Scaffolds were fabricated layer by layer using a powder layer thickness of 0.1mm (Fig.1B). After SLS processing was completed, the excess powder surrounding the scaffolds was brushed off and unsintered powder was removed from the scaffold interstices by insertion of 1mm diameter needle. Finally, the green scaffolds were calcined by the high temperature (Fig.1C).

Fig. 1. Pictures of scaffold molds (A) 3D solid modeling designed by UG NX4. (B) Before calcined bionic scaffold fabricated by SLS (Green part). (C) After calcined bionic scaffold.

148

L. Lin et al.

2.2 Compression Testing The mechanical properties of the scaffold specimens were measured using an Instron Uniaxial Testing System (Instron 5542 UK). In this experiment, three specimens were compressed at a cross head speed of 1mm/min. Load and displacement was noted and converted to stress and stain, from which the slope was used to calculate elastic modulus (E). A stress-strain curve was plotted based on the apparent stress σ (MPa) and stain ε (%) value with the initial cross-sectional area A1 (mm2) of each test specimen and the deformation values with initial specimen height H1 (mm), respectively. 2.3 Final Element Analysis (FEA) The three models which have different porosities were meshed by tetrahedral element (10 notes), respectively. We brought 100N of pressure to bear on the top face of scaffolds in this examination. And these were added the material properties which were obtained in the compressing testing. A finite element model for strength analysis of bone scaffolds mold under compression was presented. Characteristics of stress distribution and location are determined according to the model.

3 Results and Discussions 3.1 Optimization of SLS Parameters In order to establish a set of suitable processing parameters for processing thereafter the biocomposite (β-TCP/binding) powder, various specimens were tested (Table 1). Table 1. Specimen groups for β-TCP/binding biocomposite powder Laser power (W) 6 7 8 9 10 11 14 16

Scanning (mm/s)

↑ 1500 2000 2400 ↓

speed

Molding(Y/N)

Strength

Surface quality

N N N N Y Y Y Y

fragile fine fine fine

smooth smooth adust adust

According to the Andrew’s number theory the energy density received by the power in a specimen is directly proportional to the laser power and scanning speed [11]. When the two parameters were matched suitably, a set of bone scaffolds which had good mechanical properties and fine surface quality were obtained. In sintering the β-tricalcium phosphate and binding material, particles that were rapidly scanned by the laser beam would receive free surface energy. Binding material was formatted neck at local contact point occurred. Neck growth would only occur for a short period, creating

The Mechanical Properties of Bone Tissue Engineering Scaffold Fabricating

149

a porous network due to such a formation. For the biocomposite powder, three specimens were obtained using the following parameter settings: scanning speed 1500, 2000 and 2400mm/s; laser power 6, 7, 8, 9, 10, 11, 14, 16 W (see Table 1). Processing β-TCP/binding powder at 6-9W laser power did not sinter the particles, as there was few necking. So this was improved by increasing the laser power to 10W and scanning speed to 2000mm/s, the scaffold was fragile as compared to the other test specimens. When test sintering was done at a scanning speed of 2400mm/s and laser power of 11W, it can get good properties of scaffolds. Test specimens sintered with this parameter gave a more connectivity between the particles and the necking formations were more available. Test sintering were carried out above 11W laser power and 2400mm/s scanning speed, the scaffolds’ surface were adusted obviously. 3.2 Mechanical Properties Compression tests were performed to characterize the mechanical properties of the prepared scaffolds with an Instron Uniaxial Testing System. Fracture toughness, the material’s resistively to crack propagation, is an important parameter to assess the susceptibility of a scaffold to failure. Fig.2A shows the typical stress strain curve from compression testing. Three specimens of each type were tested for mechanical properties. The Fig.2B shows that the compressive strength of the porous TCP ceramics decreases linearly with increasing macropores. According to the Hook’s law, the Young’s modulus (E) was calculated in this test. The three specimens of scaffolds’ Young’s modulus were 15.38MPa, 28.57MPa,

Fig. 2. Porosity and strength behaviors of the porous scaffold (A)Stress-strain curves of different porosity in the β-TCP scaffolds(S1, S2 and S3 were the various types of scaffolds with different porosity). (B)Young’s modulus variation calculated as a function of porosity.

150

L. Lin et al.

48.7MPa, respectively. The porosity and strength behaviors of the porous β-TCP scaffold with pore volume fraction 60.7, 70.8 and 75.8 were illustrated in Fig.2B, where different symbols refer to different macro-porosity, the compressive strength of the porous β-TCP ceramic appears to be sensitive to the pore volume, and the difference in the porosity and strength behaviors became pronounce as the porosity volume decreased. The apparent stiffness of the scaffolds (calculated as effective Young’s modulus) was found to decrease linearly with the porosity. Porosity and interconnectivity were the key determinants of the porous scaffolds. It seems important to characterize the effect of porosity together with the macropore size on the compressive strength not only for a better understanding of the porosity and strength behavior but also to help in design of the porous β-TCP ceramic scaffolds with desirable mechanical property. Although there is no clearly defined criterion in mechanical properties required by bone tissue engineering, it is generally accepted that the scaffolding material could bear the force in cell implantation experiment. The presence of cells and deposited ECM can enhance the scaffold’s stability [12]. In addition, if the scaffold is implanted in vivo, the new lively bony tissue will instead of it. So the bone scaffolds do not have to own high compressive strength as true bone. The specimens are hard enough to handle in a real surgical situation. 3.3 FEA It is important that Finite Element Analysis (FEA) was used in research of the mechanical property. By analyzing the stresses and strains of the numeric models, the optimizing the processing parameters of bone scaffolds were obtained. The three models which were meshed by tetrahedral element (10 notes) were illustrated in Fig.3 (A, B and C). Model A of around 55,361 nodes could be established as a model with sufficient convergence and limited calculation time. Model B was around 54,940 nodes, and Model C was around 53,300 nodes. This corresponded to a coarsening factor of 1.4 and was used for the creation all the meshes. Then the Young’s modulus (E) was inputted into the FEA module of UGNX4. When there were defined load with 100N and added constraints, the distortion of results were showed as Fig.3 (E, F, and G). In calcinations processing, the binding materials were burn off and formed bubbles which could be destroyed scaffolds and created micropores. The micropores couldn’t design in the CAD software. But it was weaken the influence by the properties of material which got in the compression test. In the Fig.3 (D, E and F), the red areas are the large deformation region. It can be observed that the center of the scaffolds were weaker than other regions (Fig.3). Because some pores were collected in this area. According to the results of FEA models, it clearly saw the mechanical property of scaffolds was related with the porosity and the distribution of pores in scaffold. The comparison between the three pictures (D, E and F), the first scaffold which the Young’s modulus is 28.57MPa and the porosity is 70.8 had a good mechanical property. In addition, the results were used to adjust the SLS processing parameters which would obtain a better mechanical property scaffold.

The Mechanical Properties of Bone Tissue Engineering Scaffold Fabricating

151

Fig. 3. The models of scaffolds were meshed via UGNX4 FEA module (A, B and C). Stresses and strains analysis of different scaffolds (D, E and F) (the Young’s modulus (E) was 28.57MPa, 15.38MPa, 48.7MPa, respectively).

As being a mechanical property study of β-TCP scaffold for tissue engineering this study restrict in the result of the potential biological application. The interrelation between bone scaffolds and materials has not been investigated since the simulation of cell attachment onto the material has not been included. However, in this paper the deformation of the scaffolds were calculated to predict the mechanical process favorable for cell ingrowths.

4 Conclusions Selective laser sinter (SLS), a rapid prototyping technology, was investigated and successfully applied in the research to produce 3D scaffolds with enough mechanical properties. Two main parameters of SLS, namely laser power and scanning speed and sintering material were investigated to study its effect on the integrity of the test specimens, which were fabricated for the purpose of bone scaffold. Moreover, this research confirmed the decrease in compressive strength with increased with porosity of β-TCP ceramics. Examination of the mechanical deformation indicated that the porous β-TCP scaffolds stress-strain behavior highly similar to that of a typical porous material undergoing compression. Acknowledgments. The authors would like to acknowledge the support of Shanghai Education Fund (No. 5A281) and Shanghai Academic Excellent Youth Instructor Special Foundation.

152

L. Lin et al.

References 1. Kim, S.-S., et al.: Poly (lactide-co-glycolide)/hydroxyapatite composite scaffolds for bone tissue engineering. Biomaterials 27, 1399–1409 (2006) 2. Tan, K.H., Chua, C.K., Leong, K.F.: Scaffold development using selective laser sintering of polyetheretherketone-hydroxyapatite biocomposite blends. Biomaterials 24, 3115–3123 (2003) 3. Vacanti, J.P., Vacanti, C.A.: The history and scope of tissue engineering, in principles of tissue engineering, 2nd edn. Academic Press, New York (2000) 4. Ho, S.T., Hutmacher, D.W.: A comparison of micro CT with other techniques used in the characterization of scaffolds. Biomaterials 27, 1362–1376 (2000) 5. Griffith, L.G.: Polymeric biomaterials. Acta Mater 48, 263–277 (2000) 6. Buckley, C.T., O’Kelly, K.U.: Regular scaffold fabrication techniques for investigations in tissue engineering. Topics in Bio-Mechanical Engineering, 147–166 (2004) 7. Peter, X.M., Zhang, R.: Synthetic nano-scale fibrous extracellular matrix. J. Biomed. Mater Res. 46(1), 60–72 (1998) 8. Sherwood, J.K., Riley, S.L., Palazzolo, R., Brown, S.C., et al.: A three-dimensional osteochondral composite scaffold for articular cartilage repair. Biomaterials 23(24), 4739–4751 (2002) 9. Yang, S.F., Leong, K.F., Du, Z.H., Chua, C.K.: The design of scaffolds for use in tissue engineering: Part I - Traditional factors. Tissue Eng. 7, 679–690 (2001) 10. Leukers, B., Irsen, S.H., Tille, C.: Hydroxyapatite scaffolds for bone tissue engineering made by 3D printing. Journal of Materials Science: Materials in Medicine 16, 1121–1124 (2005) 11. Nelson, J.C.: Ph.D. Thesis. Selective laser sintering of calcium phosphate materials for orthopedic implants. The University of Texas at Austin. USA (1993) 12. Malda, J., et al.: The effect of PEGT/PBT scaffold architecture on the composition of tissue engineered cartilage. Biomaterials 26, 63–72 (2005)

Informational Structure of Agrobacterium Tumefaciens C58 Genome Zhihua Liu and Xiao Sun State Key Laboratory of Bioelectronics, Southeast University, Nanjing 210096, P.R .China {zhliu,xsun}@seu.edu.cn

Abstract. Base-base correlation (BBC) method, based on information theory, translates a DNA sequence into a 16-dimensional vector. It has proven quite effective in distinguishing various functional regions on one chromosome. In this study, we explore the potential use of distinguishing different chromosomes within one species, with particular emphasis on Agrobacterium tumefaciens strain C58. Our findings show that BBC method could effectively distinguish informational structure of Agrobacterium tumefaciens strain C58 genomes. In conclusion, BBC provides a new methodology in post-genome informatics and its applications could be further explored in the further.

1 Introduction The statistical properties of DNA sequences obtained a substantial amount of scientific attention in the last few years. One of the most important findings in this respect is the relation of 10-11 bp periodicities with DNA supercoiling [1]. Several global statistical properties have been developed to analyze DNA sequence and been found to be related with biological function. The variation of word-frequency within genomes has been linked to functional variation [2]. The variation of dinucleotide relative abundance reflects interspecies differences in process such as DNA modification, replication, and repair [3]. The signature of Alu repeats has been identified as peaks in the correlation function [4]. On an evolutionary scale, more closely related species have more similar word compositions [5] and the dinucleotide biases differ between species [6]. Here we developed a novel sequence feature, named as base-base correlation (BBC), which was inspired from using mutual information function (MIF) to analyze DNA sequence. Compared with MIF, BBC emphasized the information of different base pairs within the range of k. It improved the resolving power and provided a more appropriate description of sequence dissimilarity. A sequence, regardless of its length is kilobases, megabases, or even gigabases, corresponded to a unique 16-dimensional vector. Changes in the values of 16 parameters reflected difference between genome content and length. The procedure was a normalization operation to compare genomes of different scales, which are difficult to obtain a good sequence alignment. In recent study [7], BBC was applied to analyze various functional regions of the human chromosome, including exon, intron, upstream, downstream and intergenic regions. The results showed that BBC assisted in distinguishing various functional regions of genome. K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 153–161, 2007. © Springer-Verlag Berlin Heidelberg 2007

154

Z. Liu and X. Sun

In this work, BBC was applied to distinguish different chromosomes of certain species, with particular emphasis on Agrobacterium tumefaciens strain C58. Agrobacterium tumefaciens is a plant pathogen capable of transferring a defined segment of DNA to a host plant, generating a gall tumor. The genome of Agrobacterium tumefaciens strain C58 has an unusual structure consisting of a circular chromosome, a linear chromosome and two plasmids. This ability to transfer and integrate is used for random mutagenesis and has been critical for the development of modern plant genetics and agricultural biotechnology. In 2001, the genome of Agrobacterium tumefaciens C58 was sequenced by two different research groups and two papers about the genome of Agrobacterium tumefaciens C58 were published in Science [8, 9]. NCBI Genome project released this two genome sequences and named Agrobacterium_tumefaciens_C58_Cereon and Agrobacterium_tumefaciens_C58_UWash, respectively.

2 Materials and Methods 2.1 Sequences Two genome projects of Agrobacterium tumefaciens strain C58 used in this study were retrieved from NCBI (http://www.ncbi.nlm.nih.gov). The name, accession number and length for Agrobacterium tumefaciens strain C58 genomes were shown in Table 1. Table 1. The name, accession number and length for Agrobacterium tumefaciens strain C58 genome Strain Agrobacterium tumefaciens C58 Cereon

Agrobacterium tumefaciens C58 UWash

Genome chromosome circular chromosome linear plasmid AT plasmid Ti chromosome circular chromosome linear plasmid AT plasmid Ti

Accession NC_003062 NC_003063 NC_003064 NC_003065 NC_003304 NC_003305 NC_003306 NC_003308

Length (nt) 2,841,581 2,074,782 542,869 214,233 2,841,490 2,075,560 542,780 214,234

2.2 Base-Base Correlation (BBC) DNA sequences can be viewed as symbolic strings composed of the four letters (B1 , B2 , B3 , B4 ) ≡ ( A, C , G , T ) . The probability of finding the base Bi is denoted

by pi (i = 1, 2, 3, 4) . Then BBC is defined as the following: k

Tij (k ) =

∑p l =1

ij (l ) ⋅ log 2 (

pij (l ) pi p j

) i, j ∈ {1, 2, 3, 4}

(1)

Here, pij (l ) means the joint probabilities of bases i and j at a distance of l. Tij (k ) represents the average relevance of the two-base combination with different gaps from 1 to k. It reflects a local feature of two bases with a range of k. Here we take

Informational Structure of Agrobacterium Tumefaciens C58 Genome

155

k=2 in BBC calculation in the present study. For each sequences m, BBC has 16 parameters and constitutes a 16-dimensional vector Vmz ( z = 1, 2, L , 16) . Statistical independence of two bases in a distance l is defined by pij (l ) = pi p j . Thus, deviations from statistical independence is defined by Dij (l ) = pij (l ) − pi p j

(2)

We expand Tij (k ) using a Taylor series in terms of equation 2 k

Tij (k ) =

∑

pij (l ) ⋅ log 2 (

l =1

pij (l ) pi p j

＝∑ [D (l ) + p p ]⋅ ln ⎡⎢1 + Dp p(l )⎤⎥ k

ij

)

ij

i

j

⎣⎢

l =1

i

j

⎦⎥

＝∑ [D (l ) + p p ]⋅ ⎡⎢⎢ Dp p(l ) − 2Dp (pl ) + L⎤⎥⎥ k

i

j

l =1

＝∑ D (l ) + k

ij

l =1

2 ij

ij

ij

()

Dij2 l 2 pi2 p 2j

⎣

i

j

[

+ ο Dij3 (l )

i

j

(3)

⎦

]

This mathematical transformation further increases the calculation speed and solves effectively the problem of 0 ⋅ log 2 0 (i.e. pij (l ) = 0 in equation 1). 2.3 The Distance Matrix

Given two sequences m and n, the distance H mn between two sequences m and n is defined as the following: 16

H mn =

∑ (V

z m

z =1

− Vnz ) 2

m, n = 1, 2, L, N

(4)

Here, Vm and Vn represent the 16-dimensional vectors of sequences m and n. N is the total number of all sequences analyzed. According to equation 4, H mn obviously satisfies the definition of distance: (ⅰ) H mn > 0 for m ≠ n; (ⅱ) H mn = 0; (ⅲ) H mn = H nm (symmetric); (ⅳ) H mn ≤ H mq + H nq (triangle inequality). For N sequences, a real symmetric N × N distance matrix is then obtained. 2.4 Clustering

Accordingly, a real symmetric N × N matrix is used to reflect the distance between N sequences. Then, the clustering trees are constructed using original Neighbor-Joining (NJ) algorithm [10], a note on the NJ algorithm [11], BIONJ algorithm [12] and UPGMA algorithm [13], respectively. The reliability of the branches is assessed by performing 100 resamplings. Bootstrap values are shown on nodes.

156

3

Z. Liu and X. Sun

Results

3.1 GC Content of Agrobacterium Tumefaciens Strain C58 Genome

GC content of Agrobacterium tumefaciens strain C58 genome was shown in Figure 1. GC content values for the corresponding chromosome or plasmid appeared almost equal between Agrobacterium_tumefaciens_C58_Cereon and Agrobacterium_tumefaciens_C58_UWash. In addition, GC content value of chromosome was different from that of plasmid. GC content values appeared relatively large difference between plasmid AT and plasmid Ti. While, GC content values for chromosome circular and chromosome linear showed minor difference. Thereforely, it was very difficult to distinguish different chromosomes of Agrobacterium tumefaciens strain C58 genome only by difference on GC content. 0.6

Agrobacterium_tumefaciens _C58_Cereon

Agrobacterium_tumefaciens _C58_UWash

GC content

0.59

0.58

0.57

0.56

pl

as m

id

Ti

AT pl

as m

id

ea r li n os om e

ch ro m

pl

os om e ch ro m

as m

id

ci rc ul a

r

Ti

AT pl

as m

id

ea r li n os om e

ch ro m

ch ro m

os om e

ci rc ul a

r

0.55

Fig. 1. GC content of Agrobacterium tumefaciens strain C58 genome. Agrobacterium_tumefaciens_C58_Cereon and Agrobacterium_tumefaciens_C58_UWash were indicated by red and blue, respectively

3.2

BBC Curves of Agrobacterium Tumefaciens Strain C58 Genome

For each genome sequence, 16 parameters of BBC were calculated and linked to a continuous curve, which was designated as BBC curve. BBC curve was then represented as a unique feature for a given sequence, providing an intuitionistic and general description for DNA sequence. BBC curves of Agrobacterium tumefaciens strain C58 genome were displayed in Figure 2. Each curve represented a chromosome or plasmid of Agrobacterium tumefaciens strain C58 genome. It was found that BBC curve of the same type of chromosome or plasmid between Agrobacterium tumefaciens strain C58 Cereon and

Informational Structure of Agrobacterium Tumefaciens C58 Genome

157

Agrobacterium tumefaciens strain C58 UWash, tended to cluster together very closely. It was shown that the same color BBC curves bring into coincidence regarding the tendance. On the other hand, it was found that BBC curves of chromosome circular and chromosome linear tended to cluster together, and BBC curves of plasmid AT and plasmid Ti tended to cluster together. An interesting observation was the relatively large difference between plasmid AT and plasmid Ti in BBC values of A---T, C---A, G---A and G---C. To further illustrate the difference in informational structure of Agrobacterium tumefaciens strain C58 genome, the distance matrix was calculated and the clustering tree was constructed by several clustering algorithm. 0.3 chromosome circular chromosome linear plasmid AT plasmid Ti

0.2

BBC Value

0.1

0

-0.1

-0.2

-0.3

-0.4 A---A

A---C

A---G

A---T

C---A

C---C

C---G

C---T

G---A

G---C

G---G

G---T

T---A

T---C

T---G

T---T

Fig. 2. BBC curves of Agrobacterium tumefaciens strain C58 genome. Chromosome circular, chromosome linear, plasmid AT and plasmid Ti were indicated by red, green, blue and magenta, respectively.

3.3 Clustering Tree of Agrobacterium Tumefaciens Strain C58 Genomes

Comparison of four reconstruction methods, it had been found that the four phylograms had the same topology structure. Two major groups (plasmid group and chromosome group) could be seen from these four figures (Figure 3-6). In the first branch, plasmid Ti of Agrobacterium tumefaciens strain C58 UWash and that of Agrobacterium tumefaciens strain C58 Cereon tended to cluster together, with a bootstrap value of 99%. Plasmid AT of Agrobacterium tumefaciens strain C58 UWash and that of Agrobacterium tumefaciens strain C58 Cereon tended to cluster together, with a bootstrap value of 99%. This two groups clustered together and formed a bigger group (plasmid group), with a bootstrap value of 100%. In another branch, chromosome linear of Agrobacterium tumefaciens strain C58 UWash and that of Agrobacterium tumefaciens strain C58 Cereon tended to cluster together, with a bootstrap value of 97%. Chromosome circular of Agrobacterium tumefaciens strain C58 UWash and that of Agrobacterium tumefaciens strain C58 Cereon tended to cluster together, with a bootstrap value of 98%. This two groups clustered together and formed a bigger group (chromosome group), with a bootstrap value of 100%.

158

Z. Liu and X. Sun

Fig. 3. The clustering tree of Agrobacterium tumefaciens strain C58 based on original NJ algorithm. Bootstrap values were shown on nodes

Fig. 4. The clustering tree of Agrobacterium tumefaciens strain C58 based on a note on the NJ algorithm. Bootstrap values were shown on nodes.

Agrobacterium tumefaciens strain C58 has an unusual genome structure consisting of a circular chromosome, a linear chromosome, and two plasmids: the tumorinducing plasmid pTiC58 and a second plasmid pAtC58 [14, 15]. An interesting observation of the study was that the clustering-phylogram based on BBC, whether in terms of NJ algorithm or UPGMA algorithm, could not only distinguish between chromosome and plasmid, but also discriminate two chromosomes (chromosome linear, chromosome circular) and two plasmids (plasmid Ti, plasmid AT), respectively. In addition, the corresponding chromosome or plasmid of Agrobacterium tumefaciens strain C58 Cereon and Agrobacterium tumefaciens strain C58 UWash tended to cluster together.

Informational Structure of Agrobacterium Tumefaciens C58 Genome

159

Fig. 5. The clustering tree of Agrobacterium tumefaciens strain C58 based on BIONJ algorithm. Bootstrap values were shown on nodes.

Fig. 6. The clustering tree of Agrobacterium tumefaciens strain C58 based on UPGMA

algorithm. Bootstrap values were shown on nodes.

4 Discussion The biological origin of genome information being present on a large-scale statistical level is far from being understood. Short-range correlations in DNA sequences have proven informative during recent decades. Starting from the early finding that coding and noncoding sequence segments possess mutual information function with striking differences due to codon usage in the coding regions [16], an ever more detailed look at short-range correlation properties is to be related with biological function, such as the relation of 10-11 bp periodicities with DNA supercoiled structures [1]. BBC had proven quite effective in distinguishing coding and noncoding sequence segments. It could be further used to classify various functional regions on the chromosome, such as exon, intron, upstream, downstream and intergenic regions [7]. Here BBC was

160

Z. Liu and X. Sun

applied to distinguish different types of chromosomes and plasmids of Agrobacterium tumefaciens strain C58 genome, including chromosome circular, chromosome linear, plasmid AT and plasmid Ti. Our main finding supporting this view was that BBC, as a sequence feature, could distinguish not only various functional regions on one chromosome, but also different types of chromosomes and plasmids within one species. Usually, one species has more than one chromosome. It is very difficult to distinguish chromosomes or plasmids by certain property, such as GC content. Multiple sequence alignment is an alternative approach to identify different chromosomes within one species. While the procedure of alignment for whole chromosome sequences appear to be time-consuming, and even impossible. In addition, a good sequence alignment is very difficult to be obtained in the case of large sequence divergence among different chromosomes. Moreover, gaps in the sequences will be ignored in sequence alignment. This procedure will throw away the most ambiguous parts of the alignment, which may be very important to distinguish different chromosomes. In contrast to traditional alignment methods, the advantage of BBC method is low computational complexity and easy to implement. A sequence, regardless of its length is kilobases, megabases, or even gigabases, corresponds to a unique 16-dimensional vector. Changes in the values of 16 parameters reflect difference between genome content and length. Intriguingly, BBC curve provides a fast and intuitionistic tool for sequence comparison analysis. BBC was inspired from using MIF to analyze DNA sequence. Compared with MIF, BBC emphasized the information of different base pairs within the range of k. It improved the resolving power and provided a more appropriate description of sequence dissimilarity.

5 Conclusions BBC method, based on information theory, translates sequence data into a 16dimensional vector. In recent work, BBC method has proven quite effective in distinguishing various functional regions on one chromosome. In this study, we explore the potential use of distinguishing different chromosomes within one species. Our findings show that BBC method is capable of revealing the identity of different chromosomes and plasmids of Agrobacterium tumefaciens strain C58 genome. In conclusion, BBC provides a new methodology in post-genome informatics and its applications can be further explored in the further. Acknowledgments. This work is supported by the National High-Tech Research and Development Program (863 Program) of China (No. 2002AA231071), the Natural Science Foundation of China (No. 60671018; 60121101).

References 1. Schieg, P., Herzel, H.: Periodicities of 10-11bp as indicators of the supercoiled state of genomic DNA. Journal of molecular biology 343(4), 891–901 (2004) 2. Nikolaou, C., Almirantis, Y.: Word preference in the genomic text and genome evolution: different modes of n-tuplet usage in coding and noncoding sequences. Journal of molecular evolution 61(1), 23–35 (2005)

Informational Structure of Agrobacterium Tumefaciens C58 Genome

161

3. Karlin, S., Burge, C.: Dinucleotide relative abundance extremes: a genomic signature. Trends Genet. 11(7), 283–290 (1995) 4. Holste, D., Grosse, I., Beirer, S., Schieg, P., Herzel, H.: Repeats and correlations in human DNA sequences. Physical review 67(6 Pt 1), 061913 (2003) 5. Bush, E.C., Lahn, B.T.: The evolution of word composition in metazoan promoter sequence. PLoS computational biology 2(11), e150 (2006) 6. Gentles, A.J., Karlin, S.: Genome-scale compositional comparisons in eukaryotes. Genome research 11(4), 540–546 (2001) 7. Liu, Z.H., Jiao, D., Sun, X.: Classifying genomic sequences by sequence feature analysis. Genomics, proteomics & bioinformatics 3(4), 201–205 (2005) 8. Goodner, B., Hinkle, G., Gattung, S., Miller, N., Blanchard, M., Qurollo, B., Goldman, B.S., Cao, Y., Askenazi, M., Halling, C., et al.: Genome sequence of the plant pathogen and biotechnology agent Agrobacterium tumefaciens C58. Science 294(5550), 2323–2328 (2001) 9. Wood, D.W., Setubal, J.C., Kaul, R., Monks, D.E., Kitajima, J.P., Okura, V.K., Zhou, Y., Chen, L., Wood, G.E., Almeida Jr., N.F., et al.: The genome of the natural genetic engineer Agrobacterium tumefaciens C58. Science 294(5550), 2317–2323 (2001) 10. Saitou, N., Nei, M.: The neighbor-joining method: a new method for reconstructing phylogenetic trees. Molecular biology and evolution 4(4), 406–425 (1987) 11. Studier, J.A., Keppler, K.J.: A note on the neighbor-joining algorithm of Saitou and Nei. Molecular biology and evolution 5(6), 729–731 (1988) 12. Gascuel, O.: BIONJ: an improved version of the NJ algorithm based on a simple model of sequence data. Molecular biology and evolution 14(7), 685–695 (1997) 13. Highton, R.: The relationship between the number of loci and the statistical support for the topology of UPGMA trees obtained from genetic distance data. Molecular phylogenetics and evolution 2(4), 337–343 (1993) 14. Allardet-Servent, A., Michaux-Charachon, S., Jumas-Bilak, E., Karayan, L., Ramuz, M.: Presence of one linear and one circular chromosome in the Agrobacterium tumefaciens C58 genome. Journal of bacteriology 175(24), 7869–7874 (1993) 15. Goodner, B.W., Markelz, B.P., Flanagan, M.C., Crowell Jr., C.B., Racette, J.L., Schilling, B.A., Halfon, L.M., Mellors, J.S., Grabowski, G.: Combined genetic and physical map of the complex genome of Agrobacterium tumefaciens. Journal of bacteriology 181(17), 5160–5166 (1999) 16. Grosse, I., Herzel, H., Buldyrev, S.V., Stanley, H.E.: Species independence of mutual information in coding and noncoding DNA. Physical review 61(5 Pt B), 5624–5629 (2000)

Feature Extraction for Cancer Classification Using Kernel-Based Methods Shutao Li and Chen Liao College of Electrical and Information Engineering, Hunan University, 410082 Changsha, China [email protected]

Abstract. In this paper, kernel-based feature extraction method from gene expression data is proposed for cancer classification. The performances of four kernel algorithms, namely, kernel Fisher discriminant analysis (KFDA), kernel principal component analysis (KPCA), kernel partial least squares (KPLS), and kernel independent component analysis (KICA), are compared on three benchmarked datasets: breast cancer, leukemia and colon cancer. Experimental results show that the proposed kernel-based feature extraction methods work well for three benchmark gene dataset. Overall, the KPLS and KFDA show the best performance, and KPCA and KICA follow them.

1 Introduction Gene expression studies on DNA microarray data provide unprecedented chances for disease prediction and classification. However, gene datasets usually include a huge number of genes, and many of them may be irrelevant to the analysis. This poses a great difficulty to many classifiers. By performing dimensionality reduction, feature extraction is thus often critical in improving both the accuracy and speed of the prediction systems. A good feature extraction method should extract most informative features and construct a new subset with lower dimension. In recent years, several useful kernel-based learning machines, e.g. kernel fisher discriminant analysis (KFDA) [1], kernel principal component analysis (KPCA) [2] etc., have been proposed. These methods have shown practical relevance for classification, regression problem and in unsupervised learning. Well applications of kernel-based algorithms have been applied for a number of fields, such as in the context of optical pattern and object recognition, text categorization, time-series prediction, and so on [3]. The purpose of this study is to propose the kernel-based feature extraction method for cancer classification, as well as evaluate and compare the performance of four kernel-based methods: kernel Fisher discriminant analysis (KFDA) [1], kernel principal component analysis (KPCA) [2], kernel partial least squares (KPLS) [4], and kernel independent component analysis (KICA) [5]. First, the genes are preprocessed by the method of T-Test to filter irrelevant and noisy genes. Then, these kernel-based methods are used to extract highly informative and discriminative features. Finally, the new training set, with the extracted features, is input to a support vector machine (SVM) for classification. K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 162–171, 2007. © Springer-Verlag Berlin Heidelberg 2007

Feature Extraction for Cancer Classification Using Kernel-Based Methods

163

2 Kernel Methods 2.1 Kernel Fisher Discriminant Analysis (KFDA) KFDA is the method using the model defined by a linear combination of some specified kernel bases as N

y = ∑ ai K (ωi , x)

(1)

i =1

Here K is the kernel function, and ai is the coefficient vector for the ith kernel base. An

ωi − x ] is typically used as the kernel 2σ 2 2

isotropic Gaussian function K (ωi , x) = exp[−

function. The location of the kernel base ωi is fixed to one of the training samples and the number of kernel bases N equals to the number of training samples. Let k ( x) = ( K (ω1 , x),..., K (ω N , x))T be the vector corresponding to the feature vector x.

Then equation (3) can be written as y = AT k ( x ) , where AT = [a1 ,..., aN ] is the coefficient matrix. The optimal coefficient matrix A is obtained by solving the eigen-equation

∑B

(K )

T A = ∑ W AΛ ( A ∑ W A = I ) . (K )

(K )

Here, Λ is a diagonal matrix of eigenvalues, and I denotes the unit matrix. The matrics

∑W

(K )

and

∑B

(K )

are the within-class covariance matrix and the between-class

covariance matrix of the kernel bases vectors k(x). The dimension of the new feature vector y is limited to min (K-1, N). However, the setting is ill-posed as a result of estimating the N × L coefficients of the matrix A from N samples. So some regularization technique needs to be introduced. One of the simplest methods is to simply add a multiple of the identity matrix to

∑W

(K )

as ~

∑ W( K ) = ∑ W

(K)

+β I

It makes the problem numerically more stable due to the within-class covariance matrix ~

∑ W( K )

becomes positive definite as for large β . This is roughly equivalent to adding

independent noise to each of the kernel bases [1]. 2.2 Kernel Principal Component Analysis (KPCA)

PCA can be expressed as the diagonalization of an n-sample estimate of the covariance n matrix Cˆ = 1 ∑φ ( xi )φ ( xi )T , which represents a transformation of the original data to the n

i =1

new coordinates as defined by the orthogonal eigenvectors V. These eigenvectors V ˆ . This problem is (and the corresponding eigenvalue λ ) are obtained from λ V = CV

164

S. Li and C. Liao

equivalent to nλα = Kα , where α is the column vector with coefficients n

α1 ,..., α n such that V = ∑αiφ ( xi ) , and K is the kernel matrix. Normalizing the solution i =1

v k corresponding to the non-zero eigenvalue λ%k = nλ k of the matrix K translates into the condition λk (α k ⋅ α k ) = 1 . Finally, the projection of φ ( x) onto the eigenvector V k can be computed as [2]: n

β ( x ) k := (V k ⋅ φ ( x)) = ∑ α ik K ( xi , x ) . i =1

2.3 Kernel Partial Least Squares (KPLS) Let φ be the n × m ' matrix of input samples in F, where m ' is the dimensionality of F. Denote its ith row by the vector φ ( xi )T . Let φ ' be the n × m ' deflated dataset and Y ' the n × 1 deflated class label. The rule of deflation is

φ ' = φ − t (t T φ )

(2)

Y ' = Y − t (t T Y )

where t is a score vector (component) which is obtained as follows. Let w and c be the weight vectors. The process starts with random initialization of the Y-score u and then iterates the following steps until convergence: (1) w = XTu/(uTu); (2) ||w|| → 1; (3) t = Xw; (4)c = YTt/tTt;(5) u = Yc/(cTc); repeat steps 1.-5. The process is iterated Fac times. As a result, the deflated dataset can be obtained from the original dataset and the PLS component, while the deflated class labels can be obtained from the original class labels and the PLS component. Denote the obtained sequences of t’s and u’s by the n × 1 vectors t1 , t2 ,...t Fac and u1 , u2 ,...uFac , respectively. Moreover, let T = [t1 , t2 ,...tFac ] and U = [u1 , u2 ,...uFac ] . The “kernel trick” can be utilized and results in K = φφ T , where K stands for the n × n kernel matrix: K (i, j ) = k ( xi , x j ) and k is the kernel function. K can now be directly used in the deflation instead of φ , as K ' = ( I n − tt T ) K ( I n − tt T )

(3)

Here, K ' is the deflated kernel matrix and I n is the n-dimensional identity matrix. Now Eq.(2) takes the place of Eq.(1). So the deflated kernel matrix is obtained from the original kernel matrix and the PLS component. In kernel PLS, the assumption that the variables of X have zero mean in linear PLS should also be held. One can center the mapped data in the feature space F as:

1 1 K = ( I n − l n ln T ) K ( I n − l n ln T ) n n

Feature Extraction for Cancer Classification Using Kernel-Based Methods

165

Here, l n is the n × 1 vector with all elements equal to one. Given a set of test samples

{ zi }i =1 n

(where zi ∈

m

), its projection onto the feature space is D p = KtU (T T KU ) −1

where D p = [d1 , d 2 ,..., d nt ]T is a nt × p matrix, with the p KPLS components as its columns and the nt test samples in the reduced-dimensional space as its rows, K t is the nt × n kernel matrix defined on the test set such that K t (i, j ) = K ( zi , x j ) , T T KU is an upper triangular matrix and thus invertible. The centered kernel matrix K t defined on the test set can be calculated as [4]

1 1 K t = ( K t − l n lnT K )( I n − l n lnT ) n n 2.4 Kernel Independent Component Analysis (KICA)

KICA produces a set of nonlinear features of the input data by performing ICA in the kernel-induced feature space F. The input data X is first whitened in F by using the −

1

whitening matrix WPφ = (Λφ ) 2 (V φ )T , where Λφ and V φ contain the eigenvalues and 1 n eigenvectors of the covariance matrix Cˆ = ∑ i =1φ ( xi )φ ( xi )T . The whitened data is n then obtained as: X Wφ = (WPφ )T φ ( X ) = (Λφ ) −1α T K , where K is the kernel matrix, and α is the eigenvector matrix of K. After the whitening transform, we iterate the ICA learning iteration algorithm: U Iφ = WIφ X Wφ

ΔWWφ = [ I + ( I −

2 1+ e

−U Iφ

)(U Iφ )T ]WIφ

WˆIφ = WIφ + ρΔWIφ → WIφ

until WIφ converges. Here ρ is the learning rate. On testing, the new feature representation of a test pattern y can be computed as:

s = WIφ (Λφ ) −1α T K ( X , y ) where K ( X , y ) = [k ( x1 , y ), k ( x2 , y ),..., k ( xn , y )]T [5].

3 Proposed Method Denote the number of genes (features) and the number of samples (observations) in the gene expression dataset by M and N respectively. The whole data set can also be represented by the matrix:

166

S. Li and C. Liao x12 ⎡ x11 ⎢x x 21 22 X=⎢ ⎢ M M ⎢ x x ⎣ M1 M 2

L x1N ⎤ L x2 N ⎥⎥ O M ⎥ ⎥ L xMN ⎦

where xij is the measurement of the expression level of gene i in sample j. Let xj = (x1j, x2j, ..., xMj) denote the ith sample of X, and yj the corresponding class label (e.g., tumor type or clinical outcome). In the following, we assume that there are only two classes (positive class and negative class) in the sample. The proposed method is as follows: Step 1. Preprocessing using T-test: Large dimensionality increases the complexity and computation load, so the dataset is preprocessed by T-test at first. For each gene i, we compute the mean μi+ (respectively, μi− ) and standard deviation δ i+ (respectively, δ i− ) for the positive (respectively, negative) samples. Then a score T ( xi ) can be obtained as: μi+ − μi− T ( xi ) =

(δ i+ )2 (δ i− ) 2 + n+ n−

where n+ and n− are the numbers of samples in the positive and negative classes respectively. Genes are ranked according to their T, and the top p genes are selected to form a reduced dataset. Step 2. Kernel-based methods, as reviewed in section 2, are used to further extract highly informative and discriminative features to form a new training set.

Fig. 1. Schematic diagram of the whole process

Step 3. Training and classification using the SVM: The SVM is an efficient binary classification algorithm. It computes the hyperplane that maximizes the margin between the training examples and the class boundary in a high-dimensional

Feature Extraction for Cancer Classification Using Kernel-Based Methods

167

kernel-induced feature space. Due to the nonlinear mapping between the input space and the feature space, the linear discriminant function constructed by SVM in the feature space corresponds to a nonlinear function in the original input space. Step 4. Finally, the new training dataset with the extracted features is used to train a SVM, this classifier can be used for predictions on the test set. The schematic diagram of the whole process is shown in Fig.1.

4 Experimental Results 4.1 Setup In this section, we evaluate the performance of the proposed feature extraction method on three benchmark datasets: (1) Breast cancer dataset: It contains 7,129 genes and 38 samples. 18 of these samples are ER+ (estrogen receptor) while the remaining 20 are ER- [6]. (2) Leukemia dataset: It contains 7,129 genes and 72 samples. 47 of these samples are of Acute Myeloid Leukemia (AML) and the remaining 25 are of Acute Lymphoblastic Leukemia (ALL) [7]. (3) Colon cancer dataset: It contains 2,000 genes and 62 samples. 22 of these samples are of normal colon tissues and the remaining 40 are of tumor tissues [8]. The gaussian kernel

k ( x, y ) = exp(− x − y / γ ) 2

where γ is the width parameter, is used in the four kernel methods. The adjustable parameters in the T-test and kernel-based methods are listed in the following: 1. 2. 3. 4. 5.

p associated with the T-test method; Width parameter γ in the Gaussian kernel; Number of score vectors (Fac) used in KPLS; Number of principal components (K) used in KPCA; Regularization constant (mu) added to the diagonal of the within-class scatter matrix used in KFDA.

The linear kernel is used in the SVM. Values of the soft-margin parameter (C) used on the different datasets are shown in Table 1. Table 1. Values of the soft-margin parameter (C) used in the SVM

KPLS, KPCA and KFDA KICA

breast cancer 1 1

leukemia 10 1

colon cancer 100 1

Because of the small dataset size, leave-one-out (LOO) cross-validation is used to obtain the testing accuracy. Both feature extraction and classification are put together in each LOO iteration, i.e., they are performed on the training subset and then the performance is obtained on the left-out example using the extracted features.

168

S. Li and C. Liao

4.2 Results Here, we compare the performance of the four kernel-based procedures: KPLS, KPCA, KFDA and KICA. The testing accuracies on the three benchmarked datasets obtained using different parameter settings are shown in Table 2-5, respectively. Overall, KPLS and KFDA show the best classification performance on all three datasets. Both of them achieve the best accuracy of 100% on the breast cancer and leukemia datasets. On the colon cancer dataset, both attain an accuracy of 91.9%, which is the highest in our experiment. On the other hand, KPCA attains an accuracy of 100% on the breast cancer dataset and 98.6% on the leukemia dataset, but only 88.7% on the colon cancer dataset. KICA performs less well in comparison with the other three. It attains an accuracy of 100% on the breast cancer dataset, but only 97.2% on the leukemia dataset and 88.7% on the colon cancer dataset. On the breast cancer dataset, both KPLS and KPCA attain 100% testing accuracy with only 2 features, while KFDA uses 3 features and KICA uses 7 features to achieve 100%. On the leukemia dataset, KFDA attains the best accuracy with only 3 features while KPLS uses 5 features. KPCA and KICA can not obtain 100%. On colon cancer dataset, KFDA attains the best accuracy 91.9% with only 4 features while KPLS uses 10 features, and KPCA as well as KICA can not get 91.9%. In conclusion, KPLS and KFDA outperform the other two in number of features to get the best accuracy. As can be seen in the tables, the prediction accuracy is highly dependent on the choice of the parameters. Obviously, KPLS is influenced by the parameter p. Also, the effect of Fac cannot be neglected on all three datasets. In comparison, γ has a weaker effect than p and Fac. Table 2. Testing accuracies (%) using T-test and KPLS

γ 100

200

300

100

200

300

Fac 2 5 10 2 5 10 2 5 10 2 5 10 2 5 10 2 5 10

p

50

90

breast cancer 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 97.4 97.4 100.0 97.4 97.4 100.0 97.4 97.4

p

100

500

leukemia 97.2 98.6 93.1 97.2 98.6 93.1 95.8 98.6 93.1 97.2 100.0 100.0 98.6 100.0 98.6 98.6 100.0 97.2

p

50

90

colon cancer 88.7 85.5 82.3 87.1 85.8 79.0 87.1 85.5 79.0 90.3 88.7 91.9 90.3 90.3 91.9 90.3 90.3 90.3

Feature Extraction for Cancer Classification Using Kernel-Based Methods Table 3. Testing accuracies (%) using T-test and KPCA

γ 100

200

300

100

200

300

K 2 5 10 2 5 10 2 5 10 2 5 10 2 5 10 2 5 10

p

50

90

breast cancer 100.0 100.0 100.0 52.6 52.6 52.6 52.6 52.6 52.6 100.0 100.0 100.0 94.7 94.7 92.1 52.6 55.3 52.6

p

leukemia 94.4 94.4 95.8 73.6 100 84.7 86.1 65.3 65.3 65.3 98.6 98.6 97.2 97.2 98.6 500 98.6 95.8 98.6 97.2

p

colon cancer 88.7 88.7 87.1 88.7 50 88.7 88.7 88.7 88.7 88.7 88.7 88.7 88.7 88.7 88.7 90 88.7 88.7 88.7 88.7

Table 4. Testing accuracies (%) using T-test and KFDA (with mu=10-m)

γ 5

10

15

5

10

15

m 3 4 5 3 4 5 3 4 5 3 4 5 3 4 5 3 4 5

p

50

90

breast cancer 100.0 97.3 94.7 100.0 97.4 97.4 100.0 100.0 97.4 97.4 97.4 97.4 100.0 97.4 97.4 100.0 100.0 97.4

p

leukemia 95.8 94.4 94.4 95.8 100 95.8 94.4 97.2 95.8 94.4 100.0 100.0 98.6 98.6 500 100.0 100.0 97.2 100.0 100.0

p

colon cancer 87.1 88.7 82.3 62.9 50 72.6 88.7 64.5 62.9 87.1 90.3 91.9 91.9 62.9 90 90.3 90.3 64.5 59.7 90.3

169

170

S. Li and C. Liao Table 5. Testing accuracies (%) using T-test and KICA

p 5

6

7

8

9

10

γ 100 200 300 100 200 300 100 200 300 100 200 300 100 200 300 100 200 300

breast cancer 97.4 97.4 97.4 97.4 97.4 97.4 100.0 100.0 100.0 97.4 97.4 97.4 97.4 97.4 97.4 100.0 100.0 100.0

leukemia 97.2 97.2 97.2 95.8 95.8 95.8 97.2 97.2 97.2 95.8 95.8 95.8 95.8 95.8 95.8 95.8 95.8 95.8

colon cancer 88.7 88.7 88.7 88.7 88.7 88.7 88.7 88.7 88.7 83.9 83.9 83.9 80.7 80.7 80.7 82.3 82.3 82.3

For KPCA, γ has a strong impact on the breast cancer and leukemia datasets. For example, when γ =100, and p=50, the testing accuracy is 100% on breast cancer. However, when γ changes to 200 (with the same value of p), the accuracy drops to only 52.6%. For KFDA, p has the most obvious impact on all three datasets. mu and γ also have obvious effects on the testing accuracy result. For KICA, p also has the strongest influence, while the effect of γ is less obvious. Due to the slow speed of the iterative procedure, we have to use very small values of p so that the data do not become so large. Otherwise, the running time will be very long and the iterative procedure may also have numerical problems.

5 Conclusions In this paper, we propose kernel-based feature extraction method for cancer classification and discuss the performances of four kernel methods, KPLS, KPCA, KFDA and KICA. Experiments are performed on the breast cancer, leukemia and colon cancer datasets. We also compare them with other methods reported in the literature. The proposed method shows superior classification performance on all three datasets, and thus proves to be reliable for feature extraction.

Feature Extraction for Cancer Classification Using Kernel-Based Methods

171

Acknowledgements. This paper is supported by the Program for New Century Excellent Talents in University and the Excellent Youth Foundation of Hunan Province (06JJ1010).

References 1. Kurita, T., Taguchi, T.: A Modification of Kernel-based Fisher Discriminant Analysis for Face Detection. In: Proceedings of International Conference on Automatic Face and Gesture Recognition, Washington DC, pp. 300–305 (2002) 2. Schölkopf, B., Smola, A., Müller, K.-R.: Kernel Principal Component Analysis. In: Schölkopf, B., Burges, C.J.C., Smola, A.J. (eds.) Advances in Kernel Methods - Support Vector Learning, pp. 327–352. MIT Press, Cambridge, MA (1999) 3. Müller, K., Mika, S., Rätsch, G., Tsuda, K., Schölkopf, B.: An Introduction to Kernel-Based Learning Algorithms. IEEE Trans. on Neural Networks, 180–201 (2001) 4. Rosipal, R., Trejo, L.J., Matthews, B.: Kernel PLS-SVC for Linear and Nonlinear Classification. In: Proceedings of the Twentieth International Conference on Machine Learning, Washington DC, pp. 640–647 (2003) 5. Bach, F.R., Jordan, M.I.: Kernel Independent Component Analysis. J. Machine Learning Research 3 (2002) 6. West, M., Blanchette, C., Dressman, H., Huang, E., Ishida, S., Spang, R., Zuzan, H., Marks, J.R., Nevins, J.R.: Predicting the Clinical Status of Human Breast Cancer Using Gene Expression Profiles. Proceedings of the National Academy of Science 98, 11462–11467 (2001) 7. Golub, T., Slonim, D., Tamayo, P., Huard, C., Gaasenbeek, M., Mesirov, J., Coller, H., Loh, M., Downing, J., Caligiuri, M., Bloomfield, C., Lander, E.: Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Expression Monitoring. Science 28, 531–537 (1999) 8. Alon, U., Barkai, N., Notterman, D., Gish, K., Ybarra, S., Mack, D., Levine, A.: Broad Patterns of Gene Expression Revealed by Clustering Analysis of Tumor and Normal Colon Tissues Probed by Oligonucleotide Arrays. Proceedings of the National Academy of Science 96, 6745–6750 (1999)

A New Hybrid Approach to Predict Subcellular Localization by Incorporating Protein Evolutionary Conservation Information ShaoWu Zhang1, YunLong Zhang2, JunHui Li, and HuiFeng Yang1, YongMei Cheng1, and GuoPing Zhou3 1

College of Automation, Northwestern Polytechnical University, Xi’an, 710072, China [email protected] 2 Department of Computer, First Aeronautical Institute of Air Force, Henan, 464000, China [email protected] 3 Department of Biological Chemistry and Molecular Pharmacology, Harvard Medical School, Boston, Massachusetts 02115, USA [email protected]

Abstract. The rapidly increasing number of sequence entering into the genome databank has created the need for fully automated methods to analyze them. Knowing the cellular location of a protein is a key step towards understanding its function. The development in statistical prediction of protein attributes generally consists of two cores: one is to construct a training dataset and the other is to formulate a predictive algorithm. The latter can be further separated into two subcores: one is how to give a mathematical expression to effectively represent a protein and the other is how to find a powerful algorithm to accurately perform the prediction. Here, an improved evolutionary conservation algorithm was proposed to calculate per residue conservation score. Then, each protein can be represented as a feature vector created with multi-scale energy (MSE). In addition, the protein can be represented as other feature vectors based on amino acid composition (AAC), weighted auto-correlation function and Moment descriptor methods. Finally, a novel hybrid approach was developed by fusing the four kinds of feature classifiers through a product rule system to predict 12 subcellular locations. Compared with existing methods, this new approach provides better predictive performance. High success accuracies were obtained in both jackknife cross-validation test and independent dataset test, suggesting that introducing protein evolutionary information and the concept of fusing multifeatures classifiers are quite promising, and might also hold a great potential as a useful vehicle for the other areas of molecular biology.

1 Introduction One of the fundamental goals in cell biology and proteomics is to identify the functions of proteins in the cellular environment. Determination of protein subcellular location purely using experimental approaches is both time-consuming and expensive. Particularly, the number of new protein sequences yielded by the high-throughput K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 172–179, 2007. © Springer-Verlag Berlin Heidelberg 2007

A New Hybrid Approach to Predict Subcellular Localization

173

sequencing technology in the postgenomic era has increased explosively. Facing such an avalanche of new protein sequences, it is both challenging and indispensable to develop an automated method for fast and accurately annotating the subcellular attributes of uncharacterized proteins. The knowledge thus obtained can help us timely utilize these newly found protein sequences for both basic research and drug discovery [1, 2]. During the last decade, many theoretical and computational methods were developed in an attempt to predict subcellular localization of protein [3-13]. However, all these prediction methods were established basically based on a single classifier, or based on the statistical approach and amino acid physical-chemical character to represent protein sequences. Obviously, the prediction quality would be considerably limited by using only one single classifier, statistical feature and physical-chemical feature information to deal with piled-up complicated protein sequences with extreme variation in both sequence order and length. To further improve the predictive quality, a logical and key step would be to find an effective way to represent protein information and a powerful classifier. Here, by proposing an improved method to calculate protein evolutionary conservation information, the samples of proteins were formulated by hybridizing the multisource information derived from evolutionary conservation scores, weighted auto-correlation functions [14], moment descriptors [12] and multi-scale energy [13]. Based on the hybridization representation, a novel ensemble classifier was formed by fusing many individual classifiers through a product rule system [15]. The success rates obtained by hybridizing the multi-source information of proteins and the fusion classifier in predicting protein subcellular location were significantly improved.

2 Methods 2.1 Residue Conservation The residue ranking function assigns a score to each residue, and according to which they can be sorted in the order of the presumably decreasing evolutionary pressure they experience. Out of many methods proposed in the literature [16-18], Lichtarge research group’s hybrid methods [19](real-valued evolutionary trace method and zoom method) are the two robust methods, that rank the evolutionary importance of residues in a protein family which is based on the column variation in multiple sequence alignments (MSAs) and evolutionary information extracted from the underlying phylogenetic trees. However, the hybrid methods treat the gaps in the multisequences alignment as the 21st amino acid. So, we propose an improved algorithm to estimate the residue evolutionary conservation. The processes of calculation are as follows. Firstly, the initial similarity sequences were created by using three iterations of PsiBlast[20], with the 0.001 E-value cutoff, on the UniProt [21] database of proteins. The PsiBlast resulting sets were aligned by a standard alignment method such as ClustalW 1.8 [22]. So, the multiple sequence alignments (MSA) were obtained. Secondly, an MSA is divided into sub-alignments (that is, n groups) that correspond to nodes in the tree [19]. This subdivision of an MSA into smaller alignments

174

S. Zhang et al.

reflects the tree topology, and therefore the evolutionary variation information within it. Then, the evolutionary score for a residue belong to column i in an MSA is given by the following equation. Ri = 1 +

N−1

∑w n=1

20

n

∑w

f α log ∑ α

group(g)[−

node(n)

g=1

g

i

g 20 f iα

+ fi,ggap]

=1

(1)

where wnode (n) , wgroup (g ) are weights assigned to a node n and a group g, respectively. ⎧1 wnode (n) = ⎨ ⎩0

⎧1 wgroup ( g ) = ⎨ ⎩0

if n on the path to the query protein otherwise

if g on the path to the query protein otherwise

(2)

(3)

f iαg is the frequency of amino acid of type α ( α represents one of the 20 standard amino acids, that is, A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y) within a sub-alignment corresponding to group g at the level in which the sequence similarity tree is divided into n groups. Namely, the nodes (labeled by n) are assumed to be numbered in the order of increasing distance from the root, and each one of them has associated with it a division of the tree into n groups (subtrees). N is the number of alignment sequences.

f i ,ggap is the number of non-standard amino acids (such as“−”,

“X”, “Z”, “B”) of g group in the alignment position i, divided by the number of g group alignment sequences. Further details about division of tree nodes and groups can be found in literature [19]. 2.2 Multi-scale Energy [13] Through residue conservation scores calculating, the protein sequence of English letters can be translated into a numerical sequence. The numerical sequence can be considered as digital signal. Projecting the signal onto a set of wavelet basis functions with various scales, the fine-scale and large-scale conservation information of a protein can be simultaneously investigated. Here, the wavelet basis function used is symlet wavelet [23]. Consequently, the protein can be characterized as the following multi-scale energy (MSE) feature vector. MSE = [d1 ,L, d j ,L d m , am ]

(4)

Here, m is the coarsest scale of decomposition, dj is the root mean square energy of the wavelet detail coefficients in the corresponding j-th scale, and am is the root mean square energy of the wavelet approximation coefficients in the scale m. The energy factors dj and am are defined as: N j −1

dj =

1 Nj

∑[u (n)]

2

j

n =0

am =

N m −1 1 Nm

∑[v n =0

2 m (n)]

j = 1,2, L , m

(5)

A New Hybrid Approach to Predict Subcellular Localization

175

Here, Nj is the number of the wavelet detail coefficients, Nm is the number of the wavelet approximation coefficients, uj(n) is the n-th detail coefficient in the corresponding j-th scale, and vm(n) is the n-th approximation coefficient in the scale m. For the protein sequence with length L, m equals INT(log 2L). Combing with amino acid composition (AAC)which is consisted the 20-D components of the amino acid frequencies, the protein can be represented by the following (20+m+1)-D vector.

[

x = f1 , f 2 , L , fα , L , f 20 , d1 , d 2 , L , d j , L , d m , am

]T

(6)

Here fα ( α = 1,2, L ,20 ) is the occurrence frequencies of 20 amino acid in the protein concerned, arranged alphabetically according to their signal letter codes. Conveniently, the feature set based on the residue evolutionary conservation and MSE approach can be wrote as EMSE. 2.3 Weighted Auto-correlation Functions [14]

In order to calculate the weighted auto-correlation functions, we replace each residue in the primary sequence by its amino acid index PARJ860101, which can be downloaded from http://www.genome.ad.jp/dbget. Consequently, the replacement results in a numerical sequence h1 , h2 ,L, hl ,L, hL . The weighted auto-correlation functions rj are defined as: rj =

w L− j ∑ hl hl + j , L − j l =1

j = 1,2,L, λ

(7)

Here hl is the amino acid index for the l-th residue, w is weighted factor and L is the length of protein sequence. Combing with amino acid composition which is consisted the 20-D components of the amino acid frequencies, the protein can be represented by the following (20+ λ )D vector.

[

x = f1 , f 2 , L , fα , L , f 20 , r1 , r2 , L , r j , L , rλ

]T

(8)

Conveniently, the feature set based on the weighted auto-correlation functions approach can be wrote as PARJ. 2.4 Moment Descriptor [12] According to the literature [12], the protein can be represented by the following vector: X = [ f1 , f 2 , L , f α , L f 20 , μ1 , μ 2 , L μ i , L μ 20 , ν1 , ν 2 , L , ν i L , ν 20 ]

(9)

Here,

μi =

1 L ∑ xij • j L j =1

νi =

1 L ( xij • j − μi )2 ∑ L j =1

(10)

176

S. Zhang et al.

if amino acid α i appears at position j in the sequence ⎧1 xij = ⎨ ⎩0 if amino acid α i does not appear at position j in the sequence

(11)

Conveniently, the feature set based on the Moment descriptors approach can be wrote as MD.

3 Result and Discusion 3.1 Results with Different Feature Extraction Methods The training dataset and independent dataset taken from Chou [4] were used to validate the current method. The prediction quality was examined by the standard testing procedure in statistics, which are the jackknife (JACK) and independent dataset tests (INDE). Of these two, the jackknife test is regarded as the most objective and effective [24-25]. The results of four feature extraction methods based on support vector machine (SVM) [26] and “one-versus-one” classifying policy [14] are shown in table 1. Table 1. Results (in percentage) of four feature extraction methods with SVM and “one-versusone” classification strategy

Chloroplast Cytoplasm Cytoskeleton Endoplasmic reticulum Extracellular Golgi apparat Lysosome Mitochondrial Nuclear Peroxisomal Plasma membrane Vacuoles Overall accuracy

AAC JACK IND 59.1 60.6 85.9 83.9 41.2 94.7 32.7 70.8 69.6 84.2 16 0.50 56.8 87.1 26.5 12.9 80.8 76.4 22.2 43.5 92.7 96.3 33.3 77.1 80

EMSE JACK IND 67.9 59.6 88.4 87.6 44.1 100 38.8 84.9 67.9 83.1 20 25 51.4 96.8 43.4 20.9 87.1 80.4 14.8 39.1 92.7 96.5 25 79.5 82.8

PARJ JACK IND 59.1 65.1 89 88.8 47.1 100 36.7 70.8 73.2 88.4 24 25 54.1 96.8 41 17.8 84.1 77.5 22.2 47.8 96 99 33.3 80.5 83.3

MD JACK IND 66.4 77.1 90.5 89.1 50 94.7 34.7 69.8 67.9 87.4 16 50 51.4 87.1 38.6 14.1 81.9 83.1 7.4 30.4 94.3 97.1 20.8 79.4 83.5

Table 1 shows that protein evolutionary conservation information can be used to predict subcellular location. The overall accuracies of EMSE, PARJ and MD are almost equal, but they are all higher than that of AAC in jackknife and independent tests. For EMSE, the predictive accuracy is critically dependent on the input selection of sequences and also on the breadth and the depth of the associated sequence similarity tree. That is, how many initial similarity sequences were selected, and how to prune these sequences to form multiple alignment sequences? If the optimal two parameters were selected, we can obtain better results. Considering the computer power, the cutoff of initial similarity sequences was defined as 250, and we did not prune the initial similarity sequences. These results indicate the performance of predictive system can be improved by using different feature extraction methods. EMSE, PARJ and

A New Hybrid Approach to Predict Subcellular Localization

177

MD are effective to represent protein sequence and good robust for predictiing subcellular localization.

3.2

Comparison with Other Prediction Methods

The performance of the hybrid method developed in this study was compared with existing methods such as Pan’s [6], Gao’s [9] and Xia’s [10,11], which were also developed from the same dataset. The results demonstrated that overall prediction accuracies of our hybrid method are higher than that of other four methods both in the Jackknife and independent tests. For example, the overall accuracy of the hybrid method is 8.1%, 4.7% greater than that of Xia’s method [10] in the Jackknife and independent tests respectively. Table 2. Overall accuracy (in percentage) obtained by different methods Method

Jackknife test

Independent test

Pan et al [6] Gao et al [9] Xia et al [10] Xia et al [11] Hybrid (AAC+ EMSE+PARJ+MD)

67.7 69.6 73.6 72.6 81.7

73.9 79.8 74.8 85.1

4 Conclusions A new kind of protein evolutionary feature extraction method and a hybrid approach to fuse the four feature classifiers were proposed in this study. The results show that using residue evolutionary conservation and multi-scale energy to represent protein can better reflect protein evolutionary information and predict the subcellular locations. Weighted auto-correlation function and Moment descriptor methods can optimally reflect the sequence order effect. It is demonstrated that the novel hybrid approach by fusing four feature classifiers is a very intriguing and promising avenue.

Acknowledgements. This paper was supported in part by the National Natural Science Foundation of China (No. 60372085 and 60634030), the Technological Innovation Foundation of Northwestern Polytechnical University (No. KC02), the Science Technology Research and Development Program of Shaanxi Province (No. 2006k04-G14).

References 1. Chou, K.C.: Review: Structural bioinformatics and its impact to biomedical science. Curr. Med. Chem. 11, 2105–2134 (2004) 2. Lubec, G., Afjehi-Sadat, L., Yang, J.W., John, J.P.: Searching for hypothetical proteins: theory and practice based upon original data and literature. Prog. Neurobiol. 77, 90–127 (2005)

178

S. Zhang et al.

3. Chou, K.C., Elrod, D.W.: Protein subcellular location prediction. Protein Engineering 12, 107–118 (1999) 4. Chou, K.C.: Prediction of protein subcellular locations by incorporating quasi-sequenceorder effect. Biochem. Biophys. Research Commun. 278, 477–483 (2000) 5. Chou, K.C.: Prediction of protein cellular attributes using pseudo-amino acid composition. Proteins: Structure, Function, and Genetics 43, 246–255 (2001) 6. Pan, Y.X., Zhang, Z.Z., Guo, Z.M., Feng, G.Y., Huang, Z.D., He, L.: Application of pseudo amino acid composition for predicting protein subcellular location: stochastic signal processing approach. J. Protein Chem. 22, 395–402 (2003) 7. Zhou, G.P., Doctor, K.: Subcellular location prediction of apoptosis proteins. PROTEINS: Struct. Funct. Genet. 50, 44–48 (2003) 8. Park, K.J., Kanehisa, M.: Prediction of protein subcellular locations by support vector machines using compositions of amino acid and amino acid pairs. Bioinformatics 19, 1656– 1663 (2003) 9. Gao, Y., Shao, S., Xiao, X., Ding, Y., Huang, Y., Huang, Z., Chou, K.C.: Using pseudo amino acid composition to predict protein subcellular location: Approached with Lyapunov index, Bessel function, and Chebyshev filter. Amino Acid 28, 373–376 (2005) 10. Xia, X., Shao, S., Ding, Y., Huang, Z., Huang, Y., Chou, K.C.: Using complexity measure factor to predict protein subcellular location. Amino Acid 28, 57–81 (2005) 11. Xia, X., Shao, S., Ding, Y., Huang, Z., Huang, Y., Chou, K.C.: Using cellular automata images and pseudo amino acid composition to predict protein subcellular location. Amino Acid 30, 49–54 (2006) 12. Shi, J.Y., Zhang, S.W., Liang, Y., Pan, Q.: Prediction of Protein Subcellular Localizations Using Moment Descriptors and Support Vector Machine. In: PRIB: 2006, Hong Kong,China, pp. 105–114. Springer, Heidelberg (2006) 13. Shi, J.Y., Zhang, S.W., Pan, Q., Cheng, Y.M., Xie, J.: SVM-based Method for Subcellular Localization of Protein Using Multi-Scale Energy and Pseudo Amino Acid Composition. Amino Acid (2007) DOI 10.1007/s00726-006-0475-y 14. Zhang, S.W., Pan, Q., Zhang, H.C., Shao, Z.C., Shi, J.Y.: Prediction Protein Homooligomer Types by Pesudo Amino Acid Composition: Approached with an Improved Feature Extraction and Naive Bayes Feature Fusion. Amino Acid 30, 461–468 (2006) 15. Kittler, J., Hatef, M., Duin, R.P.W., Matas, J.: On Combining Classifiers. IEEE Trans. Pattern Analysis and Machine Intelligence 20, 226–239 (1998) 16. Lichtarge, O., Bourne, H., Cohen, F.: An evolutionary trace method defines binding surfaces common to protein families. J. Mol. Biol. 257, 342–358 (1996) 17. Valdar, W.S.: Scoring residue conservation. Proteins 48, 227–241 (2002) 18. Soyer, O.S., Goldstein, R.A.: Predicting functional sites in proteins: Site-specific evolutionary models and their application to neurotransmitter transporters. J. Mol. Biol. 339, 227–242 (2004) 19. Mihalek, I., Reš, I., Lichtarge, O.: A Family of Evolution–Entropy Hybrid Methods for Ranking Protein Residues by Importance. J. Mol. Biol. 336, 1265–1282 (2004) 20. Altschul, S., Madden, T., Schffer, A., Zhang, J., Zhang, Z., Miller, W., Lipman, D.: Gapped blast and psi-blast: a new generation of protein database search programs. Nucleic Acids Research 25, 3389–3402 (1997) 21. UniProt (2005), http://www.expasy.org/ 22. Thompson, J., Higgins, D., Gibson, T.: Clustal w: improving the sensitivity of progressive multiple sequence alignment through sequence weighting, position-specific gap penalties and weight matrix choice. Nucleic Acids Research 22, 4673–4680 (1994)

A New Hybrid Approach to Predict Subcellular Localization

179

23. Pittner, S., Kamarthi, S.V.: Feature extraction from wavelet coeffi-cients for pattern recognition tasks. IEEE Trans. Pattern Anal. Mach. Intell. 21, 83–88 (1999) 24. Zhou, G.P.: An intriguing controversy over protein structural class prediction. J. Protein Chem. 17, 729–738 (1998) 25. Zhou, G.P., Assa-Munt, N.: Some insights into protein structural class prediction. Proteins: Structure, Function, and Genetics 44, 57–59 (2001) 26. Vapnik, V.: Statistical Learning Theory. Wiley, New York (1998)

Support Vector Machine for Prediction of DNA-Binding Domains in Protein-DNA Complexes Jiansheng Wu, Hongtao Wu, Hongde Liu, Haoyan Zhou, and Xiao Sun* State Key Laboratory of Bioelectronics Southeast University, Nanjing 210096, China [email protected], [email protected]

Abstract. In this study, we present a classifier which takes an amino acid sequence as input and predicts potential DNA-binding domains with support vector machines (SVMs). We got amino acid sequences with known DNAbinding domains from the Protein Data Bank (PDB), and SVM models were designed integrating with four normalized sequence features(the side chain pKa value, hydrophobicity index , molecular mass of the amino acid and the number of isolated electron pairs) and a normalized feature on evolutionary information of amino acid sequences. The results show that DNA-binding domains can be predicted at 74.28% accuracy, 68.39% sensitivity and 79.76% specificity, in addition , at 0.822 ROC AUC value and 0.549 Pearson’s correlation coefficient.

1 Introduction Essential functions are performed by many proteins through interactions with nucleic acid molecules. For instance, transcription factors binding to specific cis-acting elements in the promoters result in regulation of gene expression[1]. Therefore, it is important for understanding a serial of biological processes to identification of the amino acid residues binding DNA or RNA. It is much helpful to understand the mechanisms of protein–nucleic acid interactions by analyzing structural data. The information provided from analyzing structural data has been used to predict DNAbinding residues in solved protein structures from amino acid sequence data, which have rapid increasement from many organisms [2,3,4,5,6]. In the past, artificial neural networks was been constructed coding with sequence information and residue solvent accessibility for prediction of DNA-binding residues, and got the performance which is 40.3% sensitivity and 81.8% specificity [2]. Evolutionary information, that is a position-specific scoring matrix (PSSM), was been shown to improve the predictive performance to 68.2% sensitivity and 66.0% specificity [5]. Recently, support vector machines (SVMs) combining with three simple sequence features was for the prediction of DNA and RNA-binding residues ,and the performance was at 70.31% accuracy,69.40% sensitivity,70.47% specificity and 0.7542 ROC AUC value [6]. As we known, protein–nucleic acid interactions are indeed that nucleic acid interact with some specified domains of proteins ,not just residues. It is actually more important to predict DNA-binding domains than DNA-binding residues. But till *

Corresponding author.

K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 180–187, 2007. © Springer-Verlag Berlin Heidelberg 2007

Support Vector Machine for Prediction of DNA-Binding Domains

181

now, no work have been done to predict nucleic acid -binding domains. In the present study , we developed a support vector machine based algorithm to predict the DNA-binding domains and got significantly high performance. We show that our SVM models can predict DNA-binding domains at 74.28% accuracy, 68.39% sensitivity and 79.76% specificity, in addition , at 0.822 ROC AUC value and 0.549 Pearson’s correlation coefficient.

2 Materials and Methods 2.1 Data Set PDNA-62, an amino acid sequence dataset, was utilized to construct SVM models for predicting DNA-binding domains. The PDNA-62 dataset was from 62 structure files of typical protein–DNA complexes and had less than 25% identity among the sequences [2,5,6].A segment of residues was appointed as a binding domain if it existed one or more amino acid residues in which any atoms had the distance less than a cutoff of 3.5 Å from any atoms of the DNA molecule in the complex [2, 5,6]. All the other segments of residues were designated as non-binding domains. we developed a Perl program which input a set of structure files and output a result file of amino acid sequences in which each segment of residues were labeled as a binding or non-binding domain according to the above cutoff scale. The PDNA-62 dataset contains 3667 DNA-binding domains and 3966 non-binding domains. 2.2 Feature of DNA-Binding Domains As in the previous study [6], the length of each domain was assigned as 11 in this study. The sum of (n-10) domains were extracted from an amino acid sequence with n residues. A domain that was DNA-binding was labeled with 1 (positive), or -1 (negative) was labeled if the target domain was non-binding. Each residue among every domain was coded with four biochemical features, where three were described in preference [6](the side chain pKa value, hydrophobicity index , molecular mass of the amino acid) and another new feature (the number of isolated electron pairs) was presented in this study. The three features described in preference [6] were normalized to get that the average is 0 and the standard deviation is 1 by the following standard logistic functions [7]:

Sα(i ) =

σα= P

( Pα(i ) − Pα)

σα

(1)

P

20 ⎛ 2 0 ∑ Pα ( i ) 2 − ⎜ i=1 ⎝ 400

20

∑

i=1

⎞ Pα ( i ) ⎟ ⎠

2

(2)

where S is normalized property values, α is the index of the property and i stands for the amino acid. P is the property value, and σPα are the average and the standard deviation of property α.

182

J. Wu et al. Table 1. The number of isolated electron pairs of each amino acid

Amino acid Isolated electron pairs Amino acid Isolated electron pairs

ALA

CYS

ASP

GLU

PHE

GLY

HIS

ILE

LYS

LEU

0

1

2

2

0

0

2

0

1

0

MET

ASN

PRO

GLN

ARG

SER

THR

VAL

TRP

TYR

1

2

1

2

3

1

1

0

1

1

The number of isolated electron pairs is related to the potential for hydrogen bond , which is the main force for DNA-protein binding, by residue in protein-DNA complex. The list of the sum of isolated electron pairs of each kind of amino acid are provided in table 1. In this work, PSI-BLAST program was utilized to get multiple sequence alignment profiles. We firstly downloaded the updated non-redundant (NR) protein sequence database (ftp://ftp.ncbi.nlm.nih.gov/blast/db/). Position-specific score matrices (PSSMs) were obtained using PSI-BLAST with three rounds and a cutoff E-value of 0.001 against the filtered NR database through masking out low-complexity regions and coiled coil segments. The PSSMs elements were scaled to 0–1 range by the standard logistic function [8]:

f ( x) =

1 1 + exp( x)

(3)

For a domain with 11 residues, the input vector consists of 264 values, including 44 biochemical feature values and 220 PSSMs values. 2.3 Support Vector Machine The SVM, introduced by Vapnik [9], is a learning algorithm for two- or multi-class classification problems and is known for its good performance. The basic principle of SVM is: for a given data set xi Rn (i = 1,... N) with corresponding labels yi (yi = +1 or -1, representing the two classes to be classified), SVM gives a decision function (classifier):

f(x) = sgn(

N

∑

i=1

yα i iK ( x, xi ) + b )

(4)

where αi are the coefficients to be learned and K is a kernel function. Parameters αi are trained through maximizing function : N

1 N ai − ∑ ai a j yi y j K ( xi , x j ) ∑ 2 i , j =1 i =1 N

where subject to 0≤ai≤C(i=1,…N) and

∑a i=1

i

yi = 0 .

(5)

Support Vector Machine for Prediction of DNA-Binding Domains

183

For this study, libsvm (http://www.csie.ntu.edu.tw/~cjlin/libsvm) was used for data training and classifying[10].To ensure that the parameter estimation and model generation of SVM are independent of the test data, a 5-fold cross-validation approach was used to evaluate the classifier performance. The original data set were randomly divided into five parts, then alternately use one subset for testing and the other four sets for training in each of the five iterative steps. We used different kernel functions(linear function, polynomial function and radial basis function) and different values for libsvm parameters to optimize the prediction accuracy in our experiments. The best results were obtained by using the radial basis function kernel with C = 1 and γ=0.009. 2.4 Measurement of Algorithms Performance The predictions for the test data instances are compared with the corresponding class labels (binding or non-binding) to evaluate the classifiers. The overall accuracy, sensitivity , specificity for assessment of the prediction system are defined as

TP + TN TP + TN + FP + FN

(6)

Sensitivity=

TP TP+FN

(7)

Specificity=

TN TN+FP

(8)

Accuracy =

， ，

where TP TN FP and FN are the number of true positives, true negatives, false positives and false negatives respectively. To give a better comparison and balance of sensitivity and specificity of the models , the net prediction [11]is defined as

Net Prediction=

Sensitivity +Specificity 2

(9)

The algorithms performance in this study is also evaluated by receive operating characteristic curve (ROC curve). ROC curve is a useful technique for organizing classifiers and visualizing their performance[11]. ROC graphs are two-dimensional graphs in which true positive rates (sensitivity) is plotted on the Y axis and false positive rate (1- specificity) is plotted on the X axis. Random guessing would generate identical false positive and true positive rates on average. Therefore, the diagonal (y = x) in the ROC plot is the performance of random guessing. The ROC curves move towards the upper left corner, indicating rising accuracy of performance. The area under the ROC curve (Area under curve, AUC) can be used to characterize the performance of a classifier for siRNA sequences. The AUC value ranges from 0.5 to 1, and a rising AUC value indicates higher accuracy of performance.

184

J. Wu et al.

Pearson’s correlation coefficient (r value) based on the classifiers’ output values and the labels (1 or -1) were also applied to evaluate algorithms performance. The higher r value indicates the better algorithms performance [12].

3 Results and Discussion The performance of our SVM classifiers in 5-fold cross validations were shown in the classifier named features 1 of table 2. The results showed that our classifier named features 1 for DNA-binding domains achieves 74.28% overall accuracy (SD=0.87)with 68.39% sensitivity(SD=0.39) ,70.47% specificity( SD=1.89) and 74.07% net prediction(SD=0.97)(table 2). Table 2. Performance of difference features by SVM for prediction of DNA binding residues in proteins Classifier Features 1 a Features 2 b

Accuracy±SD c (%) 74.28±0.87 64.95±1.10

Sensitivity±SD (%) 68.39±0.39 60.39±2.18

Specificity±SD (%) 79.76±1.89 69.16±1.57

Net prediction d ±SD(%) 74.07±0.97 64.77±1.14

Features 1: including the side chain pKa value, hydrophobicity index , molecular mass of the amino acid , the number of isolated electron pairs and PSSMs; Features 2: considering the same feature (the side chain pKa value, hydrophobicity index , molecular mass of the amino acids) as preference [6]; SD: Standard deviation of five iteratives’ performances in 5-fold cross-validation; Net prediction: the average of sensitivity and specificity; The ROC curve our classifier named features 1 was shown in Figure 1. The AUC value is 0.822 and the Pearson’s correlation coefficient is 0.549 of our classifier named features 1 for prediction of DNA-binding domains ,and both of them were shown in table 3. These AUC values are significantly higher than that of random guessing (0.5). The algorithm output has significant correlation with the sample labels (1 or -1) (p 0.9 ⎩

(5)

Comprehensive Fuzzy Evaluation Model for Body Physical Exercise Risk

233

(2) v2, v3, v4 membership function:

⎧0, x < 0.5, ⎪ 1.2 ⎪⎛ x − 0.5 ⎞ ,0.5 ≤ x ≤ 0.675, ⎜ ⎟ ⎪⎝ 0.175 ⎠ . ⎪ F2 ( x ) = ⎨1,0.675 ≤ x ≤ 0.725, ⎪ 1.2 ⎪⎛⎜ 0.9 − x ⎞⎟ ,0.725 ≤ x ≤ 0.9 ⎪⎝ 0.175 ⎠ ⎪ ⎩0, x ≥ 0.9 ⎧0, x < 0.3, ⎪ 1.2 ⎪⎛ x − 0.3 ⎞ ,0.3 ≤ x ≤ 0.475, ⎪⎜⎝ 0.175 ⎟⎠ . ⎪ F3 ( x ) = ⎨1,0.475 ≤ x ≤ 0.525, ⎪ 1.2 ⎪⎛⎜ 0.7 − x ⎞⎟ ,0.525 ≤ x ≤ 0.7 ⎪⎝ 0.175 ⎠ ⎪ ⎩0, x ≥ 0.7

⎧0, x < 0.1, ⎪ 1. 2 ⎪⎛ x − 0.1 ⎞ ,0.1 ≤ x ≤ 0.275, ⎪⎜⎝ 0.175 ⎟⎠ . ⎪ F4 ( x) = ⎨1,0.275 ≤ x ≤ 0.325, ⎪ 1.2 ⎪⎛⎜ 0.5 − x ⎞⎟ ,0.375 ≤ x ≤ 0.5 ⎪⎝ 0.175 ⎠ ⎪ ⎩0, x ≥ 0.5

(6)

(7)

(8)

(3) v5 membership function:

⎧1, x < 0.1, ⎪ 1 .2 ⎪⎛ 0.3 − x ⎞ F 5( x) = ⎨⎜ ⎟ ,0.1 ≤ x ≤ 0.3 . ⎪⎝ 0.2 ⎠ ⎪0, x > 0.3 ⎩

(9)

Table 2 shows a calculation illustration for the example we used above. The fuzzy values of factors have been calculated. The weight matrix of various subject have been gotten in step 2. We use the membership function above to compute the evaluation matrix of FCE, R.

Ⅳ

Step : Overall physical exercise risk remark evaluation. The composite weights of the risk remarks are then determined by aggregating the weights through the

234

Y.Wu, Y. Ding, and H. Xu

hierarchy. Using the Eq. (2) from bottom level to the top, we get the overall weights of exercise risk remark, B={0.02,0.5,0.4,0.1,0.01}. Then, the normalized overall evaluation value is computed by Eq. (3) to be: 0.6. The value is between the remark values of moderately danger and that of medium danger, so this person doing current exercise condition has a slight danger. The conclusion is consistent with the evaluation from the subject’s medical expert. Table 2. Evaluation matrix of human exercise risk calculation example R

Factor

Weights

Sub factor

Weights

Factor value

v1

Medical Condition

0.569

Dis-con

0.311

0.6

0.00

Phy-sta

0.539

0.6

0.00

Bas-inf

0.150

0.3

0.00

Intens.

0.557

0.73

0.10

Dur.

0.320

0.5

0.00

Freq.

0.123

0.67

0.00

T/H

0.623

0.6

0.00

T/W

0.239

0.2

0.00

Noise

0.137

0.5

0.00

0.334 Activity load

Environmental condition

0.097

v2 0.5 1 0.5 1 0.0 0 0.9 7 0.0 0 0.9 7 0.5 1 0.0 0 0.0 0

v3 0.5 1 0.5 1 0.0 0 0.0 0 1.0 0 0.1 2 0.5 1 0.0 0 1.0 0

v4 0.0 0 0.0 0 1.0 0 0.0 0 0.0 0 0.0 0 0.0 0 0.5 1 0.0 0

v5 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.44 0.00

4 Conclusions This is the first time to combine AHP and FCE to analysis and assess the risk level to people’s health when doing exercises. It is full of significant theoretical and practical meaning in healthcare monitoring. There are great advantages using these two techniques together. AHP deals with complex problem hierarchically and give a concrete method to acquire the comparison matrices. On the other hand, FCE takes into account vague and imprecise medical expert’s knowledge, patient information, exercise load and environment condition scientifically and express them concretely in mathematical form. Compared with other study in exercise risk research [3-5], mostly by testing or by field experiment, our comprehensive model can be implemented in wearable embedded system to monitor body’s current status, to assess the risk and to sound an alarm in time when needed. Of course, the theory and methodology presented in the paper need to be further studied and verified in practice. Acknowledgements. This work was supported in part by Program for New Century Excellent Talents in University from Ministry of Education of China (No. NCET-04415), the Cultivation Fund of the Key Scientific and Technical Innovation Project

Comprehensive Fuzzy Evaluation Model for Body Physical Exercise Risk

235

from Ministry of Education of China (No. 706024), International Science Cooperation Foundation of Shanghai (No. 061307041), and Specialized Research Fund for the Doctoral Program of Higher Education from Ministry of Education of China (No. 20060255006).

References 1. Jang, S.J., Park, S.R., Jang, Y.G., et al.: Automated Individual Prescription of Exercise with an XML-based Expert System. In: IEEE-EMBS, Shanghai, China, pp. 882–885 (2005) 2. Hara, M., Mori, M., Nishizumi, M.: Differences in Lifestyle-related Risk Factors for Death by Occupational Groups: A Prospective Study. Journal of Occupational Health 41, 137– 143 (1999) 3. Jouven, X., Empana, J.P., Schwartz, P.J., et al.: Heart rate profile during exercise as a predictor of sudden death. The new England journal of medicine 352, 1951–1958 (2005) 4. Yun-jian, Z., Ji-rao, W., Song-bo, Z.: The Application of HRV in the Healthy and Sports’ Field. Sichuan Sports Science 2, 47–49 (2004) 5. Singh, M.A.F.: Exercise comes of age: rationale and recommendations for a geriatric exercise prescription. J. Gerontol. A Biol. Sci. Med. Sci. 57, 262–282 (2002) 6. Kangary, R., Riggs, L.S.: Construction risk assessment by linguistics. IEEE Trans. Eng. Manag. 36, 126–131 (1989) 7. Saaty, T.L.: The Analytic Hierarcy Process. McGraw-Hill, New York (1980) 8. Feng, S., Xu, L.D.: Fuzzy Sets and Systems. 105, 1–12 (1999) 9. Speed, C.A., Shapiro, L.M.: Exercise prescription in cardiac disease. THE LANCET 356, 1208–1209 (2000) 10. Heyward, V.H.: Advanced fitness assessment & exercise prescription. Human Kinetics, 1– 6 (2006) 11. Donatelle, R.T., et al.: Access to Health, Benjamin Cummings, pp. 20–30 (1996)

The Effect of Map Information on Brain Activation During a Driving Task Tao Shang1, Shuoyu Wang2, and Shengnan Zhang1 1 School of Information Science and Engineering, Shenyang University of Technology No.58, Xinghuanan Street, Tiexi District, Shenyang, 110023, Liaoning Province, P.R. China [email protected] 2 Department of Intelligent Mechanical System Engineering, Kochi University of Technology, 185 Miyanokuchi, Kami, Kochi 782-8502, Japan [email protected]

Abstract. Until now, GPRS/GPS/GIS based on vehicle navigation and monitoring systems have been popularly developed to satisfy the demand for the intelligent transportation system (ITS). Such systems provide the large traffic convenience to drivers, but at the same time attach more burdens to drivers for learning about map information. Hence it is worth further verifying the negative effect of vehicle navigation and monitoring systems on drivers. Considering that human driving behavior is strongly relevant to cognitive characteristics, this study will address to the effect of vehicle navigation systems on drivers by means of measuring and analyzing the cognitive state inside brain. In this paper, a relatively new method of multi-channel nearinfrared spectroscopy (NIRS) was used to investigate the brain activation by independently manipulating the cognitive demand in the different cases of a driving simulator. Experimental results indicated that, compared with the case of no map information available, there is no more obvious priority of activation for left brain and right brain in the case of map information available. Meanwhile, there seems to be a complete activation for the prefrontal cortex of left and right brain, suggesting that GPRS/GPS/GIS based vehicle navigation systems may exhaust drivers more easily so as to bring about more danger than traffic convenience under driving environment.

1 Introduction The Intelligent Transport System (ITS) has been comprehensively paid attention to so far. Many modern technologies and products have been popularly developed to satisfy the demand for ITS [1-2]. As a kind of typical product, GPRS/GPS/GIS based on vehicle navigation and monitoring systems now have been applied on automobile for assisting safe and comfortable driving especially within the developed countries such as USA and Japan. Such systems provide the large traffic convenience to drivers, but at the same time attach more burdens to drivers for learning about map information. Hence it is worth further verifying the negative effect of vehicle navigation and monitoring systems on drivers. Considering that human driving

，

K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 236–245, 2007. © Springer-Verlag Berlin Heidelberg 2007

The Effect of Map Information on Brain Activation During a Driving Task

237

behavior is strongly relevant to cognitive characteristics, this study will address to the effect of vehicle navigation systems on drivers from the viewpoint of human cognitive characteristics. If such characteristics could be clarified, it will contribute to the disclosure of the negative effect of vehicle navigation systems on drivers and further the development of modern traffic tools. A few attempts involving human’s driving model [3-6] have been reported so far. However, most of those researches focused on external modeling for driving behavior and remained in lower level of human cognition process [3-5]. Comparatively, the literature [6] adopted predictive method to explore cognitive process of driving from the viewpoint of vision model, but it was difficult to verify the resulting model due to the limitation of owning measure device. Now with the rapid improvement of neurophysiology and electronic technology, some advanced measure devices are developed and the access to finding the internal signal of brain becomes possible. For example, multi-channel near-infrared spectroscopy (NIRS) is a relatively new method to investigate the brain activation, based on changes in oxygenated haemoglobin (O2Hb) and deoxygenated haemoglobin (HHb). Recently, it has been shown that NIRS seems to be able to detect even small changes in O2Hb and HHb concentration due to cognitive demands[7]. With respect to higher cognitive functions, NIRS has been successfully used to assess prefrontal brain activation during various tasks[8][9]. Based on these positive results and on the fact that NIRS is relatively easy to apply, this study will adopt the changes in O2Hb and HHb concentrations of the cerebral cortex to analyze human cognitive characteristics during a driving task. In order to investigate the negative effect of vehicle navigation systems on drivers from the viewpoint of human cognitive characteristics, here, we shall not only construct a driving simulator to induce human’s adaptive driving action, but also measure the changes in O2Hb and HHb concentrations of the cerebral cortex under the different environments of no map available and map available, respectively. The remaining part of this paper is organized as follows: in the section 2, the introduction to experimental system, including a developed driving simulator and used spectrometer. In the section 3, the experimental procedure is proposed. In the section 4, the experimental result and analysis is summarized. The paper finishes with a conclusion in the section 5.

2 Experimental System During the driving process, drivers need to avoid obstacle and arrive at a goal. As a result, driving behavior usually involves huge danger for human life. Considering the minimized economic damage and maximized safe guarantee, alternatively, a driving simulator was developed by using virtual reality technology. Not only driving behavior data can be acquired for further analysis, but also the limit state of human driving can be measured, which is the most distinctive feature compared with real driving. Here, we used the driving simulator as visual stimulus source to elicit a specific cerebral activation pattern. Besides, the measure device of brain activation is also necessary. Near-infrared spectroscopy (NIRS) is a non-invasive technique to measure concentration changes of oxygenated (O2Hb) and deoxygenated hemoglobin (HHb). With respect to higher

238

T. Shang, S. Wang, and S. Zhang

cognitive functions, NIRS has been adopted in many research [7-9], however, no specific NIRS activation paradigm has been reported for the measurement of frontal regions during a driving task. 2.1 Driving Simulator A driving simulator was developed just as shown in Figure 1. Virtual driving environment is implemented on the 17-inch LCD (Resolution 1280×1024). Keyboard is mounted as interaction control device for the driving direction and speed. The specification for computing environment is described as follows: Pentium4-2.66GHz CPU, 512MB memory, windows XP Professional, and the program developed by Visual C++ 6.0 and OpenGL. Refresh time is 100ms for animation.

(a)Driving scene

(b) Keyboard control

(c) Reference route (circle point denotes starting point, arrow denotes end point) Fig. 1. Driving simulator

The Effect of Map Information on Brain Activation During a Driving Task

239

Map information was designed on the right-bottom side of the driving simulator, just as shown in Figure 1-(a). When the middle scroll bar of main window is drawn to the right margin, the driving case becomes the case of no map information available for experiment. Meanwhile considering the predictive performance for practical traffic construction, the traffic scene was designed according to the 216 Route of Shenyang Public Transportation Company of Liaoning Province. Figure 1-(c) shows the reference route. All objects are drawn according to the proportion of 1 pixel: 2 meters. The initial velocity of a virtual car is zero. According to the motion equation (1), the motion trajectory of virtual car can be calculated. ⋅⋅ ⋅ m Y + D Y = F ( force, direction )

(1)

Where m : the quality of the car, 10 kilogram; Y = ( y1 , y 2 )T : the position of the car; ⋅

⋅⋅

Y , Y : the velocity and acceleration of the car, respectively;

⎡5 0 ⎤ D : the coefficient matrix ⎢ ⎥; ⎣0 5⎦ F ( force, direction ) : the composition control based on force and direction ; force : the control value from forward key, 10N; o o direction : the control value from left and right keys, -30 and +30 .

2.2 Hitachi ETG-7100 For near-infrared optical topography, we used 66(22*3) channel spectrometer (Hitachi ETG-7100). A 3*5 array of 8 laser diodes and 7 light detectors was applied resulting in 22 channels of a probe. This was realized by the fact that the near-infrared laser diodes were coded by different wavelengths each in two ranges around 695 and

Fig. 2. ETG-7100

240

T. Shang, S. Wang, and S. Zhang

830 nm (ranges: 675–715 nm and 810–850 nm). The photodiodes (light detectors) were located 30mm from an emitter. Fibers were mounted on a plastic helmet that was held by adjustable straps on the experimenter’s head. The measurement covered an area of 6.75 cm*11.25 cm centered over the electrode position. Light emitting was time coded with 10 kHz. The ETG-7100 monitor measured changes of O2Hb and HHb from a starting baseline. Data were measured with a sampling rate of 10 Hz and further analyzed using the ETG-7100 software. The scale of the hemoglobin quantity is mmol*mm, which means that changes in the concentration of O2Hb and HHb depend on the path length of the near-infrared light in the brain.

3 Experimental Procedure As the frontal lobe and occipital lobe seemed to be relevant for desired cognitive function, three probes were localized over the left prefrontal cortex, the right prefrontal cortex and occipital visual cortex, respectively, just as shown in Figure 3. Oxygenation changes with 22 channels of each probe can be measured during the driving process.

Fig. 3. Position of three probes

Stimuli were presented with the above experimental driving simulator. The stimuli come from a computer monitor placed 60 cm in front of an experimenter. The experimenter has to respond adaptively. Then the response data was recorded and used as a result of behavioral performance. With the ETG-7100 software, the changes of the concentrations of HHb and O2Hb were calculated over the experimental session. For task repetition, we defined a 30 s “baseline”, preceding each active task period (lasting 180 s) and a 30 s “rest” time period following each active task period. For our data, the time course of the measured data was corrected by the ETG-7100 software. After this correction, the resulting data was exported by the ETG-7100 program into ASCII data format and

The Effect of Map Information on Brain Activation During a Driving Task

241

video format. The concentration changes of HHb and O2Hb in the three experimental phases are further analyzed for each channel by means of two-tailed cut. In more detail, we first compared the initial condition separately to the corresponding baseline for each channel. This is done for [O2Hb] and [HHb] to ensure that the conditions lead to signs of cortical activation, that is, to an increase of [O2Hb] and a corresponding decrease of [HHb].

4 Measure Experiments 4.1 Subject One healthy and right-handed experimenter (male, age = 30 years old) was arranged to drive the car from the starting point to end point according to the specific 216 Route twice. The experimental cases include the case of no map information and map information available. Experimenter was free of medication, with no former or actual neurological or psychiatric disorder. 4.2 NIRS Data for No Map Information Available For the prefrontal cortex, the typical tracings for the changes in concentrations of O2Hb during the driving task are displayed in Figure 4. The concentration of O2Hb Time:30s

Time:44s

Time:88s

Time:74s

Time:92s

Fig. 4. Activation change of O2Hb for left and right prefrontal cortex

242

T. Shang, S. Wang, and S. Zhang

increased in partial channels during the active phase compared to the baseline, and subsequently the concentration area of O2Hb for left prefrontal cortex expanded in a smaller scope, while the concentration area of O2Hb for right prefrontal cortex changed the position of concentration area. Finally [O2Hb] declined over the time course of the rest phase compared to the active phase. Time:120s

Time:150s

Fig. 5. Activation change of O2Hb for occipital visual cortex

For the occipital visual cortex, the typical tracings for the changes in concentrations of O2Hb during the driving task are displayed in Figure 5. The concentration of O2Hb varied in partial channels during the active phase. 4.1 NIRS Data for Map Information Available For the prefrontal cortex, the typical tracings for the changes in concentrations of O2Hb during the driving task are displayed in Figure 6. The concentration of O2Hb

Fig. 6. Activation change of O2Hb for left and right prefrontal cortex

Fig. 7. Activation change of O2Hb for occipital visual cortex

The Effect of Map Information on Brain Activation During a Driving Task

243

increased in partial channels during the active phase compared to the baseline, and subsequently the concentration area of O2Hb for left prefrontal cortex expanded in a larger scope, while the concentration area of O2Hb for right prefrontal cortex expanded to a complete scope. Finally [O2Hb] still kept active over the time course of the rest phase compared to the active phase. For the occipital visual cortex, the typical tracings for the changes in concentrations of O2Hb during the driving task are displayed in Figure 7. The concentration of O2Hb keeps high in almost all channels. 4.2 Discussion As expected, in both cases of no map information available and map information available, we found a significant increase in O2Hb over measured frontal brain areas during the active phases compared to the baseline. Most importantly, both of left and right brain were activated in the active phase, with significantly higher concentrations of O2Hb. Meanwhile, the results indicated a missing activation of the prefrontal cortex sometimes. Therefore we can conclude that the driving task of the specific route mainly activates the partial prefrontal cortex. The activation map of brain illustrates specific effects of brain areas of the prefrontal cortex. But these differences between left and right prefrontal cortex did not reach significance. One conclusion can be drawn that in spite of changing the computational load imposed by a given traffic scene, there seems to a tendency of stable and symmetrical activation for left and right brain. In order to compare different cases, more details are illustrated as below: 1)

2)

For the case of no map available, according to the activation maps of O2Hb of prefrontal cortex, the effect that we also found significantly higher [O2Hb] in several points during the active phase compared to the neighboring area was probably caused by the fact of the hemodynamic response. At least two kinds of activation pattern can be found, just as shown in Figure 4. Since the pattern suggests the functional relevance of the frontal lobe for driving tasks, one conclusion is that it is possible to analyze the process of cognitive changes according to the physical position of prefrontal cortex. According to the activation maps of O2Hb of occipital visual cortex, two kinds of activation pattern can be found. Furthermore, one conclusion can be drawn that as environment becomes complex, left brain plays an initiative pole, while right brain closely follows towards the activation degree of left brain. Left brain focuses on problem solving, while right brain promotes active level. There seems to be collaborating areas of brain which each has multiple relative specializations and engages in extensive inter-area collaborations. For the case of map information available, according to the activation maps of O2Hb of prefrontal cortex, right brain completely activates in almost same degree, whereas the left brain uses the only partial brain. Only one kind of activation pattern can be found. Meanwhile, there is no more obvious priority of activation for left brain and right brain. According to the activation maps of O2Hb of occipital visual cortex, the occipital visual cortex keeps active almost in all areas of brain with the same

244

T. Shang, S. Wang, and S. Zhang

degree. Based on the above facts, it can be concluded that the case of map information available costs more energy than that of no map information available. Consequently drivers will become tired more quickly. If the case of map information available is looked on as the case with GPRS/GPS/GIS based on vehicle navigation and monitoring systems, the case of no map information available is looked on as the case without corresponding system, the kind of system will exhaust drivers more quickly so that the driver may face more potential danger, although it helps drivers search path conveniently. One limitation of our study is the missing specific effect for the HHb concentrations. In contrast to the specific concentration changes in O2Hb, concentration of HHb decreased in both conditions during the active phase compared to the baseline. In contrast to other studies claiming that HHb is more sensitive than O2Hb, our explanation for the absence of HHb is that, given the wavelengths used by the ETG-7100 system, [O2Hb] estimations are considerably more precise than estimation of [HHb], so that weaker effects in this parameter might not become statistically significant. Of course, this proposed analysis should be considered in further studies. At the same time, this shortcoming of low spatial resolution of the ETG-7100 equipment has been verified.

5 Conclusions In this paper, based on developed driving simulator, we used the NIRS method to investigate the functional activity of the cerebral cortex during the different driving tasks and discussed the negative effect of vehicle navigation and monitoring systems on drivers. After establishing a consistency with earlier research, the study produced three conclusions: firstly, in spite of changing the computational load imposed by a given traffic scene there seems to be a stable effect in a number of collaborating areas of brain, including left brain and right brain. It suggests that driving cognitive process that operates on different levels of environment may nevertheless draw on a shared infrastructure of cortical resource. Secondly, for the case of no map information available, left brain plays an initiative pole, while right brain closely follows towards the activation degree of left brain so that there is a tendency of left and right brain for symmetrical activation, suggesting that it may contribute to the construction of an easier model for driving cognitive process. Thirdly, for the case of map information available, there is no more obvious priority of activation for left brain and right brain, but with high activation degree. Meanwhile, there is a complete activation for the prefrontal cortex of left and right brain. Since map information attaches more burdens on brain, it suggests that the GPRS/GPS/GIS based on vehicle navigation and monitoring systems may exhaust drivers more easily so as to bring about more danger than traffic convenience under driving environment. We believe such conclusions will provide guide to explore well to those issues of developing computational theories for cognitive process, but also contribute to rehabilitation of those with cognitive deficits.

The Effect of Map Information on Brain Activation During a Driving Task

245

References 1. Tsugawa, S.: Automobile Driving Support Adjustable for Drivers. AISTTODAY 4, 12 (2004) 2. Akamatsu, M.: Driving Support System Based on Driving Action Data. AISTTODAY 14, 11 (2004) 3. Nechyba, M.C., Xu, Y.: Human Control Strategy: Abstraction, Verification and Replication. IEEE Control Systems Magazine 17, 48–61 (1997) 4. Koike, Y., Doya, K.: A Driver Model Based on Reinforcement Learning with Multi-Step State Estimation. Trans. IEICE Japan D-II 84, 370–379 (2001) 5. Shihabi, A., Mourant, R.R.: A Framework For Modeling Human-Like Driving Behavior For Autonomous Vehicles in Driving Simulators. In: Proc. The Fifth International Conference on Autonomous Agents, pp. 286–291 (2001) 6. Mizutani, K., Saito, G., Omori, T., Ogawa, A.: A Feasibility study of Cognitive Computation Model for Driver’s Process Estimation from driving Behavior. The Transactions of the Institute of Electrical Engineers of Japan, 967–975 (2005) 7. Jasdzewski, G., Strangman, G., Wagner, J., Kwong, K.K., Poldrack, R.A., Boas, D.A.: Differences in the hemodynamic responseto event-related motor and visual paradigms as measured by nearinfrared spectroscopy. Neuroimage 20, 479–488 (2003) 8. Fallgatter, A.J., Strik, W.K.: Reduced frontal functional asymmetry in schizophrenia during a cued continuous performance test assessed with near-infrared spectroscopy. Schizophrenia Bulletin 26, 913–919 (2000) 9. Herrmann ∗, M.J., Ehlis, A.-C., Wagener, A., Jacob, C.P., Fallgatter, A.J.: Near-infrared optical topography to assess activation of the parietal cortex during a visuo-spatial task. Neuropsychologia 43, 1713–1720 (2005)

Worm 5: Pseudo-organics Computer and Natural Live System Yick Kuen Lee1 and Ying Ying Lee2 1

Sun Yat-sen University, Software School [email protected] 2 University of Leeds, School of Biological Science [email protected]

Abstract. Life began hundred million years ago; it started from simple inorganic substances to intricate multi-cellular organics. Gradually, the brains of higher animals developed emotions and intelligence. These are well illustrated by the learning abilities and social behaviors of man. Those intelligent activities, progress from simple to complicate, primitive to sophisticated processes from incarnate to abstract. Man started to create artificial intelligent to enhance their brain capabilities sixty years ago. Here we are making a comparison between the natural and artificial intelligent, and see what we can emulate more from Nature. And disclose the author’s point of view about the creation of natural lives.

1 Introduction In a repetitive environment, events will happen with a sequenced order, with some rules govern them. Organisms can remember what happen before, and use knowledge from past experiences to predict what will happen in the future, give proper reaction. This enhances the survival of those organisms. For instance, Earth rotates around the solar system and spins about its own axis just provide such cycles are the fundamental dogma of physics and astrology. Homeostasis is the basis of survival in organisms. This delicate balance is maintained by constant interaction of the environment and organism. The organism picks up information from the ambient environment, and a proper cellular response is then elicited in the system. This instinct prolongs survival. Another characteristic of life is reproduction. Organisms can produce offsprings with similar genetic material. So even with the death of the older generation, their next generation survives. This life cycle continues the proliferation and survival of a species. The two characteristics discussed above involve information amplification. These processes require energy, which is obtained from the organism’s surrounding. However, the resources from the ambient environment are finite. Moreover, it is the nature of life to end. The equation of life entails development, reproduction and, inevitably, death. K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 246–253, 2007. © Springer-Verlag Berlin Heidelberg 2007

Worm 5: Pseudo-organics Computer and Natural Live

247

Mutation occurs during reproduction of organisms. In addition, influence exerted by external environment enhanced the mutation. Such mutates create a diversified gene pool, resulting differing phenotypes to manifest. Darwinism states that organisms with the most desirable phenotypes would survive and proliferate. This process of natural selection constitutes to evolution of organisms. Before the 20th century, science was divided into a multitude of different domains such as, Mathematics, Physics, Chemistry, Biology, Pathology, and Philosophy. Life sciences were more closely associated with philosophy and religion studies by many. Little did they know that genetics has a molecular basis when genetic theory basis. The study of human genome has united the different fields of studies. Men already possess the ability to create artificial intelligence. Compared to Nature, which went through millions of years of evolution, this new advancement has only a brief history of less than a century. There are more to be explored in this exciting field. The age of information revolution has already begun. Where do we begin? In the 21st century, theoretical physics developed further into String Theory1 while time began when the ‘Big Bang’ occurred. If life is defined as maintaining homeostasis and reproduction, it could encompass different energy levels in atomic structure to black holes in the galaxy, or to silicon polymer chain. Under different temperature, pressure and time frame, these seemingly ‘lifeless’ materials may display properties pertaining to life. However, to simplify matters, we shall stick to carbon chain, which constitutes all organisms.

2 Genetic Memories - Steady and Slow Mechanism Genetic information is encoded using four different types of bases. They are namely adenine (A), thymine (T), cytosine (C) and guanine (G). A series of DNA transcription and translation should code for proteins that are essential for life. From basic structure of cells to intricate cellular mechanisms, proteins are always, some way or another, involved. Due to the highly stable structure of DNA, the whole process of mutation is slow and dreary. Therefore, evolution does not happen in a split second; it takes many, many generations to occur. It takes a long time to interact with its environment.

3 Prokaryotes Versus Von-Neumann Machines Prokaryotes, like bacteria, are the simplest life forms that exist. They are made up of a single cell. Only a bi-phospholipid layer separates the internal and external environment of the organism. Like the Maxwell’s devil that can distinguish black and white particles in statistic theory, the protein channels embed in lipid wall can distinguish molecules and determine which can go pass the bilipid layer. DNA in the chromosome codes the inheritance information. During transcription, RNA polymerase use the template strand of DNA from to assembled a strand of mRNA2. Those mRNA will than translated into its corresponding polypeptide or enzyme by ribosomes and RNA

248

Y.K. Lee and Y.Y. Lee

polymerase during translation. Proteins (e.g. enzymes) initiate and regulate cellular activities. Organisms assimilate materials in the outside environment, and start the reproduction mechanism at maturity to produce more individuals to proliferate. Using the Von-Neumann machine, as a normal computer. Steady binary signals recorded in magnetic or optical media, coding the program and data information, CPU use Cluster of records as a template transcript them into high speed cache, from which program execute CPU instructions and manipulate electric signals as well as program instruction itself, and perform data input and output. Transformation of electric signals initiates intelligent activity, and control those intelligent activities themselves. Under control of human, program may be regurgitated and develop continuously. The two mechanisms mentioned above have many similarities. The core information is stored in a simple digital format in sequence and can be retrieved segment by segment. After spreading, the information transforms into a more power status, adjusting its environment, and feedback its signals, control complicate activities flow. Natural organisms maintain their lives, and reproduce offspring generation by generations. The born, growth, reproduction mutation become a complete cycle. And go through by the organism itself, no external interference will necessary except absorb energy from the environment. On the other hand, the name Von-Neumann Machine implies incompleteness, Man make computer machine hardware, software are written by man, turn on, turn off, and management all done by human. The computer development and evolution depends on human too. As a tool for extension human intelligent it cannot independently exist without interference from mankind.

4 Clustering of Lives - Eukaryotes, Multi-cell Organisms Prokaryotes exchange information through chemical reaction between molecules. The free path length and moving speed limits their sizes. So organism with simple structure is usually small. Individuals come together and form a larger body by communication. High coherence, low coupling function group usually clustered together, more often than not, the organization proves to be adaptable to the environment Eukaryotes could be treated as a cluster of lysosome, ribosome, mitochondrion, nucleus, etc. similar to prokaryotes packed together[1]. For higher animals, the structure of organization complicates to give rise to different hierarchies within the system. Cells in multi-cellular organism differentiate into different cells types to form tissues, organs and systems. Organisms can come together to form a community, like a herd of lions. In the field of artificial intelligent, single instructions usually combine into subroutines, and put together to be routine library, and high coherence, low-coupled structures will put together to form objects. Then come operating systems and application systems. Hierarchy is an important aspect in big systems. For hardware, single PC usually connected into a network, enterprises network most probably include function groups of file server, data base, mail server etc. then, linked together by internet.

Worm 5: Pseudo-organics Computer and Natural Live

249

5 Data Carrier - Nerve Network and Internet To maintain homeostasis in an ever-changing environment, one mechanism is negative feedback. For example, cells selectively synthesize lactase to break down lactose when there is no glucose present for glycolysis. Under the regulation of cAMP, dictyostelium will transform from amoebae to fruiting body[4]. Incretion, exocrine and inducer during embryo fetation are example of chemical signals. Chemical signal promulgate only in a small range, as the space enlarged, the speed of molecule diffusion limits propagation efficiency. Nervous system then developed in multi-cellular organism during evolution. Sensory receptor detects signals by inducing a change in the electric potential across cell membrane, transmitted along axons of nerve cells, to central nervous system (CNS) for interpretation and response. Electrical signal can be conducted more rapidly than chemical signals. Organism with nerve system survives better in a competitive environment. Together with the development of CNS, Man could communicate with voice or writing hundreds of years before, nations distributed in separated area depends on geographical location. Nowadays, by using electric signals, telegram, telephone, radio, shorten the nation’s distances, unifies the culture of different area. During the last ten years, optical fiber and Internet technology enhance the communication, increase efficacy and efficiency; making it more possible for globalization. Will Internet become a higher lever nervous system of human being?

6 Logical Network - Another Form of Memory Inchoate organism, like jellyfish, their nerves are divided into two types - sensory and motor, without CNS[2]. The batter like annelid, neurons cell bodies cluster in ganglion, have fixed format reaction to stimulations signals. The numbers of cells in eelworms are fixed[4] determined their genetic materials. Nerve memory is not a feature in them. For higher animals, CNS are highly developed. Sensations such as touch, smell, vision, hungry, cool, warm and pain can be detected. The sensory receptors of the peripheral nervous system transit impulse from sense organ to CNS, exciting the corresponding area. CNS processes such signals in sensory cortex of the brain, relays information to the motor cortex to transmit impulse to the appropriate effecter organs2. These reactions effect their survival. Nervous conduction includes chemical signals, such as direct acting substances like neurotransmitter (e.g. adrenaline and dopamine), neuropepties and Adenosine Triphosphate (ATP). In the early stage of the life, genetic information controls the structure and function of the nervous system. Most of the brain structure is an empty cavity4. However, as the system develops, sensory signals occur simultaneously, when the stimulation is above threshold level, or it happens frequently enough, synaptic connections form between the dendrites and axons in the brain cells. They connect and build a memory system based on the topology of the nervous network.

250

Y.K. Lee and Y.Y. Lee

There are some theories states that: Regulated by hippocampus in the brain, the excited cells would be connected by new nerves5, after that, stimulate only part of the cells that would transmit to others, thus, conditioned reflex, or memory forms. Conditioned reflex changes the physical structure of the brain, creating character of an individual. Compare nerve reaction with artificial process controller, the leaky channels, ligand channels, and voltage-gated channels in the dendrite and axon, are quire similar to the logical gate in electronic circuit where input signals, after transformation by the gate array, transmitted to control output devices. Field Programmable Gate Array (FPGA) is just like the nervous network in the brain[3]. However, programming of such device has to be performed by human at this stage, they cannot change their structure according to the input strength or frequency. Search engines in the Internet find the target address, redirect the signal, and get the necessary information from there. This type of operation, works in cooperation of Von-Neumann machine and logical gates. Pseudo-organic computer, like “WORM” (see below), however, has appropriate hardware, change communication connection between elements according to their structure instructions[7]; simulate the operation mode in the brain, making a foundation to the simulation of conditioned reflex.

7 Formation of Knowledge - Abstract Abilities Higher organisms have nervous system, connecting stimulate and reactions for body survival. Emotion expressed in pleasure, anger, sadness, and cheer, etc. That is combination a series of activities, including body movements, expression in face, sounds, incretion and secretion etc5. Comfortable feeling initiate organism’s intention; Avoiding pain is a basic instinct of most organisms. Those reactions, shows sequence or association of ordered events. Higher animals can acquire knowledge from past experiences, concludes rules that may predict what might happen in the future, making decisions earlier. As the brain developed in different stage, different species, having their means of reaction in quire a different ways, direct or indirect, sort terms and long terms, simple to complicate, incarnate and abstract. A lot of memories may have similarities and forms more abstract concepts. Abstract concepts can reduce memories element required, enlarge the range of concept, increases information process efficiency. For example, from many incarnate concepts like cow, sheep, grass, flower, tree, bird, door, window, we can distinguish the more abstract concepts of creature, plant, building structure, etc. The concept of abstract can be organized by hierarchy. The higher animals have the more abstract concept forming ability. Language is one of such abstract connecting concepts and voice. An ability to predict is intelligent or knowledge. A huge quantity of data, after cleaning, integration, selection, transformation, mining, and pattern evaluation, can be converted to useful information, concluded to rules to make predictions. Most work in developing artificial intelligent is knowledge discovery (mining) from large database6, search for frequent item sets; calculate the degree of association, degree of support, degree of confidence, finding associated relations. This way of knowledge discovery is similar to what the Natural does.

Worm 5: Pseudo-organics Computer and Natural Live

251

However, most of the artificial intelligent developed depends on human programming in a Von-Neumann machine. Some of which can keep the training data in permanent memories, but it is still relatively primary, cannot find their goal automatically, no emotion, no self-recognize, no reproduction, and cannot develop independently.

8 Formation of Characteristic, Self-cognize Genes records the information of species, reproduce similar organism, creating the innate instincts. However, the individuals of higher organism have their own characteristics. The characteristics of each individual are generated by memories during their growth. Genes stores the common traits of a species. Memories and experiences influence an individual’s characteristics. Organic clone, like identical twins, may have identical genes, but the personal experiences may not be the same when they grow up. Albeit the identical genetic constituents, twins may very well have different personality traits. Who has seen two trees that are replica of each other? Who am I? When you look into a mirror and think who is that, you may be making a soul-searching journey. Yourself, just are a sequence of memories start from your childhood, perceive of outside world, goals of tracing and evasiveness brought from your genes and experiences.

9 Man, Civilization and Abstract Lives Out of all the life forms that exist, Man stood out from the rest with advanced abstract intellect and imagination. This gift of intelligence allows people to acquire sophisticated language abilities. They were able to express emotions and distinguish themselves from others. Early civilization relied on verbal interactions to impart knowledge and history. As Man progressed, they were able to use writings to record knowledge and history. This more tangible form of documentation allowed Man to accumulate year’s experiences and pass down to the next generation. In 16th century Europe, the emergence of printing supplements the spread of knowledge and information. With that, it boosted the development of ancient science to modern technology. Drawing and writing are forms of memory of individuals, shared by a group of people. It is an improvement beyond innate memory. Most organisms’ memories disappear as they die. However, writing can impart human knowledge from generations to generations. We know the cataclysmic stories of ancient times and learn Euclid geometry because our ancestors have written down their knowledge so that we can learn. Writing creates human culture, experiences of a nation for thousands of years, can be recorded in books. Printing technology allows spread of knowledge to be easier. Schools and universities regulated education bring up more systematic sciences. Culture is an abstract form of life, Abstract life can be defined as any form of existence have memory ability, can reproduce itself. A nation, a society, an

252

Y.K. Lee and Y.Y. Lee Table 1. Comparison of Natural and Artificial Organism

Existing Years Simplest Form Hierarchy Examples

Nonneural Organism 3.8 billion

Neural Organism 400 million

Bacteria

Jellyfish

Cell fungi protist

Von-Neumann Machine

Computer Net-work

100

60

40

Electricrelay

Personalcomputer

Worm Fish Mammal human Neural gate infrastructure

TTL IC ProcessControl FPGA Electric gate network

Minicomputer Mainframe Seq-uence of binary code

Local area network Internet

Organ Function groups Neural pulses in parallel

Function groups

Files Program Objects Electric pulse words in Series 500 mb/s per bus channel

Logic Gate

Memory Format

Sequence of tetrad code

Structure Units

Organelle

Internal Data Carrier

Chemical Molecule

Propagate Speed

Molecular diffusion

100 ms/m in parallel

1 us/gate in parallel

Data Processor

Multiple ribosome

Multiple axon gate

Multiple local gate

Process Manner

Segment and parallel

Hierarchy parallel

parallel

Process Cycle Efficiency

seconds

0.05 seconds high

1 us Medium

Medium

Electric pulse in parallel

PseudoOrganic Computer --WORM

Binary code with infrastructure Hosts Nodes Servers Electric pulse packets in Series 100-1000 mb/s per line

Binary code with infrastructure Hardwaresoftware objects Optical pulse in packets of words 10-40 gb/s per optical channel Multiple CPU and routers Hierarchy parallel of sequential elements

30 ps

Multiple CPU and routers Hierarchy parallel of sequential elements 30 ps

Low

Very high

Very high

Single or multiple CPU Segment and sequential

30 ps

organization, a company or even software can be treated as an abstract life. VanNeumann Machine is a basic form of artificial life; Internet is also an abstract artificial life.

10 A Sample of Pseudo-organic Computer – “WORM” After billions of years wash out, the natural organism evolved many brilliant methods that are worth imitating, generation of artificial intelligence.

Worm 5: Pseudo-organics Computer and Natural Live

253

Base on a sequence of binary code, Von-Neumann machine has a tight structure, easy to program. However, logic gate similar to the brain has more processing efficiency. Objects oriented programming technique, save information in distributed objects, which communicate each other by passing massage. Network of hardware or network of software together with its infrastructure constituted as part of memory information. Communication efficiency becomes an important issue. A pseudo-organic computer project, named “WORM”^, clustering a lot of computing resources by photo-electric coupling data bus to form an object oriented networked task services system. Jump out of the frame restricted by the VonNeumann Machines. Beside the instruction flow control, use structure control to combine the multiple processors7. New computer will improve the internal communication system; dynamically control the common bus and other resources.

11 Conclusions As a tool of prediction, Pseudo-organic computer imitate only the brain manipulating information, and does not involved in reproduction itself. Can man create an artificial intelligent organism with reproduction ability? The possibility of imitate organic seems no doubt; however, there are practical problems in technology. Furthermore, it is a very difficult problem. To make computer, or even something simpler, just like paper, metal wire, plastic sheet, etc. needs the entire industry cooperation system. Nowadays, only a few countries can produce full set of computer. It is far from making a single machine that can replace the whole industry system. Philosophers argue whether God creates man, or Man creates God, according to their own image. However it can be sure that men create computer according to their image. Although the artificial intelligent is still simple and crude, Pandora's box has already opened. Successful breaking through and translate the genetic code in 21st century prelude the information revolution. The secret of lives and intelligent would be disclosed, gradually.

References 1. Baaquies, B.E., Kwek, L.C.: Superstrings, Gauge fields and Black holes 2. Miller, L.: Biology. Pearson Education Inc. (2006) 3. Zeidman, B.: Designing with FPGAs & CPLDS. Mei Guo CMP Gong Si Jiao Chuen Beijing Hang Kong Hang Tian Da Xue Chu Ban She Chu Ban (2002) 4. Muller, W.A.: Translated by Huang Xiu Yin. Fa Yu Shen Wu Xue. Beijing Gao Deng Jiao Yu Chu Ban She (2000) 5. William, G.J., Stanfield, C.L.: Principle of Human Physiology, 2nd edn. Pearson Benjamin Cummings (2005) 6. Han, J., Kamber, M.: Data Mining: Concept and Techniques 7. Shen, X.B., Zhang, F.C., Feng, G.C., Che, D.L., Wang, G.: The Classification Model of Computer Architectures. Chinese Journal of Computers 26 (2005) ^

For more details WORM, do refer to Y.K Lee. 2006. Worm 1: A Pseudo-organic Computer’s System Structure.

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics of Coupled Neurons Jiang Wang, Xiumin Li, and Dong Feng Tianjin University, School of Electrical Engineering and Automation, 300072 Tianjin, China [email protected]

Abstract. We study the stochastic dynamics of three FitzHugh-Nagumo neurons with chemical coupling and electrical coupling (gap junction) respectively. For both of the coupling cases, optimal coherence resonance and weak signal propagation can be achieved with intermediate noise intensity. Through comparisons and analysis, we can make conclusions that chemical synaptic coupling is more efficient than the well known linear electrical coupling for both coherence resonance and weak signal propagation. We also find that neurons with parameters locate near the bifurcation point (canard regime) can exhibit the best response of coherence resonance and weak signal propagation.

1 Introduction Noise-induced complex dynamics in excitable neurons have attracted great interest in recent years. The random synaptic input from other neurons, random switching of ion channels and the quasi-random release of neurotransmitter by synapses contributes to the randomicity in neurons [1]. While in contrast to the destructive role of noise, such as disorder or destabilize the systems, in some cases noise play an important and constructive role for the amplification of information transfer. Particularly, in the presence of noise, special attentions have been paid to the complex behaviors of neurons that locate near the canard regime [2-7], where neurons can exhibit great sensitive to external signal. This is important and meaningful for weak signal processing which guarantees low energy consumption in biological systems. As investigated in [2;3;6], such neurons possess two internal frequencies which correspond to the standard spiking and small amplitude oscillations (Canard orbits) respectively. For the former, it is just the frequency of the most regular spiking behavior purely induced by intermediate noise intensity, which is known as Coherence Resonance (CR). For the latter, the subthreshold oscillations are critical in the famous Stochastic Resonance (SR) phenomenon, which describes the cooperative effect between a weak signal and noise in a nonlinear system, leading to an enhanced response to the periodic force [8]. K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 254–263, 2007. © Springer-Verlag Berlin Heidelberg 2007

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics

255

Recently, E.Ullner gave detailed descriptions of several new noise–induced phenomenon in the FHN neuron in [1]. They showed that optimal amplitude of highfrequency driving enhances the response of an excitable system to a low-frequency signal [9]. They also investigated the Canard-enhanced SR [3], the effect of noiseinduced signal processing in systems with complex attractors [10] and a new noiseinduced phase transition from a self-sustained oscillatory regime to an excitable behavior [11]. And in [12], C. Zhou etc. have demonstrated the effect of CR in a heterogeneous array of coupled FHN neurons. They find that both the decrease of spatial correlation of the noise and the inhomogeneity in the parameters of the array can enhance the coherence. However, most of the relevant studies considered the single neuron [3;5;13] or neurons with linear electrical coupling (gap junctions) [4;12;14;15]. Only in [16], another very important case—nonlinear pulsed coupled neurons with noise were investigated. In the case of chemical (nonlinear) coupling, they observed a substantial increase in the CR of Morris-Lecar models, in comparison with the (linear) electrical coupling. Therefore, inspired by [16] and based on our previous work on canard dynamics of chemical coupled neurons [17], we study the effects of chemical synapses on CR and the enhancement of signal propagation in three coupled FHN neurons, which locate near the canard regime and are subjected to noisy environment. In particular, in order to investigate the signal propagation, only one of the neurons is subjected to external periodic signal. With the optimal noise intensity, chemical coupled neurons, due to the selective couplings between individuals, can enhance CR and exhibit much better response of external signal than the electrical coupled ones. This paper is arranged as follows: in Sec. II, we give descriptions of the neuron model and two kinds of coupling; In Sec. III and , comparisons are made between the chemical coupled neurons and electrical coupled neurons for coherence resonance and information transfer respectively; finally, we make conclusions and discussions in Sec.V.

Ⅳ

2 Neuron Model and Coupling Description We consider three bidirectional coupled Fitz Hugh-Nagumo (FHN) neurons which is described by

1 3 ⎧  app syn −I ⎪ε Vi = Vi − 3 Vi − Wi + I i ⎨ ⎪W = V + a − bW + B cos(ω t ) + Aξ (t ) i i i i ⎩ i

(1)

where i = 1,..., N index the neurons, a, b and ε are dimensionless parameters with ε  1 that makes membrane potential Vi as fast variable and recovery variable Wi as slow variable. ξ i is an independent white Gaussian noise with zero mean and intensity A for each element. Bi cos( wt ) is the forcing periodic signal. I app and I isyn is the external applied current and the synaptic current through neuron i respectively. For the linear diffusive coupling (gap junctions),

256

J. Wang, X. Li, and D. Feng

I isyn =

∑

g syn (Vi − V j )

j∈neigh ( i )

(2)

where g syn is the conductance of synaptic channel. For the nonlinear pulsed coupling (chemical synapses), we refer to [18;19] and I isyn is defined as I isyn =

∑

g syn s j (Vi − Vsyn ),

j∈neigh ( i )

(3)

where g syn is the synaptic coupling strength and Vsyn is the synaptic reversal potential which determines the type of synapse. In this paper, considering the excitatory synapse, we take Vsyn = 0 . The dynamics of synapse variable s j is governed by V j , and it is defined as ⎧ s j = α (V j )(1 − s j ) / ε − s j / τ syn ⎪ α0 ⎨ ⎪α (V j ) = 1 + exp(− V V ) j shp ⎩

(4)

where synaptic decay rate τ syn is written as τ syn = 1 δ . The synaptic recovery function α (V j ) can be taken as the Heaviside function. When the neuron is in silent state V < 0 , s is slowly decreasing, the first equation of (5) can be taken as s j = − s j / τ syn ; while in the other case, s fast jumps to 1 and thus makes action to the postsynaptic cells. The parameters used in this paper are respectively a = 0.7 , ε = 0.08 , Vsyn = 0 , α 0 = 2 , Vshp = 0.05 , I app = 0 and the rest parameters are given in each case. In this model, b is one of the critical parameters that can significantly influence the dynamics of the system. For the single neuron in the absence of noise, AndronovHopf bifurcation happens at b = 0.45 . As b > 0.45 , it is excitable and corresponding to the rest state; while as b < 0.45 , the system possesses a stable periodic solution generating a periodic sequence of spikes. Between these two states there exists an intermediate behavior, known as canard explosion [19]. In a small vicinity of b = 0.45 , there are small oscillations near the unstable fix point before the sudden growth of the oscillatory amplitude. This canard regime tends to zero as the parameter ε → 0 . Here we take ε = 0.08 as used in [20], and in this case canard regime exists for b ∈ [0.425, 0.45] . This regime is much sensitive to external perturbations and thus plays a significant role in the signal propagation which will be further discussed below. We numerically integrate the system by the explicit EulerMaruyama algorithm [21].

3 Coherence Resonance Coherence Resonance (CR) is a noise-induced effect and describes the occurrence and optimization of periodic oscillatory behavior due to noise perturbations [1]. In

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics

257

this section, we study the effect of chemical synapses on the coherence resonance of coupled neurons, where Bi = 0, i = 1, 2,3 . As is discussed in [16], for large enough coupling strength g syn , time traces of electrical coupled neurons are basically identical, while in the chemical coupling case, there exists a slight delay between spikes and the subthreshold oscillations are different form each other (see Fig.1). Therefore, we only examine the coherence of 2 nd neuron instead of the mean field. (b) Electrical Coupling

2

(a) Chemical Coupling

2

1.5 1.5

1

V1&V2&V3

V1&V2&V3

1 0.5 0

0.5 0 -0.5

-0.5

-1 -1

-1.5 -1.5

-2 -2 0

20

40

60

80

100

0

10

20

30

40

50

60

70

80

90

t

t

Fig. 1. Time series of Vi ( i = 1, 2,3. ). (a) b = 0.45 , g syn = 0.27 , A = 0.11 ; (b) b = 0.45 , g syn = 0.1 , A = 0.19 .

We take S i = Tki

Tmean = Tki

t

t

Var (Tki ) , i = 2

and the average interspike interval

as the coherence factor of the firing events, where Tki is the pulse

internal: Tki = τ ki +1 − τ ki , τ ki is the time of the k th firing of the i th cell. ⋅ t denotes average over time. S describes the timing precision of information processing in neural systems. We study CR for these two kinds of couplings when the neurons locate near the bifurcation point b = 0.45 , where all the cells are in subthreshold regime in the absence of noise. In order to investigate the influence of coupling strength we calculate the maxim of S ( Sm ) at the corresponding optimal noise intensity for different values of g syn in two coupling cases respectively (see Fig.2 (a) (b)), where g syn = 0.15 in (a) and g syn = 0.1 in (b) are the smallest value for neurons to fire synchronously. It is obvious that both too weak and too strong coupling can decrease CR in each case. Therefore, we choose the optimal coupling strength g syn = 0.27 for chemical coupling and g syn = 0.1 for electrical coupling. Fig.2 (c) (d) shows the coherence resonance for two coupling cases. In both cases, Tmean decays quickly and tends to the period of normal spiking with the increase of noise intensity. While chemical coupling exhibits a significant increase of CR and need smaller noise intensity to achieve the optimal periodic oscillatory behavior than electrical coupling. Similar as discussed in [16], the interpretation of this phenomenon is that chemical synapses only act while the presynaptic neuron is spiking, whereas electrical coupling connect the voltage of neurons at all times. This can be observed in Fig.1 which shows the optimal case for each coupling. Chemical coupling ensures that small oscillatory neurons are free from each other and give more opportunities for

258

J. Wang, X. Li, and D. Feng

the individuals to fire induced by noise, compared with the electrical coupling case. And once one spikes, it will stir the others to spike synchronously. While for electrical coupling, the strong synchronizations between subthreshold oscillatory neurons result in the decrease of the oscillatory amplitude and thus the increase of the threshold for firing. And from this phenomenon we can learn that subthreshold oscillations are very important for the firing of large spikes. As b increases, where neurons locate far from the canard regime, CR declined in both of the two coupling cases (see Fig.2 (d)). (a) Chemical Coupling

(b) Electrical Coupling

5

6.2 4.8

6 4.6

S

5.6

m

m

S

5.8 4.4

4.2

5.4 4

5.2 3.8

5 0.15

0.2

0.25

0.3

g0.35 syn

0.4

0.45

0.5

3.6 0.1

0.55

0.15

0.2

0.25

(c)

0.35

0.4

0.45

0.5

CC EC

m

6

20

S

m ean

T

CC EC

30

10 0 0

0.3

gsyn

(d)

40

5 4

0.1

0.2

A

0.3

0.4

0.5

0.44

0.46

b

0.48

0.5

0.52

8 CC EC

0.2

2 0 0

0.1

0.2

0.3

A

0.4

0.5

m

0.15

4

A

S

6

CC EC

0.1 0.05

0.44

0.46

b

0.48

0.5

0.52

Fig. 2. (a)(b) The maxim of S for different g syn in two kinds of coupled neurons respectively, where b = 0.45 ; (c) CR factor S and Tmean versus the noise intensity A for the coupled system in different cases, where b = 0.45 , CC: g syn = 0.27 , EC: g syn = 0.1 ; (d) The maxim of S and the corresponding noise intensity Am for different parameter b in two kinds of coupled neurons respectively, where CC: g syn = 0.27 , EC: g syn = 0.1

4 Stochastic Resonance As mentioned above, Stochastic Resonance (SR) describes the optimal synchronization of the neuron output with the weak external input signal due to intermediate noise intensity. In this section, we take the parameters of input periodic signal Bi = 0.05 and ω = 0.3 so that there are no spiking for all the neurons in the absence of noise. In order to investigate the information transfer in these coupled neurons, we consider the local stimulus, that is, only one element is subjected to external periodic signal. So except for particular statement, Bi is taken as: B1 =0.05, B2 = 0, B3 = 0 . As is shown in Fig.3, there exists an optimal response of the neurons to input signal with intermediate intensity of noise. To evaluate the response of output frequency to the input frequency, we calculate the Fourier coefficient Q for the input signal. The definition of Q [1] is

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics

2

(a) Chemical Coupling (CC)

2

1

1

V1&V2&V3

V1&V2&V3

(b) Electrical Coupling (EC)

1.5

1.5

0.5 0 -0.5 -1

0.5 0 -0.5 -1 -1.5

-1.5 -2 0

259

-2 20

40

60

80

100

-2.5 0

20

t

40

t

60

80

100

Fig. 3. Time series of Vi ( i = 1, 2,3. ) and the input signal with a ten times higher amplitude than in the model (black line) B1 =0.05, B2,3 = 0, (a) b = 0.45 , g syn = 0.15 , A = 0.015 ; (b)

b = 0.45 , g syn = 0.12 , A = 0.045

ω 2π n / ω 2Vi (t )sin(ωt )dt , 2π n ∫0 ω 2π n / ω = 2Vi (t ) cos(ω t )dt 2π n ∫0

Qsin = Qcos

(6)

Q = Qsin 2 + Qcos 2 . Where n is the number of periods 2π / ω covered by the integration time. Similar as in Sec. we only examine the response of 2nd neuron to external input instead of the mean field, that is Vi = V2 in Eq.(6). And in neuron systems, for information is carried through the large spikes but not the subthreshold oscillations, we are only interested in the frequency of spikes. So following [3], we set the threshold Vs = 0 in the calculation of Q. If V < Vs , we replace V by the value of the fix point V f ; if V > Vs , we use the original value of V . We consider the differences between these two kinds of couplings for the signal processing we when the neurons locate near the bifurcation point b = 0.45 . Following Sec. choose the optimal coupling strength g syn = 0.15 for chemical coupling and

Ⅲ

Ⅲ

g syn = 0.12 for electrical coupling (Fig.4 (a) (b)). For local stimulus B1 =0.05, B2 = 0, B3 = 0 , chemical coupling is much efficient than electrical coupling for signal processing (Fig.4 (c) ~ (e)). As is discussed above, chemical coupled neurons can make better explorations of the internal sensitive dynamics and need smaller noise intensity to fire and complete signal processing than electrical coupled ones. And the subthreshold oscillations are still very important in this case. As b increases the ability of information transfer declined greatly in both of the two coupling cases (Fig.4 (d) (e)).

260

J. Wang, X. Li, and D. Feng

(a) CC

(b) EC

0.12

0.15

0.115

0.14 0.11

0.13

Qm

0.105

Qm

0.12

0.1

0.11

0.095

0.1

0.09

0.09 0.08 0.15

0.2

0.25

0.3

0.35 gsyn

0.4

0.45

0.5

0.085 0.1

0.55

0.15

0.2

0.3 gsyn

0.25

0.35

0.4

0.45

0.5

(c)

0.35

CC B2,3=0 EC B2,3=0 EC B2,3=0.05 CC B2,3=0.05

0.3 0.25

Q

0.2 0.15 0.1 0.05 0 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

A

(d)

0.09 CC EC

0.18

0.08 0.07

0.14

0.06

Qm

Am

0.16

0.12

(e) CC EC

0.05 0.04

0.1

0.03 0.08

0.02 0.06 0.04

0.01 0.44

0.46

b

0.48

0.5

0.52

0.44

0.46

0.48

b

0.5

0.52

Fig. 4. (a)(b) The maxim of Q for different g syn in two kinds of coupled neurons respectively, where b = 0.45 ; (c) Signal processing at the input signal versus the noise intensity for the coupled system in different cases, where b = 0.45 , B1 = 0.05 , CC: g syn = 0.15 , EC: g syn = 0.12 ; (d) (e) The maxim of Q and the corresponding noise intensity Am for different parameter b in two kinds of coupled neurons respectively, where CC: g syn = 0.15 , EC: g syn = 0.12

Besides, we also investigate the global stimulus B1,2,3 =0.05 , where each neuron is forced by the forcing signal. Here the chemical coupled system is not as efficient as the electrical coupled one for the response to input signal (Fig.4 (c)). In this case,

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics

261

neurons are more active and can fire easily induced by the external signal and noise. The continuous connection in electrical coupled neurons lead to high synchronization and can make better control of the firing rate than the selective connection in chemical coupled neurons see (Fig.5). However, this case is not common in real systems, where input signals are always weak and added to only a small amount of neurons for the sake of low energy consumption. (b) Electrical Coupling (EC)

(a) Chemical Coupling (CC) 1.5 1.5

1

V1&V2&V3

V1&V2&V3

1 0.5 0 -0.5 -1

0.5 0 -0.5 -1 -1.5

-1.5

-2

-2 380

400

420

440

t

460

480

500

520

-2.5 380

400

420

440

t

460

480

500

520

Fig. 5. Time series of Vi ( i = 1, 2,3. ) and the input signal with a ten times higher amplitude than in the model (black line). (a) b = 0.45 , g syn = 0.15 , A = 0.025 , B1,2,3 = 0.05 ; (b) b = 0.45 , g syn = 0.12 , A = 0.045 , B1,2,3 = 0.05 .

5 Conclusions In this paper, we have made comparisons of coherence resonance and the response to weak external signal between chemical coupled and electrical coupled noisy neurons. Chemical coupled neurons are prone to stir spikes due to its selective coupling, while electrical coupling is beneficial for synchronization. Therefore, as subjected to noisy environment and weak forcing signal, chemical coupling is more flexible and can increase the mutual excitations between cells, which enhance the coherence resonance and weak signal propagation. Also, the canard regime where system dynamics are complex and sensitive to external perturbations plays significant roles for the information transmission in neural systems. Besides, it should be noted that canard dynamics, which had been detailedly discussed in [19;22], is critical for signal processing. The number of subthreshold oscillations between two closest large spikes has close relationships to the firing rate, which carries the information during signal propagation. We will further this study and extend it to larger size of networks with different topological connections. Acknowledgements. The authors gratefully thank the valuable discussions with Wuhua Hu. And this paper is supported by the NSFC (No.50537030).

262

J. Wang, X. Li, and D. Feng

References 1. Ullner, E.: Noise-induced Phenomena of Signal Transmission in Excitable Neural Models. DISSERTATION (2004) 2. Perc, M., Marhl, M.: Amplification of information transfer in excitable systems that reside in a steady state near a bifurcation point to complex oscillatory behavior. Physical Review E 71(2), 26229 (2005) 3. Volkov, E.I., Ullner, E., Zaikin, A.A., Kurths, J.: Oscillatory amplification of stochastic resonance in excitable systems. Physical Review E 68(2), 26214 (2003) 4. Zhao, G., Hou, Z., Xin, H.: Frequency-selective response of FitzHugh-Nagumo neuron networks via changing random edges. Chaos: An Interdisciplinary Journal of Nonlinear Science 16, 043107 (2006) 5. Zaks, M.A., Sailer, X., Schimansky-Geier, L., Neiman, A.B.: Noise induced complexity: From subthreshold oscillations to spiking in coupled excitable systems. Chaos: An Interdisciplinary Journal of Nonlinear Science 15, 026117 (2005) 6. Makarov, V.A., Nekorkin, V.I., Velarde, M.G.: Spiking Behavior in a Noise-Driven System Combining Oscillatory and Excitatory Properties. Physical Review Letters 86(15), 3431–3434 (2001) 7. Shishkin, A., Postnov, D.: Stochastic dynamics of FitzHugh-Nagumo model near the canard explosion, Physics and Control, 2003. In: Proceedings. 2003 International Conference, vol. 2 (2003) 8. Wellens, T., Shatokhin, V., Buchleitner, A.: Stochastic resonance. Reports on Progress in Physics 67(1), 45–105 (2004) 9. Ullner, E., Zaikin, A., García-Ojalvo, J., Báscones, R., Kurths, J.: Vibrational resonance and vibrational propagation in excitable systems. Physics Letters A 312(5-6), 348–354 (2003) 10. Volkov, E.I., Ullner, E., Zaikin, A.A., Kurths, J.: Frequency-dependent stochastic resonance in inhibitory coupled excitable systems. Physical Review E 68(6), 61112 (2003) 11. Ullner, E., Zaikin, A., García-Ojalvo, J., Kurths, J.: Noise-Induced Excitability in Oscillatory Media. Physical Review Letters 91(18), 180601 (2003) 12. Zhou, C., Kurths, J., Hu, B.: Array-Enhanced Coherence Resonance: Nontrivial Effects of Heterogeneity and Spatial Independence of Noise. Physical Review Letters 87(9), 98101 (2001) 13. Gong, P.L., Xu, J.X.: Global dynamics and stochastic resonance of the forced FitzHughNagumo neuron model. Physical Review E 63(3), 31906 (2001) 14. Toral, R., Mirasso, C.R., Gunton, J.D.: System size coherence resonance in coupled FitzHugh-Nagumo models. Europhysics Letters 61(2), 162–167 (2003) 15. Casado, J.M., Baltanás, J.P.: Phase switching in a system of two noisy Hodgkin-Huxley neurons coupled by a diffusive interaction. Physical Review E 68(6), 61917 (2003) 16. Balenzuela, P., Garcia-Ojalvo, J.: On the role of chemical synapses in coupled neurons with noise. Arxiv preprint q-bio. NC/0502025 (2005) 17. Wang, J., Li, X., Hu, W.: Canards and Bifurcations in the Chemical Synaptic Coupled FHN Neurons (2006) 18. Drover, J., Rubin, J., Su, J., Ermentrout, B.: Analysis of a canard mechanism by which excitatory synaptic coupling can synchronize neurons at low firing frequencies. SIAM J. Appl. Math. 65, 69–92 (2004) 19. Wechselberger, M.: Existence and bifurcation of canards in R3 in the case of a folded node. SIAM J. Applied Dynamical Systems 4, 101–139 (2005)

Comparisons of Chemical Synapses and Gap Junctions in the Stochastic Dynamics

263

20. Cronin, J.: Mathematical Aspects of Hodgkin-Huxley Neural Theory. Cambridge University Press, Cambridge (1987) 21. Higham, D.J.: An algorithmic introduction to numerical simulation of stochastic differential equations. SIAM REVIEW 43, 525–546 (2001) 22. Szmolyan, P., Wechselberger, M.: Canards in R3. Journal of Differential Equations 177(2), 419–453 (2001)

Distinguish Different Acupuncture Manipulations by Using Idea of ISI Jiang Wang1, Wenjie Si1, Limei Zhong2, and Feng Dong1 1

School of Electrical engineering and automation, Tianjin University, 300072, Tianjin, P.R. China [email protected] 2 School of Information engineering, Northeast Dianli University, 132012, Jilin, P.R. China

Abstract. As well-known, the science of acupuncture and moxibustion is an important component of Traditional Chinese Medicine with a long history. Although there are a number of different acupuncture manipulations, the method for distinguishing them is rarely investigated. With the idea of the interspike interval (ISI), we study the electrical signal time series at the spinal dorsal horn produced by three different acupuncture manipulations in Zusanli point and present an effective way to distinguish them. Comparing with the traditional analysis methods, like phase space reconstruction and largest Lyapunov exponents, this new method is more efficiently and effective.

1 Introduction The neural systems have strong nonlinear characters and will display different dynamics due to different system parameters or external inputs. Usually the dynamics of these systems experience little change when the parameters are slightly modified, but in vicinity of a critical point, the situation will be totally different. The systems would be driven from chaotic pattern to periodic pattern, one periodic pattern to another periodic pattern or from periodic pattern to chaotic pattern [1] [2]. Although there is an examination of intracellular membrane potential, most of the study is aimed at an easily obtained physiological measure, in particular ISI, to facilitate comparison with experimental data. ISIs play an important role in encoding the neuronal information which is conveyed along nerve fibres in the form of series of propagating action potentials. Continuing researches focus on the ISI sequence [3-12]. According to their work, an ISI can be seen as a state variable by which the temporal dynamics of the neuron can be characterized. In analogy with the Takens theorem [13] for discrete- or continuous-time dynamical systems, later generalized in [14], it should then be possible to reconstruct the essential features of the attractor dynamics of a neuron from measurements of only one variable, using e.g. delay embeddings on ISI sequences. Acupuncture is an important part of Chinese medicine theory and it is approved to be highly effective in treatment of more than 300 diseases [15]. Since the middle period of the 20th century, the applications of acupuncture have advanced in abirritation [16], quitting drug [17] and so on. Acupuncture at the Zusanli point is not only K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 264–273, 2007. © Springer-Verlag Berlin Heidelberg 2007

Distinguish Different Acupuncture Manipulations by Using Idea of ISI

265

utilized to treat diseases of common digestive system such as duodenum ulcer, acute gastritis and gastroptosis etc, but also has auxiliary efficiency on enteritis, dysentery, constipation, hepatitis, gallstone, kidney stone, diabetes and hypertension [18]. When acupuncture is applied to the Zusanli point, electrical activity can be recorded from the spinal dorsal horn. Different kinds of acupuncture manipulations can evoke various electrical time series and achieve different curative effect. Our paper is organized as follows. Section 2 provides the transmission path of the acupuncture signals background while section3 gives the time series evoked by different acupuncture manipulations. Section 5 introduces the ISI method for distinguishing the three different acupuncture manipulations to compare with the methods mentioned in section 4. The last part of this work is conclusion.

2 Transmission Path of the Acupuncture Signals According to the previous studies, the acupuncture signals follow a certain route from the acupuncture point to the spinal dorsal horn [19]. The corresponding transmission path for acupuncture signals at the Zusanli point is shown in Fig.1. Then the electrical signal time series at the spinal dorsal horn can be recorded.

Fig. 1. The transmission path of the acupuncture signals at the Zusanli point

3 The Time Series Evoked by Acupuncture There are twelve alternative manipulations used in acupuncturing at the Zusanli point. This paper selects three of them; they are the twist manipulation, the drag-plug manipulation and the gradual manipulation [20]. The time series at the spinal dorsal horn evoked by these three methods are shown in Fig.2.

266

J. Wang et al.

(a) twist manipulation

(b) gradual manipulation

(c) drag-plug manipulation Fig. 2. The time series evoked by three acupuncture manipulations

Distinguish Different Acupuncture Manipulations by Using Idea of ISI

267

4 Time Series Analysis 4.1 Phase Space Reconstruction Theory The phase space is reconstructed according to the delay coordinate method proposed by Takens [21] and Packard [22]. Here, define a discrete time array obtained by measurement or simulation, reconstruct the m-dimension state vector Xn by the delay coordinate method: (1)

.

where T is called time delay and m is embedding dimension. T and m are two important parameters in the phase space reconstruction. Values of T, m are obtained by the mutual information method [23] and Cao’s method [24], respectively. 4.2 Largest Lyapunov Exponents Based on the reconstructed phase space, we analyze the spatio-temporal behavior of the time series. The Lyapunov exponent is an important parameter for describing the non-linear system behavior. It states the rate of exponential divergence from initial perturbed conditions. Consider a one-dimensional map .

(2)

Assume the difference of the initial value x0 is δx 0 . The n-time iteration value is (3) . is Lyapunov exponent. The magnitude of the Lyapunov exponent is a measwhere urement of the sensitivity to initial conditions. The system is chaotic and unstable when the Lyapunov exponent is positive. For an n-dimensional map, the largest Lyapunov exponent (LLE) is preferred to estimate whether the system is chaotic or not. This paper adopts the method introduced by Wolf [25] to calculate the LLE of the time series at the spinal dorsal horn.html. 4.3 Experimental Data Processing We select 80,000 data points within 20s and reconstruct the phase space for these experimental data [26-29]. As to the data of the twist method shown in Fig.3, the time delay T=3 and the embedded dimension m=3, while T=3, m=3 and T=4, m=3 are chosen for the drag-plug method shown in Fig.4 and the gradual method shown in Fig.5, respectively.

268

J. Wang et al.

Fig. 3. The embedding parameters of the twist method

Fig. 4. The embedding parameters of the drag-plug method

The attractors of the reconstructed phase space are shown in Fig.6. They are all strange attractors for all the three methods even with different shapes according to the figures. So we confirm these signals are chaotic preliminarily. In addition, the LLEs according to the Wolf’s algorithm are calculated to quantitatively describe the time series. The calculation results of the LLEs are shown in Fig.7. The LLEs of the three methods are 1.7333 , 1.7676 , 1.7635 , respectively. Obviously, the difference among these LLEs is too small to help us to distinguish them clearly. Based on the attractors of the reconstructed phase space, as the situation of the former one, we can not differentiate which is which.

±0.0018

±0.0015

±0.0010

Distinguish Different Acupuncture Manipulations by Using Idea of ISI

Fig. 5. The embedding parameters of the gradual method

(a) twist method

(b) drag-plug method

(c) gradual method Fig. 6. The reconstructed attractors of the three methods

269

270

J. Wang et al.

(a) twist method

(b) drag-plug method

(c) gradual method Fig. 7. The LLEs of the three methods

5 Time Series Analysis Using the Idea of ISI As Seen from Fig.2, we could find that the amplitude in the time series almost ranges from -20mV to 20mV except for some points. In this work, we consider the point whose amplitude is larger than 30mV or smaller than -30mV as a quasi-spike point. All of the successive quasi-spike points with the same sign together make a quasispike, and it is apparent that a quasi-spike has one quasi-spike point at least. Then the authors take the analysis method of ISIs to study the three time series in Fig.2 to distinguish them. For the sake of convenience, we classify the quasi-spikes into two kinds: one with positive quasi-spike points as Pquasi-spike and the other with negative ones as Nquasi-spike. Using Kn denoting the number of the nth quasi-spike, the interquasi-spike intervals (quasi-ISIs) are given by

Distinguish Different Acupuncture Manipulations by Using Idea of ISI

271

(4)

.

For convenience, Pquasi-ISI is taken to denote the quasi-ISI of Pquasi-spikes while Nquasi-ISI for that of Nquasi-spikes. The graphs of quasi-ISIs are shown in Fig.3. Because the interval of two adjacent points in Fig.2 is a constant, the quasi-ISIs can also be measured by the time.

(a) Pquasi-ISI according to n

(b) Nquasi-ISI according to n

Fig. 8. quasi-ISI according to n

From Fig.8, we could easily distinguish the twist manipulation and the drag-plug manipulation from the third one, but can’t tell the differences between the former two correctly. So the means of Pquasi-ISIs and Nquasi-ISIs are evaluated respectively and the results are shown in Table 1. Table 1. Means of Pquasi-ISIs and Nquasi-ISIs Twist Mean of Pquasi-ISIs 2091.5 Mean of Nquasi-ISIs 2091.5

Drag-plug 2019.3 2503.4

Gradual 737.58 737.58

From Table 1, it is easy to differentiate the former two methods, because the means of Pquasi-ISI and Nquasi-ISI are different in the drag-plug method but the situation is reverse in the twist manipulation and the gradual manipulation. Thus this analysis method is an effective way to distinguish the different acupuncture manipulations.

6 Conclusions In this work, the authors developed a new method basing the idea of ISI to differentiate the different acupuncture manipulations. Compared with the traditional methods like the phase space reconstruction and the largest lyapunov exponent method, this

272

J. Wang et al.

method can distinguish the signals at the spinal dorsal horn evoked by the three acupuncture manipulations clearly. Through using it, the different manipulations produced by different doctors can be quantized, for example, a specialist finishes acupuncture for a sufferer, and the mean of Pquasi-ISIs and Nquasi-ISIs is measured as 2030 and 2400, respectively, then we can conclude that this specialist has used the twist and drag-plug manipulations and used the drag-plug manipulation more frequently. Ulteriorly, these data can be used to instruct doctors to acupuncture. Some common grounds have been discovered between the entrainment patterns under the external voltage situation and the signal time series produced by the acupuncture manipulations, even thought their models are different. For example, spikes or quasi-spikes can be found in both of them. There may be some relation between them, so our future work will focus on the problem of whether the acupuncture manipulations can be quantized into the external electric stimulus. And that will definitely lead the acupuncture theory to a more advanced level. Acknowledgments. The authors gratefully acknowledge the support of the NSFC (No.50537030).

References 1. Jianxue, X., Yunfan, G., Wei, R., Sanjue, H., Fuzhou, W.: Propagation of periodic and chaotic action potential trains along nerve fibers. Physica D: Nonlinear Phenomena 100(12), 212–224 (1997) 2. Eugene, M.: Izhikevich, Resonate-and-fire neurons. Neural Networks 14(6-7), 883–894 (2001) 3. Masuda, N., Aihara, K.: Filtered interspike interval encoding by class. Neurons Physics Letters A 311, 485–490 (2003) 4. Gedeon, T., Holzer, M., Pernarowski, M.: Attractor reconstruction from interspike intervals is incomplete. Physica D: Nonlinear Phenomena 178(3-4), 149–172 (2003) 5. Racicot, D.M., Longtin, A.: Interspike interval attractors from chaotically driven neuron models. Physica D: Nonlinear Phenomena 104(2), 184–204 (1997) 6. Jin, W.-y., Xu, J.-x., Wu, Y., Hong, L., Wei, Y.-b.: Crisis of interspike intervals in Hodgkin-Huxley model. Chaos, Solitons & Fractals 27(4), 952–958 (2006) 7. Tuckwell Henry, C.: Spike trains in a stochastic Hodgkin-Huxley system. Biosystems 80(1), 25–36 (2005) 8. Horikawa, Y.: A spike train with a step change in the interspike intervals in the FitzHughNagumo model. Physica D: Nonlinear Phenomena 82(4), 365–370 (1995) 9. Rasouli, G., Rasouli, M., Lenz, F.A., Verhagen, L., Borrett, D.S., Kwan, H.C.: Fractal characteristics of human parkinsonian neuronal spike trains. Neuroscience 139(3), 1153– 1158 (2006) 10. Canavier, C.C., Perla, S.R., Shepard, P.D.: Scaling of prediction error does not confirm chaotic dynamics underlying irregular firing using interspike intervals from midbrain dopamine neurons. Neuroscience 129(2), 491–502 (2004) 11. Gu, H., Ren, W., Lu, Q., Wu, S., Yang, M., Chen, W.: Integer multiple spiking in neuronal pacemakers without external periodic stimulation. Physics Letters A 285(1-2), 63–68 (2001) 12. Yang, Z., Lu, Q., Gu, H., Ren, W.: Integer multiple spiking in the stochastic Chay model and its dynamical generation mechanism. Physics Letters A 299(5-6), 499–506 (2002)

Distinguish Different Acupuncture Manipulations by Using Idea of ISI

273

13. Takens, F.: Detecting strange attractors in turbulence. Lecture Notes in Mathematicsb, vol. 898, pp. 336–381 (1981) 14. Sauer, T., Yorke, J.A., Casdagli, M.: Embedology. Journal of Statistical Physics 65, 579– 616 (1991) 15. Xuemin, S.: Acupuncture. China Press of Traditional Chinese Medicine, Beijing (2004) 16. Ke, Q., Wang, Y., Zhao, Y.: Acupuncture abirritation and its mechanism. Sichuan Journal of Anatomy 10(4), 224–230 (2003) 17. Yin, L., Jun, H., Qizhong, M.: Advance in Research on Abstinence from Narcotin Drugs by Acupuncture. Shanghai J. Acu-mox. 18(3), 43–45 (1999) 18. Zhang, J., Jin, Z., Lu, B., Chen, S., Cai, H., Jing, X.: Responses of Spinal Dorsal-horn Neurons to Gastric Distention and Electroacupuncture of ”Zusanli” Point. Acupuncture Research 26(4), 268–273 (2001) 19. Wan, Y.-H., Jian, Z., Wen, Z.-H., Wang, Y.-Y., Han, S., Duan, Y.-B., Xing, J.-L., Zhu, J.L., Hu, S.-J.: Synaptic transmission of chaotic spike trains between primary afferent fiber and spinal dorsal horn neuron in the rat. Neuroscience 125(4), 1051–1060 (2004) 20. Cheng, S.: Chinese Acupuncture 1998. People’s Medical Publishing House 21. Takens, F.: Detecting Strange Attractors in Turbulence. Lecture Notes in Mathematics, vol. 898, pp. 366–381 (1981) 22. Packard, N.H., Crutchfield, J.P., Farmer, J.D., Shaw, R.S.: Geometry from a Time Series. Phys. Rev. Lett. 45, 712–716 (1980) 23. Fraser, A.M., Swinney, H.L.: Independent coordinates for strange attractors from mutual information. Phys. Rev. A 33, 1134–1140 (1986) 24. Liangyue, C.: Practical method for determining the minimum embedding dimension of a scalar time series. Physica 110D, 43–50 (1997) 25. Wolf, A., Swift, J.B., Swinney, H.L., Vastano, J.A.: Determining Lyapunov exponents from a time series. Physica 16D, 285–317 (1985) 26. Yong, X., Jian-Xue, X.: Phase-space reconstruction of ECoG time sequences and extraction of nonlinear characteristic quantities. Acta Phys Sinica 51(2), 205–214 (2002) 27. Wang, Z.S., Zhenya, H., Chen, J.D.Z.: Chaotic behavior of gastric migrating myoelectrical complex. IEEE Trans. on Biome. Eng. 51(8), 1401–1406 (2004) 28. Matjaz, P.: Nonlinear time series analysis of the human electrocardiogram. Eur. J. Phys. 26, 757–768 (2005) 29. Small, M., Yu, D.J., Simonotto, J., Harrison, R.G., et al.: Uncovering non-linear structure in human ECG recordings. Chaos, Solitons & Fractals 13(8), 1755–1762 (2002)

The Study on Internet-Based Face Recognition System Using PCA and MMD Jong-Min Kim Computer Science and Statistic Graduate School, Chosun University, Korea [email protected]

Abstract. The purpose of this study was to propose the real time face recognition system using multiple image sequences for network users. The algorithm used in this study aimed to optimize the overall time required for recognition process by reducing transmission delay and image processing by image compression and minification. At the same time, this study proposed a method that can improve recognition performance of the system by exploring the correlation between image compression and size and recognition capability of the face recognition system. The performance of the system and algorithm proposed in this study were evaluated through testing.

1 Introduction The rapidly growing information technology has fueled the development in multimedia technique. However, demand for techniques involving searching multimedia data in alarge scale database efficiently and promptly is still high. Among physical characteristics, face image is used as one of the reliable means of identifying individuals. Face recognition system has a wide range of applications such as face-based access control system, security system and system automation based on computer vision. Face recognition system can be applied to a large number of databases but requires a large amount of calculations. There are three different methods used for face recognition: template matching approach, statistical classification approach and neural network approach[1].Elastic template matching, LDA and PCA based on statistical classification approach are widely used for face recognition[2, 3]. Among these methods, statistical classification-based methods that require a small amount of calculations are most commonly used for face recognition. The PCA-based face recognition method identifies feature vectors using a Kahunen-Loeve transform. Given the proven feasibility of PCA as face recognition method, this study used PCA along with Kenelbased PCA[4, 5] and 2D-PCA[6]. The real-time face recognition system proposed in this study will be available in a network environment such as Network. Each client is able to detect face images and forward detected images to remote server by compressing the images to reduce file size. However, the compression of facial images poses a critical risk because of the possibility of undermining image quality. This study investigated the effects of image K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 274–283, 2007. © Springer-Verlag Berlin Heidelberg 2007

The Study on Internet-Based Face Recognition System Using PCA and MMD

275

compression and image size on recognition accuracy of the face recognition systems based on PCA, KPCA, 2D_PCA algorithms and came up with the most effective realtime face recognition system that can be accessed across the Network.

2 Network-Based Face Recognition System Based on the assumption that multiple variations of the face improves recognition accuracy of face recognition system, multiple image sequences were used. To reduce transmission delay, the images were compressed and minimized in the proposed system Fig. 1.

Fig. 1. Composition of the Proposed Face Recognition System

3 Face Recognition Algorithms The real-time recognition accuracy was evaluated using PCA, KPCA and 2DPCAbased algorithms. 3.1 PCA(Principal Component Analysis) The PCA-based face recognition algorithm calculates basis vectors of covariance matrix ( C ) of images in the following equation.

C=

1 M

M

∑ (X i =1

i

− m)( X i − m)T

(1)

Where xi represents 1D vector converted from the i th image in a sequence of images in the size of m × n . m indicates average of total M images of training face.

276

J.-M. Kim

Maximum number of eigenvectors ( m × n ) of covariance matrix ( C ) of images are also calculated. Top K number of eigenvectors are selected according to descending eigenvalues and defined as basisvector ( U )[7]. Feature vectors( w ) of input image( x ) are distributed as basis vectors in the vector space according to the following equation (2):

w = U T ( x − m)

(2)

3.2 2DPCA While covariance matrix is computed from 1D images converted from input images for PCA, covariance matrix ( G ) is computed from 2D images and the average image for 2DPCA in the following equation (3) [6].

C=

1 M

M

∑ ( A − E ( A))( A − E ( A)) i =1

i

i

T

(3)

Eigenvalues/eigenvectors of covariance matrix of images are calculated. Top k number of eigenvectors according to descending values are defined as basis vectors ( U ). Feature vector ( wi ) of the i th image in the image sequence of face ( A ) are extracted in the equation (4). Characteristics of the face B = [ wi ,......, wk ] can be extracted from wi .

wi = Aui ,

i = 1,2,..., K

(4)

Compared with covariance matrix used for PCA analysis, the covariance matrix derived from input images for 2DPCA analysis is smaller. This means that 2DPCA has the advantage of requiring less learning time [6]. 3.3 KPCA(Kernel Principal Component Analysis) KPCA face recognition algorithm involves converting input data on a face image into an image using nonlinear functions Φ . The converted images are reproduced as eigenvectors of the covariance matrix calculated for a set of nonlinear functions Φ and coefficients obtained during this process are used for face recognition in KPCA analysis. For PCA, the covariance matrix can be efficiently computed by using kernel internal functions as the elements of the matrix [8,9]. In the equation (5), nonlinear function Φ (x) is substituted for input image x , and F was substituted for the feature space R N .

Φ : R N → F , xk → Φ ( xk )

(5)

The training matrix and covariance matrix of images in the nonlinear space are pre~ sented in the equations (6) and (7). The nonlinear function Φ in the equation (6) must be mean zero by meeting the requirements for normalization.

The Study on Internet-Based Face Recognition System Using PCA and MMD

C

Φ

=

1 l l

∑

l

∑

k =1

k =1

~ ~ Φ ( x k )Φ ( x k ) T

277

(6)

~ Φ (xk ) = 0

(7)

4 Face Recognition Rate 4.1 Changes in Recognition Rates with Image Compression Image compression is essential to shorten transmission time through the Network. However, compression approach also has a downside as it may hinder image quality. As presented in Fig. 2, data file size was reduced but subjective image quality deteriorated as quantization parameters of compressed images increased. As a result, the recognition performance of the system is expected to decline. It is however found that recognition rate was not varied by the value of quantization parameters at the original

(a)Original image

(b)QP=10

(c)QP=20

(d)QP=30

Fig. 2. Changes in visual image with QP value

(a) Changes in data size with QP value (b) Changes in recognition rate with QP value Fig. 3. Effects of image compression on data size and recognition performance

278

J.-M. Kim

(a)QP value of 5

(b) QP value of 15

(c) QP value of 30

Fig. 4. Changes to the distance between the original image and compressed image with QP value. (Triangle in an image with the minimum distance, rectangular indicates the closest image to other class, green represents the original image and blue represents an compressed image.)

image size of 92*122 pixels, as shown in Fig. 4 (b). Such a phenomenon was also confirmed by the distance between the original image and compressed image. Changes to the distance between the original image and compressed image with the value of quantization parameters are presented in Fig. 4. There was a positive correlation between the distance and QP value. In other words, the distance between the original image and compressed image ( d 3 ) increased to 59, 305 and 689 as the value of QP reached 5, 15 and 30, respectively. However, changes to the distance is meager, so the effects of compression can be eliminated, meaning almost no changes in recognition performance but a significant reduction in data file size to be forwarded. In conclusion, transmission time can be reduced without affecting recognition performance of the system. 4.2 Changes in Recognition Rates with Image Size The size of image has an impact on transmission time and computational complexity during the recognition process. Images were passed through a filtering stage to get the low-low band using wavelet transform. Image size ( S R ) is defined in the following equation:

SR =

S origin 4( R )

(8)

For instance, the original image is reduced to 25% of its actual size when R equals to 1. Effects of image filtering are presented in Fig. 5. Effects of image size on time required for learning and recognizing images and recognition performance are presented in Fig 6. As shown in Fig 6 (a) and (b), the time required for learning and recognizing images drastically fell as the size of the image was reduced. The recognition rate also dropped when R was less than 4 but stopped its decline and remained almost unchanged when R was 4 or above. In fact, it

The Study on Internet-Based Face Recognition System Using PCA and MMD

(a)original

(b)R=1

(c)R=2

279

(d)R=3 (e)R=4

Fig. 5. Effects of image filtering

(a) Learning time

(b) recognition time

(c) recognition rate

Fig. 6. Effects of image size on time required for learning and recognizing images and recognition rate

is difficult to recognize features of the image with eyes when the image size became smaller. This is due to the fact that image size reduction involves reducing the number of faces in original images and the size of coefficient vectors.

5 Majority-Making-Decision Rule The present study found that recognition rates remained almost unchanged in response to certain degrees of compression and minification of images. Based on these findings, the study proposes a face recognition algorithm capable of improving recognition performance of the system. The algorithm is designed to calculate a recognition rate based on the majority, composition and decision-make rules when multiple input images are used. A theoretical estimation of recognition rate ( Pm ) can be calculated in the following equation on the condition that more than half of transmitted images were matched with image models.

280

J.-M. Kim

Pm =

⎛n⎞ k ⎜⎜ ⎟⎟ p s (1 − p s ) n−k k = ⎣n / 2 ⎦ ⎝ k ⎠ n

∑

(9)

Where n is the number of images forwarded, Ps is the average probability of face

⎛n⎞

recognition ⎜ ⎟ is possible number when k number of features are recognized among ⎜k ⎟

⎝ ⎠

n number of features. ⎣x ⎦ is a fixed number that is smaller than x but the largest value. For instance, when Ps is 0.94 and three images are forwarded, the value of Pm is 0.99 based on the majority, composition and decision-making rules. The proposed algorithm was tested with 3, 5 and 7 images in PCA-, KPCA- and 2DPCA-based real-time face recognition systems. According to the equation (9), the saturation estimation was achieved when Ps was roughly larger than 0.7 and n equaled to 5. Five images were therefore used for the test of the proposed system. Test results are presented in Fig 7.

6 Experimental Results The composition of the proposed system is presented in Fig 1. For the test, ChosunDB (50 class, 12 images in the size of 60*1 20) and Yaile DB were used. The test was performed in a network environment, and the optimized value of quantization parameters was applied. Experimental results are presented in Table 1. The performance of real-time face recognition system is measured by the length of time required for learning and recognizing face images, total amount of data transmitted and recognition rate. The KPCA-based proposed system increased recognition rate by 14% and reduced the time required for recognizing images by 86%. The time required for learning images was reduced when smaller sizes of images were used. The 2DPCA-based proposed system showed the recognition rate of 96.4%, compared with 90.3% of the existing 2DPCA-based system. Besides, a 78% decrease was observed in learning time and a 24% decrease in recognition time in the same system. The amount of data transmitted was reduced to 3610 bytes from 19200 bytes, leading to a 81 reduction in transmission delay. Table 1. Comparison of performance between proposed and existing systems Algorithm PCA Proposed system(PCA) 2D PCA Proposed system(2D PCA ) KPCA Proposed system(KPCA)

Recognition Rate(%) 88.7 92.0 91.3 96.4 79.0 93.5

Training Time(min) 28 16 27 6 2.4(hour) 0.33(hour)

Recognition Time(sec) 1.5 1.0 0.5 0.38 36 5

The Study on Internet-Based Face Recognition System Using PCA and MMD

281

(a) Recognition Rate

(b) Training Time Fig. 7. Effects of image compression and minification on recognition rate and time required for the recognition process of five images

282

J.-M. Kim

(c) Recognition Time Fig. 7. (continued)

7 Conclusions This study proposed a real-time face recognition system that can be accessed across the network. The test of the proposed system demonstrated that image filtering and image compression algorithms reduced transmission delay and the time required for learning and recognizing images without hindering recognition accuracy of the system. This study used multiple input images in order to improve the recognition performance of the system, and the proposed real-time face recognition system proved robust on the network. Although the system was based on PCA algorithms, it can be integrated with other face recognition algorithms for real-time detection and recognition of face images.

References 1. Jain, A.K., Duin, R.W., Mao, J.: Statistical pattern recognition: a review. IEEE Trans. on Pattern Analysis and Machine Intelligence 22(1), 4–37 (2000) 2. Yambor, W.: Analysis of PCA based and Fisher Discriminant-based Image Recognition Algorithms. Technical Report CS-00-103, Computer Science Department Colorado State University (2000) 3. Murase, H., Shree K.N.: Visual Learning and Recogntion 3-Dobject from appearance. International journal of Computer Vision 14 (1995)

The Study on Internet-Based Face Recognition System Using PCA and MMD

283

4. Zhang, Y., Liu, C.: Face recognition using kernel principal component analysis and genetic algorithms. In: Proceedings of the, 12th IEEE Workshop on Neural Networks for Signal Processing, pp. 337–343 (2002) 5. Yang, J., Zhang, D., Frangi, A.F., Yang, J.Y.: Two-Dimensional PCA: A New Approach to Appearance-Based Face Representation and Recognition. IEEE Trans. Pattern Analysis and Machine Intelligence 26(1) (January 2004) 6. Bourel, F., Chibelushi, C.C., Low, A.A.:Robust facial expression recognition using a statebased model of spatially localised facial dynamics. In: Proceedings of Fifth IEEE International Conference on Automatic Face andGesture Recognition, pp.106–111 (2002) 7. Georghiades, A.S., Belhumeur, P.N., Kriegman, D.J.: From Few to Many: Illumination Cone Models for Face Recognition under Variable Lighting and Pose. IEEE Trans. Pattern Analysis and Machine Intelligence 23(6), 643–660 (2001) 8. Viola, P., Jones, M.: Rapid Object Detection using a Boosted Cascade of Simple Features. Computer Vision and Pattern Recognition 1, 511–518 (2001) 9. Yang, H.-S., Kim, J.-M., Park, S.-K.: Three Dimensional Gesture Recognition Using Modified Matching Algorithm. In: Wang, L., Chen, K., Ong, Y.S. (eds.) ICNC 2005. LNCS, vol. 3611, pp. 224–233. Springer, Heidelberg (2005) 10. Belhumeur, P.N., Hepanha, J.P., Kriegman, D.J.: Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection. IEEE Trans. Pattern Analysis and Machine Intelligence 19(7), 711–720 (1997) 11. Georghiades, A.S., Belhumeur, P.N., Kriegman, D.J.: From Few to Many: Illumination Cone Models for Face Recognition under Variable Lighting and Pose. IEEE Trans. Pattern Analysis and Machine Intelligence 23(6), 643–660 (2001) 12. Kim, J.-M., Yang, H.-S.: A Study on Object Recognition Technology using PCA in Variable Illumination. In: Li, X., Zaïane, O.R., Li, Z. (eds.) ADMA 2006. LNCS (LNAI), vol. 4093, pp. 911–918. Springer, Heidelberg (2006)

Simulation of Virtual Human's Mental State in Behavior Animation Zhen Liu Faculty of Information Science and Technology, Ningbo University, 315211, China [email protected]

Abstract. Human mental state is related to outer stimuli and inner cognitive appraisal, and it mainly includes emotion, motivation, personality and social norm. Modeling mental state of virtual humans is very important in many fields. Simulating virtual humans with mental state is a challenging branch of computer animation, where virtual humans are regarded as agents with sense, perception, emotion, personality, motivation, behavior and action. 3D virtual humans are constructed with sensors for perceiving external stimuli and are able to express emotions autonomously. Mental state-based animation is demonstrated in a prototype system.

1 Introduction Life simulation is the dream that people have been pursuing all the time. The development of artificial intelligence has promoted the advancement of traditional computer animation. The combination of computer animation and artificial intelligence is closer in recent years, and modeling 3D virtual humans have already caused extensive concerns from many fields. Artificial life is the research field that tries to describe and simulate life by setting up virtual artificial systems with the properties of life. Artificial life becomes a new method in computer animation [1][2], entertainment industry need cleverer virtual human with built-in artificial life model. Computer animation and artificial life are blended each other, intelligent virtual life, a new research field is born in the intersection between artificial life and computer animation. We can get more understanding from the developing history of computer animation. Early computer animation only includes shape and movement of geometry model, it is very difficult to draw complex natural landscape, and artificial life can help to solve these problems. In general, artificial life model is based on bottom-up strategy. Emergence is the key concept of artificial life. It means complex system is from the simple location interactions of individuals. Another key concept of artificial life is adaptation, which means evolution. In 80 years of the 20th century, many models of computer animation were presented, such as particle model, L-system, kinematics and dynamics, facial animation, etc. In 90 years of the 20th century, artificial life influenced the development of the computer animation greatly, many intelligent animation characters with perception were realized on computer systems, and behavior animation and cognitive model were the milestones of development of computer animation. K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 284–291, 2007. © Springer-Verlag Berlin Heidelberg 2007

Simulation of Virtual Human's Mental State in Behavior Animation

285

Behavior animation of virtual humans is becoming more and more important in computer graphics. People want to find a flexible and parameterized method to control locomotion of virtual human. Badler developed Jack software for virtual human animation [3]. Jack software is designed for simulating human factors in industrial engineering, they also built a system called Emote to parameterized virtual human and to add personality model for virtual humans [4]. N.M.Thalmann suggested that virtual human should not only look visual, they must have behavior, perception, memory and some reasoning intelligence [5], they also presented a personality model for avatars [6]. Gratch et al. presented a domain-independent framework for modeling emotion, they thought that people have beliefs about past events, emotions about those events and can alter those emotions by altering the beliefs [7]. Cassell et.al realized a behavior animation toolkit [8], and Pelachaud et.al presented a method to create facial expression for avatars [9]. A model of virtual human's mental state is presented in this paper, a believable 3D virtual human should be provided with the mechanism of mental variables that include emotion, personality, motivation and social norm (see Fig.1). Social norm includes status, interaction information and interaction rules, it controls the process of a nonverbal social interaction, it provides the social knowledge for virtual human. Based on Freud theory [10], ID is the inborn, unconscious portion of the personality where instincts reside, and it operates on the pleasure principle. Libiduo is a source of psychic energy. Ego is responsible for organizing ways in the real world, and it operates on the reality principle. Superego is the rulers that control what a virtual human should do or not. The research is mainly based on behavior animation, and the goal is setting up a mental state-based animation model for 3D virtual humans.

Fig. 1. Structure of a virtual human

2 Perception of Virtual Human In this paper, we only discuss the visual perception. Synthetic vision is an important method for visual perception [5], which can accurately simulate the vision from view of a virtual human, the method synthesis vision on PC. When a virtual human needs to observe the virtual environment, the demo system can render the scene in invisible

286

Z. Liu

windows with no texture, and get a synthesis vision image. The virtual human can decide what he (she) could see from the values in color buffer and depth buffer. The purpose of using color buffer is to distinguish objects with different color code. The purpose of using depth buffer is to get the space position of a pixel in the window, if the coordinate of a pixel is ( inv_x, inv_y), and zscreen is depth value. Let VTM is view transformation matrix, PTM is projection transformation matrix, and VM is viewport matrix. Let PW is the world coordinate corresponding to the pixel, and IPW ={ inv_x, inv_y, zscreen}, PW is calculated by formula(1). PW= IPW × (VTM × PTM × VM)-1.

(1)

In order to simulate perception for space, we use static partition of scene octree that is a hierarchical variant of spatial-occupancy enumeration [5]. We partition the static part of the scene in advance and record octree in data base module. We can use octree to solve path searching problem, as scene octree and the edges among them compose a graph, and so the path searching problem can be transformed to the problem of searching for a shortest path from one empty node to another in the graph.(See Fig.1).

Fig. 2. A* path searching (the left is no obstacle and the right is near a house)

In a complex virtual environment in which there are a lot of virtual human, synthetic vision will be costly. Furthermore, this method cannot get the detail semantic information of objects. Therefore, another efficient method for simulation of visual perception is presented. The visual perception of virtual human is limited to a sphere [1], with a radius of R and angle scope of θ max. The vision sensor is at point Oeyes (the midpoint between the two eyes), and sets up local left-handed coordinate system. Oeyes is the origin and X axis is along front orientation (See Fig.2). To determine whether the object Pob is visible, the first step is to judge whether Pob is in the vision scope. If || Pob - Oeyes ||< R and the angle between the ray and X axis is less than θ max/2, the object Pob is in the vision scope. The second step is to detect whether other obstacle occlude Pob. We can shoot a ray OP from the Oeyes to Pob, cylinders can serve as the bounding boxes of obstacles. In order to check the intersection of OP with an obstacle’s bounding box, we can check whether OP intersects with a circle that is a projection of the obstacle’s bounding box, and

Simulation of Virtual Human's Mental State in Behavior Animation

287

further to check whether OP intersects with the obstacle’s bounding box. In a 3D virtual environment, there are a lot of dynamic objects, on which we set up feature points (such as geometric center). If one feature point is visible, the object is regarded as visible. In the demo system, all obstacles are building. Based on the Gibson’s theory of affordances [11], affordances are relations among space, time and action. A virtual human can perceive these affordances directly. An affordance is invariance for environment. In this paper, we use the Gibson’s theory to guide navigation, affordances of objects hints navigation information. We set up some navigation information in database for special area or objects in the 3D virtual environment. For example, when a virtual human wants to walk across a road, we set navigation information of the zebra crossing is accessible, so that the virtual human will select zebra crossing. We use scene octree to simulate the virtual human’s perception for static object in 3D virtual environment. The locations of all dynamic objects are recorded in memory module in animation time step. If an object is visible, we suppose that a virtual human moves on a 2D plane, let Dovc is detection radius, dmin is avoiding distance for the virtual human, if Dovc60%. d) Pore: 500-800μm diameter size and 10-30% software constructed porosity. All of this will be considered when modeling the repair bioscaffold.

3 Modeling and Fabricate the Defective Bone Repair Bioscaffold In this part, the method of the modeling and fabrication of the defective bone repair bioscaffold is presented. There’re four main stages to build the model. 3.1 Reconstructing Defective Bone Model To construct the model of the defective part of the bone, the 3d model of the skull should be obtained. This stage includes three steps. Image Processing: The reconstruction of the defective bone model begins with the process of the medical CT images. Because of the precision and the electronic parts, noises of images can not be avoided. As the quality of the images is an important factor in reconstruction, the linear transform and median filter are adopted to improve the quality of the images. 3D-Reconstruction: After the process of the images, 3d reconstructing is required to build the defective bone model. The general algorithm is the MC method [6]. This algorithm is a kind of isosurface distilling method. It has a nicer speed and it is efficient on normal PC. An improved MC algorithm is adopted here [7]. Triangle Reduction: A large number of triangles that represent the surface of the defective bone model are created after 3d reconstruction. It has negative effect, so a triangle reduction is needed. In the paper the reduction method of Stan Melax [8-9] is extended. When compute the collapse cost of an edge to delete a vertex, the edge length, normal changes[9], and two new rules are considered: Sharp triangle: A sharp triangle is defined by the minimum angle (θ) of the new created triangles. Ifθis smaller than the pre- specified angle, this collapse will not be done and the sharp triangles will not be created. Area of the new created triangles: The rule of area is the same with the factor of the sharp triangle rule. Area is the parameter of this rule. 3.2 Repairing the Defective Bone In this stage, the main purpose is to repair the defective skull, and get the macro-shape of the defective part of the skull. As defect can occur with both symmetrical

A Software Method to Model and Fabricate the Defective Bone Repair Bioscaffold

(a)

447

(b)

Fig. 1. (a) Symmetrical defect (b) Unsymmetrical defect

(see figure 1 a) and unsymmetrical situation (figure 1 b), to solve any defective types, that is to say, to repair the defective bone with defects occur in any situations, a hole filling method should be developed. Peter Liepa proposed a geometric method for filling holes in triangular meshes [10]. But this method is used in a single-layer, non-closed triangular mesh. Here, a closedsurface model’s hole filling method is proposed. The aim of this stage is to build a meaningful surface, which matches the macro-shape of the defective part of the skull. The main steps are described as fellow. Boundary Identification To improve the identification speed, enlightening from Ping’s triangular mesh feature classification[11] a boundary identification method is proposed, which is base on a hypothesis: An edge that begins with a vertex is a boundary edge, if the following two rules are contented: (1) The direction of the boundary edge should be most consistent with the direction of the vector that begins with start vertex and the destination vector. (2) The normals of its two adjacent triangles have the least angle. If more than one vertex could be selected, the edge length will be considered. Rule (2) is more prior than rule (1). 1. Some guide-points are specified by users. Put G ← {G1, G2, …, Gn}. G is a set of the guide-points. It can be a vector or an array, etc. 2. For i = 1, 2, …, n, do BoundarySearch(i, (i+1) % n) to get the boundary edges of the defective bone between vertex i and i+1. Put these boundary edges to a set P. The function BoundarySearch is a recursive function. Its initial input is two guidepoints. Each recursion will select an edge as the boundary edge. For example, G1 and G2 are the input (see figure 2). First, the angles between G1’s edges (e1, e2, …, e8) and vector G1G2 will be calculated. The less two, e4 and e5 will be chosen as the candidates. Then Calculate the angles of T1&T2 (share edge e4), and T2&T3 (share edge e5). This can be got by there normal. Angle (T2&T3) is smaller than Angle (T1&T2), so e5 will be the result of this recursion. And the next recursion’s input will be V1 and G2. Other situations also follow the rules of this algorithm. Triangulation The Delaunay triangulation building method will be implemented on the boundary edges, by which a polygon is constructed. A Delaunay triangulator is adopted, which is a part of a C program Triangle[12]. By translating the boundary into the specified

448

Q. Hu, H. Yang, and Y. Yao

Fig. 2. Boundary identification

format of the adopted triangulator, it will output the desired triangular mesh. Here, the boundary points will be projected on the coordinate Cf (constructed in the next step, see Mapping Fairing) to be convenient to the triangulator’s process. This step will finish the surface construction and get a planar triangular mesh T, which will be adjusted to match the curvature of the defective part of the bone by the next step: Mapping Fairing To keep consistent with the curvature of the defective part of the skull, the planar triangular mesh created in operation Triangulation should be adjusted. A mapping fairing method is developed here. There are three steps: 1. Coordinates construction: A three-dimensional coordinate Cf will be created by specifying three points among the boundary points. The surface equation can be calculated. Let it be f(x, y, z) = 0. Then a series of plane will be built, each of which is vertical to Cf. Let this planes is S = {S1, S2, …, Sn}. 2. Reference points construction: The planar triangular mesh’s adjustment will be according to a set of points, which is named reference points. These referenced points should be on an implicit surface, which is minimally distinguishable from the surrounding mesh. A set of curves will be constructed to simulate the implicit surface. The more curves be constructed, the better accuracy will be got. When constructing the curves, planes S created in step 1 will be used. Extract the intersecting line (represented by a sampled points), and get a series of lines: L={L1, L2, …, Ln}. Each line is composed by sampled points. Let Li = {Pi1, Pi2, …,Qi1, Qi2, …, Pin}. Qi1 and Qi2 are points on the boundary of the defective bone. Spline curve will be used to create the points on the implicit surface. And a set of spline curves will be created: LQ = {LQ1, L Q2, …, L Qn}. The same as L, L Qi is composed by points. Let L Qi = {Qi1, Ri1, Ri2, …, Rin, Qi1}. Here, Cardinal curve is adopted. The reference points will be stored in LQ. 3. Mapping relation setup Searching Mapping: From the reference points LQ, the mapping relation (R: LQ → T) will be constructed. Each point of the reference points will be mapped to a point from the triangular mesh T. Both LQ and T will be projected to the coordinate Cf. The new set of these points (L’Q, T’) can be calculated by coordinate transforming. For each source point of L’Q, a target point from T’ will be determined by the rules of mapping: (1) For a source point, if a same point PT can be found in T’, PT will be the target point. (2) If can not found a same point, the source point will drop into the interior of a triangle, and the target point will be determined by the minimum distance

A Software Method to Model and Fabricate the Defective Bone Repair Bioscaffold

449

from the source point to the three point of the surround triangle. If the source point is on the edge of a triangle, the repeated calculation can be avoided in the process of programming. Interpolating mapping: The reference points LQ are created by created curves. According to the positions (index in a curve sample set) of these points, a series of average coordinate closed contour lines is constructed. The value of a contour is the average position of the points on the contour line. Then the other points of T will be mapped to the implicit surface by interpolating operation between two contour lines whose area the points drop into. Here, a linear interpolating is adopted. If point (x, y, z) drops into the area of contour line 1 (CL1, value = (x1, y1, z1)) and contour line 2 (CL2, value = (x2, y2, z2)), then the point (x, y, z) will mapped to (x, y, z + (z2 – z1) / k). k is an average proportion of x, y to x1, y1.

D ’

Fig. 3. Mapping of the points. Red points are selected points of T. A’ has the same x and y value with A. B drops into a triangle Tri and the distance between B’ and C is the shortest one. D’ is the interpolation operation on point D, contour CL1 and CL2.

4. Fairing: When finished the processes above, a triangular mesh will be generated, which is not very smooth. So a fairing operation will be done to make the mesh smooth. Taubin’s smoothing approach[13, 14] is adopted here.

Δ Pi =

∑

w ij ( v

j

− vi ) ,

j∈ i *

∑

w

ij

=1

(1)

j∈ i *

Pi’ = (1 + λΔ) (1 + μΔ) Pi

kPB = 1/λ + 1/μ

(2) (3)

wij is weight value. i* is a set that includes the neighbors of point vi. P’ is the new point after smoothing. P is the old point. λ is the weight and its value is between 0 and 1. μ is the negative scale factor and μ < -λ. kPB is cutoff frequency. Values from 0.01 to 0.1 will produce good results[14]. To preserve the shape of the mesh, fixed points are introduced. The boundary points and mapped points will be regarded as constraints and these points are fixed when smoothing. Extruding After repairing the defective bone, a solid model will be obtained by extruding the surface along the normal direction of the related triangles. The triangles are created in Triangulation operation, and have been done a mapping operation. The model of the

450

Q. Hu, H. Yang, and Y. Yao

defective bone repair is obtained at end of this step, which has a macro-shape of the repair. 3.3 Constructing the Macro-pores The macro-pores are necessary for the growth of the cells. So according to the requirements of the bioscaffold, proper structure and a well designed size and the porosity of the macro-pores are very important. This can be done by a Boolean operation between the repair and a designed interior structure model. 3.4 Manufacture the Bioscaffold Via RP Apparatus At present, there are mainly five RP technologies. That is SLS, SLA, FDM, LOM, and 3DP. Here, SLS is chosen to process the polymeric blends.

Fig. 4. (a) Hole of the defective bone (b) Repaired bone (c) Extruded (d) The Model of defective bone repair bioscaffold. Pore size is 600μm. Porosity is bout 25%. (e) The Bioscaffold fabricated by SLS. Materials are polymeric blends.

4 Experiment Result and Discussion 4.1 Experiment Result Using the proposed method, the repair bioscaffold is constructed. And the result is shown in figure 4. In this paper, kPB is set to 0.1. It’s the same with Taubin’s[14]. 4.2 Discussion Many geometric methods to fill holes are special for a single layer triangular mesh, which is not suitable for a closed triangular mesh. In this paper, the hole filling method is a method of mapping approach. A set of curves are constructed and the

A Software Method to Model and Fabricate the Defective Bone Repair Bioscaffold

451

points of a planer mesh are mapped to an implicit surface by two kinds of mapping operation. The sampled points on the constructed curves in fact are constraints in the process of fairing. This will assure the curvature of the objective. In the proposed method, by user selecting limited number of guide-points, the boundary search could be done quickly. However, the study of a fully automatic boundary identification method is of significance for accelerating the process of hole filling. In the procedure of mapping fairing, there are two kinds of mapped points. One is the points on the constructed curves, and the other is the interpolating points. The former is the result of direct mapping, and the latter is the z interpolation of two contour lines. The direct mapped points in fact are the constraints when fairing. While the interpolation created points are objects to be smoothed. When constructing the macro-pores, the interior structure model is very important for the growing of cells. Here, the interior model is a simple 3d interconnect structure, which is 90o in three directions (x, y, z). So studying the relation between the interior structure and the cells’ growth is very necessary. Limited to the process of SLS, some of the macro-pores were not well fabricated. So process of RP apparatus is also to be considered in further study, which can provide useful information for building more accurate defective bone repair bioscaffold model.

5 Conclusion In this paper, biologic property and physical property of bioscaffold is studied. A new modeling method is proposed, which can construct a defective bone repair bioscaffold 3d digital model that has the macro-shape and macro-pores. This method combines the image processing technology, 3d-reconstructing technology, and a new hole filling method. It can be used in both symmetrical and unsymmetrical defective bone. By programming, this method was successfully implemented and the repair bioscaffold 3d digital model was constructed. Through RP apparatus, using polymeric blends, a nicer physical model was obtained, which meets the requirements of the bioscaffold. In our proposed method, the boundary identification method is not fully automatic, and the interior model is just a simple 3d interconnect structure. So, the automatic boundary identification method, the relation between interior structure and the cell seeding, and the process of RP apparatus will be the main focus in our further study. Acknowledgments. Thanks to Jiang Ying and Zhang Quan for their help in this research. This project is supported by Shanghai Education Development Foundation Fund (No. 06AZ029) and Shanghai Splendid Youth Teachers Special Research Foundation Fund (No. B. 7-0109-07-011).

References 1. Chuanglong, H., Yuanliang, W., Lihua, Y., Jun, Z., Liewen, X.: Recent Advances in Natural Derived Extracellular Matrix Materials in Bone Tissue Engineering. Chinese Biotechnology 23(8), 11–17 (2003) 2. Liu, H., Hu, Q., Li, L., Fang, M.: A Study of the Method of Reconstructing the Bionic Scaffold for Repairing Defective Bone Based on Tissue Engineering. IFIP 207, 650–657 (2006)

452

Q. Hu, H. Yang, and Y. Yao

3. Ho, S.T., Hutmacher, D.W.: A comparison of micro CT with other techniques used in the characterization of scaffolds. Biomaterials 27, 1362–1376 (2006) 4. Deville, S., Saiz, E., Tomsia, A.P.: Freeze casting of hydroxyapatite scaffolds for bone tissue engineering. Biomaterials 27, 5480–5489 (2006) 5. Qingxi, H., Xianxu, H., Liulan, L., Minglun, F.: Design and Fabrication of Manual Bone Scaffolds via Rapid Prototyping. ITIC 1, 280–283 (2006) 6. Lorensen, W.E.: Marching Cubes: A High Resolution 3D Surface Construction Algorithm. Computer Graphics 21(4), 163–169 (1987) 7. Delibasis, K.S., Matsopoulos, G.K., Mouravliansky, N.A., Nikita, K.S.: A novel and efficient implementation of the marching cubes algorithm. Computerized Medical Imaging and Graphics 25, 343–352 (2001) 8. Melax, S.: A Simple, Fast, and Effective Polygon Reduction Algorithm, Game Developer Magazine (November 1998), http://www.melax.com 9. Shi-xiang, J., Jian-xin, Y.: Model Simplification Algorithm Based on Weighed Normal Changes. Journal of system simulation 17(9) (September 2005) 10. Liepa, P.: Filling holes in meshes. In: Proceedings of the 2003 eurographics/ACM SIGGRAPH symposium on geometry processing (SGP’03), pp. 200–205 (2003) 11. Xueliang, P., Laishui, Z., Shenglan, L.: Triangle Mesh Smoothing Method with Feature Preservation. Computer engineering and application 12 (2006) 12. Shewchuk, J.R.: Triangle: Engineering a 2D Quality Mesh Generator and Delaunay Triangulator, http://www.cs.cmu.edu/ quake/tripaper/triangle0.html 13. Guangming, L., Jie, T., Huiguang, H., Mingchang, Z.: A Mesh Smoothing Algorithm Based on Distance Equalization. Journal of Computer-Aided Design & Computer Graphics 14(9) (September 2002) 14. Taubin, G.: A signal processing approach to fair surface design. In: ACM, Proceedings of the 22nd annual conference on Computer graphics and interactive techniques (September 1995)

Using Qualitative Technology for Modeling the Process of Virus Infection* Hailin Feng1,2 and Chenxi Shao1,3,4 1

Computer Science Dept. of University of Science and Technology of China, Hefei 230027, China 2 School of Information Science and Technology, ZheJiang Forestry University, Lin’an 311300, China [email protected], [email protected] 3 Simulation Center, Haerbin Institute of Technology, Haerbin, 150001, China 4 Anhui Province Key Laboratory of Software in Computing and Communication, Anhui, Hefei 230027, China

Abstract. The quantitative analysis of viral infection dynamical model can’t be processed easily due to the lack of complex quantitative knowledge in such biological system; therefore, the methods based on qualitative analysis become an alternative solution to researches in the complicated biological process. In this paper the qualitative technology is introduced to model and analyze the process of virus entry. A rough model is proposed first to be the foundation of further research. With more knowledge in the process, the framework is expanded by inserting the qualitative description of different kinds of factors that have interactive influence on the process of virus entry. The factors are described qualitatively in influencing degree, and the qualitative model is built based on the interaction among these influencing factors and the viruses and cells. The judging matrices are constructed according to the qualitative model and the coherence of these matrices is verified. A qualitative analysis about the process is given finally.

1 Introduction In the field of virus infection, the researching methods can be roughly divided into three categories: the theoretical study, the experimental study and the emulational study. Generally speaking, in the traditional researching processes, complete quantitative knowledge is mostly needed to be the prerequisite to more advanced studies. Unfortunately, quantitative information necessary to evaluate and analyze the process of viral entry is usually hard to gain. Thus, with only incomplete quantitative knowledge the traditional qualitative methods to model the process are less effective. For example, a number of studies on influenza virus have shown that virus particles enter the cell by endocytosis[1-2]. Moreover, amantadine, an inhibitor of influenza virus infection[9], has been shown to act at an intracellular location[10]. Other studies suggest that influenza virus, like Sendai virus, may fuse directly with the plasma *

Supported by Key Project of Natural Science Foundation of China (No.60434010).

K. Li et al. (Eds.): LSMS 2007, LNBI 4689, pp. 453–461, 2007. © Springer-Verlag Berlin Heidelberg 2007

454

H. Feng and C. Shao

membrane of the cell[11]. The process of viral entry can’t be fully quantitatively described even now[5-6]. One reason is that the entry mechanisms used by many of them remain unclear. Another important reason is that lots of quantitative knowledge in the process can’t be gained easily. In the process of viral entry, there are lots of factors that have influence on viruses and cells[7-8], e.g. the percentage of virus binding to cells increases when temperature gets higher [12]. Karl s.Matlin[13] found that when PH>5, the percentage of virus binding to cells decreases if PH decreases. The existence of positive ion in the environment may also affect the process of viral binding and entry. However, as discussed above, the influence in the environment is difficult or impossible to express and analyze fully quantitatively. What we know is often the qualitative interaction among the variables and factors. So the qualitative technology comes to become our alternative solution to researches in the complicated biological process. Researchers have also shown much interest in such complex systems with incompletely known knowledge. In reference [3] the authors use a qualitative method to describe the interaction among the factors existing in viral infection system and build qualitative viral infection model. If the system is given as ordinary differential equation, B.Kuipers writes its qualitative model in the formula of qualitative differential equation (QDE) and reasons the process qualitatively[14]. To address the problem, in this paper we introduce a qualitative modeling frame to promote the research in the field. The basic idea is to model the process of virus entry, taking the influencing factors into account qualitatively. We propose a rough framework to be the foundation of further research. With more knowledge in the process, the framework is expanded by inserting the qualitative description of different kinds of factors that have interactive influence on the process of virus entry. The factors are described qualitatively in the influencing degree, and the qualitative model is built based on the interaction among these influencing factors and the viruses and cells. The judging matrices are constructed according to the qualitative model and the coherence of these matrices is verified finally. Based on the technology connecting qualitative to quantitative technology, we joint the behavior of viruses and the environment and the cells together.

2 The Preliminary Knowledge To qualitatively model the interaction between viruses and cells and the influence factors in the environment, the first difficult is to qualitatively specify the involved factors among the system, and the specification of interaction terms should be in favor of the reasoning of qualitative simulation. We propose the model that can describe the initial framework of the process, as shown in Fig.1.

Fig. 1. The framework of the interactive process

Using Qualitative Technology for Modeling the Process of Virus Infection

455

Here the V indicates the individual situation of the virus, and E is the description of whole environment, C is the situation about the cell, B is the prediction of the possible behaviors of the virus. However, the model is so rough that it can’t play much role as a real model. So we should expatiate on the detailed individual situation and environment of the process to make the model more realistic. An advantage is that the model can be worked further and step by step with more quantitative and qualitative knowledge about the process. At first, it is possible that we may have little information about the object need to be modeled. Then more quantitative and qualitative knowledge can be inserted into the initial framework. And by keeping adding information to it, the model can be more and more complicated and precise. Then we can give the specification of influence. In the process of virus entry, the first step is the interaction between the virus attachment protein (VAP) on the virus and the viral receptor on the surface of cells[1], which is the pivotal step to determine whether the virus would infect the cells. There are lots of factors that will influence the process of attachment[2]. The percentage of virus binding to cells increases when temperature gets higher between the range from 0℃ to 37℃[12]. Karl s.Matlin[13] found that when PH>5, the percentage of virus binding to cells decrease if PH decreases. The existing of positive ion in the environment may also affect the process of viral binding and entry. The static gravitation between the virus attachment virus and the receptor on the cells play a central role at the primary stage of attachment. The positive ion may accelerate the process of attachment and entry. After the step of attachment, the virus penetration comes to be the next stage, which is the process that viruses penetrate into the host cells by using different ways. There are four different ways for viruses to penetrate into the host cells: injection, viropexis, envelope fuse, and other ways to penetrate directly. The virus uncoating is the process that the virus infective nucleic acid release from the coat of viruses, and the ways to uncoat are different depending on what kind of virus it is and whether it contains the envelope. The existence of enzyme and the structure of the cellular skeleton play a vital role in this step. There are lots of factors that influence the behaviors of virus in the process of viral entry, and these factors can be divided into three categories: 1) Individual factors: The individual factors include some situations about the viruses themselves, and some evaluating standard such as activity and size and structure character. 2) Environment factors: The environment factors include temperature, PH value and positive ions. 3) Cellular factors: The cellular factors include the disturbing of receptor sites and content of enzyme. The factors listed above may have different importance of influence on the entry process. Here we use the formula to express the relationship between the factors: H i = f ( Pi ) , Pi indicates the factors that have influence on the process; H i

456

H. Feng and C. Shao

indicates the possible of the virus to bind and entry cells, and f ( ) denotes the function of influence degree. According to the available knowledge in the process of viral entry, we can give a table to show the influence degree of the factors. All the degrees are expressed in certain number based on quantitative and qualitative information. Table 1. The table of influence degree The individual factors 1 2 3 4 5 6 7 8

Virus activity the structure of virus attachment protein The environment factors temperature PH value Positive ion The Cell factors Receptor site enzyme The structure of cellular skeleton

Influence degree 9 8 5 5 5 7 5 6

Of course, the value can also be certain one other than those in the table, and here we just qualitatively simplify the expression. According to the description above, we can gain the figure of relation among the system of cells and viruses and the environmental influence factors.

Fig. 2. Model framework of virus behavior in the entry process

After we have discussed the possible influencing factors that may play their role in the process of virus entry, and the model framework is shown in figure, we will build the qualitative model in the next section based on the interaction among these influencing factors and the viruses and cells.

Using Qualitative Technology for Modeling the Process of Virus Infection

457

3 The Qualitative Modeling of Virus Behavior We model the process qualitatively combined with the Analytic Hierarchy Process(AHP), which is an effective approach proposed by T.L.Saaty in 1970’s that can handle the subjective decision-making questions. The building of hierarchical structure is the key step. Generally speaking, the preconcerted goal of the problem is defined as the goal-hierarchy, the middle level is usually the rule and sub-rule, and the lowest level is the behavior one. According to the table of influence degree, we can get the initial model: The behavior level

The rule level

The goal level

Virus Activity Structure of VAP

The success rate of attachment

Temperature PH Value The behaviors of virus

Positive Ion Cellular Receptor Enzyme

The success rate of peneration

The success rate of uncoating

Cellular Skeleton

Fig. 3. Qualitative model of the entry process

When the hiberarchy model of the process is created, we should consider how to analyze the weight and verify the coherence. The judging matrixes should be constructed to do the weight analysis[4]. The evaluating matrixes indicate the contrast of relative weightiness between the items in certain level and in its above level. The values of elements in the evaluating matrixes imply the knowledge of relative weightiness of every element in the researching field, so the participance of relative experts and the investigation of the field are needed to ensure the quality. The relative weightiness can shown the relative importance between two factors based on biological knowledge. Aij indicates the relative importance degree of the factor i to factor j. The bigger the value of Aij is, the more important the factor i is than j. For example, when the value is 1, it means the two factors may have the same importance to the object. If the value is defined as 3, it means the factor i is a little more important than factor j. Analogically, the value of Aij increases as the relative importance of factor i over j increases.

458

H. Feng and C. Shao

We define the meaning of reciprocal is: ai j = 1/ a ji , which implies the corresponding elements in the matrix are reciprocal. Subsequently, we can gain the evaluating matrixes: ⎛ 1 ⎜ 1/ 3 ⎜ ⎜ 1/ 7 ⎜ 1/ 6 A=⎜ ⎜ 1/ 7 ⎜ ⎜ 1/ 5 ⎜ 1/ 8 ⎜⎜ ⎝ 1/ 7

7 ⎞ 5 ⎟⎟ 1/ 6 1 2 3 1/ 9 1/ 6 1/ 7 ⎟ ⎟ 1/ 5 1/ 2 1 3 1/ 8 1/ 5 1/ 6 ⎟ 1/ 7 1/ 3 1/ 3 1 1/ 7 1/ 6 1/ 8 ⎟ ⎟ 1 9 8 7 1 5 3 ⎟ 1/ 3 6 5 6 1/ 5 1 1/ 2 ⎟ ⎟ 1/ 5 7 6 8 1/ 3 2 1 ⎟⎠ 3 1

7 6

6 5

7 7

5 1

8 3

In the matrix A, the element Aij indicates the importance grade that the factor i is higher than j, the importance grade is described and evaluated qualitatively by using certain number. 5 7⎞ 7 8⎞ ⎛ 1 7 9 ⎞ ⎛ 1 ⎛ 1 2 3⎞ ⎛ 1 ⎜ ⎟ B = ⎜1/ 7 1 1/ 2 ⎟ B = ⎜ 1/ 5 1 2 ⎟ B = ⎜1/ 7 1 2 ⎟ B1 = ⎜ 1/ 2 1 2 ⎟ 2 ⎜ ⎟ 3 ⎜ ⎟ 4 ⎜ ⎟ ⎜ 1/ 9 2 1 ⎟ ⎜ 1/ 7 1/ 2 1 ⎟ ⎜ 1/ 8 1/ 2 1 ⎟ ⎜ 1/ 3 1/ 2 1 ⎟ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠

⎛ 1 1/ 3 1/ 7 ⎞ ⎛ 1 1/ 3 1/ 7 ⎞ 5 7⎞ ⎛ 1 ⎛ 1 6 9 ⎞ ⎜ ⎟ B = ⎜ 1/ 5 1 3 ⎟ B = ⎜ 3 1 1/ 7 ⎟ B = ⎜ 3 1 1/ 5 ⎟ B5 = ⎜ 1/ 6 1 1/ 3 ⎟ 6 ⎜ ⎟ 8 ⎜ ⎟ ⎟ 7 ⎜ ⎜7 5 ⎟ ⎜7 7 ⎟ ⎜ 1/ 7 1/ 3 1 ⎟ ⎜ 1/ 9 3 1 ⎟ 1 1 ⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠ In the matrix B, the element Bij implies the qualitatively relative value of impact that the individual activity of virus has on the element i over the element j in the next level.

4 The Weight Vector and the Verifying of Coherence The coherence of elements in the matrixes can be used to evaluate the quality of the matrix and model. However, a full coherence is usually unavailable due to the complexity during the viral entry process and the lack of quantitative knowledge in the process, so an approximate coherence can also match the requirement. 4.1 The Weight Vector

To verify the coherence of the matrixes, the definition of relative computation should be given. The formula to compute the coherence is: C.I = (λmax − n) /(n − 1) . The unitary matrix A* can be obtained according to the rows: A * = a ij ij n ∑ aij i =1

Using Qualitative Technology for Modeling the Process of Virus Infection ⎛ 0.444 ⎜ ⎜ 0.148 ⎜ 0.063 ⎜ 0.074 A *ij = ⎜ ⎜ 0.063 ⎜ ⎜ 0.089 ⎜ 0.056 ⎜⎜ ⎝ 0.063

459

0.496 0.190 0.180 0.167 0.632 0.410 0.367 ⎞ ⎟ 0.165 0.163 0.150 0.167 0.126 0.154 0.262 ⎟ 0.028 0.027 0.060 0.071 0.014 0.009 0.007 ⎟ ⎟ 0.033 0.014 0.030 0.071 0.016 0.102 0.008 ⎟ 0.024 0.009 0.010 0.024 0.018 0.009 0.007 ⎟ ⎟ 0.165 0.244 0.240 0.167 0.126 0.256 0.157 ⎟ 0.055 0.163 0.150 0.143 0.025 0.051 0.026 ⎟ ⎟ 0.033 0.190 0.180 0.190 0.042 0.102 0.052 ⎟⎠

Then get the sum of the whole lines. ⎛ 2.886 ⎞ ⎜ ⎟ ⎜ 1.335 ⎟ handler ⎜ 0.279 ⎟ ⎜ ⎟ 0.348 ⎟ V =⎜ ⎜ 0.164 ⎟ ⎜ ⎟ ⎜ 1.444 ⎟ ⎜ 0.669 ⎟ ⎜⎜ ⎟⎟ ⎝ 0.852 ⎠

⎛ 0.361 ⎞ ⎜ 0.167 ⎟ ⎜ ⎟ ⎜ 0.035 ⎟ it unitarily: ⎜ ⎟ 0.044 ⎟ W =⎜ ⎜ 0.021 ⎟ ⎜ ⎟ ⎜ 0.181 ⎟ ⎜ 0.084 ⎟ ⎜⎜ ⎟⎟ ⎝ 0.107 ⎠

To verify the coherence: λ = 1 ∑ ( AW )i max n i wi 1 ⎛ 3.8440 1.8323

0.3148

0.2753

⎛ 3.8440 ⎞ ⎜1.8323 ⎟ ⎜ ⎟ ⎜ 0.3148⎟ ⎜ ⎟ 0.2753⎟ C = AW = ⎜ ⎜ 0.1760⎟ ⎜ ⎟ ⎜1.9752 ⎟ ⎜ 0.8305⎟ ⎜⎜ ⎟⎟ ⎝1.0973 ⎠

0.1760 1.9752

0.8305 1.0973 ⎞

λmax = ⎜ + + + + + + + ⎟ 8 ⎝ 0.361 0.167 0.035 0.044 0.021 0.181 0.084 0.107 ⎠ = 9.538 The unitary eigenvector corresponding to the eigenvalue λ = 9.538 is: W = {0.361 0.167 0.035 0.044 0.021 0181 0.084 0.107}

So we get: C .I . =

λmax − n n −1

= ( 9.538 − 8 ) /(8 − 1)=0.219

The standard to verify the coherence is: C.R = C.I / R.I . Here the average random coherence standard R.I. can be gained as the arithmetic mean of random eigenvalue of judging matrix, which should be taken for lager numbers of times randomly. Table 2. The coherence table of RI value

n

1

2

3

4

5

6

7

8

R.I

0

0

1.58

1.90

2.12

2.24

2.32

2.41

C.R=0.219/2.41=0.090. Generally speaking, the approximate coherence of the judging matrix can be accepted if C.R