Computational Neurogenetic Modeling
TOPICS IN BIOMEDICAL ENGINEERING INTERNATIONAL BOOK SERIES Series Editor: Evangelia Micheli-Tzanakou Rutgers University Piscataway, New Jersey
Signals and Systems in Biomedical Engineering: Signal Processing and Physiological Systems Modeling Suresh R. Devasahayam
Models of the Visual System Edited by George K. Hung and Kenneth J. Ciuffreda
PDE and Level Sets: Algorithmic Approaches to Static and Motion Imagery Edited by Jasjit S. Suri and Swamy Laxminarayan
Frontiers in Biomedical Engineering: Edited by Ned H.C. Hwang and Savio L-Y. Woo
Handbook of Biomedical Image Analysis: Volume I: Segmentation Models Part A Edited by Jasjit S. Suri, David L. Wilson, and Swamy Laxminarayan
Handbook of Biomedical Image Analysis: Volume II: Segmentation Models Part B Edited by Jasjit S. Suri, David L. Wilson, and Swamy Laxminarayan
Handbook of Biomedical Image Analysis: Volume III: Registration Models Edited by Jasjit S. Suri, David L. Wilson, and Swamy Laxminarayan
Complex Systems Science in Biomedicine Edited by Thomas S. Deisboeck and J. Yasha Kresh Computational Neurogenetic Modeling Lubica Benuskova and Nikola Kasabov
A Continuation Order Plan is available for this series. A continuation order will bring delivery of each new volume immediately upon publication. Volumes are billed only upon actual shipment. For further information please contact the publisher.
Computational Neurogenetic Modeling
Dr. Lubica Benuskova Senior Research Fellow Knowledge Engineering and Discovery Research Institute AUT, Auckland, New Zealand
and
Professor Nikola Kasabov Founding Director and Chief Scientist Knowledge Engineering and Discovery Research Institute AUT, Auckland, New Zealand
~ Springer
Dr. Lubica Benuskova Senior Research Fellow Knowledge Engineering and Discovery Research Institute, www.kedri.info AUT, Auckland, New Zealand
[email protected] Professor Nikola Kasabov Founding Director and Chief Scientist Knowledge Engineering and Discovery Research Institute, www.kedri.info AUT, Auckland, New Zealand
[email protected] Library of Congress Control Number: 200693690 I
ISBN-lO: 0-387-48353-5 ISBN-13: 978-0-387-48353-5
eISBN-lO: 0-387-48355-1 eISBN-13: 978-0-387-48355-9
Printed on acid-free paper. © 2007 Springer Science + Business Media, LLC All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science + Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights.
9 8 765 4 3 2 1 springer.com
Dedication
To the memory ofmy parents.
Lubica
To my mother and the memory ofmy father.
Nikola
Preface
It is likely that future progress in many important areas of science (e.g.
brain science, bioinformatics, information science, physics, communication engineering and social sciences) can be achieved only if the areas of computational intelligence, brain science and bioinformatics share and integrate their methods and knowledge. This book offers some steps in this direction. The book presents the background knowledge and methods for the integration of gene information and brain activity information with the purpose of the creation of biologically plausible computational models aiming at modeling and understanding the brain. The book is aiming at encouraging research in information sciences in the direction of human-like and human-oriented information processing. In this context, "human-like" means that principles from the brain and genetics are used for the creation of new computational methods. "Humanoriented" means that these machines can be used to discover and understand more about the functioning of the brain and the genes, about memory and learning, about speech and language, about image and vision, and about ourselves. This work was partially supported by the research grant AUTX0200l "Connectionist-based intelligent information systems", funded by the New Zealand Foundation for Research, Science, and Technology - FRST, through the New Economy Research Fund - NERF. There are a number of people whom we would like to thank for their encouragement and contribution to the book. These are several colleagues, research associates and postgraduate students we have worked with at the Knowledge Engineering and Discovery Research Institute in Auckland, New Zealand, in the period 2002- 2006: Dr Qun Song, Dr Zeke S. Chan, Dr Paul S. Pang, Dr Liang Goh, Vishal Jain, Tian-Min Ma (Maggie), Peter Hwang, Paulo Gottgtroy, Natalia Bedran, Joyce D'Mello and especially Simei Gomes Wysoski, who did the technical edition of the book and developed specialized simulation tools we used for our computer experiments. We appreciate the discussions we had with a number of colleagues from different laboratories and countries. Among them are Walter Freeman from University of California at Berkeley; Takeshi Yamakawa - Kyushu
VIII
Preface
Institute of Technology ; John G. Taylor - Kings College, London; Cees van Leuwen - RIKEN, Japan; Michael Arbib - University of Southern California ; Dimiter Dimitrov - National Cancer Institute in Washington DC; Alessandro E. P. Villa -University of Lausanne, Bassem Hassan Catholic University of Leuven, Gary Markus - New York. We thank the Springer, New York team and especially Aaron Johnson and Beverly Rivero for the encouragement and patience. When one of the authors, N. Kasabov, presented his talk at the ICONIP 2002 in Shanghai and suggested that gene interaction information should be used in biologically plausible neural network models, Walter Freeman commented "Yes, that makes sense, but how do we do that?" Michael Arbib when visiting Auckland in 2004 made a remark that integrating genes (molecular level) into neural networks may require to go to atom (quantum) level. This book is presenting some initial answers to these questions . It presents the foundations and the concepts of computational neurogenetic modeling (CNGM), initially introduced by the authors in 2004 (Kasabov and Benuskova 2004). The book was written by the two authors in a close collaboration, where Lubica Benuskova wrote chapters 2,3 ,8,9, 10 and compiled Appendix I , while Nikola Kasabov wrote chapters 1,4,5,6, 7 and compiled Appendix 2 and 3. Each of the authors also had a smaller contribution to the other chapters as well. The book is intended for postgraduate students and researchers in the areas of information sciences, artificial intelligence, neurosciences , bioinformatics, and cognitive sciences. The book is structured so that every chapter can be used as a reading material for research oriented courses at a postgraduate level. Additional materials, including: data, simulation programs, lecture notes, color figures, etc. can be found on the web site www.kedri.info. Dr Lubica Benuskova Prof. Dr. Nikola Kasabov Knowledge Engineering and Discovery Research (www.kedri.info), Auckland University of Technology Auckland, New Zealand I September 2006
Institute
Contents
Preface 1 Computational Neurogenetic Modeling (CNGM): A Brief Introduction 1.1 Motivation - The Evolving Brain 1.2 Computational Models of the Brain 1.3 Brain-Gene Data, Information and Know ledge 1.4CNGM: How to Integrate Neuronal and Gene Dynamics? 1.5 What Computational Methods to Use for CNGM? 1.6 About the Book 1.7 Summary 2 Organization and Functions of the Brain 2.1 Methods of Brain Study 2.2 Overall Organization of the Brain and Motor Control 2.3 Learning and Memory 2.4 Language and Other Cognitive Functions 2.4.1 Innate or Learned? 2.4.2 Neural Basis of Language 2.4.3 Evolution of Language, Thinking and the Language Gene 2.5 Neural Representation ofInformation 2.6 Perception 2.7 Consciousness 2.7.1 Neural Correlates of Sensory Awareness 2.7.2 Neural Correlates of Reflective Consciousness 2.8 Summary and Discussion
VII
1 1 .4 6 12 14 15 16 17 18 23 25 29 29 30 33 36 37 41 .41 .44 .49
3 Neuro-Information Processing in the Brain 53 53 3.1 Generation and Transmission of Signals by Neurons 3.2 Learning Takes Place in Synapses: Toward the Smartness Gene 56 58 3.3 The Role of Spines in Learning 3.4 Neocortical Plasticity 61
X
Contents 3.4.1 Developmental Cortical Plasticity 61 3.4.2 Adult Cortical Plasticity 64 3.4.3 Insights into Cortical Plasticity via a Computational Model... 66 3.5 Neural Coding: the Brain is Fast, Neurons are Slow 74 3.5.1 Ultra-Fast Visual Classification 74 3.5.2 Hypotheses About a Neural Code 77 Coding Based on Spike Timing 77 The Rate Code 77 3.6 Summary 78
4 Artificial Neural Networks (ANN) 4.1 General Principles 4.2 Models of Learning in Connectionist Systems 4.3 Unsupervised Learning (Self Organizing Maps - SOM) 4.3.1 The SOM Algorithm 4.3.2 SOM Output Sample Distribution Clustering Information Visualization of Input Variables Relationship Between Multiple Descriptors The Connection Weights Interpretation by the Fuzzy Set Theory 4.3.3 SOM for Brain and Gene Data Clustering 4.4 Supervised Learnin g 4.4.1 Multilayer Perceptron (MLP) 4.4.2 MLP for Brain and Gene Data Classification Example 4.5 Spiking Neural Networks (SNN) 4.6 Summary
81 81 84 93 93 95 95 95 96 96 96 97 97 98 98 99 99 102 105
5 Evolving Connectionist Systems (ECOS) 5.1 Local Learning in ECOS 5.2 Evolving Fuzzy Neural Networks EFuNN 5.3 The Basic EFuNN Algorithm 5.4 DENFIS 5.4.1 Dynamic Takagi-Su geno Fuzzy Inference Engine 5.4.2 Fuzzy Rule Set, Rule Insertion and Rule Extraction 5.5 Transductive Reasoning for Personalized Modeling 5.5.1 Weighted Data Normalization 5.6 ECOS for Brain and Gene Data Modeling 5.6.1 ECOS for EEG Data Modeling, Classification and Signal Transition Rule Extraction
107 107 108 112 116 118 119 120 122 122 122
XI
5.6.2 ECOS for Gene Expression Profilin g 5.7 Summary
124 126
6 Evolutionary Computation for Model and Feature Optimization .127 6.1 Lifelong Learning and Evolution in Biological Species: Nurture vs. Nature 127 128 6.2 Principl es of Evolutionary Computation 6.3 Genetic Algorithms 128 6.4 EC for Model and Parameter Optimization 133 6.4. 1 Example 133 6.5 Summary 136
7 Gene/Protein Interactions - Modeling Gene Regulatory Networks (GRN) 137 7.1 The Central Dogma of Molecular Biology 7.2 Gene and Protein Expression Data Analy sis and Modeling 7.2.1 Example 7.3 Modeling Gene/Prote in Regulato ry Networks (GPRN) 7.4 Evolving Connectionist Systems (ECOS) for GRN Modeling 7.4.1 General Principles 7.4.2 A Case Study on a Small GRN Modeling with the Use of ECOS 7.5 Summary
8 CNGM as Integration of GPRN , ANN and Evolving Processes 8.1 Modeling Genetic Control of Neural Development 8.2 Abstract Computational Neurogenetic Model 8.3 Continuous Model of Gene-Protein Dynamics 8.4 Towards the Integration of CNGM and Bioinformatics 8.5 Summary
9 Application of CNGM to Learning and Memory 9.1 Rules of Synaptic Plasticity and Metaplasticity 9.2 Toward a GPRN of Synaptic Plasticity 9.3 Putative Molecular Mechanisms of Metaplasticity 9.4 A Simple One Protei n-One Neuronal Funct ion CNGM 9.5 Application to Modeling ofL-LTP 9.6 Summary and Discussion 10 Applications of CNGM and Future Development 10.1 CNGM of Epilepsy 10.1.1 Genetically Caused Epilepsies
137 141 143 145 150 150 151 153
155 156 161 165 171 175
177 177 185 193 196 198 202
205 206 206
XII
Contents
10.1.2 Discussion and Future Developments 10.2 CNGM of Schizophrenia 10.2.1 Neurotransmitter Systems Affected in Schizophrenia 10.2.2 Gene Mutations in Schizophrenia 10.2.3 Discussion and Future Developments 10.3 CNGM of Mental Retardation 10.3.1 Genetic Causes of Mental Retardation 10.3.2 Discussion and Future Developments 10.4 CNGM of Brain Aging and Alzheimer Disease 10.5 CNGM of Parkinson Disease 10.6 Brain-Gene Ontology 10.7 Summary
209 210 212 214 217 2 18 219 223 224 229 232 235
Appendix 1.•................................................•..............•....•••..............•...... 237 237 A.I Table of Genes and Related Brain Functions and Diseases Appendix 2 A.2 A Brief Overview of Computational Intelligen ce Methods A.2.1 Probabil istic and Statistical Methods Stochastic Models A.2.2 Boolean and Fuzzy Logic Models Boolean Models Fuzzy Logic Models A.2.3 Artificial Neural Networks Evolving Classifier Function (ECF) A.2.4 Methods of Evolutionary Computation (EC)
247 247 247 250 250 250 251 253 254 256
Appendix 3 257 A.3 Some Sources of Brain-Gene Data, Information, Knowledge and Computational Models 257
References
259
Index
287
1 Computational Neurogenetic Modeling (CNGM): A Brief Introduction
This chapter introduces the motivation and the main concepts of computational neurogenetic modeling (CNGM). It argues that with the presence of a large amount of both brain and gene data related to brain functions and diseases, it is required that sophisticated computational models are created to facilitate new knowledge discovery that helps understanding the brain in its complex interaction between genetic and neuronal processes. The chapter points to sources of data, information and knowledge related to neuronal and genetic processes in the brain. CNGM is concerned with the integration of all these diverse information into a computational model that can be used for modeling and prediction purposes. The models integrate knowledge from mathematical and information sciences (e.g. computational intelligence - CI), neurosciences, and genetics. The chapter also discusses what methods can be used for CNGM and how. The concepts and principles introduced in this chapter are presented in detail and illustrated in the rest of the book.
1.1 Motivation - The Evolving Brain According to the Concise Oxford English Dictionary (1983), "evolving" means "revealing", "developing". It also means "unfolding, changing". The term "evolving" is used here in a broader sense than the term "evolutionary". The latter is related to a population of individual systems traced over generations (Darwin 1859, Holland 1975, Goldberg 1989), while the former, as it is used in this book, is mainly concerned with a continual change of the structure and the functionality of an individual system during its lifetime (Kasabov 2003, Kasabov 2006). In living systems and in the human brain in particular, evolving processes are observed at different levels (Fig. 1.1) (Kasabov 2006). At the quantum level, particles are in a complex evolving state all the time, being at several locations at the same time, which is defined by probabilities.
2
I Computational Neurogenetic Modeling(CNGM): A BriefIntroduction
At a molecular level, DNA, RNA and protein molecules, for example, evolve and interact in a continuous way. The area of science that deals with the information processing and data manipulation at this level is Bioinformatics. At the cellular level (e.g. a neuronal cell) all the metabolic processes, the cell growth, cell division etc., are evolving processes. 6. Evolutionary (population/generation) processes 5. Brain cognitiveprocesses (learning,thinking, etc.) 4. System level information processing(e.g. auditory system) 3. Information processing in a cell (neuron) 2. Molecularlevel of information processing(genes, proteins) 1. Quantum level of information processing Fig. 1.1. Six levels of evolving processes in the brain: evolution, cognitive brain processes,brain functions in neural networks, a single neuron functions, molecular processes, and quantumprocesses
At the level of cell ensembles, or at the neural network level, an ensemble of cells (neurons) operates in a concert, defining the function of the ensemble or the network, for instance perception of sound. In the human brain, complex dynamic interactions between groups of neurons can be observed when certain cognitive functions are performed, e.g. speech and language processing, visual pattern recognition, reasoning and decision making. At the level of population of individuals, species evolve through evolution (Darwin 1859)- the top level in Fig. 1.1. Evolutionary processes have inspired the creation of computational modeling techniques called evolutionary computing (EC) (Holland 1975, Goldberg 1989). A biological system evolves its structure and functionality through both lifelong learning by an individual and the evolution of populations of many such individuals. In other words, an individual is a result of the evolution of many generations of populations, as well as a result of its own developmental lifelong learning processes. There are many physical and information processes of dynamic interaction within each of the six levels from Fig. 1.1 and across the levels. Inter-
1.1 Motivation - The Evolving Brain
3
actions are what make an organism a living one, and that is also a challenge for computational modeling. For example, there are complex interactions between DNA, RNA and protein molecules. There are complex interactions between the genes and the functioning of each neuron, a neural network, and the whole brain. Some of these interactions are known to have caused brain diseases, but most of them are unknown at present. An example of interactions between genes and neuronal functions is the dependence of development of brain with human characteristics on expression of genes like FOXP2, the gene involved in speech production (Enard et al. 2002), ASPM and Microcephalin that affect the brain size (Evans et al. 2005, Mekel-Bobrov et al. 2005), and HARIF, that is of fundamental importance in specifying the six-layer structure of the human cortex (Pollard et al. 2006). Another example is the observed dependence between long-term potentiation (learning) in the synapses and the expression of the immediate early genes and their corresponding proteins such as Zif/268 (Abraham et al. 1994). Yet another example are putative genetic mutations for many brain diseases that have been already discovered (see Appendix 1). Generally speaking, neurons from different parts of the brain, associated with different functions, such as memory, learning, control, hearing and vision, function in a similar way. Their functioning is defined by several factors, one of them being the level of neurotransmitters. These factors are controlled both through genetics and external inputs. There are genes that are known to regulate the level of neurotransmitters for different types of neurons from different areas of the brain. The functioning of these genes and the proteins produced can be controlled through nutrition and drugs. This is a general principle that can be exploited for different models of the processes from Fig. 1.1 and for different systems performing different tasks (learning, hearing etc.). We will refer to the above as neurogenetics (Kasabov and Benuskova 2004). The evolving processes in the brain are based on several major principles (Arbib 1972, Grossberg 1982, Arbib 1987, Taylor 1999, Freeman 2000, Arbib 2003, van Ooyen 2003, Marcus 2004a), such as: • Evolving is achieved through both genetically defined information and learning. • The evolved neurons have a spatial-temporal representation where similar stimuli activate close neurons. • The evolving processes lead to a large number of neurons involved in each task, where many neurons are allocated to respond to a single stimulus or to perform a single task; e.g. when a word is heard, there are millions of neurons that are immediately activated.
4
1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction
• Memory-based learning, i.e. the brain stores exemplars of facts that can be recalled at a later stage. • Evolving is achieved through interaction with the environment and other systems. • Inner processes take place, such as sleep memory consolidation. • The evolving processes are continuous and lifelong. • Through evolving brain structures, higher-level functions emerge which are embodied in the structure, and can be represented as a level of abstraction (e.g. the acquisition and the development of speech and language). The advancement in brain science, molecular biology and computational intelligence, results in a large amount of data, information and knowledge on brain functioning, brain-related genetics, brain diseases and new computational intelligence methods. All these constitute a strong motivation for the creation of a new area of science that we call computational neurogenetic modeling (CNGM), with the following general objectives: 1. To create biologically plausible neuronal models. 2. To facilitate a better understanding of the principles of the human brain, the genetic code, and life in general. 3. To enable modeling of the brain. 4. To create new generic methods of computational intelligence and a new generation of intelligent machines.
1.2 Computational Models of the Brain A project, called The Blue Brain Project, marks the beginning of a study how the brain works by building very large scale models of neural networks (http://bluebrainproject.epfl.ch/index.html). This endeavor follows a century of experimental "wet" neuroscience and development of many theoretical insights of how neurons and neural networks function (Arbib 2003). The Blue Brain Project was launched by the Brain Mind Institute, EPFL, Switzerland and IBM, USA in May, 2005. Scientists from both organizations will work together using the huge computational capacity of IBM's Blue Gene supercomputer to create a detailed model of the circuitry in the neocortex - the largest and most complex part of the human brain. The neocortex constitutes about 85% of the human brain's total mass and is thought to be responsible for the cognitive functions of language, learning, memory and complex thought. The Blue Brain Project will also build models of other cortical and subcortical parts of the brain and models of
1.2 Computational Models of the Brain
5
sensory and motor organs. By expanding the project to model other areas of the brain, scientists hope to eventually build an accurate, computerbased model of the entire brain. The project is a massive undertaking because of the hundreds of thousands of parameters that need to be taken into account. EPFL' s Brain and Mind Institute's world most comprehensive set of empirical data on the micro-architecture of the neocortex will be turned into a working 3-dimensional model recreating the high-speed electrochemical interactions of the brain's interior. The first objective is to create a software replica of the neocortical column at a cellular level for real-time simulations. An accurate replica of the neocortical column is the essential first step to simulating the whole brain. The second and subsequent phases will be to expand the simulation to include circuitry from other brain regions and eventually the whole brain. In the Blue Column, a nickname for the software replica of the neocortical column, not only the cells but an entire microcircuit of cells will be replicated (like the duplication of a tissue). Neocortical column is stereotypical in many respects from mouse to man with subtle variations in different ages, brain regions and species. The Blue Column will first be based on the data obtained from rat somatosensory cortex at 2 weeks of age because these data are the most abundant. Once built and calibrated with iterative simulations and experiments, comparative data will be used to build columns in different brain regions, ages and species, including humans. The Blue Column will be composed of 104 morphologically complex neurons with active ionic channels to enable generation of electrical currents and potentials. The neurons will be interconnected in a 3dimensional (3D) space with 107 -108 dynamic synapses. The Blue Neuron will receive about 103 -104 external input synapses and generate about 103 -104 external output synapses. Neurons will transmit information according to dynamic and stochastic synaptic transmission rules. The Blue Column will self-adapt according to synaptic learning algorithms running on 107 _10 8 synapses, and according to metaplasticity, supervised and reward learning algorithms running on all synapses. The column project will also involve the database of 3D reconstructed model neurons, synapses, synaptic pathways, microcircuit statistics, and computer model neurons. Single synapses and whole neurons will be modeled with the molecular level details however a neocortical column will be modeled at the cellular level. In the future, the research will go in two directions simultaneously: 1. The first direction will be the simplification of the column and its software or hardware duplication to build larger parts of the neocortex and eventually the entire neocortex.
6
1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction
2. The second direction will stay with a single neocortical column, moving down to the molecular level of description and simulation. This step will be aimed at moving towards genetic level simulations of the neocortical column. A very important reason for going to the molecular level is to link gene activity with electrical activity, as the director of the project Henry Markram reckons. A molecular level model of the neocortical column will provide the substrate for interfacing gene expression with the network structure and function. The neocortical column lies at the interface between the genes and complex cognitive functions. Establishing this link will allow predictions of the cognitive consequences of genetic disorders and allow reverse engineering of cognitive deficits to determine the genetic and molecular causes. It is expected that this level of simulation will become a reality with the most advanced phases of Blue Gene and Blue Brain Project development. As is the model replica of the cortical column based on many computational models of neurons, channel kinetics, learning, etc., we can ask whether there are any models of the computational neurogenetic type to be employed to model the interaction between the genes and neural networks. In other words, are there already corresponding neurogenetic models for brain functions and processes? The goal of this book is to address this question and to introduce the development of such models, which we call computational neurogenetic models (CNGMs).
1.3 Brain-Gene Data, Information and Knowledge The brain is a complex system that evolves its structure and functionality over time. It is an information-processing and control system, collaborating with the spinal cord and peripheral nerves. Each part of the brain is responsible for a particular function, for example: the Cerebrum integrates information from all sense organs, motor functions, emotions, memory and thought processes; the Cerebellum coordinates movements, walking, speech, learning and behavior; the Brain stem is involved in controlling the eyes, in swallowing, breathing, blood pressure, pupil size, alertness and sleep. A simplified view of the outer structure of the human brain is given in Fig 2.1. The structure and the organization of the brain and how it works at a higher and a lower level are explained in Chaps. 2 and 3, respectively. Since the 50-ties of the 20th century, experimental brain data gathering has been accompanied by the development of explanatory computational brain models. Many models have been created so far, for example:
1.3 Brain-Gene Data, Information and Knowledge
7
- Brain models created at USC by a team lead by Michael Arbib, at http://www-hbp.usc.edu/Projects/bmw.htm; - Mathematical brain function models maintained by the European Bioinformatics Institute (EBI): http://www.ebi.ac.uk; - Wayne State Institute Brain Injury Models at http://rtb.eng.wayne.edu/braini; - The Neural Micro Circuits Software: www.lsm.tugraz.at; - Neuroscience databases (Koetter 2003); - Genetic data related to brain (Chin and Moldin 2001) and many more (see Appendix 3). None of these brain models incorporate genetic information despite of the growing volume of data, information and knowledge on the importance and the impact of particular genes and genetic processes on brain functions. Brain functions, such as learning and memory, brain processes, such as aging, and brain diseases, such as Alzheimer, are strongly related to the level of expression of genes and proteins in the neurons (see Appendix 1, Appendix 3 and also Chaps. 9 and 10). Both popular science books and world brain research projects, such as the NCB I (the National Center for Biomedical Information), the Allen Brain Institute, the Blue Brain Project, the Sanger Centre in Cambridge, and many more, have already revealed important and complex interactions between neuronal and genetic processes in the brain, creating a massive world repository of brain-gene data, information and knowledge. Some of the information or/and references to it, are given in Appendix 1 and 3 and Chaps. 9 and 10. The central dogma ofmolecular biology states that DNA, which resides in the nucleus of a cell or a neuron, transcribes into RNA, and then translates into proteins. This process is continuous, evolving, so that proteins in turn cause genes to transcribe, etc. (Fig. 1.2). The DNA is a long, double stranded sequence (a double helix) of millions or billions of 4 base molecules (nucleotides) denoted as A, C, T and G, that are chemically and physically connected to each other through other molecules. In the double helix, they make pairs such that every A from one strand is connected to a corresponding C on the opposite strand, and every G is connected to aT. A gene is a sequence of hundreds and thousands of bases as part of the DNA, that is translated into a protein or several proteins. Only less than 5% of the DNA of the human genome contains protein-coding genes, the other part is a non-coding region that may contain useful information as well. For instance, it contains the RNA genes, regulatory regions, but mostly its function is not currently well understood.
8
1 Computational Neurogenetic Modeling (CNGM): A BriefIntroduction
The DNA of each organism is unique and resides in the nucleus of each of its cells. But what makes a cell alive are the proteins that are expressed from the genes, and define the function of the cell. The genes and proteins in each cell are connected in a dynamic regulatory network consisting of regulatory pathways - see Chap. 7. Normally, only few hundreds of genes are expressed as proteins in a particular cell. At the transcription phase, one gene is transcribed in many RNA copies and their number defines the expression level of this gene. Some genes may be "over-expressed", resulting in too much protein in the cell whereas some genes may be "under-expressed" resulting in too little protein. In both cases the cell may be functioning in a wrong way that may be causing a disease. Abnormal expression of a gene can be caused by a gene mutation - a random change in the code of the gene, where a base molecule is either inserted, or - deleted, or - altered into another base molecule. Drugs can be used to stimulate or to suppress the expression of certain genes and proteins, but how that will affect indirectly the other genes related to the targeted one, has to be evaluated and that is where computational modeling of gene regulatory networks (GRN) and CNGM can help. Output Cell Function Translation mRNA into protein production Transcription Genes copied as mRNA
DNAgenes
RNA
Proteins
Protein-gene feedback loop through Transcription Factors
Fig. 1.2. The genes in the DNA transcribe into RNA and then translate into proteins that define the function of a cell (The central dogma of molecular biology). Gene information processing in moredetails is presented in Chap. 7
It is always difficult to establish the interaction between genes and proteins. The question "What will happen with a cell or the whole organism if one gene is under-expressed or missing?" is now being answered by using a technology called "Knockout gene technology" (Chin and Moldin 2001). This technology is based on a removal of a gene sequence from the DNA
1.3 Brain-Gene Data, Information and Knowledge
9
and letting the cell/organism to develop, where parameters are measured and compared with these parameters when the gene was not missing. The obtained data can be further used to create a CNGM as described in Chap. 8. Information about the relationship between genes and brain functions is given in many sources (see Appendices 1 and 3). In the on-line published book "Genes and Diseases" (www.ncbi.nlm.nih.gov/books) the National Center for Biological Information (NCB!) has made available a large amount of gene information related to brain diseases, for example: - Epilepsy: One of several types of epilepsy, Lafora disease (progressive myoclonic, type 2), has been linked to a mutation of the gene EMP2A and EMP2B found on chromosome 6 (see Chap. 10). - Parkinson disease: Several genes: Parkin, PARK7, PTEN, alpha synuclein and others, have been related to Parkinson disease, described in 1817 by James Parkinson (see Chap. 10). - Huntington disease: Mutation in the HD gene on chromosome 4 was linked to this disease. - Sclerosis: A gene SODI was found to be related to familial amyotropic lateral sclerosis (see Chap. 10). - Rett syndrome: A gene MeCP2, on the long arm of chromosome X (Xq28), has been found to be related to this disease. The gene is expressed differently in different parts of the brain (see Fig. 1.3). The Allen Brain Institute has completed a map of most of the genes expressed in different sections of the brain of a mouse and has published it free as the Allen Brain Atlas (www.alleninstitute.org). In addition to the gene ontology (GO) of the NCBI, a brain-gene ontology (BGO) of the Knowledge Engineering and Discovery Research Institute KEDRI (www.kedri.info) contains genes related to brain functions and brain diseases, along with computational simulation methods and systems (Fig. 1.4). The BGO allows users to "navigate" in the brain areas and find genes expressed in different parts of it, or for a particular gene - to find which proteins are expressed in which cells of the brain. Example is given in Fig. 1.5. Gene expression data of thousands of genes measured in tens of samples collected from two categories of patients - control (class 1) and cancer (class 2) using micro-array equipment in relation to the brain and the central nervous system have been published by (Ramaswamy et al. 2001) and (Pomeroy et al. 2002). The first question is how to select the most discriminating genes for the two classes that can possibly be used as drug targets. The second question is to build a classifier system that can correctly
10
I Computational Neurogenetic Modeling (CNGM): A BriefIntroduction
classify (predict) for a new sample which class it is likely to belong to that can be used as an early diagnostic test. The answer to the latter question is illustrated in Chaps. 4, 5 and 6, where classification and prediction computational models are presented. Average Difference Value
:
•
.
I
~
l
..
I I I I
I
!
Fig. 1.3. The gene MeCP2, related to Rett syndrome, is expressed differently in different parts of the human brain (left vertical axis), the highest expression level being in the Cerebellum (source: Gene Expression Atlas at http://expression.gnf.org/cgi-bin/index.cgi)
Fig. 1.4. A snapshot of a structure of the Brain-Gene Ontology (BGO) of the KEDRI Institute (httpz/www.kedri.info/)
1.3 Brain-Gene Data, Information and Knowledge
11
Average 01fference Value
c....br
D"
01
~
01.
•
(i,llD)JA
ow
;~1:»1l5..~1~
VM;';';;-
'::<J J
Fig. 1.6. 12 genes selected as top discriminating genes from the Central Nervous System (CNS) cancer data that discriminates two classes - survivals and not responding to treatment (Pomeroy et al. 2002). The NeuCom software system is used for the analysis (www.theneucom.com) and the method is called "Signal-toNoise ratio". See Color Plate 1
1.4 CNGM: How to Integrate Neuronal and Gene Dynamics? A CNGM integrates genetic, proteomic and brain activity data and performs data analysis, modeling, prognosis and knowledge extraction that reveals relationship between brain functions and genetic information. Let us look at this process as a process of building mathematical function or a computational algorithm as follows. A future state of a molecule M' or a group of molecules (e.g. genes, proteins) depends on its current state M, and on an external signal Em: (1.2) A future state N' of a neuron, or an ensemble of neurons, will depend on its current state N and on the state of the molecules M (e.g. genes) and on external signals En: (1.3) And finally, a future cognitive state C' of the brain will depend on its current state C and also on the neuronal- N, and the molecular- M state and on the external stimuli E;
1.4 CNGM: How to Integrate Neuronal and Gene Dynamics?
13
(1.4)
C'=Fc (C,N,M, Ec )
The above set of equations (or algorithms) is a general one and in different cases it can be implemented differently, e.g.: • One gene - one neuron/brain function (see Chaps. 9 and 10). • Multiple genes - one neuronlbrain function, no interaction between genes (see Chaps. 9 and 10). • Multiple genes - multiple neuron/brain functions where genes interact in a gene regulatory network (GRN) and neurons also interact in neural network architecture (see Chap. 8 and also Fig. I. 7). • Thousands of genes - complex brain/cognitive function/s where genes interact within GRN and neurons interact in several hierarchical neural networks (discussed in Chap. 8). GRN
ANN
output
.. - - - - - - - - - ,
Fig. 1.7. A more complex case ofCNGM, where a GRN of many genes is used to represent the interaction of genes, and an ANN is employed to model a brain function. The model output is compared against real brain data for validation of the model and for verifying the derived gene interaction GRN after model optimization is applied (see Chap. 8)
The most common models of brain functions so far are based on the methods of artificial neural networks (ANN) (Rolls and Treves 1998), which is not surprising, as ANN are design using principles of the brain. Other methods, such as differential equations, statistical regression, evolutionary computation (EC), etc. are also used to model the brain. Generally speaking, different methods of mathematical and information sciences can be applied for the development of CNGM. In the next section we briefly discuss these methods, their principles, advantages, limitations and roles in building CNGMs. A short description of the methods is given in Appendix 2 and a more detailed description of the main methods - ANN and evolutionary computation (EC) is given in Chaps. 4, 5 and 6.
14
1 Computational Neurogenetic Modeling (CNGM): A Brieflntroduction
1.5 What Computational Methods to Use for CNGM? In order to implement Eq. 1.2 and Eq. 1.3 and to build a CNGM for classification or prediction and for knowledge discovery from brain-gene data, different mathematical and information science methods can be used. Here we will present a list of methods, most of them falling in the area of computational intelligence (CI) and knowledge engineering (KE), with a brief comment about their applicability to build CNGM. • Experimentally derived analytical functions and statistically derived regression functions (Pevzner 2000) are applicable for instance to derivation of GRN when sufficient experimental data and expert knowledge are available. • Probabilistic learning methods, e.g. Hidden Markov Models HMM (Hunter 1994, Pevzner 2000, Somogyi et al. 2001) are applicable when a priori information is known as prior probabilities and the distribution of the data is also known in advance. • Statistical learning methods, e.g. Support Vector Machines (SVM), Bayesian classifiers (Vapnik 1998, Baldi and Brunak 2001) are applicable when statistically significant amount of data is available and certain probability distribution may be required. • Case-based reasoning (e.g. k-I\TN; transductive reasoning) (Baldi and Brunak 2001) are applicable when a smaller number of variables is used and a relatively small number of samples, that have reliable variable values. The method is adaptable to new data and applicable for personalized modeling (Song and Kasabov 2006). • Decision trees (Hunter 1994) are applicable for classification tasks and are not adaptive to new data. They are good to extract structured domain information. • Rule-based systems (propositional logic dated back to Aristotle) and fuzzy systems (introduced by (Zadeh 1965)) are applicable if domain knowledge on the problem in hand is available even in an imprecise or incomplete form. Usually fuzzy logic methods are combined with ANN methods for the creation of fuzzy-ANN (Kasabov 1996a). • Artificial neural networks (ANN) (Grossberg 1982, Kohonen 1984, Bishop 1995, Kasabov 1996a) in their different models such as self organizing maps (SOM), multilayer perceptrons (MLP), radial basis function neural networks (RBF), are "model -free" and are applicable when some data is available, but there is no knowledge on what analytical function would be appropriate to use (Chap. 4). Some of the ANN models are adaptive to new data, such as the evolving connectionist systems (ECOS) (Kasabov 2001).These methods are described in Chap. 5.
1.6Aboutthe Book
15
• Evolutionary computation (EC) methods, such as genetic algorithms (GA) (Holland 1975, Goldberg 1989) are applicable when neither data nor much knowledge is available on the problem. They are based on the "generate and test" approach using fitness (goodness) criteria to evaluate how good a generated solution is. For applications of the EC methods, see (D'Haeseleer et al. 2000, Fogel and Come 2003). For a more detailed description of the principles of the EC and some of their applications, see Chap. 6. • Hybrid systems, e.g. knowledge-based neural networks; neuro-fuzzy systems; neuro-fuzzy-genetic systems; evolving connectionist systems (Kasabov 1996a, Kasabov and Song 2002) have the advantages of both ANN and rule based systems (see Chap. 5).
1.6 About the Book This book consists of 5 main parts as shown in Fig. 1.8: 1. Background knowledge on brain information processing (Chaps. 2, 3); 2. Background knowledge on gene information processing (Chap. 7); 3. Background knowledge on computational methods and methods of computational intelligence (CI), mainly ANN and EC (Chaps. 4, 5, 6); 4. Methodologies for building CNGM (Chaps. 8, 9); 5. Applications of CNGM for modeling brain functions and diseases and as novel generic techniques of Cl (Chaps. 9, 10). Brain information processing (Chapters 2, 3, Appendices I and [ \ , 3) Gene information processing (Chapter 7 and Appendix 3) Computational modeling techniques (Chapters 4, 5, 6, and Appendix 2)
f----+
/
Methodologies for building CNGM (Chapters 8 and 9)
r-+
Applications of CNGM for modeling brain functions and diseases (Chapters 9 and 10)
Fig. 1.8. The five partsof the book and theirconnections
16
1 Computational Neurogenetic Modeling (CNGM): A Brief Introduction
1.7 Summary This chapter presents the motivations and the rationale behind the CNGM, along with some introduction to the main concepts, expectations and organization of the book. To conclude, we raise several questions that will be addressed in the book. Hopefully the readers will be able to answer them on the completion of the reading: • Is it possible to create a truly adequate CNGM of the whole brain? Would gene-brain maps help in this respect (see http://alleninstitute.org)? • How can a dynamic CNGM be used to trace over time and predict the progression of brain diseases, such as epilepsy and Parkinson disease? • Is it possible to use CNGM to model gene mutation effects? • Is it possible to use CNGM to predict drug effects? • How can CNGM facilitate better understanding of brain functions, such as memory and learning? • Whether and which generic problems of artificial intelligence (AI), such as classification, prediction, feature selection, pattern discovery, adaptation and visualization, etc. can be efficiently solved with the use of a brain-gene inspired CNGM (Bentley 2004)?
2 Organization and Functions of the Brain
This chapter gives an overview of the brain organization and functions performed by different parts of the brain. We will try to answer the following questions: How is the human brain organized at the macroscopic and microscopic levels? Which functions are performed by the brain? How is the organization of the human brain related to its functions? These and many more other questions about the brain are still under investigation of thousands of neuroscientists all over the world. The first Nobel Prize for pioneering discoveries related to the brain microscopic organization was given to the Spanish scientist Santiago Ramon y Cajal (1852-1934) and Italian Camillo Golgi (1843-1926). These scientists are considered to be the founders of neuroscience and modem brain study. German physicist Herman Ludwig Ferdinand von Helmholtz (1821-1894) is the founder of psychophysics, that is a quantitative experimental and theoretical research of relations between mental and brain functions. Up to these days, the division of cerebral cortex based on its microstructure introduced by a neuroanatomist Korbinian Brodmann (1868-1918), is being used. Frenchman Paul Broca (1824-1880) and Russian Alexander Romanovich Luriya (1902-1977) pioneered research on brain localization of cognitive functions based upon cognitive deficits caused by brain lesions. Even nowadays, invaluable information on brain functions comes from the study of patients with mental and neurological deficits caused by injuries of particular brain areas. Currently however, these approaches are refined with noninvasive imaging techniques like functional magnetic resonance (fMRI), positron emission tomography (PET), electroencephalogram (EEG), and others, which provide a rich source of information about the dynamics and organization of the brain. We will begin this chapter with the description of methods of brain study. Next we will introduce the reader into the basics of how human brain works, how different parts specialize and how they are interconnected, how different parts form hierarchically organized neural networks, how they cooperate and lead to behavior. Although the focus will be on human brain, many features are shared with animals especially mammals and the closest evolutionary relatives (great apes) in particular. We will mention what is happening in the human brain during cognitive processes such as perception, learning, memory
18
2 Organization and Functions of the Brain
storage and recall, thinking and language processing. We will conclude with description of relationships between different neural levels of complexity in order to discover conditions that supposedly lead to consciousness that is the capability of reflecting the outer and inner world.
2.1 Methods of Brain Study At present, a number of techniques are available to investigate where in the brain particular cognitive and other kinds of functions are based. In general, these methods are divided as being invasive or noninvasive. In medicine the term invasive relates to a technique in which the body is entered by puncture, incision or other intrusion. Noninvasive means the opposite, the technique that does not intrude into the body. We will provide an overview of both kinds of methods of brain study starting with the invasive techniques. Information about functions of the brain is still being gathered from brain damaged subjects. Deficits in cognitive processing are observed in people who have suffered some kind of brain damage, due to an accident, stroke, tumor, etc. The damaged areas indicate their involvement in those mental processes or brain functions, which became disturbed. The main problem with this method is that observations are made after the event and therefore lack the proper experimental control before the accident. Very similar in nature to this first method are lesion studies. Comparison is made between cognitive performance before and after the deliberate removal or lesion of part of the brain. These types of studies are usually performed on animals. In humans they are performed only for therapeutic reasons like dissecting the corpus callosum connecting the two hemispheres to treat the life-threatening epilepsy or removal of a lifethreatening tumor. The problem with this approach is that lesions may damage other systems which happen to be next to or pass through the target part being damaged and thus the proper involvement of the target part in a given function may be misjudged. Another invasive method of the brain study is the direct stimulation. Researchers perform electrical, magnetic or chemical stimulation of some neural circuit or part of it, and observe the consequences. Electrical stimulation is delivered through microelectrodes inserted into the brain. This type of research is done routinely on animals. It can be done on human subjects during the brain surgery when the skull has to be opened anyway and surgeons have to map the functions of the operated area and its surrounded parts. Electrical stimulation ofthe brain (ESB) can be also used to
2.1 Methods of BrainStudy
19
treat chronic tremors associated with Parkinson disease, chronic pain of patients suffering from back problems and other chronic injuries and illnesses. ESB is administered by passing an electrical current through a microelectrode implanted in the brain. With chemical stimulation, a particular chemical compound is administered into a chosen part of the brain that is supposed either to stimulate or inhibit neurons within it. The least invasive methods of the stimulation methods is magnetic stimulation, called the Transcranial Magnetic Stimulation (TMS). TMS and rTMS (repetitive TMS) are simply the applications of the principle of electromagnetic induction to get electric currents across the insulating tissues of the scalp and skull without the tissue damage. The electric current induced in the surface structure of the brain, the cortex, activates nerve cells in much the same way as if the currents were applied directly to the cortical surface. However, the path of this current is complex to model because the brain is a non-uniform conductor with an irregular shape. With stereotactic, MRIbased control (see below), the precision of targeting TMS can be as good as a few millimeters. However, besides the invasiveness there are other problems with the methods of direct stimulation. Intensity of an artificial stimulation can be stronger or weaker than the level of spontaneous activity in the target circuit. Therefore artificial stimulation can engage more or respectively less of brain circuitry than is normally involved in the studied function. Thus, there are difficulties in determining which brain circuitries have been actually affected by the stimulation and thus which brain structures actually mediate the studied function. Often used methods of brain study in animal research are the single- and multi unit recordings. Microelectrode recordings from individual neurons or from an array of neighboring neurons indicate specific neural networks dedicated to processing of particular stimuli (e.g. bars of a certain orientation, movement in a particular direction, particular objects like faces, and so on). The problem with this method albeit very precise is that it is an invasive method, i.e. requires an invasion into the brain and into the brain cells. Moreover, without post-mortem histology, it is almost impossible to tell where exactly the recordings were actually made from. Classical anatomical methods, by means of which Cajal and other pioneers made their discoveries, are histology and staining. Anatomists still use to dissect dead brains, stain their cells with different dyes (Golgi stain, lucifer yellow, etc), and study them under the microscope. Thus they can reveal the microscopic structure of the brain in terms of cell types and neural connectivity between cells. The biggest disadvantage is that this study can be performed on dead specimen only.
20
2 Organization and Functions of the Brain
The oldest noninvasive method to measure electrical activity of the brain is the electroencephalography (EEG). An EEG is a recording of electrical signals from the brain made by attaching the surface electrodes to the subject's scalp. These electrodes record electric signals naturally produced by the brain, called brainwaves. EEGs allow researchers to follow electrical potentials across the surface of the brain and observe changes over split seconds of time. An EEG can show what state a person is in (e.g., asleep, awake, epileptic seizure, etc.) because the characteristic patterns of brainwaves differ for each of these states. One important use of EEGs has been to show how long it takes the brain to process various stimuli. A major drawback of EEGs, however, is that they cannot show us the structures and anatomy of the brain and tell us which specific regions of the brain do what. In recent years, EEG has undergone technological advances that have increased its ability to read brain activity from the entire head from up to 128 sites simultaneously. The greatest advantage of EEG is that it can record changes in the brain activity almost instantaneously. On the other hand, the spatial resolution is poor, and thus should be combined with CT or MRI (sec below). Related method to EEG, called magnetoencephalography (MEG) measures millisecond-long changes in magnetic fields created by the brain's electrical currents. MEG is a rare, complex and expensive neuroimaging technique. MEG machine uses a non-invasive, whole-head, 248-channel, super-conducting-quantum-interference-device (SQUID) to measure small magnetic signals reflecting changes in the electrical signals in the human brain. The incorporation of liquid helium creates the incredibly-cold conditions (4.2 degrees of Kelvin) necessary for the MEG's SQUIDS to be able to measure these brain magnetic fields that are billions of times weaker than the earth's magnetic force. Investigators use MEG to measure magnetic changes in the active, functioning brain in the speed of milliseconds. Besides its precision another advantage of MEG is that the biosignals it measures are not distorted by the body as in EEG. Used in conjunction with MRI or fMRI (see below), to relate the MEG sources to brain anatomical structures, researchers can localize brain activity and measure it in the same temporal dimension as the functioning brain itself. This allows investigators to measure, in real-time, the integration and activity of neuronal populations while either working on a task, or at rest. The brains of healthy subjects and those suffering from dysfunction or disease are imaged and analyzed . The oldest among the noninvasive methods to study brain anatomy is Comput er Tomography (CT). It is based on the classical X-ray principle. X-rays reflect the relative density of the tissue through which they pass. If a narrow X-ray beam is passed through the same point at many different
2.1 Methods of BrainStudy
21
angles, it is possible to construct a cross-sectional visual image of the brain. A 3D X-ray technique is called the CAT (Computerized Axial Tomography). CT is noninvasive and shows only the anatomical structure of the brain, not its function. Positron Emission Tomography (PET) is used for studying the living brain activity. This noninvasive method involves an on-site use of a machine called cyclotron to label specific drugs or analogues of natural body compounds (such as glucose or oxygen) with small amounts of radioactivity. The labeled compound (a radiotracer) is then injected into the bloodstream which carries it into the brain. Radiotracers break down, giving off sub-atomic particles (positrons). By surrounding the subject's head with a detector array, it is possible to build up images of the brain showing different levels of radioactivity, and therefore, cortical activity. Thus, depending on whether we used glucose (oxygen) or some drug, PET can provide images of ongoing cortical or biochemical activity, respectively. Among the problems with this method are expense including the on-site cyclotron and also technical parameters like the lack of temporal (40 seconds) and spatial (4 mm - 1 em) resolution. Usually the PET scan is combined either with CT or MRI to correlate the activity with brain anatomy. Single-Photon Emission Computed Tomography (SPECT) uses gamma radioactive rays. Similar to PET, this noninvasive procedure also uses radiotracers and a scanner to record different levels of radioactivity over the brain. SPECT imaging is performed by using a gamma camera to acquire multiple images (also called projections) from multiple angles. A computer can then be used to apply a tomographic reconstruction algorithm to the multiple projections, yielding a 3D dataset (like in CT). Special SPECT tracers have long decay time, thus no on-site cyclotron is needed, which makes this method much less expensive than PET. However, the temporal and spatial resolution of brain activity is even smaller than in PET. Magnetic Resonance Imaging (MRI) uses the properties of magnetism instead of injecting the radioactive tracers into the bloodstream to reveal the anatomical structure of the brain. A large (and loud) cylindrical magnet creates a magnetic field around the subject's head. Detectors measure local magnetic fields caused by alignment of atoms in the brain with the externally applied magnetic field. The degree of alignment depends upon the structural properties of the scanned tissue. MRI provides a precise anatomical image of both surface and deep brain structures, and thus can be combined with PET. MRI images provide greater detail than CT images. Problems: Expense, cannot be used in patients with metallic devices, patient must hold still for 40-90 min.
22
2 Organization and Functions of the Brain
Functional MRJ (fMRI) combines visualization of brain anatomy with the dynamic image of brain activity into one comprehensive scan. This noninvasive technique measures the ratio of oxygenated to deoxygenated hemoglobin which has different magnetic properties. Active brain areas have higher levels of oxygenated hemoglobin than less active areas. An fMRI can produce images of brain activity as fast as every 1-2 seconds, with very precise spatial resolution of about 1-2 mm. Thus, fMRI provides both an anatomical and functional view of the brain and is very precise. FMRI is a technique for determining which parts of the brain are activated by different types of brain activity, such as sight, speech, imagery, memory processes, etc. This brain mapping is achieved by setting up an advanced MRI scanner in a special way so that the increased blood flow to the activated areas of the brain shows up on fMRI scans. The subject in a typical experiment lies in the magnet and a particular form of stimulation is set up (auditory, visual, etc). For example, the subject may wear special glasses so that pictures can be shown during the experiment. Then, MRI images of the subject's brain are taken. Firstly, a high resolution single scan is taken. This is used later as a background for highlighting the brain areas which were activated by the stimulus. Next, a series of low resolution scans are taken over time, for example, 150 scans, one every few seconds. For some of these scans, the stimulus (sound, picture) will be presented, and for some of the scans, the stimulus will be absent. The low resolution brain images in the two cases can be compared, to see which parts of the brain were activated by the stimulus. After the experiment has finished, the set of images is analyzed. Firstly, the raw input images from the MRI scanner require mathematical transformation to reconstruct the images into space, so that the images look like brains. The rest of the analysis is done using a series of tools which correct for distortions in the images, remove the effect of the subject moving their head during the experiment, and compare the low resolution images taken when the stimulus was off with those taken when it was on. The final statistical image shows up bright in those parts of the brain which were activated by this experiment. These activated areas are then shown as colored blobs on top of the original high resolution MRI scan, for interpretation of the experiment. This combined activation image can be rendered in 3D, and the rendering can be calculated from any angle. By means of fMRI a very comprehensive picture of brain in action can be derived. Comparisons between healthy and ill brains can be made and correlated with structural changes if these are present. There is a serious interpretation problem with all the methods that measure cortical activity (PET, fMRI, EEG, MEG). How does an experimenter decide which cortical activity is specifically related to the psychological process in question? It is done by the so-called subtraction method.
2.2 Overall Organization of the Brain and Motor Control
23
The experimenter calculates the difference image between that of the process and that of a control situation. The difference images from individual subjects are averaged to produce a group mean difference image. It can be quite problematic to determine which situation should represent the control background. It is also problematic to align individual or average functional images with anatomical structures, when comparing the coordinates of activation to a standard atlas which by no means has been proven to reflect the borders of particular areas for all people. In fact, the converse is true; most Brodmann areas differ between individuals. Nevertheless, comparing images taken during some cognitive processing to those taken before or after it, scientists are gaining many new insights about the brain structure and function.
2.2 Overall Organization of the Brain and Motor Control It is estimated that there are io'' - 1012 of neurons in the human brain (Kandel et al. 2000). Three quarters of neurons form a 4-6 mm thick cerebral cortex that constitutes a heavily folded brain surface. Cerebral cortex is thought to be a seat of cognitive functions, like perception, imagery, memory, learning, thinking, etc. The cortex cooperates with evolutionary older subcortical nuclei that are located in the middle of the brain, in and around the so-called brain stem (Fig. 2.1). Subcortical structures and nuclei are comprised for instance of basal ganglia, thalamus, hypothalamus, limbic system and dozens of other groups of neurons with more or less specific functions in operations of the whole brain. For example, the input from all sensory organs comes to the cortex preprocessed in thalamus. Emotions and memory functions depend upon an intact limbic system. When one of its crucial parts, hippocampus, is damaged, humans (and animals) loose their ability to store new events and form new memories. When a particular cortical area has been damaged, a particular cognitive deficit follows. However, all the brain parts, either cortical or subcortical, are directly or indirectly heavily interconnected, thus forming a huge recurrent neural network (in the terminology of artificial neural networks). Thus, we cannot speak of totally isolated neuroanatomic modules. Fig. 2.1 shows a schematic functional division of the human cerebral cortex. One third of the cortex is devoted to processing of visual information in the primary visual cortex and higher-order visual areas in the parietal cortex and in the inferotemporal cortex. Association cortices take about one half of the whole cortical surface. In the parietal-temporal-occipital as-
24
2 Organization and Functions of the Brain
sociation cortex, sensory and language information are being associated. Memory and emotional information are associated in the limbic association cortex (internal and bottom portion of hemispheres). The prefrontal association cortex takes care of all associations, evaluation, planning ahead and attention. Language processing takes place within the temporal cortex, parietal-temporal-occipital association cortex, and frontal cortex.
limbic association cortaX
Fig. 2.1. Gross anatomical and functional division of the human cerebral cortex. The same division applies for the right hemisphere. Dashed curves mark the position of evolutionary older subcortical nuclei in and around the brainstem of the brain. Each of the depicted areas has far more subdivisions
At the border between the frontal and parietal lobes, there is a somatic sensory cortex, which processes touch and other somatosensory signals (temperature, pain, etc.) from the body surface and interior. In the front of it, there is a primary motor cortex, which issues signals for voluntary muscle movements including speech. These signals are preceded by the preparation and anticipation of movements that takes place in the premotor cortex. The plan of actions and their consequences, inclusion and exclusion of motor actions into and from the overall goal of an organism, are performed within the prefrontal association cortex. Subcortical basal ganglia participate in preparation and tuning of motor outputs, in the sense of initiation and the extent of movements. Cerebellum executes routine automatic movements like walking, biking, driving, etc. We want to point out that
2.3 Learning and Memory
25
there are far more anatomical and functional subdivisions within each of the mentioned areas. Functions, or better, dominances of the right and left hemispheres in different cognitive functions are different (Kandel et al. 2000). It was shown by Roger Sperry and Michael Gazzaniga in the studies of the so-called "split-brain" patients to whom connections between the two hemispheres were cut because of the therapeutic reasons. The dominant hemisphere (usually the left one) is specialized for language, logical reasoning, awareness of cognitive processes and awareness of the results of cognitive processes. Although the non-dominant hemisphere (usually the right one) is able to carry out cognitive tasks, it is not aware of them or their results. It is specialized for emotional and holistic processing, intra- and extrapersonal representation of space. Its intactness is crucial for the awareness of the body integrity (Damasio 1994). Lesion of the parietal cortex including the somatosensory cortex leads to the so-called anosognosia. The limbs and the body are intact but the cortical and mental representations become missing. Patients who have undergone a stroke to the right parietal lobe, neglect the left half of their body, in spite they can see it. It is not a consequence of the left hemiparalysis. Mirror damage to the left parietal lobe does not lead to anosognosia. It seems that the right hemisphere is dominant in mental representations of intra- and extrapersonal space. In other words, subjective experience of the body self depends upon specific brain mechanisms, namely an integrity of primary and higher-order somatosensory cortical areas in the right hemisphere (Damasio 1994). Although, right part of the body is represented in the right somatosensory cortex and the left half of the body in the right hemisphere, the latter seems to have a special role in the integral self-awareness.
2.3 Learning and Memory Capability of learning and memory formation is one of the most important cognitive functions. Our identity largely depends upon what we have learned and what we can remember. We can divide the study of learning and memory into two levels: 1. The system level (where?) that attempts to answer the question which brain parts and pathways the memory trace is stored in - the top-down approach, which will be the topic of this section, and 2. Molecular level (how?), which is devoted to investigation of the ways of coding and storage of information at the cellular and molecular level the bottom-up approach, which will be introduced in the next chapter.
26
2 Organization and Functions of the Brain
It is interesting that learning and memory is independent from other cognitive functions and as such can be studied independently. It was
shown for the first time at a famous patient from 1953, which underwent a bilateral removal of the middle portion of his temporal lobe including hippocampus to treat severe epilepsy. He has lost the ability to store new memories (the so-called anterograde amnesia). He could not keep a memory of new people, objects, facts, or places no longer than few minutes. Moreover, he lost all memories from about two years before the operation (the so-called retrograde amnesia), the period which scientists consider to be the consolidation period during which memories enter the long-term memory. However, all these heavy cognitive deficits did not affect other mental functions like language processing and thinking, including IQ and implicit memory (see below).
Facts
-j-
Conditioning
Events
Priming
!
Neocortex
hippocampus and neocortex
Habituation and sensitisation
Skills
j Reflex pathways
Emotional
Classical
j
j
Amygdala
Cerebellum
Basal ganglia
Fig. 2.2. Different kinds of long-term memory fall under two general categories: explicit and implicit
It has been long recognized that there is a short-term memory and a long-term memory. Short-term memory lasts for a few minutes and is also called the working memory. It occurs in the prefrontal cortex, although
2.3 Learning and Memory
27
other parts of the cortex relevant to the memory content are activated too (Roberts et al. 1998). The learning process and the process of long-term memory formation can be divided into these four stages: 1. Encoding. Attention focus and entering of new information into the working memory. Finding associations with already stored memories. 2. Consolidation. The process of stabilization of new information, transformation into a long-term memory by means of rehearsal. 3. Storage. Long-term storing of information in memory. 4. Recall. Retrieval of information into the working memory. Based on clinical, imaging and animal studies we can divide long-term memory into two main categories that have different subtypes with different mechanisms and different localizations in the brain (Fig. 2.1). Explicit (declarative) memory is a memory of facts (semantic memory) and a memory of events (episodic memory). Recall from explicit memory requires conscious effort and stored items can be expressed in language. Hippocampus is a crucial but only a transitory stage in the explicit memory. How is the explicit memory formed? Information comes to brain through the sensory organs (visual, auditory, olfactory, tactile), and proceeds through subcortical sensory nuclei and sensory cortical areas into multimodal association areas, like for instance the parieto-temporooccipital association cortex, limbic association cortex and the prefrontal association cortex. From there the information is relayed through parahippocampal cortex, perirhinal cortex and entorhinal cortex into the hippocampus. From hippocampus the information is relayed to subiculum from where it returns back to entorhinal cortex and all the way back to association cortical areas. Thus the brain circuit for the long-term explicit memory storage forms a re-entrant closed loop. According to experimental data, the "synaptic re-entry reinforcement" or SRR hypothesis and the corresponding computational model have been formulated and simulated (Wittenberg et al. 2002, Wittenberg and Tsien 2002). According to this hypothesis, after initial learning, reactivation of hippocampal memory traces repeatedly drives cortical learning. Thus, a memory trace (engram) is stored after many repetitions. Repeated reinforcement of synapses during the reactivation of memory traces could lead to a situation in which memory traces compete, such that the strengthening of one memory is always at the expense of others, which are either weakened or lost entirely. In other words, a single memory stored in a neural network is either lost (owing to synaptic decay) or strengthened and maintained by repeated rounds of synaptic potentiation each time the memory is reactivated. Once cortical connections are fully consolidated and stabilized, the hippocampus itself becomes dispensable. Differences in the frequency with which memory traces are
28
2 Organization and Functions of the Brain
either consciously or subconsciously recalled could be another factor affecting the selection of which memories are consolidated. An increasing amount of evidence suggests a role of sleep in memory consolidation by means of learning-induced correlations in the spontaneous activity of neurons and replaying the patterns of wake neural activities during sleep (Maquet 2001, Stickgold et al. 2001). Although others point out that people lacking REM sleep do not show memory deficits and that a major role of sleep in memory consolidation is unproven (Siegel 2001). An interesting question is how the degradation of out-dated hippocampal memory traces occurs after memory consolidation is finished. The most recent hypothesis is that memory clearance may actually involve newborn neurons. Neurogenesis in the dentate gyrus of the hippocampus persists throughout life in many vertebrates, including humans. The progenitors of these new neurons reside in the subgranular layer of the dentate gyrus (Seri et al. 2001). Deletion of the Presenilin-l gene in excitatory neurons of the adult mouse forebrain led to a pronounced deficiency in enrichment-induced neurogenesis in the dentate gyrus (Feng et al. 2001). This reduction in neurogenesis did not result in appreciable learning deficits, indicating that addition of new neurons is not required for memory formation. However, the postlearning enrichment experiments led to a postulate that adult dentate neurogenesis may playa role in the periodic clearance of outdated hippocampal memory traces. The clearance can happen because these adult-born neurons are short-lived, with a life span of several weeks in rodents. The implicit or nondeclarative memory serves to store the perceptual and motor skills and conditioned reactions. Recall of stored implicit information occurs without a conscious effort, automatically and the information is not expressed verbally. Basal ganglia and cerebellum are important for acquisition of motor habits and skills that are characterized by precise patterns of movements and fast automatic reactions. Cerebellum is the key structure for conditioning. Conditioned emotional reactions require amygdala in the limbic system. Nonassociative learning like habituation and sensitization occur in primary sensory and reflex pathways. Priming is an increase in the speed or accuracy of a decision that occurs as a consequence of a prior exposure to some of the information in the decision context, without any intention or task related motivation, and occurs in neocortex. Although implicit and explicit learning concern different memory contents, they share cellular and molecular mechanisms (Bailey et al. 2004). These mechanisms will be one of the topics of the next chapter. Later we also introduce the genetics of learning and memory and the neurogenetic computational model.
2.4 Language and OtherCognitive Functions
29
2.4 Language and Other Cognitive Functions Language is distinctive from other forms of communications by its form, content, use and creativity (Mayeux and Kandel 1991): • Form. Language is formed from a limited set of "nonsense" elementary sounds (i.e. phonemes) that are arranged into various predictable sequences that signal content. • Content. Language provides means of communicating contents whose meanings are independent of the immediate situation. • Use. Through language we organize our sensory experience and express our thoughts, feelings, and expectations. • Creativity. With every new thought we speak we create original sentences. We readily interpret original sentences spoken by others.
2.4.1 Innate or Learned? Although the acquisition of language undoubtedly involves learning, developmental and neurobiological studies indicate a large innate part to it. First, both natural and sign language functions are dominantly localized to the left hemisphere, as revealed by lesion studies and brain imaging. Second, there are universal regularities in the acquisition of language across all human cultures (see Table 2.1). Children learn the words and rules of language effortlessly by simply listening to the speech around them. There exist about 6000 live languages and all of them have principal characteristics in common. Currently, there exist strong evidence from comparative linguistics and genetic anthropology that language has appeared just once, in Africa, and all other languages descend from it (Cavalli-Sforza 2001). Noam Chomsky assumes that humans have some innate program (generative grammar) that prepares them to learn language in general. In more modem terms, these innate neural mechanisms determine possible characteristics of acquired language, and the process of its acquisition. In addition, there is a critical period during postnatal development for learning language. Investigation of feral children has showed that it is impossible to acquire a fully developed language (especially syntax-vice and grammar-vice) when this critical period is over, which is at about 8 years. In summary, psychologists and linguists now believe that the mechanisms for the universal aspects of language acquisition are determined by the structures in the human brain. Thus, the human brain is innately prepared to learn and use language. The particular language spoken and the
30
2 Organizationand Functions of the Brain
dialect and accent are determined by the social environment. The questions now being debated are which language characteristics derive from neural structures specifically related to language acquisition and which from cognitive characteristics that are more general. Table 2.1. Stages of development in the acquisition oflanguage in humans Average Age
Language ability
6 months 1 year 1 VI years
Beginning of distinct babbling. Beginning of language understanding, one-word utterances. Words used singly, child uses 30-50 words (simple nouns, adjectives, action words) one at a time but cannot link them to make phrases, does not use functors (the, and, can, be) necessary for syntax. Two-word (telegraphic) speaker, 50 to several hundred words in the vocabulary, much use of two-word phrases that are ordered according to syntactic rules, child understands propositional rules. Three or more words in many combinations, functors begin to appear, many grammatical errors and idiosyncratic expressions, good understanding of language. Full sentences, few errors, vocabulary of around 1000 words. Close to adult speech competence.
2 years
2 VI years
3 years 4 years
2.4.2 Neural Basis of Language The basic model of language processing during the simple task of repeating the word that has been heard is the Wernicke-Geschwind model (Mayeux and Kandel 1991) (Fig. 2.3). According to this model, the language task involves transfer of information from the inner ear through the auditory nucleus in thalamus to the primary auditory cortex (Brodmann's area 41), then to the higher-order auditory cortex (area 42), before it is relayed to the angular gyrus (area 39). Angular gyrus is a specific region of the parietal-temporal-occipital association cortex, which is thought to be concerned with the association of incoming auditory, visual and tactile information. From here, the information is projected to Wernicke's area (area 22) and then, by means ofthe arcuate fasciculus, to Broca's area (44, 45), where the perception of language is translated into the grammatical structure of a phrase and where the memory for word articulation is stored. This information about the sound pattern of the phrase is then relayed to the facial area of the motor cortex that controls articulation so that the word can be spoken. It turned out that a similar pathway is involved in naming an object that has been visually recognized. This time, the input
2.4 Language and Other Cognitive Functions
31
proceeds form retina and LGN (lateral geniculate nucleus) to the primary visual cortex, then to area 18, before it arrives to the angular gyrus, from where it is relayed by a particular component of arcuate fasciculus directly to Broca's area, bypassing Wernicke's area. Lesions in different parts of the cerebral cortex cause selective language disturbances , called aphasias, instead of an overall reduction in language ability. Normal language is dependent not only on cortical but also on subcortical structures and connections. Lesions that do not affect the cerebral cortex, typically vascular lesions in the basal ganglia and/or thalamus, can also result in aphasia. Basal ganglia take part in motor output and thalamus in perception. Furthermore , the damage to the brain language areas often affects also other cognitive and intellectual skills to some degree.
Occipital cortex
Eye
Temporal cortex
Fig. 2.3. Lateral view of the exterior surface of the left hemisphere with the main language processing areas. Broca's area (Brodmanu's areas 44/45) is adjacent to the regions of premotor (6) and motor (4) cortices that control the movements of facial expressions, articulation and phonation. Wernicke's area (area 22) lies in the posterior superior temporal lobe near the primary and higher-order auditory cortices in the superior temporal lobe (areas 41/42). Wernicke's and Broca's areas are joined by a fiber tract called the arcuate fasciculus
Wernicke's aphasia is characterized by a prominent deficit in language comprehension. The lesion primarily affects area 22 (Wernicke's area), and often extends to the superior portions of the temporal lobe (areas 39/40), and inferiorly to area 37. Comprehension of both auditory and visuallanguage inputs is severely impaired, accompanied with severe reading
32
2 Organization and Functions of the Brain
and writing disabilities. Speech is fluent, grammatical, but lacks meaning. Phenomena like empty speech, neologisms, and logorrhea occur. Patients are generally unaware of these speech failures, probably because of the lack of their own language comprehension. Occasionally, a right visual field defect is encountered. Conduction aphasia is the result of damage to the arcuate fasciculus. Symptoms resemble those of the Wernicke's aphasia. Many patients with conduction aphasia have some degree of impairment of voluntary movement.
j
Speech
Writing
Visual processing Visual cortex (areas 17, 18, 19)
Auditory processing Auditory cortex (areas 41/42)
j Language comprehension Temporo-parietal areas (areas 39, 22, 37, 40) Semantic association Left anterior inferior frontal cortex Motor control Left prem otor cortex (areas 44, 45, 6)
Speech
1
Writing
Fig. 2.4. Recent model of the neural processing of language build upon the Wernicke-Geschwind original model. Simplified scheme shows the relationships between various anatomical structures and functional components of language. Connections are actuallyreciprocal
Broca's aphasia is characterized by a prominent deficit in language production. Lesions affect areas 44 and 45 (Broca's area), and in severe cases also other prefrontal regions (8, 9, 10, 46) and premotor regions (area 6). The most severe case is the complete muteness. Usually, speech contains
2.4 Language and OtherCognitive Functions
33
only key words, nouns are expressed in singular, verbs in the infinitive. Articles, adjectives, adverbs and grammar are missing altogether. Unlike Wernicke's aphasia, patients with Broca's aphasia are generally aware of these errors. Reading and writing are also impaired, because they include also motor components. Some defects in comprehension related to syntax may be encountered. Right hemiparesis and loss of vision is almost always present in this type of aphasia. Lesions to the prefrontal cortical regions other than Broca's area or to the parietal-temporal cortical regions other than Wernicke's area can result in various language deficits in production or comprehension, respectively. When certain portions of higher-order visual areas are damaged, specific disorders of reading and/or writing follow (dyslexias, alexias and agraphias). Homologous language areas in the right hemisphere process affective components of language like musical intonation (prosody) and emotional gesturing. Disturbances in affective components of language associated with damage to the right hemisphere are called aprosodias. The organization of prosody in the right hemisphere seems to mirror the anatomical organization of the cognitive aspects of language in the left hemisphere. Thus, patients with posterior lesions do not comprehend the affective content of other people's language. On the other hand, lesion to the anterior portion of the right hemisphere leads to a flat tone of voice whether one is happy or sad. To sum up, recent cognitive and imaging studies have revealed that language processing involves a larger number of areas and a more complex set of interconnections than just a serial interconnection of Wernicke's area to Broca's area. Thus, a more realistic scheme illustrating the neural processing of language is shown in Fig. 2.4. 2.4.3 Evolution of Language, Thinking and the Language Gene
In most individuals the left hemisphere is dominant for language and the cortical speech area of the temporal lobe (the planum temporale) is larger in the left than in the right hemisphere. Since important gyri and sulci often leave impression upon the skull, it is possible to examine human fossils in search for such impressions. Marjorie LeMay searched the fossil skulls for the morphological asymmetries associated with speech and has found them besides in the modem Homo sapiens also in the Neanderthal man (Homo sapiens neanderthalensis, dating back 30,000 to 50,000 years) and in Peking man (Homo erectus pekinensis, dating back 300,000 to 500,000 years). The left hemisphere is also dominant for the recognition of species-
34
2 Organization and Functions of the Brain
species cries in Japanese macaque monkeys, and asymmetries similar to those of humans are present in brains of modem-day great apes. G. Rizzolatti et al. (DiPellegrino et aI. 1992, Rizzolatti et aI. 1996) have found out that neurons in the ventral premotor cortex of macaque monkeys are active not only when the monkey executes motor actions, but also when she watches others, either monkeys or humans, to perform the same actions. Thus, these mirror neurons follow or imitate what others are doing. They may form a neural basis for learning by imitation, which is very important for language acquisition. In humans, the ventral premotor area includes Broca's area (areas 44/45), which is a specific cortical area associated with expressive and syntactical aspects of language. Thus maybe, evolution of the ventral premotor area with its mirror neurons played an important role in evolution of neural basis for language (Rizzolatti and Arbib 1998). It is also intriguing to see areas responsible for contemplation of (motor) actions and areas processing language being at the same place in the brain . Thus thinking can be hypothesized to be a contemplation of actions in the real or abstract spaces. According to Rizzolatti and Arbib (Rizzolatti and Arbib 1998) and Corballis (Corballis 2003) speech and language have evolved from the communication gestures and not from the vocal communication of primates. Vocal production of primates is controlled in the emotional cortical and subcortical centers and serves mainly to communicate emotional state (anger, fear, content, etc.). On the other hand, communication modality of gestures is visual and motor and involves the premotor cortex. Our ancestors probably started to use the gestures and mirror neurons to communicate nonemotional contents. Brachial and manual gestures were probably accompanied with the oro-facial movements and differentiating sounds. Communication gestures started to be associated with these accompanying sounds on a regular basis, which has led to the liberation of these sounds in language. Although the anatomical structures that are prerequisites for language may have arisen early, many linguists believe that language per se emerged rather late in the prehistoric period of human evolution (about 100,000 years ago). There exist strong evidence from comparative linguistics and genetic anthropology that language arose only once, in Africa, and all other languages descend from a single original language (CavalliSforza 2001). Modem Homo sapiens, which evolved in Africa, began to leave Africa some 100,000-80,000 years ago, while taking this first language with them. This theory fits well with the estimated time of fixation of the so-called language gene FOX?2 in modem humans (Enard et aI. 2002). FOX?2 (fork-head box P2) is located on human chromosome 7q31. It codes a protein containing 715 amino-acid bases, which belongs to the class of the so-called fork-head transcription factors that control transcrip-
2.4 Language andOther Cognitive Functions
35
tion of DNA. So far it is not known which transcription is controlled by FOXP2. This gene is present in all mammals however humans gained two mutations compared to chimpanzees, gorillas and macaques, which have identical proteins FOXP2. It seems that two functional copies of the human gene FOXP2 must be present to acquire a full language. People with the point mutation of one of the gene copies have serious problems with articulation, grammar, expression and comprehension of language. One of the consequences of this mutation is the disorder of sequencing of subtle oro-facial movements. This motor disorder is accompanied with serious mental problems with sequencing the syllables into words and words into grammatically correct sentences. Over the years much evidence has accumulated to support the idea that aspects of our genetic makeup are critical for acquisition of language (Marcus and Fisher 2003). To gain insights into the evolution of language and also into its neural basis, studies of communication systems and language abilities of great apes like chimpanzees and bonobo have been very helpful (SavageRumbaugh and Lewin 1994). In short summary, bonobos' capacity for processing grammar even after many years of proper training remains limited. With respect to the language production, they are at the level of a two-year child (see Table 2.1). With respect to the language understanding, they can be at the level of 2.5-3-year child. Thus, so far it seems that fully developed language is an exclusive form of human communication. An important topic in the study of language is the relation of language and its evolution to other cognitive functions and their evolution, respectively (Gardenfors 2000, Marcus 2004b). Language is to communicate about something that is not here and not now. Thus, a more general cognitive abilities such as being able to create detached representations and being able to make anticipatory planning (planning about future needs, goals, events, etc.) can be necessary (but not sufficient) cognitive prerequisites for language. Grammar is an enhanced formal mean for organization of language. Except the evolution of general cognitive abilities, the evolution of language may go hand in hand with the development of advanced forms of cooperation. Without the aid of symbolic communication about detached contents, we would not be able to share visions about the future. We need language in order to convince each other that a future goal is worth striving for (Gardenfors 2000).
36
2 Organization and Functions of the Brain
2.5 Neural Representation of Information The first principle of representation of information in the brain is redundancy. Redundancy means that every information (meant in any sense) is stored, transmitted and processed by a redundant number of neurons and synapses so that it does not become lost when neural networks undergo damage, for instance due to aging. When neural networks get damaged, their performance does not drop down to zero abruptly, like in a computer, but instead it degrades gracefully. Computer models of neural networks also confirm the idea that a degradation of performance with the loss of neurons and synapses is not linear but instead neural networks can withstand quite substantial damage, and still perform well. Next, the contemporary view on the nature of neural representation is such that information (in the sense of content or meaning) is represented by place in the cortex (or in general in the brain). However, this placing is a result of anatomical framework and shaping by input, i.e. by experience-dependent plasticity. For instance, a sound pattern for the word "apple" is represented in the auditory areas of the temporal cortex. It is represented as a spatial pattern of active versus inactive neurons. This neural representation is associated (connected) through synaptic weights with the neural representation of a visual image of apple in the parietal cortex, with the neural representation of an apple odor in the olfactory cortex, with memories on the grandma garden and facts about apples, being represented in some other areas of the cortex, etc. Neural representations (that is distributions or patterns of active neurons) within particular areas and their associations between areas appear as a result of learning (i.e. synaptic plasticity). Different objects are represented by means of different patterns or distributions of active neurons within cortical areas. Therefore we speak about the so-called distributed representations. Current hypothesis states that recall from memory is an active process. Instead of passive processing of all electrical signals that arrive from hierarchically lower processing levels, cortical neural networks should be able to use fragments of activity patterns to fill in the gaps, and thus quickly recreate the whole neural representation. The filling-in process can be nicely modeled by means of model neural networks (Fig. 2. 5). Neural representations (patterns of activities) are stored in the matrix of synaptic weights through which neurons in the network are interconnected. The weight distribution storing a particular object representation is created due to an experience-dependent synaptic plasticity (learning). When a sufficiently large portion of this neural representation is activated from outside the network,
2.6 Perception
37
few electric signals along all the synapses in the network quickly switch on the correct remaining neurons in the representation. Neural representations in the sense of patterns of activity have a holistic character. Patterns of activity are being recalled (restored) as a whole. Thus, we can see a nice relation between the character of neural representations and gestalts. Gestalt psychology was developed at the beginning of the 20th century by Max Wertheimer, Kurt Koffka and Wolfgang Kohle in Germany. Gestalt psychology considers holistic mental gestalts (shapes, forms) to be the basic mental elements. For the gestalt to be stored and recalled, certain rules must be fulfilled, like the rules of proximity, good continuation, symmetry, etc. These rules have been experimentally verified.
.. • ....••
lIiiii••
• :=iii=:::
-·-·:-1
11111111· .Iiiiil•
•;:iiii:: .
iiiiiiii :. Fig. 2.5. Illustration of spontaneous re-creation of neural representation after few input impulses(figure in the uppermost left corner). Black pixel represents a firing neuron while blank pixel represents a silent neuron. Between each pattern of activity from left to right (I ms time frame), neurons in the network exchange only one impulse. Thus, basically after exchanging only two-three spikes, the memory pattern is re-created. Network remains in the restored memory pattern until a different external input arrives
To conclude, neural representations of objects are stored in the matrix of synaptic weights as a whole. We are not able to trace down a sequence of steps leading to the holistic percept. Synaptic weights implicitly bind together parts of the pattern.
2.6 Perception Perception is accompanied by sensory awareness, and therefore we will describe the underlying neural processes in relation to the next section on consciousness. We will concentrate on visual perception and visual awareness since similar principles apply to all sensations. Neurons in different areas of the visual cortex respond to various elementary features, like ori-
38
2 Organization and Functions of the Brain
ented edges of light intensity (bars), binocular disparity, movement, color, etc. (Kandel et al. 2000). Visual areas in the occipital, parietal and inferior temporal cortex, though reciprocally connected, are hierarchically organized. Results of processing at lower hierarchical levels are relayed to higher-order areas. Neurons in higher-order areas respond to various combinations of elementary features from lower-order areas. In primates, based on matching psychophysical and physiological data, three main visual systems, relatively independent but mutually heavily interconnected, have been identified: the "magno", "parvo" and the color system (Livingstone and Hubel 1988). The "magno" system is responsible for perception of movement, depth and space, and separation of objects. Several cues leading to the depth perception have been identified: stereopsy, depth from perspective, depth from mutual movement and occlusion, etc. The "parvo" system is responsible for shape recognition. For separation and recognition of objects, we use separation based on movement, separation from background, filling in of borders, shape from shading, etc. The color system is responsible for color perception. With respect to cortical neurons belonging to these three systems, they possess different combinations and ranges of these four physiological properties: sensitivity to color (small/ large), sensitivity to the light contrast (small/large), temporal resolution (small/large), spatial resolution (small/large). These are the so-called elementary features of visual objects. Elementary features belonging to one visual object activate different and spatially separated groups of neurons within the cerebral cortex. Scientists at the Max Planck Institute in Germany under the leadership of Wolf Singer analyzed trains of spikes of neurons within the visual cortex. They have proposed an intriguing hypothesis about the neural correlate of perception and sensory awareness (Gray et al. 1989, Singer 1994, Roelfsema et al. 1997). Binding of spatially separated neurons coding for features belonging to one visual object could be performed by transient synchronization of firing of these neurons (Fig. 2.6). Similar synchronous oscillations of neurons were detected also in auditory, somatosensory, parietal, motor, and prefrontal cortices in the case of auditory, tactile and other perceptions, respectively (Traub et al. 1996). Oscillations of neurons with frequencies around and above 40 Hz (long known as gamma oscillations) have been detected in the cerebral cortex of humans, primates and other investigated mammals, in particular as a result of sensory stimulation. This synchronization occurs over relatively long distances (mm to em), between different cortical areas, between cortex and thalamus, between the two hemispheres. Synchronization means that neurons discharge with the same frequency and the same phase (Fig. 2.6). This results in a distributed pattern of simul-
2.6 Perception
39
taneously firing neurons. Neural correlates of different objects can differ in (a) which neurons are members of the pattern, (b) which is the particular frequency of their synchronization, and (c) which is the phase of their synchronization. Thus, transient synchronous gamma oscillations have been suggested as a possible candidate for the mechanism of binding many elementary features belonging to one object to one transient whole corresponding to a percept. Establishment of transient synchrony is based upon the underlying synaptic connectivity. In the laboratory ofW. Singer, an interesting experiment was performed to demonstrate that when these synchronizations are disturbed, perception is also disturbed (Konig et al. 1996). In normal rearing, kittens develop normal sharp vision. Neurons in their primary visual cortex are sharply orientationally selective. Distribution of ocular dominance favors binocular neurons; however there are still also monocular neurons present. Firing of neurons responding to the left and right eye is synchronized when they are visually stimulated with the same object. After birth, the eyesight of kittens was disturbed in such a way that surgically they were made strabismic to one eye. They developed a syndrome typical for a strong uncorrected strabismus. Objects are fixated only with the good eye, whereas the strabismic eye is connected with a perceptual deficit called strabismic amblyopia (something like a blurred vision). Analysis of spike trains in response to visual stimuli in the primary cortex of these cats have revealed several facts: (i) neurons that respond only to the stimulation of the strabismic eye have normal physiological characteristics in terms of orientation selectivity, (ii) however synchronization of discharges of these neurons is very weak compared to synchronization of discharges of neurons responding to the stimulation of a good eye. (iii) There is no synchronization between firing of neurons responding to the good and strabismic eye. Thus, these animals have a perceptual deficit connected to the strabismic eye and this perceptual deficit is accompanied by the desynchronization in the visual cortex. Absence of synchronization means troubles in simultaneous binding of features belonging to one object which results in a blurred vision. Another experimental phenomenon strongly suggesting a one-to-one correspondence between transient synchronizations and perception is binocular rivalry. During binocular rivalry, each eye is constantly stimulated with a different pattern. Visual percept is neither an average of these two patterns nor their sum. Instead, a random alternation between the two percepts occurs as if they were competing with each other, hence the term binocular rivalry. Fries et al. (Fries et al. 1997) discovered that neurons which respond to one or the other pattern are synchronized only during the corresponding percept. Thus, although the pattern is constantly stimulating
40
2 Organization and Functions of the Brain
an eye, cortical neurons get synchronized only when the pattern is perceived.
a
• • c
e
b
' . ·/;::····I ·~~>· · ·
. ', _. \', : • •
'- '
d
..
:
I
.:.
-::
....
. ( , .. • •
\:
. ,
.....
-
Fig. 2.6. (a) Scheme illustrating spatially separated specialized neurons in the visual cortex. Each of them responds to a different elementary feature of one object; (b) Record of firing of these neurons before, during and after the presentation of that object. During the presence of object that activates a given set of neurons, their spikes are synchronized. These synchronizations repeat several times with some period, thus we speak about synchronous oscillations. Period of object presentation is denoted by a thick bar under the spikes. Activities before and after the object presentation are desynchronized spontaneous discharges of neurons; (c) The same as in (a), however this time a different object is presented, which activates a different set of neurons (features); (d) The same as in (b), the spikes of a different set of neurons are synchronized thus binding together features belonging to their object; (e) Illustration of different frequencies and phases of oscillations. Full and dashed curves denote oscillations of the same frequencies but occurring with different phases. The dotted curve denotes an oscillation with the frequency three times frequency of the first two oscillations
An important study of Rodriguez et a1. (Rodriguez et a1. 1999) has demonstrated that perception of faces in humans is accompanied by a transient (~180 ms) synchronization of gamma activity in hierarchically highest visual areas in the parietal cortex and premotor areas in the frontal cortex. They used sophisticated computational procedures for analysis of EEG
2.7 Consciousness
41
signals recorded during the task of face recognition. Humans were presented with the so-called Mooney faces either in the canonic upright position or in an upside down position. These black-and-white pictures are made out of the very high contrast photographs of human faces. It is very difficult to recognize a human face when they are presented upside down. Viewing these images, either in a normal or turned position is always accompanied with the increase of gamma activity in the visual areas. However, precise transient synchronization of gamma oscillations occurs only when the subject perceives a face. It is intriguing that this synchronization occurs only in the left hemisphere which is the so-called conscious hemisphere. When the subject did not perceive a face, but instead only a nonsense arrangement of black-and-white patches, no synchronization happened in the cortex. The second transient episode of synchronization occurred in the premotor and motor areas of both hemispheres during the motor response of subjects by which they indicated whether they saw a face or a no-face. Thus, transient synchronizations may accompany also other cognitive processes not only perception. W. Miltner et al. (Miltner et al. 1999) indeed detected synchronization of gamma oscillations during an associative learning. Humans were supposed to learn to associate a visual stimulus with the tactile stimulus . A selective synchronization occurred between the visual cortex and that part of somatosensory cortex which represented the stimulated hand, during and after the learning. When people forgot the learned association, synchronization between these two stimuli, or rather between neural responses to these two stimuli, disappeared.
2.7 Consciousness When we speak of consciousness we usually mean reflective or secondary consciousness, "the inner eye of our mind". Present neuroscience has a good reason to assume that neural mechanisms of reflective consciousness are derived from the mechanisms of sensory awareness that is related to perception (Singer 1999b). Thus, in building the picture of neural correlates of reflective consciousness we will proceed through assumed neural correlates of sensory awareness that is sometimes referred to as primary or phenomenal consciousness.
2.7.1 Neural Correlates of Sensory Awareness Currently , transient (l00-200 ms) synchronous gamma oscillations are being studied as a promising candidate for the mechanism of binding many
42
2 Organization and Functions of the Brain
elementary features belonging to one object to one transient whole corresponding to a percept of that object (Engel et al. 1999, Singer 1999a). Such synchronized activity summates more effectively than nonsynchronized activity in the target cells at subsequent processing stages, and the activity can spread to a longer distances. If so, synchronization could increase the effect that a selected population of neurons has on other populations with great temporal specificity (in the range of milliseconds). There is also evidence that synchrony is important for inducing changes in synaptic efficacies and hence facilitate transfer of information into memory. Different objects in one scene may be associated with different phase-locked synchronous oscillations within the gamma frequency band. Thus, increased coherence between brain areas confined to a narrow band around 40 Hz may denote a holistic perception of a complex stimulus. Based on experimental findings, crucial neural conditions for a conscious percept to be experienced is (Koch and Crick 1994, Singer 1994, Crick and Koch 1995, Koch 1996, Rodriguez et al. 1999): • Over the chain of primary and higher-order sensory areas with the areas that have direct connections to the frontal cortex being at the end of this chain (e.g. the posterior parietal cortex), and over the evolutionary youngest cortical areas, i.e. the frontal and prefrontal cortex, certain suprathreshold quantity (number) of neurons must be coherently active for a certain time of 100-200 milliseconds (see Fig. 2.7).
Fig. 2.7. Corticocortical connections between the posterior parietal cortex and the main subdivisions of the frontal cortex. Illustrated areas showed increased coherence within the 40 Hz band in the Rodriguez's et al. experiment on recognition of Mooney faces. When a human face was recognized, transient coherence occurred in the time window of 180-360 ms after the beginning of the picturepresentation
2.7 Consciousness
43
Let us go through this condition in greater detail. Why higher-order sensory areas? Because they code for invariant object features, and thus come closer to the invariant object identification (Engel et al. 1999). Primary sensory areas also take part in the chain and their activity must take part in a reciprocal reverberant interaction with higher-order areas (Silvanto et al. 2005). With respect to the quantitative condition, that a certain suprathreshold number of neurons must be synchronized within relevant cortical areas: Electrophysiological measurements on blind-sighted monkeys and fMRI on blind-sighted humans have shown that besides the superior colliculus, also the hierarchically higher visual cortical areas remain responsive to visual stimuli when VI is inactivated or damaged (Sahraie et al. 1997). Thus, blindsight seems to be mediated by both, intact relays within the extra-geniculostriate pathway which leads to superior colliculus, and also by the sparse and spared relays within the retino-geniculatecortical pathways themselves. However, neither subcortical structures nor an insufficient number of active cortical neurons can lead to a conscious percept. There is also an intuition from theory: in order for a large synchronization to occur in some physical system, a certain threshold number of elements must start the process otherwise synchronization does not spread over distance. Generating sensory awareness involves the process of attention. Several areas in the prefrontal cortex are crucially involved in attention, namely areas 8Av (major connections with the visual system), 8Ad (major connections with the auditory system) and 8B (major connections with the limbic system) (Roberts et al. 1998). Attentional selection may depend on appropriate binding (coherence) of neuronal discharges in sensory areas in two simultaneously active directions: an attentional mechanism in prefrontal cortex could induce synchronous oscillations in selected neuronal populations (top-down interaction) , and strongly synchronized cell assemblies could engage attentional areas into coherence (bottom-up interaction) (Singer 1994). Another prefrontal areas activated during sensory perception include areas 9, 10,45,46,47 (see Fig. 2.8). These prefrontal areas are known to be involved in an extended action planning. In addition, these prefrontal areas plus the posterior parietal cortex are known to be involved in the working memory. Posterior parietal cortex is also known to be involved in mental imagery. For planning of actions it is necessary to keep track of at least one sequence of partial actions, hence the overlap between planning and memory mechanisms . It might be that sensory contents reach awareness only if they are bound to prefrontal areas via the posterior parietal cortex and thus have a possibility to become part of the working memory and action planning (Engel et al. 1999). In tum, action planning may influence
44
2 Organization and Functions of the Brain
organization of attentional mechanisms and thus what is being perceived. Actually, the underlying action planning can occur at a subconscious level (Libet 1985,1999). Coherences in the involved areas are generated internally within the cortex and although they are phase-locked, they are not stimulus locked. They are superimposed upon global thalamocortical gamma oscillations which are generated and maintained during cognitive tasks (Ribary et al. 1991). Thalamocortical oscillations may provide the basic oscillatory modulation of cortical oscillations. Other cortical mechanisms are then responsible for a precise phase-locking of internal cortical synchronous oscillations. In particular, these are lateral inhibitory and excitatory interactions, regularly bursting layer V pyramidal cells, and spike-timing dependent rapid synaptic plasticity. In the latest, synapses and thus the inputs which do not drive the postsynaptic cell in synchrony are temporary weakened . Lateral
.~.!3." " " "''' '''' ' '''' ' 7'....~...''
..····8Ad
6
'..
.........·; ;: 6d / ..·8·;..:;;;·
..······)
..... ( ..··· 46
9/46v
88
.......................... , ......i:':
! 44\6
.................. 45A 458
Medi al
cc
~.~
.
9
,
.
/ ..
--'( ;;_ :;=!=3;=:'~ ~O
.
47;:j'2·..·..·
45A 47/12
C·..· ~·3. ·...
~
·11........ 10
....···14
Orb ito-frontal
..../
Fig. 2.8. Human prefrontal cortex. Lateral view (from outside), medial view (from inside) and the orbito-frontal view (from below) at the left hemisphere. The same divisions hold also for the right hemisphere. Numbers denote the corresponding Brodmann's areas. CC means corpus callosum
2.7.2 Neural Correlates of Reflective Consciousness
Since early childhood, we are engaged in learning, first through nonverbal and later through verbal communication, to assume what is going on inside of other people. Our reflective consciousness and our self-reflection develop gradually, step by step. Due to learning to assume what is going on
2.7 Consciousness
45
inside of other people, we can learn to assume what is going on inside of ourselves. Self-reflection is possible only thanks to communication and social interaction. However, it seems that our brains already possess certain structures that have been prepared and selected for this task - these are mirror neurons (Rizzolatti et al. 1996) and mentalization module (Frith 2001). In 1996, scientific community got stirred by the discovery of mirror neurons. G. Rizzolatti et al. (1996) recorded activity of neurons in the ventral premotor cortex of macaque monkeys. They have found out that these neurons are active not only when the monkey executes its own actions, but also when she watches others, either monkeys or humans, to perform the same action. Thus these neurons fire in the same way when others perform a given action as when a monkey performs the same action. The observation/execution matching system represents a given action irrespectively who performs it. Mirror neurons may be a part of the mind reading system. People have the ability to explain and predict behavior of others in terms of their presumed thoughts and feelings. The ability to attribute mental states to others and to ourselves is called "mentalization" or "theory of mind". Noninvasive brain imaging has shown that the ability to attribute various mental states, desires and beliefs to others and also to ourselves depends upon full functioning of a specific neurocognitive module (Frith 2001). The mentalization module includes in both hemispheres: (i) the medial prefrontal cortex (area 32), in particular the most anterior part of paracingulate cortex, a region on the border between anterior cingulate and medial prefrontal cortex (very medial), (ii) the temporal-parietal junction at the top of the superior temporal gyrus (stronger on the right), and (iii) the temporal poles adjacent to the amygdala (somewhat stronger on the left). Neural activity in all three or at least in the prefrontal part of this mentalization module as revealed by the brain imaging is significantly lower in autistic people (Frith 2001). Autistic people are not able to "read out" neither the mind of others nor the mind of themselves, while there is a whole spectrum of severity of autistic disorder. Mentalization module overlaps substantially with the brain higher-order emotional system. Medial parts of the prefrontal cortex (area 32) and orbitofrontal parts (i.e. areas 10, 11, 12, 13, 14) are evolutionary younger parts of the brain emotional system (Damasio 1994). These medial and orbital prefrontal areas are thought to be responsible for the so called secondary emotions. Secondary emotions are emotional feelings based on learned variety of associations between primary emotions and life situations. Hierarchy of these associations involves planning and strategies related to one's social role and personal goals in relation to the past and future. Evaluation and planning and feelings in the social and emotional spheres are therefore linked to be processed by the same structures in the prefrontal cortex, and
46
2 Organization and Functions of the Brain
these overlap with the mentalization module. According to Damasio (1994), the feeling of self, and consequently the awareness of self, would depend also on the intactness of the somatosensory system, on the signaling from the cortex down to the body and back. As we have said, it might be that sensory contents reach awareness only if they are temporarily synchronized with activity in the prefrontal areas, thus displaying a highly coherent joint activity. In such a way they can become part of conscious working memory and action planning. According to Singer (l999a) reflective consciousness would be based upon the same processes, i.e. highly coherent activity, happening over the prefrontal areas involved in planning and working memory and between areas devoted to representations of our inner world. Secondary consciousness or metaawareness would result from iteration of the very same processes that support primary consciousness, except that they are not applied to the signals arriving from the sensory organs, i.e. from the outer world, but to the outputs of previous cognitive operations (Singer 1999b). By means of recording electromagnetic activity of the brain it is possible to capture and visualize the fast semi-global coherent activity of the brain that accompanies conscious perception of a stimulus (Tononi and Edelman 1998). It is almost a magical view, because this semi global coherent activity changes with time as a burning fire which is boosted from the centre of the brain and its flames transiently engage currently synchronized brain areas (see Fig. 2.9). Why is the fire boosted from the centre of the brain? Clinical research has revealed that damaging intralaminar nuclei in the thalamus leads to the loss of consciousness and to coma. Neurons in intralaminar nuclei possess dense reciprocal connections to and from the brain cortex. Intralaminar nuclei are the source of arousal, without which the cortex cannot function. G. Edelman and G. Tononi (Edelman and Tononi 2000) call this ever changing semi global coherent activity, the dynamic core. The dynamic core corresponds to a large (semi global) continuous cluster of neuronal groups that are coherently active on a time scale of hundreds of milliseconds. Its participating neuronal groups are much more strongly interactive among themselves than with the rest of the brain. The dynamic core must also have an extremely high complexity as opposed to for instance convulsions. Each roughly 150 ms, a pattern of semi-global activity must be selected within less than a second out of a very large, almost infinite, repertoire of options. Thus, the dynamic core changes in composition over time. As suggested by imaging, exact composition of the core varies significantly not only over time within one individual, but also vary significantly across individuals.
2.7 Consciousness
47
According to Edelman and Tononi (2000), the dynamic core consists of a large number of distributed groups of neurons which enter the core temporarily based on their mutual coherence. Connecting groups of neurons into temporarily synchronized whole requires dense recurrent connections between brain areas, along which a reiterated reentry of signals occurs. Neural reference space for any conscious state may be viewed as an abstract N-dimensional space, where each axis (dimension) stands for some participating group of neurons that code for (represent) a given aspect of the conscious experience. There can be hundreds of thousands of dimensions. The distance from the beginning of the axis represents the salience of that aspect. It may, for instance, correspond to the number of firing neurons within a given group. We would like to point out the interesting similarity between this abstract N-dimensional neural space and the conceptual spaces introduced by Gardenfors (2000). Heanng
Imagery
Fig. 2.9. (a) Illustration of the dynamic core, a changing coherent semi global activity of the brain, which is supposed to be a neural correlate of consciousness. One configuration of the core lasts for about 150 ms; (b) Interpretation of the dynamic core as an N-dimensional neuronal reference space, where each axis (dimension) denotes some group of neurons which encodes (represents) a given aspect of the conscious experience. Each axis can be broken down into more elementary axes. There can be hundreds of thousands of dimensions
What would be, in this theory, a neural basis for subconsciousness? The same group of neurons may at times be part of the dynamic core and underlie conscious experience, while at other times it may not be part of it and thus be involved in subconscious processing. Koch and Crick (1994) have proposed that those active neurons which are not at the moment tak-
48
2 Organization and Functions of the Brain
ing part in the semi global activity keep processing their inputs, and results of this processing may still affect behavior. We would like to mention also the explanation of neural correlate of qualia or the hard problem of consciousness, according to Edelman and Tononi (2000). Qualia are specific qualities of subjective experiences, like redness, blueness, warmth, pain, and so on. According to the dynamic core hypothesis, pure redness would be represented by one particular state of the dynamic core that is by one and only one point in the N-dimensional neural space. This core state would certainly include large participation of neurons that code for the red color and a small participation of neurons that code for other colors and for anything else. Coordinates of a point in the Ndimensional reference neural space are determined by activities of all neuronal groups that are at the moment part of the core. And these activities vary in time and across individuals. Thus, the subjective experience of redness will be different in different people and can be different for the same individual for instance in the morning and in the evening. Another mystery is why consciousness fades as we fall asleep even when we nowadays know that the brain and especially the cortex remain highly active. Sleep research has revealed that during sleep, humans normally go through two-three cycles of two sleep phases. One of these two phases is the so called REM sleep, according to the accompanying Rapid Eye Movements. EEG activity of the brain during the REM phase is very similar to the EEG activity of the awake brain during cognitive activity. Hence the term paradoxical sleep for the REM sleep phase, as it was not sleep at all. We dream mostly during REM sleep phases. When awakened during the REM phase, we can recall the content of a dream. We experience self-awareness when we dream but not when we are in the deep sleep (Llinas and Ribary 1994). Thus "1" is preserved during dreaming as well as the awake-like EEG activity of the brain. When awakened around at the end of the REM phase, we can remember that we dreamt, not knowing about what. When awakened during the non-REM sleep phase, we mostly deny any experience of dreaming. The non-REM sleep phase is also called the deep sleep, and the brain activity occurs in typical slow large regular waves. Recently, experiment with the spread of activity within neocortex during sleep have revealed that different cortical areas stop communicating over distance with each other during the non-REM sleep - a stage of sleep for which people mostly report no or very little conscious experience on waking (Massimini et al. 2005). Thus, it seems that the coherent semi global activity is disrupted during the non-REM sleep, and so is the conSCIOUS awareness.
2.8 Summary and Discussion
49
2.8 Summary and Discussion Neurobiologists have made a great progress in trying to find the so-called neural correlates of consciousness. However, not every scientist finds this research compelling . In his influential book Shadows of the Mind, physicist Roger Penrose brings the problem of explaining consciousness to the domain of elementary laws of physics (Penrose 1994). According to him, at present any scientific (including physical) theory does not help us to come to terms with the puzzle of mentality including consciousness within such a physically determined universe. He has been calling for a radical upheaval in the very basis of physical theory. He sees the consciousness link in a new physical theory based upon the union of Einstein's general relativity with quantum theory. His critics question the competence of physics ever having anything of importance to say about mental phenomena in general, and consciousness in particular (http://psyche.cs.monash.edu.au/psyche-index-v2 .html#som). The grounds for this criticism are different, ranging from computational to neurobiological arguments (Koch and Hepp 2006). Nevertheless , some neurobiologists find quantum states of microtubules, tiny elongated organelles spanning the interior of neurons, to be the gate to consciousness (Hameroff and Penrose 1996). There is indeed a puzzling temporal aspect to the consciousness. Famous experiments of Benjamin Libet on human subjects showed that it takes a while, about 0.5s, for a conscious awareness to develop (Libet 1985). Acts initiated by a subject's free will are preceded by a specific electrical change in the brain, the so-called readiness potential (RP), which begins 550 ms before the act itself. Human subjects became aware of their own intention to act 350-400 ms after RP started, but 200 ms before the act execution (which is about the time of the process of initiation and generation of motor movements) . These data pose many questions including philosophical about the nature of free will, whether it exists at all (Libet 1999). While the author remains optimistic about this issue, the question still remains why it takes so long for a conscious awareness to develop. Why it develops at all? Which processes are responsible for it and how they lead to consciousness? Even concerning the mystery of mentality, there is not a general agreement among scientists. Some think that there is no mystery at all, and consciousness and other mental phenomena are equivalent to a particular underlying process, be it the specific computations (Dennett 1991), brainspecific processes (Searle 2002), or complexity (Edelman and Tononi 2000). In another influential book written on consciousness, philosopher
50
2 Organization and Functions of the Brain
and mathematician, David J. Chalmers clearly argues that consciousness and mentality are indeed genuinely puzzling and not explainable by present theories (Chalmers 1996). Chalmers asks: why should quantum processes (or any other specific physical processes) in microtubules (or any other brain substructures) or any specific computational processes give rise to consciousness? If one takes consciousness seriously, Chalmers says, one has to go beyond a strict materialist framework. The fundamental laws linking the physical and the experiential are yet to be discovered. The exercise for the reader can be to think about the nature of such fundamental laws. Although a lot is known about the brain, issues about its functioning, representation and processing of information are still subjects of an intense research. The nature of brain dynamics is still unknown. Some researchers find evidence of chaos, whereas some are doubtful (Theiler 1995). Main proponents of a chaotic dynamics, W.J. Freeman (Freeman 2003) and I. Tsuda (Tsuda 2001), argue in favor of chaotic itinerancy based on EEG and other neurophysiological data. According to the picture of chaotic itinerancy, a complex system such as the (human) brain evolves by steps along a trajectory in the state space. Each step corresponds to a shift from one basin of attraction to another. Attractors represent classes for abstraction and generalization. Thus, the brain states evolve aperiodically through sequences of attractors. In a closed system the next attractor would be chosen solely by internal dynamics. In an open system, such as the brain, external inputs interfere with internal dynamics. Moreover, due to the changes induced by learning, trajectories continually change. Chaotic itinerancy occurs in sequence of cortical states marked by state transitions that appear in temporal discontinuities in neural activity patterns (Freeman 2003). Experimental EEG data show that the entire cerebral cortex is constantly wandering in the fractal distributions of phase transitions that give the 1/j form of the temporal and spatial frequency spectra (with a E (1,3), (Freeman 2003)). From this type of frequency spectra it appears that the brain maintains a state of self-organized criticality (Bak et al. 1987). The self-organized criticality state can form the basis of the brain capacity to rapidly adjust to new external and internal stimuli. State changes resembling phase transitions occur continually everywhere in cortex at scales ranging from millimeters to ~O.l m. Local neural activity can trigger a massive state change. However, several issues of caution should be pointed out. In spite the compelling evidence for self-organized criticality in the brain, the nature of the critical state is still unknown in neurobiological interpretation. The spatial and temporal power spectral densities often show the l/j form, however more often this form is broken down due
2.8 Summary and Discussion
51
to the distortions by clinically defined peaks. Therefore the measurements of a vary widely. Aperiodic oscillations giving the llj power spectral densities are commonly referred to as chaotic. However the brain activity is not at all consistent with low-dimensional deterministic chaos (Theiler 1995, Benuskova, Kanich et al. 2001). It is high dimensional, noisy, nonGaussian, and nonstationary (Freeman 2003). Tremendous physical complexity of the brain arises also from the fact that it is not a homogenous tissue. Each part of the brain is morphologically different and has its own genetic profile as can be seen by analysis of large-scale human and mouse transcriptomes. Therefore the conditions for assessment of the type of dynamics are difficult to be met. Moreover brains are open systems driven by stochastic input. Thus it seems that the brain activity hardly can conform to the mathematical definitions of chaos. Whether the term chaotic itinerancy (or any other term from the chaotic vocabulary) is appropriate to describe state transitions in brain and cortex in particular remains open to challenge. Thus, the complex spatio-temporal activity data from the brain still awaits explanation.
3 Neuro-Information Processing in the Brain
While Chapter 2 presents the higher-level brain organization, this chapter presents a view on the low level information processing in the brain. Neuro-infonnation processing in the brain depends not only on the organization of the brain and properties of brain neural networks, but also on the properties of processing units - neurons and signal processing networks within neurons. These internal networks involved in signal processing are comprised of second and third messengers, enzymes, transcription factors and genes.
3.1 Generation and Transmission of Signals by Neurons A neuron (Fig. 3.1) receives and sends out electric and chemical signals. The place of signal transmission is a synapse. In the synapse, the signal can be nonlinearly strengthened or weakened . The efficacy of synaptic transmission is also called a synaptic weight or synaptic strength. One neuron receives and sends out signals through 103 to 105 synapses. Dendrites and soma, i.e. the body of a neuron, constitute the input surface for receiving signals from other neurons.
axon
basal dendrites dendritic tree
synapse
Fig. 3.1. Schematic illustration of a neuron and its parts. There is a synapse at every dendritic spine. Synapses are also formed on the dendritic shafts and on the soma
54
3 Neuro-Infonnation Processing in the Brain
Dendritic tree consists of thousands of dendrites which are covered by tiny extensions called spines. Most of synapses are formed on dendrites, particularly on spines. Spines are very important devices in relation to learning and memory, as we will see later. Electrical signals transmitted by synapses can have a positive or negative electric sign. In the former case, we speak about excitatory synapses and in the latter case about inhibitory synapses. When the sum of positive and negative contributions (signals) weighted by synaptic weights gets bigger than a particular value, called the excitatory threshold, a neuron fires, that is, emits an output signal called a spike. A spike is also called an action potential or a nerve impulse. Usually, as a result of synaptic stimulation and summation of positive and negative signals, a neuron fires a whole series (train) of spikes (Fig. 3.2). Mean frequencies of these spike trains range from 1 to102 Hz. The output frequency is proportional to the overall sum of positive and negative synaptic contributions. Spikes are produced at the initial segment of an axon, the only neuronal output extension. Then they propagate very quickly along the axon towards other neurons within a network. Propagation speed of nerve impulses is 5-100m/s. At its distant end, an axon makes thousands of branches, each of which is ended by a synaptic terminal (bouton). Spike train
11111 - -
Fig. 3.2. Generation and propagation of spikes in neurons. EPSP = excitatory postsynaptic potential, IPSP = inhibitory postsynaptic potential, LPSP = total postsynaptic potential equal to EPSP - IPSP, & = excitatory threshold
Transmission of signals from one neuron to another occurs in synapses. A synapse consists of a presynaptic terminal (bouton), synaptic cleft and postsynaptic membrane (Fig. 3.3). In the presynaptic terminal, there are dozens of vesicles filled with molecules of neurotransmitter (NT) ready to be released. When a presynaptic spike arrives into a terminal, calcium ions rush in and cause the fusion of vesicles with the presynaptic membrane. This process is also called exocytosis. Molecules of NT are released into the synaptic cleft (Fig. 3.3b), and diffuse towards the receptors within a postsynaptic membrane. Molecules of NT form a transient bond with the
3.1 Generation and Transmission of Signals by Neurons
55
molecules of receptors. This causes opening of ion channels associated with postsynaptic receptors. In the excitatory synapse, receptors are associated with sodium (Na+) ion channels, and a positive excitatory postsynaptic potential (EPSP) is generated. In the inhibitory synapse, receptors are associated with chlorine «(1) ion channels, and a negative inhibitory postsynaptic potential is generated. Alternatively, there can be ion channels for potassium (K+), which flows out and thereby lowers the interior potential. Eventually, NT releases its bond with receptors and diffuses back to the presynaptic membrane and out of the synaptic cleft. Special molecular transporters within a presynaptic membrane take molecules of NT back inside the terminal, where they are recycled into new vesicles. This process is called a reuptake of NT. The whole synaptic transmission lasts for about I millisecond. Such a synapse is called a chemical synapse, because the transmission of an electric signal is performed in a chemical way.
10 " m
pos tsynaptic membrane
Fig. 3.3. Scheme of synaptic transmission. (a) Synapse is ready to transmit a signal. NT = neurotransmitter, R = postsynaptic receptor. (b) Transmission of an electric signal in a chemical synapse upon arrival of an action potential into the terminal (plus signs around the terminal). AMPAR = AMPA-receptor-gated ion channel for sodium, NMDAR = NMDA-receptor-gated ion channel for sodium and calcium
The postsynaptic potential (PSP), either excitatory or inhibitory, has some amplitude and duration. The amplitude and duration of PSP depend upon the number of activated receptor-ion channels and upon the time for how long they stay open. This may last for milliseconds, tens of milliseconds or hundreds of milliseconds. The duration of channel opening depends upon the number of released NT molecules and upon the type of receptors that are associated with ion channels. The amplitude of PSP also
56
3 Neuro-Information Processing in the Brain
depends upon the electric input resistance for ions, which in tum depends upon the size and shape of a postsynaptic spine and dendrites, and upon the distance of synapse from soma. For instance, a short and stubby dendritic spine has a much smaller electric resistance than a long and thin spine. All these pre- and postsynaptic factors determine the weight (strength, efficacy) of a particular synapse. Within a postsynaptic membrane, there are also kinds of receptors that are not associated with an ion channel, but instead with an enzyme. When the overall amount of released NT reaches some critical concentration, these receptor-enzyme complexes activate particular cytoplasmatic enzymes, the so-called second messengers. Second messengers trigger chains of various biochemical reactions which may lead to a change in synaptic weight, or even to transient changes in gene expression leading to alteration in biomolecular synthesis of receptors, neurotransmitters and enzymes. Thus, second messengers may act locally within a synapse itself, or they may activate further (third and so on) messengers that carry the message to the genome of a neuron, thus causing a change in its biochemical machinery related to signal processing. Therefore, it is now widely accepted that the activity of a neuron itself, influences its processing of information, and even its life itself, whether it survives or not.
3.2 Learning Takes Place in Synapses: Toward the Smartness Gene For major discoveries in the field of synaptic mechanisms of learning, the 2000 Nobel Prize for medicine went to the neuroscientists Eric R. Kandel and Paul Greengard. The 3rd laureate, Arvid Carlsson, got his share of the prize for discoveries of actions of neurotransmitter dopamine. At present, it is widely accepted that learning is accompanied by changes of synaptic weights in cortical neural networks (Kandel et al. 2000). Changes of synaptic weights are also called synaptic plasticity. In 1949, the Canadian psychologist Donald Hebb formulated a universal rule for these changes: "When an axon of cell A excites cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells so that A's efficiency as one of the cells firing B is increased", which has been verified in many experiments and its mechanisms elucidated (Hebb 1949). In cerebral cortex and in hippocampus of humans and animals, learning takes place in excitatory synapses formed upon dendritic spines that use glutamate as their neurotransmitter. In the regime of learning, glutamate
3.2 Learning Takes Place in Synapses: Toward the Smartness Gene
57
acts on specific postsynaptic receptors, the so-called NMDA receptors (Nmethyl-D-aspartate). NMDA receptors are associated with ion channels for sodium and calcium (see Fig. 3.3). The influx of these ions into spines is proportional to the frequency of incoming presynaptic spikes. Calcium acts as a second messenger thus triggering a cascade of biochemical reactions which lead either to the long-term potentiation of synaptic weights (LTP) or to the long-term depression (weakening) of synaptic weights (LTD). In experimental animals, it has been recorded that these changes in synaptic weights can last for hours, days, even weeks and months, up to a year. Induction of such long-term synaptic changes involves transient changes in gene expression (Mayford and Kandel 1999, Abraham et al. 2002). A subcellular switch between LTD and LTP is the concentration of calcium within spines (Shouval, Bear et al. 2002). We speak about an LTD/LTP threshold. In tum, the intra-spine calcium concentration depends upon the intensity of synaptic stimulation that is upon the frequency of presynaptic spikes. That is, more presynaptic spikes mean more glutamate within synaptic cleft. Release of glutamate must coincide with a sufficient depolarization of the postsynaptic membrane to remove the magnesium block ofthe NMDA receptor. The greater the depolarization, the more ions of calcium enters the spine. Postsynaptic depolarization is primarily achieved via AMPA (amino-methylisoxasole-propionic acid) receptors, however, recently a significant role of backpropagating postsynaptic spikes has been pointed out (Markram et al. 1997). Calcium concentrations below or above the LTD/LTP threshold, switch on different enzymatic pathways that lead either to LTD or LTP, respectively. However, the current value of the LTD/LTP threshold (i.e. the properties of these two enzymatic pathways) can be influenced by levels of other neurotransmitters, an average previous activity of a neuron, and possibly other biochemical factors as well. This phenomenon is called metaplasticity, a plasticity of synaptic plasticity (Abraham and Bear 1996). Dependence of the LTD/LTP threshold upon different postsynaptic factors is the subject of the Bienenstock, Cooper and Munro (BCM) theory of synaptic plasticity (Bienenstock et al. 1982) (for a nice overview see for instance (Jedlicka 2002)). The BCM theory of synaptic plasticity has been successfully applied in computer simulations to explain experience-dependent changes in the normal and ultrastructurally altered brain cortex of experimental animals (Benuskova et al. 1994). The easiness with which LTD and LTP can be evoked in the developing and in the adult brain are not the same. One of the factors responsible for this difference may be the genetically programmed difference in the NMDA receptor composition (Bliss 1999). The NMDA receptor is made up of an NR1 subunit, which is obligatory for channel function, and a se-
58
3 Neuro-Information Processing in the Brain
lection of developmentally and regionally regulated NR2 subunits (A to D). For example, the glutamate-evoked positive current has a longer duration in receptors containing NR2B subunits than in those containing NR2A subunits. The proportion of NR2B subunits is higher in young animals than in adults, and this probably accounts for the greater degree of synaptic plasticity seen in young animals that is accompanied by greater easiness of youngsters in formation of new memories (Bliss 1999). This conclusion was indeed confirmed in another study, in which scientists inserted an extra copy of the NR2B gene into mice (Ping et al. 1999). Mice with the NR2B gene insertion performed better than mice without the insertion on all of the tests used to evaluate their intelligence and memory. Adding a single gene to mice significantly boosted the animals' ability to solve maze tasks, learn from objects and sounds in their environment and to retain that knowledge. This strain of mice also retained into adulthood certain brain features of juvenile mice, which, like young humans, are widely believed to be better than adults at grasping large amounts of new information. The research proves that the NR2B gene is a key switch that controls the brain's ability to associate one event with another, the core feature of learning. The finding also shows that genetic improvement of intelligence and memory in mammals is now feasible, thus offering a striking example of how genetic technology may affect mankind and society in future. This research can ultimately lead to human gene therapy for use in areas such as dementia, mental retardation, etc, although more research is needed on comparison of regulation and functions of human and mouse NMDA receptors.
3.3 The Role of Spines in Learning Dendrites of cortical excitatory pyramidal neurons are abundant in tiny membrane extensions called spines. They are named so because they resemble in shape the spines on the rose stem. About 80% of all synaptic connections in the cerebral cortex are excitatory and vast majority of them is formed on the heads of synaptic spines. For many years the role of spines was a mystery. Nowadays it is accepted that they play several important roles in synaptic plasticity and learning. First, it was discovered that spines change their size, shapes and numbers during the induction and maintenance of LTP (Lee et al. 1980, Geinisman et al. 1991). There are growth changes on spines, like spine head swelling, spine neck thickening, and increase in appearance of spines with mushroom-shaped heads. Morphological properties of spines and changes
3.3 The Role of Spinesin Learning
59
in their shape were first supposed to playa role in affecting the efficacy of synaptic transmission by means of changes in the input resistance (Koch and Poggio 1983). Long, thin spines create a big input electrical resistance, while short, stubby spines create a smaller input resistance. Later, a role in sequestering and amplifying the calcium concentrations was suggested to be the main role of spines (Zador et al. 1990). Through this role a mechanisms for saturation and stopping the infinite growth of synaptic weights was proposed, as well as the role in the LTP/LTD threshold (Gold and Bear 1994). While all these effects can take place, another important role for spines was suggested in the transport of new receptors into the spine head (Benuskova 2000). This model is based on our older hypothesis that the changes in efficacy of excitatory dendritic spine synapses can result from the fusion of transport vesicles carrying new membrane material with the postsynaptic membrane of spines (Fedor et al. 1982). Spacek and Harris indeed found structural evidence for exocytotic activity within spines in hippocampal CAl pyramidal neurons (Spacek and Harris 1997). Smooth vesicles of the diameter around 50 nm occurred in the cytoplasm of spine heads, adjacent to the spine plasma membrane, and fusing with the plasma membrane. In addition, (Lledo et al. 1998) showed that inhibitors of membrane fusion blocked or strongly reduced LTP when introduced into CAl pyramidal cells. On the other hand, an increase in synaptic strength was elicited when membrane fusion was facilitated. In the CA 1 region, LTP requires the activation of the NMDA glutamate receptors and a subsequent rise in postsynaptic calcium concentration. Besides other roles, Ca2+ plays a crucial role in the final stage of vesicle fusion with the membrane, and the number of fused vesicles is proportional to [Ca2+] (Sudhof 1995). Since LTP in CA 1 neurons is accompanied by appearance of AMPA subclass of glutamate receptors (Liao et al. 1995), it is reasonable to assume that vesicles can be a mean of their insertion. Indeed, Kharazia et al. observed GluRI (a subunit of AMPA receptors) containing vesicles associated with the cytoplasmic side of some GluRI-containing cortical synapses (Kharazia et al. 1996). Moreover, tetanic stimulation induces a rapid delivery of GluRI into spines and this delivery requires activation of NMDA receptors (Shi et al. 1999). Another effect of the vesicle fusion with the spine membrane would be the shaping and growth of the spine, which were observed during the induction and maintenance of LTP. However, prior to fusion the vesicles must get very close to the plasma membrane. The main mechanism for displacement of vesicles within axons and dendrites is the fast active transport with the speeds of 0.001-0.004 um/ms (Schnapp and Reese 1986). Fast transport depends on the direct interaction of transported vesicles with microtubules via the translocator kinesin-like molecules
60
3 Neuro-Information Processing in the Brain
(Schnapp and Reese 1986). However, microtubules do not enter spines (Spacek and Harris 1997). Thus, while the fast transport can bring vesicles close to the walls of dendritic shafts, another mechanism must come into play within spines themselves. The first natural candidate for this mechanism can be the diffusion of vesicles. However, we have shown that an electrophoretically driven, directed motion of negatively charged vesicles towards the spine head, evoked by the synapse stimulation itself can be ten times faster (Benuskova 2000). We have estimated the intensity of intra-spine electric fields triggered by stimulation of excitatory spine synapses. We have shown that this electric force can cause fast electrophoretic movement of negatively charged vesicles which bring new postsynaptic receptors and membrane for insertion during the induction of LTP. Due to the direction of an intra-spine electric field, movement of vesicles is electrophoretically directed along the longitudal spine axis towards the spine head. Spinnulae, small extrusions in the middle of the spine head might be the witnesses of such a directed vesicle fusion. The number of fused vesicles may be proportional not only to the increased calcium concentration within the spine head during the induction of LTP but also to the magnitude of electric force which drives vesicles towards the postsynaptic membrane. The thicker the spine gets, the smaller the electric field becomes, thus none or far less vesicles would get to the postsynaptic membrane in time to catch up with the increased [Ca2+J during NMDAR stimulation. This effect is congruent with the effect of spine dimensions on [Ca2+] (Gold and Bear 1994), and thus may also be a part of saturation mechanism in LTP. Electrophoretic hypothesis is illustrated in Fig. 3.4.
Fig. 3.4. Schematic illustration of the electrophoretic hypothesis linking the stimu-
lation of the spine excitatory synapses with the morphological changes and insertion of new receptors during LTP. Arrows point in the direction of the intra-spine electric field E. E is getting smalleras the spine gets thickerdue to new membrane insertion. Spinulae in the middle of the spine head would be the place where the vesicles fuse most often due to the direction of E. SER = smooth endoplasmic reticulum
3.4 Neocortical Plasticity
61
Morphological changes on spines, dendrites and growth of synapses probably constitute the mechanism of long-term and permanent memory storage (Rhawn 1996, Bailey et al. 2004, Maviel et al. 2004). There is a local machinery for protein synthesi s at spines and dendrites consisting of polyribosomes, tRNAs , microRNAs, initiation factors and mRNAs for glutamate receptors, structural proteins and kinases like CaMKII (Steward 1997). MicroRNAs are small, non-coding RNAs that control the translation of target messenger RNAs (mRNAs). It has been shown that both sensory experience and synaptic stimulation leads to translation of a-CaMKII (Wu et al. 1998), the kinase that is crucially involved in the induction of LTP. In addition it has been shown that a brain-specific microRNA, miR134, is localized at the synapto-dendritic compartment of rat hippocampal neurons and regulates the size of dendritic spines. This effect is mediated by miR-134 inhibition of the translation of an mRNA encoding the protein kinase , Limkl (Lim-do main- containing protein kinase) , that controls spine development (Schratt et al. 2006).
3.4 Neocortical Plasticity The ability of the brain to wire and rewire itself in response to changes in experience has become known as experience-dependent plasticity. The brain is able to remodel its connections in order to adjust the organism's respon se to changing conditions. Experience-dependent neocortical plasticity means that cortical neurons change their response characteristics due to altered stimulation of cortical inputs (i.e. experience). Amount of experience-dependent cortical plasticity, i.e. the degree of changes, is the biggest in early development and decays with age, however it never ceases totally. Developmental plastic ity refers to the cortical changes during the early stages of postnatal brain development. 3.4.1 Developmental Cortical Plasticity With respect to the cortical developmental plasticity, the most studied and best known neural system is the visual system of mammals. It is assumed that the visual system of primates has around twenty hierarchical processing levels (areas) , starting from the eye retina, going through the LGN in thalamus and primary visual cortex in the occipital cortex, to tens of the higher-order visual areas in the parietal and temporal cortices. Neurons at every hierarchical level of the visual system respond to different abstract elementary features of visual objects. These elementary features are for
62
3 Neuro-Information Processing in the Brain
example, edges of light intensity, angles between edges, distances between points (disparity), colors, movements, direction and/or speed of the movement, and so on. Higher-order visual neurons respond to various combinations of elementary features processed at lower levels. The most studied cortical visual area is the primary visual cortex (visual one, VI), which is the hierarchically lowest visual cortical area. The Nobel laureates, David Hubel and Torsten Wiesel, discovered that neurons in VI have different orientation selectivity and ocular dominances. Orientation selectivity means that different neurons respond to the optical stimulation by the bars (edges) of different angular orientations that are like this: \ I/ / - , and so on, over the 360 0 circle. Ocular dominance means the degree to which a neuron is dominated by the right or left eye. When a neuron is dominated by both eyes equally, we say it is binocular. Hubel, Wiesel, and many other scientists have studied whether the distribution (pattern) of ocular dominances and orientation selectivity within neurons of VI is inborn or acquired by experience. Currently, it is widely accepted that these patterns are in part inborn however for the development of the normal sharp binocular vision, normal visual stimulation is necessary. Immediately after eye opening, neurons in VI are not silent but instead each one responds to an optical stimulation by bars of several close orientations. It is said that neurons are broadly tuned. There is also some initial ocular dominance distribution. Degrees of ocular dominance are equally distributed. It can be said that nature (i.e. innate factors) provides some initial "scaffold", in the sense of a "framework outlining parts to be formed on it later". In time, and experiencing normal (natural) visual stimulation, neurons become sharply tuned and the distribution of ocular dominance becomes skewed towards binocular neurons. The development of sharply tuned binocular neurons is completed during the so-called critical period of postnatal development. In other words, the critical period is a limited time period after birth, only during which the development of normal response characteristics of neurons is possible. After eye opening, the critical period of cats lasts for about 2-3 months the critical period of monkeys lasts for about 6 months, and in humans up to about 5-6 years after birth. Developmental plasticity, i.e. the degree of evoked long-lasting changes in response characteristics of cortical neurons, decays with time during the critical period and ceases by its end. Abnormal visual experience during the critical period causes the development of irreversible abnormal response characteristics of cortical neurons. Such an abnormal visual experience can change even the initial response properties of neurons. Thus, a "scaffold" itself, i.e. the innate basis for normal response
3.4Neocortical Plasticity
63
characterist ics, can be disrupted by an inappropriate experience (at least in the visual cortex). Examples of abnormal visual experience during the critical period that lead to the development of permanent and irreversible changes in response characteristics of cortical neurons in VI (the same outcomes were observed for monkeys and cats) (Clothiaux et al. 1991): • Binocular deprivation (BD). When both eyes are close shut during the whole critical period, both eyes are completely deprived of visual inputs. Neurons in VI remain broadly orientation selective and the distribution of ocular dominances remains unchanged . At the perceptual level, poor orientation selectivity results in blurred vision, which can not be corrected by any means. • Monocular deprivation (MD). One of the eyes is close shut during the whole critical period. Neurons in VI become dominated only by the open eye. Neurons dominated by the open eye develop normal, i.e. sharp tuning to oriented bars, and thus gain normal orientation selectivity properties. These response characteristics are irreversible, even when the closed eye is opened at the end of the critical period and remains open for the rest of life. At the perceptual level, an animal stays blind on the closed eye for the whole life and has a sharp vision connected only to the open eye that received a normal visual input. • Normal rearing (NR) after MD. When after cortical neurons loose their responsivity to the closed eye, this eye is opened and the critical period is not over yet, neurons in V 1 regain binocular properties and develop normal orientation selectivity. A normal sharp vision can be restored. • Reverse suture (RS). One of the eyes is close shut for some time since the beginning of the critical period. After cortical neurons loose their responsivity to the closed eye, this eye is opened and the formerly opened eye is closed (before an end of the critical period). Cortical neurons in VI loose their responsivity to the newly closed eye, and regain and retain responsivity to the newly opened eye. Properties and time course of cortical plasticity in the auditory cortex is not very well known. Only recently it has been shown that the development of normal cortical representations of sounds correlates with the capacity to hear. Prof. Klinke 's team from the Goethe University in Frankfurt has advanced our knowledge about the developmental plasticity in the auditory cortex (Klinke et al. 1999). First they demonstrated that the cats that are deaf because of the innate (congenital) degeneration of the organ of Corti do not have normal representations of sound frequencies in their auditory cortex, although there is some weak representation of sounds be-
64
3 Neuro-Information Processing in the Brain
cause some sounds can be transmitted through bones in the skull. Scientists implanted the so-called cochlear implant into the inner ear of deaf kittens which were 2 to 4 months old. The cochlear implant is an electronic device that enables the transformation of sound air waves into electric stimulation of an auditory nerve (which is spared in these cats) and thus enables to relay sounds into the auditory cortex. They used a cochlear implant that was a crude substitute for a true cochlea, in the sense it was far from the faithful replication of natural stimulation. In spite of that, after a period of I to 3 weeks the kittens, completely deaf before they received the implant, started to hear. Perceptual improvement was accompanied by the development of almost normal sound maps (representations of sound frequencies) within auditory cortices of these cats. Thus, the cats have not started to hear immediately after they received a cochlear implant. Instead, they started to hear only after almost normal neural sound representations have developed in their auditory cortex. Degree of plastic changes in the auditory cortex in response to auditory stimulation decreases with age, although it seems that there is not such a sharp critical period like in the visual system. 3.4.2 Adult Cortical Plasticity
For the first time, the experience-dependent neocortical plasticity of the adult cerebral cortex was demonstrated in the somatosensory cortex of monkeys. At any moment, millions of signals from the whole body, its internal and external surface, travel to the brain through myriads of peripheral nerves and the spinal cord. Processing of bodily signals, like touch, warmth, pain, takes place in the somatic sensory cortex, which occupies the postcentral gyrus of parietal cortex (see Fig. 3.5). Human (or monkey or any other animal) body is mapped upon the somatosensory cortex in a topographic order. Topography means that neighboring body parts occupy neighboring parts in the map. A body map is deformed, i.e. those body parts that are often actively used are represented by bigger neural areas. Jon Kaas from Vanderbilt University and Michael Merzenich from University of California were among the first to discover that none of the two somatosensory maps are the same. They have discovered that individual variations in body representations reflect variances in tactile experience of different individuals especially with respect to fingers (Kandel et al. 2000). Details of cortical maps represent individual experience. Many experiments support their conclusions. For instance, in one experiment, monkeys were trained to touch a rotating disk for one hour per day. Before the experiment and several weeks later, scientists
3.4 Neocortical Plasticity
65
measured an extent of finger representation in the somatosensory cortex. They discovered that the cortical area representing stimulated finger tips has enlarged. Each finger has its own discrete area of the hand map. Neurons in area 1 respond only when touching finger 1, and so on. After differential tactile stimulation of the tips of fingers 2, 3 and 4, their cortical area has enlarged. In another experiment researchers sutured the skin of the fingers 3 and 4. Thus, a monkey had to use these two fingers as a single one. In the somatosensory hand map, an entirely new representation has developed. On the border between finger 3 and 4 exclusive representations, a band of neurons responding only to the joint stimulation of these two fingers has emerged.
Fig. 3.5. Schematic illustration of the bodily representation in the somatosensory cortex. Body surface is represented topologically while more important parts like face, lips, hands and especially fingers take more cortical representation than other parts like back, legs, arms, etc.
Similar somatosensory plasticity has been revealed in humans by means of magnetoencephalography (Mogilner et al. 1993). Adult humans were studied before and after surgical separation of inborn webbed fingers (syndactyly). The presurgical map contained nonsomatotopic hand representation. Within weeks after surgery, cortical reorganization over distances of 3-9 mm was evident, correlating with the new functional status of their separated digits. On the other hand, 95-100 % of people who have lost their arm or leg experience the phantom limb (Melzack 1999). They feel their lost limb as if it was still there, they feel it moving when they move, and it is still part of their subjective body experience. They can feel pain in it, touches, warmth, cold, and so on. At the beginning, the phantom limb
66
3 Neuro-Information Processing in the Brain
has a normal shape and size, but in time it starts to change: it may float freely in the air, contract into to the stump, or acquire any other bizarre shape, size or connection with the rest of the body. Nevertheless, it always remains an integral part of the patient's self as if the reorganization of large somatotopic maps was not entirely possible. 3.4.3 Insights into Cortical Plasticity via a Computational Model As we have seen although some forms of experience-dependent cortical plasticity disappear at the end of a developmental critical period, the adult cortex retains a significant capacity to undergo functional changes in response to alterations in sensory input. We are interested in the rules that determine how the adult cortex changes its synaptic circuitry to adapt to changes in the pattern of afferent activity. Very well structured and suitable system for such studies is the rat whisker system. Whiskers can be thought of as tiny fingers, with which the animal palpates objects all around. The facial whiskers of rats are aligned in five rows (row A is dorsal and row E is ventral) and the whiskers within a row are numbered from caudal to rostral, like positions on a chessboard (see Fig. 3.6a). Each facial whisker projects via the trigeminal nuclei and "barreloids" in the ventral posterior medial nucleus (VPM) in the thalamus into a separate cluster of neurons in layer IV, called barrels (Jensen and Killackey 1987). Clusters of neurons in layer IV remind the barrels of neurons hence their name. Barrels form the morphological basis of a discrete one whisker-one column organization of the barrel cortex (Fig. 3.6b).
a
b
Fig. 3.6. (a) In the experiment to gain insights into an adult experience-evoked cortical plasticity, all whiskers on one side of the rat's face but two (D2 and D3) were cut close to the fur. Dots mark the positions of cut whiskers aligned in five rows, A to E; (b) Illustration of cortical one whisker-one barrel mapping and the positionof the recording electrode. Clusters of neurons in layer IV remindthe barrels of neurons hencetheir name
3.4 Neocortical Plasticity
67
To alter the pattern of sensory activity, all whiskers except two, D2 and one neighbor in the D row, were cut close to the fur on one side of the face. We can think about this procedure as a removal of all fingers but two (however neither pain nor nerve damage is involved in whisker clipping). After 3, 7-10 and 30 days of "whisker pairing" (rats could use only 2 whiskers for palpation, all other whiskers were re-clipped regularly), the activity of single neurons in barrel D2 was measured in response to controlled deflection of the two paired whiskers, D2 and "D-paired", and the three cut neighbors ("D-cut", C2, and E2). Prof. Ebner and his team recorded and documented progressive and complex changes in D2 barrel cell responses during the course of paired whisker experience (Diamond et aI. 1993, Armstrong-James et aI. 1994). The physiological study outlined above motivated us to develop a computational model of a barrel D2 neuron to gain a deeper insight into which synapses in the cortex modify, how they modify and why they modify (Benuskova et aI. 1994). The weights of the synaptic inputs to the modeled barrel neuron (see Fig. 3.7a) were modifiable according to the Bienenstock, Cooper and Munro (BCM) theory (Bienenstock et aI. 1982, Cooper et aI. 2004). The BCM theory postulates that the neuron possesses a synaptic modification threshold, ~ that changes as a nonlinear function of the time-averaged postsynaptic activity (see Fig. 3.7b). b
a
4»
0: potent iation c
I I
I
,
I
postsynapti c activity c
I I I
I I
TT r
wh iskers C20 1 0 2
n 0 3 E2
c 4> < 0: depression
Fig. 3.7. (a) Schematic illustration of the barrel cortex circuit used in our model. In our simplified circuitry, whiskers C2, DI, D2, D3, and £2 converge through polysynaptic pathways (broken lines) upon a cell in the VPM barreloid D2, and through separate polysynaptic intracortical pathways upon a model cell in the barrel D2. Synaptic weights (m) that were modifiable in our model are denoted by open triangles. (b) Schematic illustration of the BCM synaptic modification rule together with how the same level of postsynaptic activity c' can result in synapse potentiation or depression depending on the current value of Brvt. According to the BCM synaptic modification rule, active excitatory inputs d, > 0 are strengthened when postsynaptic activity c > Brvt and ¢ > O. Active inputs weaken when c < Brvt and ¢< 0
68
3 Neuro-Information Processing in the Brain
Current position of 8M dictates whether a neuron's activity at any given instant will lead to strengthening or weakening of the synapses impinging on it. Whisker pairing was simulated by setting input activities of the model barrel cell to the noise level, except two inputs that represented untrimmed whiskers. Initially low levels of cell activity, resulting from whisker trimming, led to low values of~. As certain synaptic weights potentiated, due to the activity of the paired whiskers, the values of ~ increased and after some time their mean reached an asymptotic value. This saturation of ~ led to the depression of some inputs that were originally potentiated. The changes in cell response generated by the model replicated those observed in in vivo experiments (Benuskova et al. 1994). Previously, the BCM theory has explained salient features of developmental experience-dependent plasticity in kitten visual cortex (Bienenstock et al. 1982, Clothiaux et al. 1991). Our results suggested that the idea of a dynamic synaptic modification threshold, ~, is general enough to explain plasticity in different species, in different sensory systems and at different stages of brain maturity. Another test of BCM theory of synaptic plasticity was its application upon experience-evoked plasticity in the developmentally altered neocortex (Benuskova, Rema et al. 2001). If some synaptic plasticity rule captures real relations in the biological system then it should work even if this system is altered, provided corresponding parameters in the synaptic plasticity equations are changed accordingly. Numerous experimental results indicate that prenatal ethanol exposure impairs the development of synaptic plasticity mechanisms in the rat neocortex: namely, there is a decrease in the number of NMDA receptor-gated ion channels and the development of long, ineffectual dendritic spines (Miller and Dow-Edwards 1988, AIRabiai and Miller 1989, Rema and Ebner 1999). After the offsprings were born, they were fed normally, that is without ethanol, until they had reached adulthood. Then they were tested for ability of their cerebral cortex to undergo plastic changes in response to a new sensory experience. In these experiments again whisker D2 and D3 neighbor, "D-paired", were left intact while all other whiskers were cut. A novel sensory experience had caused plastic neural changes in both, normal and impaired cortices (i.e., cortices that were exposed to ethanol in utero), but these changes differed markedly from each other (Rema and Ebner 1999). As we will show below, the BCM theory can explain both the normal and impaired neocortical plasticity. The original BCM theory is a one cell theory that is one cell in this theory represents a population of cells with the defined properties, and inputs are expressed as instant frequencies of spikes. For instance here, one
3.4 Neocortical Plasticity
69
model cell represents the whole population of excitatory cells in the barrel D2 (Fig. 3.7a). Similarly, the inputs and their weights represent rather particular relay pathways with many synaptic contacts than a one individual synapse. Thus, a change in synaptic weight on the model neuron may represent any kind of plastic changes - i.e. biochemical, and/or morphological. Synaptic plasticity is modeled according to the BCM synaptic modification rule (Cooper 1987). If we consider the case of a linear cell, the modification of the /h synapse with the weight m, at time t is proportional to the product of input activity at the /h synapse, d(t), and a function ¢, in such a way that dm(t) [ lJ = TJ¢ e(t), eM (t)}I, (t) dt
- '-
(3.1)
The "modification rate", 17, is equal to the magnitude of the synaptic modification for the i" input in one time step, when ¢ = 1 and di(t)= 1. Modification function ¢ is a parabolic function of the cell's current firing rate e(t) and modification threshold ~(t), i.e. ¢[c(t),eM(t)] = e(t)[c(t) - eM (t)]
(3.2)
The dynamic synaptic modification threshold ~ is proportional to the squared postsynaptic response averaged over some recent past time t, such that (3.3)
The positive constant a determines how far to the right on the x axis we can place the actual or effective threshold for synaptic potentiation in the equation for the modification function ¢. It will tum out that the value of the constant a is the key to the differences in simulated experience-evoked cortical plasticity between normal and faulty cortex. Cell's current firing rate, e, for simulation of the model cell is calculated as a linear sum of the thalamocortical (VPM) input and intracortical input activities, such that
e(t) = m
vpm (t) Ld?m (t) + Lm~or (t)dt or (t)
(3.4)
where m vpm is the synaptic weight of the thalamocortical input from VPM upon the cell in the barrel-column and mi" is the weight of the lateral intracortical connection from neighboring barrel-column correspond-
70
3 Neuro-Infonnation Processing in the Brain
ing to the whisker i = D2, DI, D3, C2, E2 (see also Fig. 3.7a). The thalamocortical or intracortical input activities d?m(/) ot d," (I), respectively, relayed from the lh whisker, i = D2, DI, D3, C2, E2, are comprised of the sum of the evoked response plus random noise: (3.5) (3.6) In these equations, d(t) is equal to either 1 or 0, depending on whether or pm not the lh whisker is deflected. Parameter 0 < Ir < 1 is the input strength constant of the lh whisker input relayed through VPM to the model barrelcolumn D2. Parameter 0 < I;or < 1 is the intracortical input strength constant of the lh whisker input relayed through its own barrel-column to the model barrel-column D2. Qualitative and quantitative agreement with exm
cor »Irpmfor i = DI,
perimental data was obtained when Ib2 »I?m and I i m,
D3, C2, E2. For D2, and Ie;; '" Ib2 in accordance with experimental data (Armstrong-James et al. 1991). The thalamocortical noise n?m(/) and intracortical noise n~or (I) are defined as random variables which are uniformly distributed in the interval [-Ai(noise),+AiCnoise)], where Ai < ~ is the noise amplitude. They represent the random spontaneous neural activity. Whisker pairing was simulated by setting input activities to the noise levels, except for the two inputs that represented untrimmed whiskers. To match the simulated cell response evolution in time with the experimental testing procedure, the cell's response, c, of the model barrel-column cell was calculated as a sum of the short-latency « lOms) and long-latency (10-100 ms) responses to 50 deflections of the corresponding whisker (Armstrong-James and Callahan 1991, Armstrong-James et al. 1991). The short-latency response, CSL, is the number of spikes generated in response to activation of a multi-whisker thalamocortical input (see Fig. 3.7a). The long-latency response, CLL, is the number of spikes generated in response to activation of intracortical inputs. The total cell response reads (3.7)
At testing intervals, the weight values were fixed and
cor
dr = d;or = 1 for
deflected whisker, otherwise d?m=di = O. Since the response to intracor-
3.4 Neocortical Plasticity
71
tical inputs was measured over the time interval 9 times longer that the short-latency response, we multiply the second term by this number. For the values of simulation parameters and their ranges, as well as other details of the model, we refer the reader to (Benuskova, Rema et aI. 2001). In the following two figures, Fig. 3.8 and 3.9, we show the results of computer simulations of the above described model of neocortical experiencedependent plasticity. a
60
'3 60 E
:l ....40 0 00
.. 30
"-2"',,~
'
~30
-- ,-
II Do
.20 ..."i5.10
b
35
E
i26
-~..... , ........ -.... _.... _, ...... a
2
, ~20 ,,
f ,--r----------. .....
"",
.
............. ~
-
2 f
8,16
.ll.
-
..."t10
i
-::6
"IlL
10
15
20
25
Da", of whl"ker palMoa
30
5
10
15
20
25
30
D8'" of whl"ker pa1rlna
Fig. 3.8. Evolution of the long-latency responses evoked in the real and simulated barrel-column D2 by deflection of paired whiskers, D2 and D-paired (D3). Longlatency responses (10-100 ms poststimulus) are mediated by intracortical pathways. Lines correspond to simulation results, and discrete points with S.E. bars correspond to experimental data averagedover the whole population of cells. Evolution of responses to (a) whisker D2 and (b) paired whisker D3, for the normal cortex (dashed lines and squares) and for the impaired cortex (solid lines and triangles)
Our model based on the BCM theory of synaptic plasticity, explains the plastic changes in the barrel D2 as a result of the modification of excitatory synapses between (i) neighboring cortical barrel-columns, and (ii) between VPM and the model cell representing the barrel D2. In the model of normal barrel cortex: (1) short-latency «10 ms) responses to all whiskers increase, (2) long-latency responses to the paired whiskers, D2 and its spared D row neighbor (D3) initially increase and then decrease. In the model of barrel cortex of adult rats which were prenatally exposed to ethanol (impaired cortex): (1) short-latency «10 ms) responses to all whiskers decrease, (2) long-latency responses to the paired whiskers, D2 and its spared D row neighbor (D3) initially decrease and then increase. Compared to the values of parameters of the model simulating normal cortex, in the model simulating faulty cortex, three factors made a difference: (1) Levels of evoked intracortical activity and levels of noise were much lower than in the model of normal cortex. This is in accordance with
72
3 Neuro-Information Processing in the Brain
experimental findings in (Miller and Dow-Edwards 1988, Rema and Ebner 1999). (2) The thalamocortical noise had to be increased compared to the model of normal plasticity, thus reflecting alterations of thalamocortical projections induced by prenatal ethanol (Granato et al. 1995). (3) The value of the constant a in temporal dependence of 8M upon the past activity of neuron (equation 2.4) must have been twice as large as the value of a in the model of normal cortex (Benuskova, Rema et al. 2001). In spite of that the 8M behaves differently from 8M in the model of a normal cortex and determines a different course of plasticity. a
2S
-= •
~
E20 :;
~
--------------.
",,--------- -
.
S'S
y
..c OJ
.
to
OJ
,
.5
.E
8.10
Errthr), ELSE Update: (a) the input, (b) the output, and (c) the temporal connection vectors (if such exist) of the rule node k = inda, as follow: (a)
(b)
(c)
Ds(EX, WI(k)) = EX - WI (k); WI(k) = WI(k) + lr, * (5.12) Ds(EX, WI(k)), where lr, is the learning rate for the first layer; Ds(EX, WI(k)) = EX - WI (k); WI(k) = WI(k) + lr, * (5.13 Ds(EX, WICk)), where tr, is the learning rate for the first layer; Wll, k) = W3(l, k) + lrs * AI(k) * AI(lit- 1J, here I is the (5.14 winning rule neuron at the previous time moment (t-l),
5.3 The Basic EFuNN Algorithm
115
and A I (l)(t-I) is its activation value kept in the short term memory. 13. Prune rule nodes} and their connections that satisfy the following fuzzy pruning rule to a pre-defined level: 14. IF (a rule node rj is OLD) AND (average activation A1airj) is LOW) AND (the density of the neighboring area of neurons is HIGH or MODERATE (i.e. there are other prototypical nodes that overlap with) in the input-output space; this condition apply only for some strategies of inserting rule nodes as explained in a sub-section below) THEN the probability of pruning node (rj) is HIGH. The above pruning rule is fuzzy and it requires that the fuzzy concepts of OLD, HIGH, etc., are defined in advance (as part of the EFuNN's chromosome). As a partial case, a fixed value can be used, e.g. a node is OLD if it has existed during the evolving of an EFuNN from more than 1000 examples. The use of a pruning strategy and the way the values for the pruning parameters are defined depends on the application tasks.
15. Aggregate rule nodes, if necessary, into a smaller number of nodes. 16. END of the while loop and the algorithm. 17. Repeat steps 2 to step 17 for a second presentation of the same input data. With good dynamic characteristics, the EFuNN model is a novel efficient model especially for on-line tasks. The EFuNN model has the following major strong points: • • • • •
Incremental Fast learning (possibly 'one pass') On-line adaptation 'Open'structure Allowing for time and space representation based on biological plausibility • Rule extraction and rule insertion
116
5 Evolving Connectionist Systems(ECOS)
5.4 DENFIS The Dynamic Evolving Neural-Fuzzy Inference Systems (DENFIS) is also based on the ECOS principle and motivated by EFuNNs. DENFIS has an approach similar to EFuNNs especially similar to EFuNNs' m-of-n mode. DENFIS is a kind of dynamic Takagi-Sugeno type fuzzy inference systems (Takagi and Sugeno 1985). An evolving clustering method (ECM) is used in DENFIS models to partition the input space for creating the fuzzy rules. DENFIS evolve through incremental, hybrid (supervised/unsupervised), learning and accommodate new input data, including new features, new classes, etc. through local element tuning. New fuzzy rules are created and updated during the operation of the system. At each time moment the output of DENFIS is calculated through a fuzzy inference system based on mmost activated fuzzy rules which are dynamically selected from the existing fuzzy rule set. As the knowledge, fuzzy rules can be inserted into DENFIS before, or during its learning process and, they can also be extracted during the learning process or after it. The fuzzy rules used in DENFIS are indicated as follows:
R,: if XI is F Il and X2 is F 12 and ... and Xp is F IP,
(5.15)
then y, = b» + blIxI + b12X2 +... + btpxp where "Xi is F,", 1= 1,2, ... m;j = 1,2, ... P, are M x P fuzzy propositions that form m antecedents for m fuzzy rules respectively; xj>j = 1,2, ... , P, are antecedent variables defined over universes of discourse ~,j = 1, 2, ... , P, and Fij, 1= 1,2, ... M;j = 1,2, ... , P are fuzzy sets defined by their fuzzy membership functions !!Fij: X, ~ [0, 1], I = 1, 2, M; j = 1, 2, ... , P. In the consequent parts of fuzzy rules, y" 1= 1,2, m, are the consequent variables defined by linear functions. In DENFIS, Fij are defined by the following Gaussian type membership function GaussianMF = a exp [-
(x _m)2]
(5.16)
20'2
When the model is given an input-output pair (Xi, di ) , it calculates the following output value:
Color Plate 1
...
-
-- - -
· 1-
II
.. .. .. .. ... .. ••
II
"
••
•
...
-
..
I .:J
.
~._-
.---- _....
~ c
~ ~
Fig. 1.6. 12 genes selected as top discriminating genes from the Central Nervous System (CNS) cancer data that discriminates two classes - survivals and not responding to treatment (Pomeroy et al. 2002). The NeuCom software system is used for the analysis (www.theneucom.com) and the method is called "Signal-toNoise ratio"
Fig. 4.5. SOM output derived from 60 samples and 12 input gene expression variables from the CNS data (Pomeroy et al. 2002) (See Fig. 1.6) - the top left map; in the top right map the class labels are mapped (class survival in the left blob, and class fatal - on right side); the bottom three maps show the contribution of gene G1, G3 and G4 respectively - none of them on its own can discriminate correctly the samples and a good discrimination is achieved through their interaction and pattern formation. The software system ViscoverySOMine was used for this experiment (http://www.somine.info/)
Color Plate 2
..
\ . - - . I _ .... \r..-........
.
-
Fig. 4.6. Selecting the top 12 genes from the case study CNS cancer data (Pomeroy et al. 2002) as it was shown in Fig.1.6, but here using the t-test method. The selected genes compare well with the selected in Fig.1.6 genes. (A proprietary software system SIFTWARE (www.peblnz.com) was used for the purpose)
.[rt
• -
_o-
.....
".
..e : ._
x
....
lW, --
...;..;;;.....-
--
Mol
-
-,
"JI rr.!L..J
.!J
,. ............ ,.. v
I
.!.L...J
~
-"-'
.
Fig. 4.7. T e iscnmmatrve power 0 t e se ecte genes genes mig. 4.6 is evaluated through Principal Component Analysis (PCA) method. It is seen that the first PC has a significant importance in keeping the samples still distant after the PCA transformation. (A proprietary software system SIFTWARE (www.peblnz.com) was used for the purpose of the analysis)
Color Plate 3
, ........- - JCiAI
'~ '.
•
II
..
:::J
J Fig. 4.8. MLP that has 12 inputs (12 gene expression variables from Fig. 4.6), 1 output (the class of survivals vs not responding to treatment) and 5 hidden nodes is trained on all 60 samples of the gene expression data from the CNS cancer case study data. The error decreases with the number of iterations applied (altogether 500). (The experiments are performed in a software environment NeuCom (www.theneucom.com) , _ ..=::.
Fig. 5.8. The rule nodes of an evolved ECOS model from data of a person A using 37 EEG channels as input variables, plotted in a 3D PCA space. The circles represent rule nodes allocated for class 1 (auditory stimulus), asterisks - class 2 (visual stimulus), squares - class 3 (AV- auditory and visual stimulus combined) and triangles - class 4 (no stimulus). It can be seen that rule nodes allocated to one stimulus are close in the space, which means that their input vectors are similar
Color Plate 4
_ • -
r-rr-rr-r-
Fig. 5.10. A leave-one-cross validation method is applied to validate an ECF ECOS model on the 60 CNS cancer samples (Pomeroy et al. 2002), where 60 models are created - each one on 59 samples, after one example is taken out, and then the model is validated to classify the taken out example. The average accuracy over all 60 examples is 82%, where 49 samples are classified accurately and 11 incorrectly. Class 1 is the non-responding group (21 samples) and class 2 is the group of survivals (39 samples)
-- -_
or-
-
r--r-
-
--
..,
Fig. 5.11. An ECOS classifier is evolved on the 12 CNS cancer genes from Fig. 4.6. Aggregated (across all clusters) general profiles for each of the two classes are shown. The profiles, that capture the interaction between genes, show that genes 1, 5 and 11 are differently expressed across samples of each class, gene 6 is highly expressed in both classes and the other genes - lowly. This suggests an interesting interaction between some genes that possibly define the outcome of cancer of the CNS. The analysis is performed with the use of a proprietary software system SIFTWARE (www.peblnz.com)
Color Plate 5
a
'0 11
•
a...._
.!l ...J
;: . ,-... - ....----"I
b
'0 11
•
-"
a...._
,;::--...-....
.!l ...J
•
'2
Fig. 5.12. As ECOS are local learning models based on clustering of data into clusters, it is possible to find the profiles of each cluster of the same class and see that the profiles are different that points to the heterogeneity of the gene expressions in CNS cancer samples (data from (Pomeroy et al. 2002)). (a) Class 1; (b) Class 2 (a proprietary software system SIFTWARE (www.peblnz.com))
Color Plate 6
) Siftwdre Genelic Algorithm ror Offline
GJ
rcr Oplimiwlion
e.-v_ e.-v_ r J
0...
Si'lgIo AI 100Sl Zvorll"dU""","""
::I f70
, . . - - - - E....... (; M.
100. - - - - - - , - - - - , . . - - --.,.---,--------,
~
.. MnF"ooId'.....
~
r. EpOOho......
~
'IDOl.)
r Splo...o,..o.a
_ _- - ,
F"lIIdcr... ..
~
!ll
J'----~.-/
U.. GAIorF_E_
10
25
15
G.ner.hoM F_E...actlonR....
r-:---:--- R.... - - - - , SOIls-. 3lS367 1100 '_R-.ng 0 ; 0 . 0
l...G.. ~
•
::I
mar-
II_S... S_nPqMotion
0
0
G_L.-glh
28
N"""O-OO N...... F......
~--
roR....._ _ (' R _
(' U.. nr-~
R---.g G___
1-
r
2 7
Alil+_."_IIWlnm.,,_
2
'---
l
6
B
10 12
1
~
Fig. 6.3. GA optimization of the parameters and the set of input variables (features) of an ECOS model for classification of CNS cancer samples into two classes - class of survivals and a class of non-responding to treatment (see (Pomeroy et al. 2002)). The best ECOS model, after 20 generations of populations of 20 individuals, has an accuracy of almost 94% when tested in a 3-fold cross validation procedure. The model has the following input variables: 1, 3, 4, 7, 9, 11, I2,(represented in lighter color) and variables 2, 5,6, 8 and 10 are not used (represented in a darker color). Optimal values of the ECF parameters (Rmax , Rmin , m of n, epochs) are shown in the figure
Color Plate 7
......
, \,frw.1Ir
Iff
'_
'5c
Fig. 6.4. The optimal parameter values and input gene variables are used to derive the final ECOS model that has 22 clusters (rule nodes). This figure shows aggregated profiles for the two classes while the individual cluster profiles for each class are shown in Fig. 6.5 and Fig. 6.6, respectively
Color Plate 8 ~
) ( 1.,,,,, 1
Gen. EJlp,..i4Oft CI..... Rul• • (Red•
hi~.
Grt.n. low)
" 12
•
G.nt NumbiN
.1.1 ~
Ir=--S-_-.-G..- I I
Fig. 6.6. Individual cluster profiles for class 2 (cancer survivors) obtained using 7 genes selected through GA optimization as shown in Fig. 6.3
b
Fig. 7.4. (b) Gene expression microarray contains in each cell the expression level of one gene in one sample, or - the ratio of its expression between two samples (e.g. normal and diseased). The level of gene expression in each pixel is encoded on the black and white scale with darker cells denoting lower gene expressions and lighter cells denoting higher gene expressions
Color Plate 9
'::==317.
~ j~~
~_~ J~ r: C111.! J .:JlVr _~I~J~
--~-.oo!:!j.:J
~~
.- ~ J-:'1IIr~ • ...... ~
hI
J.
J
r:
~
J.:J IV
~r
-t:;JJ~
J.:Jr1rr' J. r
-...
_~
J .:JIIIrr'
J.:PIt'
Fig. 7.5. ass pro 1 es 0 4 pes 0 cancer extracteCl rom a trameCl EFuNN on 399 inputs (gene expression values) and 14 outputs using data from (Ramaswamy et al. 2001). The profiles of each class can be modified through a threshold tuned for each individual class that defines the membership degree above which a gene should be either over-expressed (lighter sign) or under-expressed (darker sign) in all rules of this class in order for this gene to appear in the profile. The last profile is of the CNS cancer
...c.._-=-_.. . . .
e- .....
1Fig. 7.6. Among the CNS cancer group there are 3 clusters that have different gene expression profiles, as detected by an EFuNN ECOS trained system from Fig. 7.5. The highly expressed genes (lighter lines) in cluster 1,2 and 3 of CNS cancer data are different
Color Plate 10
1.5
0 .5 l::
o
"fi)
:fi ... - 0 .5
~
c..
>
.
0
L>.
0
L>.
L>. 0
L>. 0
0
0
0
0
L>.
L>.
L>.
0 0
.Xz
L>. L>.
0
0
• - a new data vector o - a sample from D L\ - a sample from M
D,
0 0
0
0
L>. 0
0 0
L>.
Dz
Fig. 5.6. In the centre of a transductive reasoning system is the new data vector (here illustrated with two of them - Xl and X2), surrounded by a fixed number of nearest data samples selected from the training data D and generated from an existing model M (Song and Kasabov 2006)
122
5 Evolving Connectionist Systems (ECOS)
5.5.1 Weighted Data Normalization In many neural network and fuzzy models and applications, raw (not normalized) data is used. This is appropriate when all the input variables are measured in the same units. Normalization, or standardization, is reasonable when the variables are in different units, or when the variance between them is substantial. However, a general normalization means that every variable is normalized in the same range, e.g. [0, I] with the assumption that they all have the same importance for the output of the system. For many practical problems, variables have different importance and make different contribution to the output(s). Therefore, it is necessary to find an optimal normalization and assign proper importance factors to the variables. Such a method can also be used for feature selection or for reducing the size of input vectors through keeping the most important ones This is especially applicable to a special class of neural networks or fuzzy models - the clustering based models (or also: distance-based; prototypebased) such as: RBF, ART, ECOS. In such systems, distance between neurons or fuzzy rule nodes and input vectors are usually measured in Euclidean distance, so that variables with a wider normalization range will have more influence on the learning process and vice versa. A method, called TWNFI (Transductive weighted neuro-fuzzy inference method) that incorporates the ideas of transductive neuro-fuzzy inference and the weighted data normalization is published in (Song and Kasabov 2006).
5.6 ECOS for Brain and Gene Data Modeling
5.6.1 ECOS for EEG Data Modeling, Classification and Signal Transition Rule Extraction In (Kasabov et al. 2006) a methodology for continuous adaptive learning and classification of human scalp electroencephalographic (EEG) data in response to multiple stimuli is introduced based on ECOS. The methodology is illustrated on a case study of human EEG data, recorded at resting-, auditory-, visual-, and mixed audio-visual stimulation conditions. It allows for incremental, continuous adaptation and for the discovery of brain signal transition rules. The method results in a good classification accuracy of EEG signals of a single individual, thus suggesting that ECOS could be successfully used in the future for the creation of intelligent per-
5.6 ECOS for Brain and Gene Data Modeling
123
sonalized human-computer interaction models, continuously adaptable over time, as well as for the adaptive learning and classification of other EEG data, representing different human conditions. The method could help understand better hidden signal transitions in the brain under certain stimuli when EEG measurement is used (see Fig. 5.7). €F'o'
0 e @ ~60eCDcweee
(!!)90ClQ0lge e 00000000
€>eOcue
eG..,
(")o8G>Ge
e> o6e e e e o G oOe
Fig. 5.7. Layout of the 64 EEG electrodes (extended IntemationallO-lO System)
Fig. 5.8 shows the rule nodes of an evolved ECOS model from data of a person A using 37 EEG channels as input variables, plotted in a 3D PCA space.
1
o. 00
0'
0. -
o. o.
0.02
,
O.
Fig. 5.8. The rule nodes of an evolved ECOS model from data of a person A using 37 EEG channels as input variables, plotted in a 3D PCA space. The circles represent rule nodes allocated for class 1 (auditory stimulus), asterisks - class 2 (visual stimulus), squares - class 3 (AV- auditory and visual stimulus combined) and triangles - class 4 (no stimulus). It can be seen that rule nodes allocated to one stimulus are close in the space, which means that their input vectors are similar. See Color Plate 3 The circles represent rule nodes allocated for class 1 (auditory stimulus), asterisks - class 2 (visual stimulus), squares - class 3 (A V- auditory and
124
5 Evolving Connectionist Systems (ECOS)
visual stimulus combined) and triangles - class 4 (no stimulus). It can be seen that rule nodes allocated to one stimulus are close in the space, which means that their input vectors are similar. The allocation of the above nodes (cluster centers) back to the EEG channels for each stimulus is shown in Fig. 5.9.
Fig. 5.9. The allocation of the cluster centers from the ECOS model in Fig. 5.7 back to the EEG channels for each of the stimuli of classes from 1 to 4 (i.e. A, Y, AY, No - from left to right,respectively)
5.6.2 ECOS for Gene Expression Profiling
ECOS can be used for building adaptive classification or prognostic systems and for extracting the rules (profiles) that characterize data in local clusters (Kasabov 2002a, Kasabov 2006). This is illustrated in Fig. 5.10 and Fig.5.11 on the 12 CNS genes from Fig. 4.6, where a classification system is evolved and the aggregated (across all clusters) general profiles for each of the two classes are shown. The profiles, that capture the interaction between genes, show that some genes are differently expressed across samples of each class. This points to an interesting interaction between genes that possibly defines cancer of the CNS, rather than a single gene only. Before the final classifier is evolved in Fig. 5.11, a leave-one-cross validation method is applied to validate the ECOS model on the 60 samples, where 60 models are created - each one on 59 samples, after one example is taken out, and then the model is validated to classify the taken out example. The average accuracy over all 60 examples is 82% as shown in Fig.5.10. 49 samples are classified accurately, out of 60. This accuracy is further improved in Chap. 6 when EC is used to optimize the feature/gene set and the parameters of the ECOS model.
5.6 ECOS for Brain and Gene Data Modeling
I ..
125
::J
.-s,--------
,
~
II
J,.
I I
J
I I
~
A
I I
"
,I
1 -------
Fig. 5.10. A leave-one-cross validation method is applied to validate an ECF ECOS model on the 60 CNS cancer samples (Pomeroy et al. 2002), where 60 models are created - each one on 59 samples, after one example is taken out, and then the model is validated to classify the taken out example. The average accuracy over all 60 examples is 82%, where 49 samples are classified accurately and 11 incorrectly. Class 1 is the non-responding group (21 samples) and class 2 is the group of survivals (39 samples). See Color Plate 4
"',..,......, ,..-
,0-
e-
0... 1
r,..-
, "
-:l110l:ll'l
t l.."
Fig. 5.11. An ECOS classifier is evolved on the 12 CNS cancer genes from Fig. 4.6. Aggregated (across all clusters) general profiles for each of the two classes are shown. The profiles, that capture the interaction between genes, show that genes 1, 5 and 11 are differently expressed across samples of each class, gene 6 is highly expressed in both classes and the other genes - lowly. This suggests an interesting interaction between some genes that possibly define the outcome of cancer of the CNS. The analysis is performed with the use of a proprietary software system SIFTWARE (www.peblnz.com). See Color Plate 4
126
5 Evolving Connectionist Systems (ECOS)
The profiles shown in Fig. 5.11 are integrated, global class profiles. As ECOS are localleaming models based on clustering of data into clusters, it is possible to find the profiles of each cluster of the same class. We can see that the profiles are different which points to the heterogeneity of the cancer CNS samples (see Fig. 5.12). a
b
. --
-'- -.J [
0- ....
1]
.,)
,...---
Fig. 5.12. As ECOS are local learning models based on clustering of data into clusters, it is possible to find the profiles of each cluster of the same class. Different profiles point to the heterogeneity of the gene expressions in CNS cancer samples (data from (Pomeroy et al. 2002)). (a) Class 1; (b) Class 2 (a proprietary software system SIFTWARE (www.peblnz.com)). See Color Plate 5
5.7 Summary This chapter gives a brief introduction to a class of ANN models, called ECOS. These techniques are illustrated for the analysis and profiling of both brain and gene expression data. Further development of the techniques is their use to combine genes and brain data, where each neuron (node) will have gene parameters that need to be adjusted for the optimal functioning of the neuron.
6 Evolutionary Computation for Model and Feature Optimization
This chapter introduces the main principles of evolutionary computation (EC) and presents a methodology for using it to optimize the parameters and the set of features (e.g. genes, brain signals) in a computational model. Evolutionary computation (EC) methods adopt principles from the evolution in Nature (Darwin 1859). EC methods are used in Chaps. 7 and 8 of the book to optimize gene interaction networks as part of a CNGM.
6.1 Lifelong Learning and Evolution in Biological Species: Nurture vs. Nature Through evolutionary processes (evolution) genes are slowly modified through many generations of populations of individuals and selection processes (e.g. natural selection). Evolutionary processes imply the development of generations of populations of individuals where crossover, mutation, selection of individuals, based on fitness (survival) criteria are applied in addition to the developmental (learning) processes of each individual. A biological system evolves its structure and functionality through both, lifelong learning of an individual, and evolution of populations of many such individuals, i.e. an individual is part of a population and is a result of evolution of many generations of populations, as well as a result of its own developmental, of its lifelong learning process. Same genes in the genotype of millions of individuals may be expressed differently in different individuals, and within an individual - in different cells of their body. The expression of these genes is a dynamic process depending not only on the types of the genes, but on the interaction between the genes, and the interaction of the individual with the environment (the Nurture versus Nature issue). Several principles are useful to take into account from evolutionary biology:
128 • • • •
6 Evolutionary Computation for Model and Feature Optimization
Evolution preserves or purges genes. Evolution is a non-random accumulation of random changes. New genes cause the creation of new proteins. Genes are passed on through evolution - generations of populations and selection processes (e.g. natural selection).
6.2 Principles of Evolutionary Computation Evolutionary computation (EC) is concerned with population-based search and optimization of individual systems through generations of populations (Goldberg 1989, Koza 1992, Holland 1998). EC has been applied so far to the optimization of different structures and processes, one of them being the connectionist structures and connectionist learning processes (Fogel et al. 1990, Yao 1993). Methods ofEC include in principal two stages: 1. Creating new population of individuals, and 2. Development of the individual systems, so that a system develops, evolves through interaction with the environment that is also based on the genetic material embodied in the system. The process of individual (internal) development has been in many EC methods ignored or neglected as insignificant from the point of view of the long process of generating hundreds generations, each of them containing hundreds and thousands of individuals.
6.3 Genetic Algorithms Genetic algorithms (GA) are EC models that have been used to solve complex combinatorial and organizational problems with many variants, by employing analogy with Nature's evolution. Genetic algorithms were introduced for the first time in the work of John Holland (Holland 1975). They were further developed by him and other researchers (Goldberg 1989, Koza 1992, Holland 1998). The most important terms used in a GA are analogous to the terms used to explain the evolution processes. They are: • Gene - a basic unit, which defines a certain characteristic (property) of an individual. • Chromosome - a string of genes; it is used to represent an individual, or a possible solution to a problem in the solution space.
6.3 GeneticAlgorithms
129
• Population - a collection of individuals. • Crossover (mating) operation - sub-strings of different individuals are taken and new strings (off-springs) are produced. • Mutation - random change of a gene in a chromosome. • Fitness (goodness) function - a criterion which evaluates how good each individual is. • Selection - a procedure of choosing a part of the population which will continue the process of searching for the best solution, while the other set of individuals "die". A simple genetic algorithm consists of steps shown in Fig. 6.1. The process over time has been 'stretched' in space.
Crossover
Crossover
Initial population Selection Crossover
.:fD U.. 6o\""'..... E..
6o\P.......... - ---,
nrnr-
G......-.. "-
Ip"G.....
'"' 3 fOli2
MlUbonRate·
Ip"G..... J;> ........
• ..:J e...s""-,,"-
f:; AIow~,epc:JOJcflCn roRri __
r -. . r u.. F
rnro
I
---_...-..-/
a:J
70
1-
60
-
50 0
3
6o\T_
UonoYer RlJle •
1RJ
10 15 GeneratIOns
So.1
Average
20
,..-_ _ Aed. _ _--,
Be1l: SCOII!
939361
TmoR..........
R.......... G......_
0 ; 0
I 100
0
0
I~
0
G........ LeC
W
- 1
-2 -2.5 L - - ' -_ _- ' - -_ _L . . . - _ - - - I ._ _- - ' -_ 4 10 2 6
Ti~e
_
- ' -_ _- ' - - _ - - - - ' ' - : - _
12
14
16
Fig. 7.8. A cluster of genes that are similarly expressed over time (17 hours). See Color Plate 10
In (Kasabov et al. 2004) a simple GRN model of 5 genes is derived from time course gene expression data of leukemia cell lines U937 treated with retinoic acid with two phenotype states - positive and negative. The
7.3 Modeling Gene/Protein Regulatory Networks (GPRN)
149
model, derived from time course data, can be used to predict future activity of genes as shown in Fig. 7.9. 1 ,5
• (?)'
1
~
• ...
33 827
•
21
.'>Y'
'-: ' 8 0 .5
"
. . ''8 o ' . '1!r _,
-0.5
~~. . ..
o
' rrC'd icli utls
O " -d '
10
20
40
30
50
60
70
Fig. 7.9. The time course data of the expression of 4 genes (#33,8,27,21) from the cell line used in (Kasabov et al. 2004). The first 4 time data points are used for training and the rest are the predicted by the model values of the genes in a future time
Another example of GRN extraction from data is presented in (Chan et al. 2006) where the human response to fibroblast serum data is used (Fig. 7.10) and a GRN is extracted from it (Fig. 7.11). lo!! l Ou-xp rc s viou )
2 ,----'------
-
-
T he Respo nse or Hu ma n Fi bro blas ts to Serum Data -.----
---'---
-
-.-----
-
-
-
-.---
-
-
-
-,---
-
-
--,
Fig. 7.10. The time course data of the expression of genes in the Human fibroblast response to serum data. See Color Plate 10
150
7 GenelProtein Interactions - Modeling Gene Regulatory Networks (GRN)
-0.3
Fig. 7.11. A GRN obtained with the use of the method from (Chan et al. 2006) on the data from Fig. 7.10
Despite of the variety of different methods used so far for modeling GRN and for systems biology in general, there is not a single method that will suit all requirements to model a complex biological system, especially to meet the requirements for adaptation, robustness and information integration.
7.4 Evolving Connectionist Systems (ECOS) for GRN Modeling
7.4.1 General Principles
Microarray data can be used to evolve an ECOS with inputs being the expression level of a certain number of selected genes (e.g.l 00) and the outputs being the expression level of the same genes at the next time moment as recorded in the data. After an ECOS is trained on time course gene expression data, rules are extracted from the ECOS and linked between each other in terms of time-arrows of their creation in the model, thus representing the GRN. The rule nodes in an ECOS capture clusters of input genes that are related to the output genes at next time moment. The extracted rules from an EFuNN model for example (see Chap. 5, Sect. 5.2) represent the relationship between the gene expression of a
7.4 Evolving Connectionist Systems (ECOS) for GRN Modeling
151
group of genes G(t) at a time moment t and the expression of the genes at the next time moment G(t+dt) , e.g.: IF gI 3(t) is High (0.87) and g23(t) is Low (0.9) (7.1) THEN g8 7 (t+dt) is High (0.6) and gI 03(t+dt) is Low
Through modifying a threshold for rule extraction one can extract stronger or weaker patterns of dynamic relationship. Adaptive training of an ECOS makes it possible for incremental learning of a GRN as well as adding new inputs/outputs (new genes) to the GRN. A set ofDENFIS models (see Chap. 5, Sect. 5.4) can be trained , one for each gene gi so that an input vector is the expression vector G(t) and the output is a single variable gi(t+dt). DENFIS allows for a dynamic partitioning of the input space. Takagi-Sugeno fuzzy rules, that represent the relationship between gene gi with the rest of the genes, are extracted from each DENFIS model, e.g.: IF gl is (0.63, 0.70, 0.76) and g2 is (0.71, 0.77, 0.84) and g3 is (0.71, 0.77, 0.84) and g4 is (0.59,0.66,0.72)
(7.2)
THEN g5 =1.84 - 1.26g1 -1.22g2 + 0.58g3 - 0.03g4 7.4.2 A Case Study on a Small GRN Modeling with the Use of ECOS
Here we used the same data of the U937 cell line treated with retinoic acid (Dimitrov et al. 2004) as shown in Fig. 7.9. The results are taken from (Kasabov and Dimitrov 2002). Retinoic acid and other reagents can induce differentiation of cancer cells leading to gradual loss of proliferation activity and in many cases death by apoptosis. Elucidation of the mechanisms of these processes may have important implications not only for our understanding of the fundamental mechanisms of cell differentiation but also for treatment of cancer. We studied differentiation of two subclones of the leukemic cell line U937 induced by retinoic acid. These subclones exhibited highly differential expression of a number of genes including c-Myc, Idl and Id2 that were correlated with their telomerase activity - the PLUS clones had about 100fold higher telomerase activity than the MINUS clones. It appears that the MINUS clones are in a more "differentiated" state. The two subclones were treated with retinoic acid and samples were taken before treatment (time 0) and then at 6 h, 1, 2, 4, 7 and 9 days for the plus clones and until day 2 for the minus clones because of their apoptotic death. The gene ex-
152
7 Gene/Protein Interactions - Modeling Gene Regulatory Networks (GRN)
pression in these samples was measured by Affymetrix gene chips that
contain probes for 12,600 genes. To specifically address the question oftelomerase regulation we selected a subset of those genes that were implicated in the telomerase regulation and used ECOS for their analysis. The task is to find the gene regulatory network G= {gl,g2,g3,grest-,grest+} of three genes gl=c-Myc, g2=Idl, g3=Id2 while taking into account the integrated influence of the rest of the changing genes over time denoted as grest- and grest+ representing respectively the integrated group of genes, expression level of which decreases over time (negative correlation with time) and the group of genes, expression of which increases over time (positive correlation with time). Groups of genes grest-, grest+ were formed for each experiment of PLUS and MINUS cell line, forming all together four group of genes. For each group of genes, the average gene expression level of all genes at each time moment was calculated to form a single aggregated variable grest. Two EFuNN models, one for the PLUS cell, and one - for the MINUS cell, were trained on 5 input vector data, the expression level of the genes G(t) at time moment t, and five output vectors - the expression level G(t+ 1) of the same genes recorded at the next time moment. Rules were extracted from the trained structure that describes the transition between the gene states in the problem space. The rules are given in as a transition graph on Fig. 7.12a and 7.12b.
a.--
--,
b
3
5
..... "'0
o I .. t! '04
p,.. .
........
. .0 '"I
I I
o C-myc
p~ 2
..... "'0
Olt 3
2
\ \
\ \ \
'0 1
1 C-myc
Fig. 7.12. (a) The genetic regulatory network extracted from a trained EFuNN on time course gene expression data of genes related to telomerase of the PLUS leukemic cell line U937. Each point represents a state of the 5 genes used in the model, the arrows representing (rules) transitions of the states. (b) The regulatory network of three time steps for the MINUS cell line represented in the 2D space of the expression level of the first two genes - c-Myc and Idl
7.5 Summary
153
Using the extracted rules, that form a gene regulatory network, one can simulate the development of the cell from initial state G(t=O), through time moments in the future, thus predicting a final state of the cell.
7.5 Summary This chapter gave some background information on gene and protein interactions in cells and neurons as GRN. These interactions were linked to phenotype processes, such as cell cancer development (the CNS cancer data), or a proliferation of a cell line (also leading to a cancerous cell). Each gene interacts with many other genes in the cell, inhibiting or promoting, directly or indirectly, the expression level of messenger RNAs and thus the amounts of corresponding proteins. Transcription factors are an important class of regulating proteins, which bind to promoters of other genes to control their expression . Thus, transcription factors and other proteins interact in a manner that is very important for determination of cell function. A major problem is to infer an accurate model for such interactions between important genes in the cell. To predict the models of gene regulatory networks it is important to identify the relevant genes. The abundant gene expression microarray data can be analyzed by clustering procedures to extract and model these regulatory networks. We have exemplified some methods of GRN discovery for a large number of genes from multiple time series of gene expression observations over irregular time intervals. One method integrates genetic algorithm (GA) to select a small number of genes and a Kalman filter to derive the GRN of these genes (Chan et al. 2006). GA is applied to search for smaller subset of genes that are probable in forming GRN using the model likelihood as an optimization objective . After GRNs of smaller number of genes are obtained, these GRNs may be integrated in order to create the GRN of a larger group of genes of interest. The method is designed to deal effectively with irregular and scarce data collected from a large number of variables (genes). GRNs are modeled as discrete-time approximations of firstorder differential equations and Kalman filter is applied to estimate the true gene trajectories from the irregular observations and to evaluate the likelihood of the GRN models. The next chapter links a GRN to a functioning (e.g. spiking) of a neuron and then - to the functioning of the whole ANN model, that can be compared with targeted behavior, e.g. using brain data, thus creating a more complex CNGM.
8 CNGM as Integration of GPRN, ANN and Evolving Processes
This chapter presents a methodology for CNGM that integrates gene regulatory networks with models of artificial neural networks to model different functions of neural system. Properties of all cell types, including neurons, are determined by proteins they contain (Lodish et al. 2000). In tum, the types and amounts of proteins are determined by differential transcription of different genes in response to internal and external signals. Eventually, the properties of neurons determine the structure and dynamics of the whole neural network they are part of. Interaction of genes in neurons affects the dynamics of the whole neural network model through neuronal parameters , which are no longer constant, but change as a function of gene expression. Through optimization of the gene interaction network, initial gene/protein expression values and neuronal parameters , particular target states of the neural network operation can be achieved , and meaningful relationships between genes, proteins and neural functions can be extracted . One particular instance where the time scale of gene expression matches and in fact determines the time scale of neural behavior is the circadian rhythm. A circadian rhythm is a roughly-24-hour cycle in the physiological processes of plants and animals. The circadian rhythm partly depends on external cues such as sunlight and temperature, but otherwise it is determined by periodic expression patterns of the so-called clock genes (Lee et al. 1998, Suri et al. 1999). Smolen et al. (Smolen et al. 2004) have developed a computational model to represent the regulation of core clock component genes in Drosophila (per, vri, Pdp-I, and Clk). To model the dynamics of gene expression, differential equations and first-order kinetics equations were employed for modeling the control of genes and their products. The model illustrates the ways in which negative and positive feedback loops within the gene regulatory network cooperate to generate oscillations of gene expression. The relative amplitudes and phases of simulated oscillations of gene expressions resemble empirical data in most of simulated situations. The model is based on transcriptional regulation of per, Clk (dclock), Pdp-I , and vri (vrille). The model postulates that histone acetylation kinetics make transcriptional activation a nonlinear function of
156
8 CNGMas Integration ofGPRN, ANN and Evolving Processes
[CLK]. Simulations suggest that two positive feedback loops involving Clk are not essential for oscillations, because oscillations of [PER] were preserved when Clk, vri, or Pdp-I expression was fixed. However, eliminating positive feedback by fixing vri expression altered the oscillation period. Eliminating the negative feedback loop, in which PER represses per expression, abolished oscillations. Simulations of per or Clk null mutations, of per overexpression, and of vri, Clk, or Pdp-I heterozygous null mutations altered model behavior in ways similar to experimental data. The model simulated a photic phase-response curve resembling experimental curves, and oscillations entrained to simulated light-dark cycles. Temperature compensation of oscillation period could be simulated if temperature elevation slowed PER nuclear entry or PER phosphorylation. The model of Smolen et al. (Smolen et al. 2004) shows that it is possible to develop detailed models of gene control of neural behavior provided enough experimental data is available to adjust the model. Models of particular gene networks need to be based on measured values of biochemical parameters, like the kinetics of activation or expression of relevant transcription factors. Use of parameter values that do not describe the in vivo situation can lead to erroneous predictions of genetic and neural dynamic behaviors (Smolen et al. 2000). In this chapter we will envisage CNGM for any brain function, namely by formulating: (1) how to model internal gene/protein dynamics, (2) how to link parameters of a neuron model to activities of genes/proteins, (3) which genes/proteins are to be included in the model, (4) how to optimize the CNGM parameters, (5) how to validate CNGM on real brain data, (6) how to discover new knowledge from CNGM, and finally (7) how to integrate CNGM with bioinformatics.
8.1 Modeling Genetic Control of Neural Development Majority of existing models on neural development are molecular and biochemical models that do not take into account the role and dynamics of genes (see e.g. (van Ooyen 2003)). Computational models were developed for early neural development, early dendritic and axonal morphogenesis, formation of dendritic branching patterns, axonal guidance and gradient detection by growth cones, activity-dependent neurite outgrowth, etc. Although these models can be taken one step further by linking proteins to genes, this step was actually performed only by Marnellos and Mjolsness (Mjolsness et al. 1991, Marnellos and Mjolsness 2003), Storjohann and Marcus (Storjohann and Marcus 2005) and (Thivierge and Marcus 2006).
8.1 Modeling Genetic Control of Neural Development
157
Mjolsness et al. (Mjolsness et al. 1991) and Mamellos and Mjolsness (Mamellos and Mjolsness 2003) have introduced a modeling framework for the study of development including neural development based upon genes and their interactions. Cells in the model are represented as overlapping cylinders in a 2-dimensional hexagonal lattice where the extent of overlap determines the strength of interaction between neighboring cells. Model cells express a small number of genes corresponding to genes that are involved in differentiation. Genes in broad terms can correspond to groups of related genes, for instance proneural genes or epithelial genes, etc. Abstracting from biochemical detail, genes interact as nodes of a recurrent network. They sum up activating and inhibitory inputs from other genes in the same cell at any given time t, the overall sum denoted as g:
e, = ITabp~(t)
(8.1)
b
where genes are indexed by a and b, Tab is the interaction between genes a and b within cell i, and pib(t) are gene product levels within that cell. The developmental model also includes interactions from neighboring cells such that ga(t)
= ITabp~(t) + IIZabPb (t) j*i
b
(8.2)
b
where Zab is the interaction between genes a and b in neighboring cells, and Pb(t) are gene product levels in the neighboring cell). Neighborhood of a cell consists of the six surrounding cells. Thus, genes in a cell interact as nodes in a fully recurrent network with connection weights depending on the kind of the interaction. Two kinds of interaction are allowed: an intracellular and an inter-cellular one. A gene a sums inputs from genes in the same cell and from the neighboring cells at time t. Level (concentration) pia(t) of the product of the gene a then changes according to (8.3)
where R; is the rate of production of gene a's product, Aa is the rate of decay of gene a product, and ha is the threshold of activation of gene a. Function o(x) E (0, 1) is a sigmoid function defined as
a(x)
= 0.5[1 +
x
~(1 + x 2 )
J
(8.4)
158
8 CNGM as Integration of GPRN, ANN and EvolvingProcesses
As authors of the developmental model state (Mamellos and Mjolsness 2003) , levels of gene products should be viewed as corresponding to gene product activities rather than actual concentrations and gene interactions should be viewed as corresponding more to genetic rather than specific biochemical (transcriptional, etc.) interactions. The gene network allows cell transformations in the model. For instance, cells may change their state (i.e., the levels of gene products or other state variables), change type, strength of interaction, can give birth to other cells , or die. These transformations are represented by a set of grammar rules, the L-grammar as in Lindenmayer systems. Rules are triggered according to the internal state of each cell (or other cells as well) and are of two kinds: discrete (leading to abrupt changes) and continuous (leading to smooth changes). A set of binary variables C keeps track of what rules are active in any particular cell at any given time, thus representing the influence of a meta-rule for the constraints as to what rules may be active in a cell at a time. Vector g for cell i is therefore described more accurately by the next equation, where if C;' = I, then the corresponding rule is active, if C;' = 0, then the rule is inactive:
gi = LC;T;P i + LC;LAijT~P j r
r
(8.5)
i
where T I ' is the interaction strength matrix for one-cell rule r, Pi is the state variable (gene product level) vector for cell i, T / is the interaction strength matrix for two-cell rule r . Variable r stands as a label for a particular rule, which can be, for instance, mitosis , cell death , interphase, and so on. Aij is a factor that modifies the influence of cell} on cell i. Models using the gene network framework can be formulated as optimization tasks that look for the model parameters so that the model optimally fits biological data or behaves in a certain desired manner. Optimization seeks the minimum of the objective (or error) function E(p), which depends on the state variable values. An example of the objective function can be the least-squares error function: E(p)
= L (P~MODEL (t) - P~DATA (t))
(8.6)
i.a.t
which is the squared difference between gene product levels in the model and those in the data, summed over all cells (i) , over all gene products (a) and over all times (t) for which data are available. The objective function in gene network models typically have a large number of variables and parameters, are highly nonlinear and cannot be solved analytically or readily optimized with deterministic methods. Therefore the more
8.1 Modeling Genetic Control of Neural Development
159
appropriate methods for optimization are stochastic optimization methods like simulated annealing (Cerny 1985) or evolutionary computation (Goldberg 1989). What is actually being optimized is the set of adjustable parameters of the gene regulatory network that is the gene interaction weights, activation thresholds, protein production and decay rates, etc. The gene network framework has been applied to modeling to the development of the Drosophila embryo at the blastoderm stage (Reinitz et al. 1995). This model included a well-characterized hierarchy of regulatory genes that control the early events of Drosophila embryogenesis by setting up their expression patterns along the embryo's length and dividing it into segments. The model yielded predictions and interpretations of experimental observations. Marnellos and Mjolsness applied this approach to modeling early neurogenesis in Drosophila and constructed models to study and make predictions about the dynamics of how neuroblasts and sensory organ precursor cells differentiate from proneural clusters (Marnellos and Mjolsness 2003). The gene interaction strengths were optimized in order to fit gene expression patterns described in experimental literature. The objective function was the least-squares one and optimization was done by means of simulated annealing. The Drosophila developmental model made predictions about how the interplay of factors such as proneural cluster shape and size, gene expression levels, and strength of cell-cell signaling determine the timing and position of neuroblasts and sensory organ precursor cells. The model also made predictions about the effect of various perturbations in gene product levels on cell differentiation. Optimization found optimal values for model parameters so that the system evolved from the initial state to the desired final one that matched experimental findings on gene expression data and developmental phenomena in Drosophila. This is a novel contribution of computational neurogenetic modeling where the optimization leads to optimal hidden parameter values, like interactions between genes that constitute the main prediction of the model. Construction of the hidden gene regulatory network enables predictions about consequences of gene mutations. Another example of a neurodevelopmental process that is dependent upon gene expression is formation of topographic maps in the brains of vertebrates. Topographic maps transmit visual, auditory, and somatosensory information from sensory organs to cortex and between the cortical hemispheres (Kaas 1997). Experimental evidence suggests that topographic organization is maintained also in sensory neural structures where learning occurs, in other words, tactile information is stored within the spatial structure of maps (Diamond et al. 2003). It is known that the topographic map formation depends on activity-independent (genetic) and activ-
160
8 CNGM as Integration of GPRN, ANN and Evolving Processes
ity-dependent processes (learning or activity-dependent synaptic plasticity) (Willshaw and Price 2003). To study the interplay between these processes a novel platform is under development called INTEGRATE (Thivierge and Marcus 2006). It is similar in nature to a novel computational programming system for integrated simulation of neural biochemistry, neurodevelopment and neural activity within a unifying framework of genetic control, called NeuroGene (Storjohann and Marcus 2005). NeuroGene is designed to simulate a wide range of neurodevelopmental processes, including gene regulation, protein expression, chemical signaling, neural activity and neuronal growth. Central is a computational model of genes, which allows protein concentrations, neural activity and cell morphology to affect, and be affected by, gene expression. Using this system, the authors have developed a novel model for the formation of topographic projections from retina to the midbrain, including activity-dependent developmental processes which underlie receptive field refinement and ocular dominance column formation. Neurons are controlled by the genes, which are evaluated in all cell components. Regulation of gene transcription and translation is simulated through the use of queries. During the evaluation of a gene within a given cell component , the gene queries the cell component, retrieving information about the biochemical , neural or morphological state of a cell component or its immediate environment. This information is used to determine the expression rate of the gene in that cell component , according to the gene's regulation section. It is the state of the individual cell component (not the cell as a whole) which determines the expression rate of the gene. Effects of the gene, including protein production, apply to the cell component, such as dendrites, postsynaptic sites and growth cones. The expression of a gene can thus be limited to certain cell component type. The properties of simulated proteins are defined as part of the corresponding gene definition. Genes' influence on cellular behavior, morphology and neural properties in nature is mediated through molecular interactions involving proteins and other molecules. In NeuroGene programming language, this relationship is modeled by actions of genes. The actions are only invoked when and where the gene is expressed (i.e., the expression rate is greater than zero), reflecting the causal relationship between gene expression and cellular changes. NeuroGene can thus represent genetic control over cellular biochemistry , morphology and neural activity. Gene expression within a particular cell component can depend on extracellular protein concentrations, concentration gradients and/or the average concentrations of membrane bound proteins bound to neighboring cell components. Neural activity can affect gene expression through queries. This can be used to model genes which are expressed in response to neural activity.
8.2 Abstract Computational Neurogenetic Model
161
A case study of modeling projection formation from retina to tectum involves genes encoding the properties and expression profiles of known proteins (ephrins and Eph receptors), genes encoding postulated proteins such as retinal and tectal cell markers, and genes causing morphological change, including growth cone formation (Storjohann and Marcus 2005). The authors also implemented the learning rule introduced by Elliott and Shadbolt (Elliott and Shadbolt 1999) to model the competition among presynaptic terminals for the postsynaptic protein. The learning rule is encoded entirely in simulated genes. NeuroGene simulations of activitydependent remodeling of synapses in topographic projections had two results in accordance with experimental data. First, retino-tectal arbors, which initially form connections to many tectal cells over a large area, become focused so that each retinal ganglion cell connects to only one or a few tectal cells. This improves the topographic ordering of the projection. Second, the tectum, which receives overlapping topographic projections from both eyes, becomes subdivided into domains (known as ocular dominance columns) which receive neural input exclusively from one or the other eye. In addition, NeuroGene successfully modeled the EphA knockin experiment in which the retinal EphA level was increased and the resulting retino-tectal projections were specifically disrupted (Brown, Yates et al. 2000). NeuroGene can be considered to be a neurogenetic model in spite it does not include interactions between genes. Genes obey the known expression profiles and these can be changed as a consequence of mutation, gene knockout or knockin, and thus the model can be used for predictions of some neurodevelopmental disorders of the visual tract in vertebrates.
8.2 Abstract Computational Neurogenetic Model This methodology was first introduced in (Kasabov and Benuskova 2004,2005). In general, we consider two sets of genes: a set G gen that relates to proteins of general cell functions and a set G spcc that codes specific neuronal information-processing proteins (e.g. receptors, ion channels, etc.). The two sets form together a set G ={G f , G2, .. .. Gn } that forms a gene regulatory network (GRN) interconnected through matrix of gene interaction weights W (see Fig. 8.1). Proteins that mediate general cellular or specific information-processing functions in neurons are usually complex molecules comprised of several subunits, each of them being coded by a separate gene (Burnashev and Rozov 2000). We assume that the expression level of each gene g/t+L1t) is a nonlinear function of expression levels
162
8 CNGMas Integration ofGPRN, ANN and Evolving Processes
of all the genes in G. Relationship can be expressed in a discrete form (Weaver et al. 1999, Wessels et al. 2001) , i.e.:
+
g/t M) = Wi'
+,,(~ w"g, (t)J
(8.7)
where: N G is the total number of genes in G, WjO ~ 0 is the basal level of expression of gene j and the gene interaction weight Wjk represents interaction weight between two genes j and k. The positive interaction, Wj k > 0, means that upregulation of gene k leads to the upregulation of gene j. The negative interaction, Wj k < 0, means that upregulation of gene k leads to the downregulation of gene j. We can work with normalized gene expression values in the interval git) E (0, I). Initial values of gene expressions can be small random values, i.e. giO) E (0,0.1). It is a common practice to derive the gene interaction matrix W= {wjd (see Fig. 8.1) based on all gene expression data being collected at the same time intervals /)"t (Kasabov et al. 2004). In a living cell, gene expression, i.e. the transcription of DNA to messenger RNA followed by translation to protein, occurs stochastically, as a consequence of the low copy number of DNA and mRNA molecules involved. It has been shown at a cell level that the protein production occurs in bursts, with the number of molecules per burst following an exponential distribution (Cai et al. 2006). However, in our approach, we take into account the average gene expression levels and average levels of proteins taken over the whole population of cells and over the whole relevant time period. We assume a linear relationship between protein levels and gene expression levels. The linear relationship in the next equation is based on findings that protein complexes, which have clearly defined interactions between their subunits, have highly correlated levels with mRNA expression levels (Jansen et al. 2002 , Greenbaum et al. 2003). Subunits of the same protein complex show significant co-expression, both in terms of similarities of absolute mRNA levels and expression profiles, e.g., subunits of a complex have correlated patterns of expression over a time course (Jansen et al. 2002). This implies that there should be a correlation between mRNA and protein concentration, as these subunits have to be available in stoichiometric amounts for the complexes to function (Greenbaum et al. 2003). Thus the protein level pit+Llt) reads Np j
p/I + /),1) = ZjO+ 2::>jkgk (I) k=!
(8.8)
8.2 AbstractComputational Neurogenetic Model
163
where: Npj is the number of protein j subunits, ZjO 2 0 is the basal concentration (level) of protein j and Zjk 2 0 is the coefficient of proportionality between subunit gene k and protein j (subunit k content). Time delay f...t corresponds to time interval when protein expression data are being gathered. Determining protein levels requires two stages of sample preparation. All proteins of interest are separated using 2-dimensional electrophoresis, followed by identification using mass spectrometry (MacBeath and Schreiber 2000). Thus in our current model the delays f...t represent the time points of gathering both gene and protein data.
Fig. 8.1. What are the coefficients of the gene interaction matrix W? Which genes and which gene interactions lead to a neural spiking activity with particular characteristics? This is the main question which we will ask in our research. For simplicity we illustrate only a small GRN. Solid (dashed) lines denote positive (negative) interactions between genes, respectively
Some protein levels are directly related to the values of neuronal parameters P, such that
PJCt) = PJCO) PJCt)
(8.9)
where: PJCO) is the initial value of the neuronal parameter at time t = 0, and PJCt) is a protein level at time t. In such a way the gene/protein dynamics is linked to the dynamics of artificial neural network (ANN). The CNGM model from Eq. 8.7 to Eq. 8.9 is a general one and can be integrated with any neural network model, depending on what kind of neural activity one wants to model. In the presented model we have made several simplifying assumptions: • Each neuron has the same GRN, i.e. the same genes and the same interaction gene matrix W.
164
8 CNGM as Integration of GPRN, ANN and Evolving Processes
• Each GRN starts from the same initial values of gene expressions. • There is no direct feedback from neuronal activity or any other external factors to gene expression levels or protein levels. This generic neurogenetic model can be run continuously over time in the following way: 1. Set initial expression values of the genes G, G(t = 0), in the neuron and the matrix W of the GRN, basal levels of all genes and proteins, and the initial values of neuronal parameters p(t = 0), if that is possible. 2. Run the GRN and calculate the next vector of expression levels of the gene set G(t+~t) using equation (8.7). 3. Calculate concentration levels of proteins that are related to the set of neuronal parameters using equation (8.8). 4. Calculate the values of neuronal parameters P from the gene state G using equation (8.9). 5. Update the activity of neural network based on new values of parameters (taking into account all external inputs to the neural network). 6. Go to step 2. The biggest challenge of our approach and the key to the predictions of CNGM is the construction of the GRN transition matrix W, which determines the dynamics of GRN and consequently the dynamics of the ANN. There are several ways how to obtain W: 1. Ideally, the values of gene interaction coefficients
are obtained from real measurements through reverse engineering performed on the microarray data (Kasabov and Dimitrov 2002, Kasabov et al. 2004). 2. The values of W elements are iteratively optimized from initial random values, for instance with the use of genetic algorithm (GA), to obtain the desired behavior of the ANN. The desired behavior of the ANN can simulate certain brain states like epilepsy, schizophrenic hypofrontality, learning, etc. This behavior would be used as a "fitness criterion" in the GA to stop the search process for an optimal interaction matrix W. 3. The matrix W is constructed heuristically based on some assumptions and insights into what result we want to obtain and why. For instance, we can use the theory of discrete dynamic systems to obtain a dynamic system with the fixed point attractor(s), limit cycle attractors or strange attractors (Katok and Hasselblat 1995). 4. The matrix W is constructed from databases and literature on geneprotein interaction. 5. The matrix W is constructed with the use of a mix of the above methods. wij
8.3 Continuous Model of Gene-Protein Dynamics
165
The above method 2 of obtaining coefficients of Wallows us to investigate and discover relationships between different GRNs and ANN states even in the case when gene expression data are not available. An optimization procedure to obtain this relationship can read: 1. Generate a population of CNGMs, each with randomly generated values of coefficients for the GRN matrix W, initial gene expression values g(O), and initial values of ANN parameters P(O); 2. For each set of parameters run the CNGM over a period of time T and record the activity of the neurons in the associated ANN; 3. Evaluate characteristics of the ANN behavior (e.g. connectivity, level of activity, spectral characteristics ofLFP, etc); 4. Compare the ANN behavior characteristics to the characteristics of the desired ANN state (e.g. normal wiring, hypoactivity, etc.); 5. Repeat steps (1) to (4) until a desired GRN and ANN model behavior is obtained. Keep the solution if it fulfils the criterion; 6. Analyze all the obtained optimal solutions of GRN and the ANN parameters for significant gene interaction patterns and parameter values that cause the target ANN model behavior. In the step 1, which is the generation of the population of CNGM, we can apply the principles of evolutionary computation (see e.g. Chap. 6 in this book) with the operations of crossover and mutations of parameter values. In such a way we can simulate the process of evolution that has led to the neural GRN with the gene interactions underlying the desired ANN behavior.
8.3 Continuous Model of Gene-Protein Dynamics Instead of the discrete gene-protein dynamics introduced in the previous section on abstract CNGM we can use the system of continuous equations. Let us formulate a set of general equations for the gene-protein dynamic system. As a first gross simplification, we will again assume that every neuron has the same gene/protein regulatory network (GPRN) - that is, interactions between genes and proteins are governed by the same rules in every neuron. This assumption is partly justified by the fact that gene and protein expression data are usually average data obtained from a pool of cells, rather than from individual cells. The following set of nonlinear delay differential equations (DDEs) was inspired by (Chen and Aihara 2002), who derived the general conditions of their local stability and bifurcation for some simplifying assumptions. Particular terms on the right-hand side
166
8 CNGM as Integrationof GPRN, ANN and Evolving Processes
of equations were inspired by the "rough" network models from (Wessels et al. 2001). Underlying GPRN is illustrated in Fig. 8.2.
8.... - ____ -,
From other nodes
P1
(t
-t p 1)
•••••••••••••
••••••••••
P J(t - 0
(9.11)
L'lw_ = A_ exp(-L'lt / r_)for L'lt < 0
where ~t = tpost - t pre is the time difference between the post- and presynaptic spikes. The novelty of our approach is that the amplitudes ofpositive and negative synaptic change, A+ and A_ , respectively, are not constant anymore, but instead they depend on the dynamic synaptic modification threshold Brvt in the following way (Benuskova and Abraham 2006): A (t)
= A+(O) BM(t)
+
A_ (t)
=
(9.12)
A_ (O)BM(t)
where A(O) are initial (constant) values and Brvt is the modification threshold. Thus, when Brvt increases, ~ increases and A+ decreases, respectively, and vice versa. If e(t) = l, when there is a postsynaptic spike, and e(t) = 0, otherwise, the rule for sliding modification threshold Brvt reads (9.13)
9.4 A Simple One Protein-One Neuronal Function CNGM
197
where ~t) depends on the slowly changing level of pCREB, the time average (e(t»,M depends on some fast activity-integration process, which 2 for instance involves the dynamics of available Ca +-sensitive CaMKII (Bear 1995, Benuskova, Rema et al. 2001), and a is the scaling constant. The time average of postsynaptic activity can be calculated as in (Benuskova, Rema et al. 2001), that is by numeric integration of the following integral:
(e(t») 'M
= _1 (fe(tl)exp( -(t - t') / T M) dt' T
M
(9.14)
-00
where TM can be on the order of minutes. Thus Btvt will have a fast component changing in the matter of minutes and a slow component ~t) that will change over hours as the level of pCREB does after NMDARs stimulation (Schulz et al. 1999, Wu et al. 2001, Leutgeb et al. 2005). PCREB induces gene expression together with a co-activator factor CBP (see e.g. Fig. 9.5). It has been shown that the CBP production reaches maximum within the first hour after NMDARs stimulation and remains highly elevated up to 24 hr afterwards ((Hong et al. 2004), supporting Table 1, item 633, group 2 genes). Thus actually the rate limiting factor for stimulationinduced genes is pCREB, which changes in a biphasic manner after NMDARs stimulation (Schulz et al. 1999). Since Btvt determines the easiness of LTP induction, function ~t) will be the inverse of the pCREB formation curve, i.e.:
1 rp(t)
(9.15)
= [pCREB(t)]
where [pCREB(t)] is the concentration of phosphorylated CREB in the postsynaptic neuron. Early CaMK-dependent CREB phosporylation occurs after any high-frequency stimulation and later, PKA-dependent phase of CREB phosporylation occurs when the presynaptic stimulation lasts longer than 1 min (Schulz et al. 1999, Leutgeb et al. 2005). Thus the duration of presynaptic HFS stimulation will provide a threshold for the switch between the first phase of CREB phosporylation and its second phase. In a more detailed biophysical model this switch should arise from the kinetics of postsynaptic enzymatic reactions. Thus our CNGM is more abstract and highly simplified, but therefore perhaps more suitable for simulating larger networks of artificial neurons. But first we would like to demonstrate its feasibility in the experimental study of one neuron in reproduction of the actual experimental results from (Schulz et al. 1999).
198
9 Application ofCNGM to Learning and Memory
9.5 Application to Modeling of L-LTP In this study, we will employ as a spiking neuron model, a simple spiking neuron model of Izhikevich (Izhikevich 2003). Let the variable v (mV) represents the membrane potential of the neuron and u represents a membrane recovery variable, which accounts for the activation of K+ ionic currents and inactivation of Na + ionic currents, and thus provides a negative feedback to v. The dynamics of these two variables is described by the following set of differential equations:
v=0.04v 2 +5v+140-u+I
(9.16)
if = a(bv-u)
(9.17)
Synaptic inputs are delivered via the variable I. After the spike reaches its apex (AP = 55 mY), the membrane voltage and the recovery variable are reset according to the equation if
v zAP,
then {
v~c
(9.18)
u e-iu v d
Values of dimensionless parameters a, b, C, d differ for different types of neurons, i.e. regularly spiking, fast spiking, bursting, etc. (Izhikevich 2003). We will assume that the total synaptic input I(t)
=Iwit )
(9.19)
where the sum runs over all active inputs and wit) is the value of synaptic weight of synapse j at time t. In order to reproduce experimental data from (Schulz et al. 1999) we construct a simple spiking model of a hippocampal dentate granule cell (GC), in which we ignore the effect of inhibitory neurons. For the schematic illustration of hippocampal formation see Fig. 9.6. Thus, a model GC has three excitatory inputs, two of them representing ipsilateral medial and ipsilateral lateral perforant paths, mpp and lpp, respectively, and one excitatory input from a contralateral entorhinal cortex (cEC) (Amaral and Witter 1989). Mpp and lpp are two separate input pathways coming from the ipsilateral entorhinal cortex (EC) and terminating on separate but adjacent distal dendritic zones of the hippocampal dentate granule cells (McNaughton et al. 1981). They together form an ipsilateral perforant pathway input (pp). Input from the contralateral entorhinal cortex (cEC) terminates on the proximal part of the granule cell dendritic tree (Amaral and Witter 1989).
9.5 Application to Modeling ofL-LTP
199
As a neuron model we employ the simple model of spiking neuron introduced by Izhikevich (Izhikevich 2003), with the parameters values corresponding to regularly spiking cell, i.e. a = 0.02, b = 0.2, c = -69 mY, d = 2, and the firing threshold equal to 24 mY (McNaughton et al. 1981). The model is simulated in real time with the time step of 1 ms. Total synaptic input corresponding to variable I reads:
I(t)
=
hmpp(t)wmpp(t) I mpp + h1pp(t) WIpp(t) I 1pp+ heEc(t) WeEc(t) I eEC
(9.20)
where wmpp (WIpp, WeEe) is the weight of the mpp (lpp, cEC) input, and I mpp ( I\pp , IeEe) is the intensity of electric stimulus delivered to mpp (lpp, cEC), respectively. The function hmpp(t), h1pp(t) and heEc(t) is equal to 1 or 0 when presynaptic spike occurs or is absent at a respective input at time t. In our simulations, the spontaneous, testing and training intensities are the same and equal to I mpp = IIpp = I eEC = 100. Actually, the interpretation of stimulus intensity in the model is the number of input fibers within a given pathway that are engaged by stimulation. Initial values of synaptic weights wmpp(O) = Wlpp(O) = WeEc(O) ;:::; 0.05, so when the three input pathways were stimulated simultaneously or in a close temporal succession, a postsynaptic spike followed.
\ Entorhinal
cortex
Ipp
Fig. 9.6. Schematic illustration of hippocampal pathways and GC inputs
To simulate synaptic plasticity, we employed the STDP rule expressed by Eqs. 9.10 - 9.11, with the sliding BCM threshold incorporated through the amplitudes of synaptic changes (Eqs. 9.12 - 9.15) with these parameters values: A+(O) = 0.02, A _(0) = 0.01, T+ = 20 ms, T _ = 100 ms, TM = 30 s, a= 3000. We simulate the experimental situation, in which to induce
200
9 Application of CNGM to Learning and Memory
LTP in dentate gyrus (round cells in Fig. 9.6), electrical stimulation was delivered to the perforant pathway, which is the mixture of lpp and mpp fibers. Nondecrementallong-lasting LTP was induced by stimulating perforant pathway with 20 trains of impulses. Each train consisted of 15 pulses. The frequency within the train was 200 Hz for high frequency stimulation (HFS). The distance between trains was 5 sec. Nondecremental LTP or L-LTP lasted for at least 24 hours (Schulz et al. 1999). In computer simulations, spontaneous spiking input from EC (ipsilateral and contralateral) was generated randomly (Poisson train) with an average frequency of 8 Hz to simulate the spontaneous theta modulation (Frank et al. 200 l). That has lead to a postsynaptic spontaneous activity of granule cells of ~ 1 Hz (Kimura and Pavlides 2000). Spontaneous input has to be synchronous between the inputs so that their weights keep approximately the same value. There is an anatomical basis for such a synchronization within EC (Biella et al. 2002). Decorrelated random spontaneous activity of frequency < 1 Hz can be superimposed upon all three input weights with no effect. Model GC received spontaneous spikes all the time. HFS of 20 pulse trains was delivered to pp at t = 2 hours. During the HFS of perforant pathway, there was an ongoing 8 Hz-spontaneous input activity from cEC input. During the 5s intertrain intervals all inputs received uncorrelated spontaneous activity of the frequency of 8Hz. After the pp HFS, 8Hz correlated spontaneous spikes at all three inputs resumed again. In the following figures, we summarize results of our computer simulation. All presented simulated curves are averages from 6 measurements, similarly like in (Schulz et al. 1999). Fig. 9.7 shows the results of simulation of induction and maintenance of nondecremental LTP in granule cells. Magnitude and duration of fEPSP change (i.e. 24 hours) in our computer simulation are the same as in the experimental study (Schulz et al. 1999). Percentual change in the field EPSP was calculated as a dimensionless linear sum either of mpp and lpp weight changes for pp input, i.e. ~fEPSP = ~wmpp + ~Wlpp or for the contralateral input as ~fEPSP = ~WcEC. As we can see in Fig. 9.7a, HFS of pp consisting of 20 trains leads to homosynaptic LTP of pp and heterosynaptic LTD of cEC input. Since the induction of LTD of cEC pathway was not tested in the simulated experiments of Schulz et al. (Schulz et al. 1999), it can be considered to be the model prediction. However, this prediction of the model is in accordance with experimental data of Levy and Steward (Levy and Steward 1983), in which the HFS of ipsilateral pp depressed the contralateral pathway when the latter was not receiving a concurrent HFS, which is the case of our study.
9.5 Applicationto ModelingofL-LTP
201
Fig. 9.7b shows the temporal courses of [pCREB(t)] that accompanies the induction and maintenance of L-LTP and has the same course and amplitude as in the experimental study (Schulz et al. 1999). Fig. 9.7c depicts temporal evolution of the modification threshold ~ in our computer simulations. Synaptic weights and therefore ~ change slowly in dependence on [pCREB(t)] and quickly in dependence on the time average of postsynaptic spiking activity over the last TM = 30 sec. To conclude, we would like to note that in the experimental study (Schulz et al. 1999), decremental or early E-LTP was also induced and [pCREB] measured but the paper does not provide sufficient details (like amplitude and detailed time course of [pCREBJ) for setting up our model for that situation. HFS
a
~
2
8
6
4
CD UJ
u .e,
12
14
18
20
22
16
18
20
22
24
16
18
20
22
24
16
time (hours )
b a:
10
~LR 2
0
4
6
8
2
12
14
tim e (h o urs )
C
o
10
4
6
8
10
12
14
time (hours)
Fig. 9.7. (a) Temporal evolution of fEPSP in our computer simulation of L-LTP. PP means perforant path, cEC means contralateral entorhinal cortex. Nondecremental L-LTP continues to last for 24 hours of simulation; (b) Biphasic course of [pCREB(t)] that accompanies the induction and maintenance of L-LTP as measured in the experiment (Schulz et al. 1999); (c) Evolution of the modification threshold ~ in the model
202
9 Application of CNGM to Learning and Memory
9.6 Summary and Discussion Our computer simulations faithfully reproduce the results of experimental study of L-LTP (Schulz et al. 1999). In our model, we have linked the temporal changes in the levels of pCREB as measured in experiment to the dynamics of the BCM synaptic modification threshold 8M that determines the magnitude of synaptic potentiation and depression in STDP, which is a novel and original contribution of this chapter. Learning rule, which we have introduced in this chapter and which we have used to model the experimental data on hippocampal synaptic plasticity leads to the following picture of relative synaptic changes during the course of the model simulation (see e.g. Fig. 9.8).
0 .05
-.
0 .04 0 .03 0 .02
::+
-e
::
O
(A.2.12)
j(x) = { O,x :O:;O
the unit is called a threshold gate and can generate only binary decisions. ANN can implement different machine learning techniques and hence the variety of the ANN architectures. Many of these architectures are known as "black boxes" as they do not facilitate revealing internal relationships between inputs and output variables of the problem in an explicit form. But for the process of knowledge discovery, having a "black box" learning machine is not sufficient. A learning system should also facilitate extracting useful information from data for the sake of a better understanding and learning of new knowledge. The knowledge-based ANNs (KBANNs) have been developed for this purpose. They combine the strengths of different AI techniques, e.g. ANN and rule-based systems, or fuzzy logic. Evolving connectionist systems (ECOS) have been recently developed to facilitate both adaptive learning in an evolving structure and knowledge discovery (Kasabov 2003). ECOS are modular connectionist-based systems that evolve their structure and functionality in a continuous, self-organized, on-line, adaptive, interactive way from incoming information; they can process both data and knowledge in a supervised and/or unsupervised way. Learning is based on clustering in the input space and on function estimation for this cluster in the output space. Prototype rules can be extracted to represent the clusters and the functions associated with them. Different types of rules are facilitated by different ECOS architectures, such as evolving fuzzy neural networks (EFuNN) (see Fig. A.2.3), dynamic neuro-fuzzy inference systems (DENFIS), etc. An ECOS structure grows and "shrinks" in a continuous way from input data streams. Feedforward and feedback connections are both used in the architectures. The ECOS are not limited in number and types of inputs, outputs, nodes, connections. A simple learning algorithm of a simplified version of EFuNN called ECF (Evolving Classifying Function) is given in next section. Evolving Classifier Function (ECF)
The learning algorithm for the ECF ANN:
A.2 A Brief Overview of Computational Intelligence Methods
255
1. Enter the current input vector from the data set (stream) and calculate the distances between this vector and all rule nodes already created using Euclidean distance (by default). If there is no node created, create the first one that has the coordinates of the first input vector attached as input connection weights. 2. If all calculated distances between the new input vector and the existing rule nodes are greater than a maximum-radius parameter Rmax , a new rule node is created. The position of the new rule node is the same as the current vector in the input data space and the radius of its receptive field is set to the minimum-radius parameter Rmin ; the algorithm goes to step 1; otherwise it goes to the next step. 3. If there is a rule node with a distance to the current input vector less then or equal to its radius and its class is the same as the class of the new vector, nothing will be changed; go to step I; otherwise: 4. If there is a rule node with a distance to the input vector less then or equal to its radius and its class is different from those of the input vector, its influence field should be reduced. The radius of the new field is set to the larger value from the two numbers: distance minus the minimum-radius; minimum-radius. New node is created as in 2 to represent the new data vector. 5. If there is a rule node with a distance to the input vector less than or equal to the maximum-radius, and its class is the same as of the input vector's, enlarge the influence field by taking the distance as a new radius if only such enlarged field does not cover any other rule nodes which belong to a different class; otherwise, create a new rule node in the same way as in step 2, and go to step 1. Recall procedure (classification of a new input vector) in a trained ECF: 1. Enter the new input vector in the ECF trained system. If the new input vector lies within the field of one or more rule nodes associated with one class, the vector is classified in this class. 2. If the input vector lies within the fields of two or more rule nodes associated with different classes, the vector will belong to the class corresponding to the closest rule node. 3. If the input vector does not lie within any field, then take m highest activated by the new vector rule nodes, and calculate the average distances from the vector to the nodes with the same class; the vector will belong to the class corresponding to the smallest average distance. ECOS have been used for different tasks, including gene expression modeling and profile discovery (see the next section), GRN modeling, protein data analysis, brain data modeling, etc. (Kasabov 2003).
256
Appendix 2
A.2.4 Methods of Evolutionary Computation (EC)
EC methods are inspired by the Darwinian theory of evolution. These are methods that search in a space of possible solutions for the best solution of a problem defined through an objective function (Goldberg 1989). EC methods have been used for parameter estimation or optimization in many engineering applications. Unlike classical derivative-based (like Newton) optimization methods, EC is more robust against noise and multi-modality in the search space. In addition, EC does not require the derivative information of the objective function and is thus applicable to complex, blackbox problems. Several techniques have been developed as part of the EC area: genetic algorithms (GA), evolutionary strategies, evolutionary programming, particle swarm optimization, artificial life, etc., the GA being the most popular technique so far. GA is an optimization technique aiming at finding the optimal values of parameters ("genes") for the "best" "individual" according to a pre-defined objective function (fitness function). A GA includes the following steps: • GAL Create a population ofN individuals, each individual being represented as a "chromosome" consisting of values (alleles) of parameters called "genes". • GA2. Evaluate the fitness of each individual towards a pre-defined objective function. If an individual achieves a desired fitness score, or alternatively - the time for running the procedure is over, the GA algorithm STOPS. • GA3. Otherwise, select a subset of "best" individuals using pre-defined selection criteria (e.g. top ranked, roulette-wheel, keep the best individuals through generations, etc. • GA4. Crossover the selected individuals using a crossover ("mating") technique to create a new generation of a population of individuals. • GA5. Apply mutation using a mutation technique. Go to GA2. GA is a heuristic and non-deterministic algorithm. It can give a close to optimal solution depending on the time of execution. For a large number of parameters ("genes in the chromosome") it is much faster than an exhaustive search and much more efficient. Representing real genes, or other biological variables (proteins, binding strengths, connection weights, etc) as GA "genes", is a natural way to solve difficult optimization tasks in CI. For this reason GAs are used for several tasks in this book and also in the proposed CNGM.
Appendix 3
A.3 Some Sources of Brain-Gene Data, Information, Knowledge and Computational Models - Allen Institute and the Allen Brain Atlas: http://www.alleninstitute.org - Alzheimer disease & frontotemporal dementia mutation database: http://www.molgen.ua.ac.be/admutations - Alzheimer research forum genetic database of candidate genes: http://www.alzforum.org/ - Blue Brain Project: http://bluebrainproject.epfl.ch/index.html - Brain-Gene Ontology: http://www.kedri.info/ - Brain models at USC: http://www-hbp.usc.edu/Projects/bmw.htm - Brain models: http://ttb.eng.wayne.edu/brain/ - Cancer gene expression data: http://wwwgenome.wi.mit.edu/MPRlGCM.html - eMedicine: http://www.emedicine.com/ - Ensemble Human Gene View: http://www.ensembl.org/Homo_sapiens/index.html - Epilepsy: http://www.epilepsy.com/epilepsy/epilepsy_brain.html - European Bioinformatics Institute EEl: http://www.ebi.ac.uk - ExPASy (Expert Protein Analysis System) Proteomics Server: http://www.expasy.org/ - Genes and disease: http://www.ncbi.nlm.nih.gov/books/ - Gene Expression Atlas: http://expression.gnf.org/cgi-bin/index.cgi - GeneCards (integrated database of human genes): http://www.genecards.org/index.html - GeneLoc (presents an integrated map for each human chromosome): http://bioinfo2.weizmann.ac.illgeneloc/index.shtml - How Stuff Works: http://health.howstuffworks.com/brainl.htm - KEGG (Kyoto Encyclopedia of Genes and Genomes): http://www.genome.jp/kegg/
258
Appendix 3
- MathWorld - A Wolfram Web Resource: http://mathworld.wolfram.com/DelayDifferentialEquation.htmI - NCBI Genbank: http://www.ncbi.nlm.nih.gov/Genbank/index.html - Neural Micro Circuits Software: http://www.1sm.tugraz.at - Neuro-Computing Decision Support Environment (NeuCom): http://www.aut.ac.nz/researchlresearch_institutes/kedri/research_centres /centre_for_novel_methods_oCcomputational_intelligence/neucom.htm - The Brain Guide: http://www.omsusa.org/pranzatelli-Brain.htm - The National Society for Epilepsy: http://www.e-epilepsy.org.ukl
References
Abbott LF, Nelson SB (2000) Synaptic plasticity: taming the beast. Nature Neurosci 3:1178-1183 Abraham WC, Christie BC, Logan B, Lawlor P, Dragunow M (1994) Immediate early gene expression associated with the persistence of heterosynaptic longterm depression in the hippocampus. Proc Nat! Acad Sci USA 91:1004910053 Abraham WC, Bear MF (1996) Metaplasticity: the plasticity of synaptic plasticity. Trends Neurosci 19(4):126-130 Abraham WC, Tate WP (1997) Metaplasticity: a new vista across the field of synaptic plasticity. Prog Neurobiol 52(4):303-323 Abraham WC, Mason-Parker SE, Bear MF, Webb S, Tate WP (2001) Heterosynaptic metaplasticity in the hippocampus in vivo: a BCM-like modifiable threshold for LTP. Proc Nat! Acad Sci USA 98(19):10924-10929 Abraham WC, Logan B, Greenwood JM, Dragunow M (2002) Induction and experience-dependent consolidation of stable long-term potentiation lasting months in the hippocampus. J Neurosci 22(21):9626-9634 Aha DW, Kibler D, Albert MK (1991) Instance-based learning algorithm. Machine Learning 6:37-66 Al-Rabiai S, Miller MW (1989) Effect of prenatal exposure to ethanol on the ultrastructure of layer V of mature rat somatosensory cortex. J Neurocytology 18:711-729 Albus JS (1975) A new approach to manipulator control: The cerebellar model articulation controller (CMAC). Trans of the ASME: Journal of Dynamic Systems, Measurement, and Control 27:220-227 Alzheimer disease & frontotemporal dementia mutation database (2006), http://www.molgen.ua.ac.be/admutations. Human Genome Variation Society Alzheimer research forum genetic database of candidate genes (2006), http://www.alzforum.org/ Amaral DG, Witter MP (1989) The three-dimensional organization of the hippocampal formation: a review of anatomical data. Neuroscience 31:571-591 Amari S (1967) A theory of adaptive pattern classifiers. IEEE Trans on Electronic Computers 16:299-307 Amari S (1990) Mathematical foundations of neuro-computing. Proc IEEE 78:1143-1163 Amari S, Kasabov N (eds) (1998) Brain-like computing and intelligent information systems, Springer, Singapore
260
References
Arbib M (1972) The metaphorical brain. An introduction to cybernetics as artificial intelligence and brain theory. John Wiley & Sons, New York Arbib M (1987) Brains, machines and mathematics. Springer, Berlin Arbib M (ed) (2003) The handbook of brain theory and neural networks, ed 2, MIT Press, Cambridge, MA Armstrong-James M, Callahan CA (1991) Tha1amo-cortica1 processing of vibrissal information in the rat. II. Spatiotemporal convergence in the thalamic ventroposterior medial nucleus (VPm) and its relevance to generation of receptive fields ofSl cortical "barrel" neurones. J Comp Neuro1303:211-224 Armstrong-James M, Callahan CA, Friedman MA (1991) Thalamo-cortical processing of vibrissal information in the rat. I. intracortical origins of surround but not centre-receptive fields of layer IV neurones in the rat S1 barrel field cortex. J Comp Neuro1303:193-210 Armstrong-James M, Diamond ME, Ebner FF (1994) An innocuous bias in whisker sensation modifies receptive fields of adult rat barrel cortex neurons. J Neurosci 11(14):6978-6991 Arnold SE, Trojanowski JQ (1996) Recent advances in defining the neuropathology of schizophrenia. Acta Neuropathol (Berl) 92(3):217-231 Artola A, Brocher S, Singer W (1990) Different voltage-dependent threshold for inducing long-term depression and long-term potentiation in slices of rat visual cortex. Nature 347:69-72 Artola A, Singer W (1993) Long-term depression of excitatory synaptic transmission and its relationship to long-term potentiation. Trends Neurosci 16(11):480-487 Bailey CH, Kandel ER, Si K (2004) The persistence of long-term memory: a molecular approach to self-sustaining changes in learning-induced synaptic growth. Neuron 44:49-57 Bak P, Tang C, Wiesenfe1d K (1987) Self-organized criticality: an explanation of l/fnoise. Phys Rev Lett 59:381-384 Baldi P, Brunak S (2001) Bioinformatics. The machine learning approach, ed 2nd. MIT Press, Cambridge, MA Baldi P, Long AD (2001) A Bayesian framework for the analysis of microarray expression data: regularized t -test and statistical inferences of gene changes. Bioinformatics 17(6):509-519 Baldi P, Hatfield GW (2002) DNA microarrays and gene expression: from experiments to data analysis and modeling. Cambridge Univ. Press, Cambridge, UK Barnett KJ, Corballis MC, Kirk IJ (2005) Symmetry of callosal information transfer in schizophrenia: a preliminary study. Schizophr Res 74(2-3):171-178 Basalyga DM, Simionescu DT, Xiong W, Baxter BT, Starcher BC, Vyavahare NR (2004) Elastin degradation and calcification in an abdominal aorta injury model: role of matrix metalloproteinases. Circulation 110(22):3480-3487 Bear MF, Cooper LN, Ebner FF (1987) A physiological basis for a theory of synapse modification. Science 237:42-48 Bear MF (1995) Mechanism for a sliding synaptic modification threshold. Neuron 15(1):1-4
261 Bear MF, Connors BW, Paradiso MA (2001) Neuroscience: exploring the brain, ed 2. Lippincott Williams & Wilkins, Baltimore, MD Beattie EC, Carroll RC, Yu X, Morishita W, Yasuda H, vonZastrow M, Malenka RC (2000) Regulation of AMPA receptor endocytosis by a signaling mechanism shared with LTD. Nature Neurosci 3(12):1291-1300 Beierlein M, Fall CP, Rinzel J, Yuste R (2002) Thalamocortical bursts trigger recurrent activity in neocortical networks: layer 4 as a frequency-dependent gate. J Neurosci 22(22):9885-9894 Benes FM (1989) Myelination of cortical-hippocampal relays during late adolescence. Schizophr Bull 15(4):585-593 Bentley PJ (2004) Controlling robots with fractal gene regulatory networks. In: deCastro L, vonZuben F (eds) Recent developments in biologically inspired computing, vol 1. Idea Group Inc, Hershey, PA, pp 320-339 Benuskova L (1988) Mechanisms of synaptic plasticity. Czechoslovak Physiology 37(5):387-400 Benuskova L, Diamond ME, Ebner FF (1994) Dynamic synaptic modification threshold: computational model of experience-dependent plasticity in adult rat barrel cortex. Proc Natl Acad Sci USA 91:4791-4795 Benuskova L (2000) The intra-spine electric force can drive vesicles for fusion: a theoretical model for long-term potentiation. Neurosci Lett 280( 1):17-20 Benuskova L, Kanich M, Krakovska A (2001) Piriform cortex model of EEG has random underlying dynamics. In: Rattay F (ed) Proc. World Congress on Neuroinformatics. ARGESIM/ASIM-Verlag, Vienna Benuskova L, Rema V, Armstrong-James M, Ebner FF (2001) Theory for normal and impaired experience-dependent plasticity in neocortex of adult rats. Proc Natl Acad Sci USA 98(5):2797-2802 Benuskova L, Abraham WC (2006) STDP rule endowed with the BCM sliding threshold accounts for hippocampal heterosynaptic plasticity. J Comp Neurosci (in press) Benuskova L, Kasabov N, Jain V, Wysoski SG (2006) Computational neurogenetic modelling: a pathway to new discoveries in genetic neuroscience. Intl J Neural Systems 16(3):215-227 Bertram L, Tanzi RE (2005) The genetic epidemiology of neurodegenerative disease. J Clin Invest 115(6):1449-1457 Bezdek JC (1981) Pattern recognition with fuzzy objective function algorithms. Plenum Press, New York Bi G-q, Poo M-m (1998) Synaptic modification in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci 18(24):10464-10472 Biella G, Uva L, Hoffmann UG, Curtis MD (2002) Associative interactions within the superficial layers of the entorhinal cortex of the guinea pig. J Neurophysiol 88(3): 1159-1165 Bienenstock EL, Cooper LN, Munro PW (1982) Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex. J Neurosci 2(1):32-48
262
References
Bienvenu T, Poirier K, Friocourt G, Bahi N, Beaumont D, Fauchereau F, Jeema LB, Zemni R, Vinet M-C, Francis F, Couvert P, Gomot M, Moraine C, Bokhoven Hv, Kalscheuer V, Frints S, Gecz J, Ohzaki K, Chaabouni H, Fryns J-P, Desportes V, Beldjord C, Chelly J (2002) ARX, a novel Prd-class-homeobox gene highly expressed in the telencephalon, is mutated in X-linked mental retardation. Human Molecular Genetics 11(8):981-991 Bishop CM (1995) Neural networks for pattern recognition. Oxford University Press, Oxford Biswa1 B, Dasgupta C (2002) Neural network model for apparent deterministic chaos in spontaneously bursting hippocampal slices. Physical Review Letters 88(8):88-102 Bito H, Deisseroth K, Tsien RW (1997) Ca2+-dependent regulation in neuronal gene expression. Curr Opin Neurobio1 7:419-429 Bliss TV, Lomo T (1973) Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbitt following stimulation of perforant path. J Physiol 232(2):331-356 Bliss TVP (1999) Young receptors make smart mice. Nature 401:25-27 Bortolotto ZA, Collingridge GL (1998) Involvement of calcium/ca1modulindependent protein kinases in the setting of a molecular switch involved in hippocampal LTP. Neuropharmacology 37:535-544 Bower JM, Beeman D (1998) The book of GENESIS: exploring realistic neural models with the GEneral NEural SImulation System, ed 2. TELOS/Springer, New York Bradley P, Louis T (1996) Bayes and empirical Bayes methods for data analysis. Chapman & Hall, London Bradshaw KD, Emptage NJ, Bliss TVP (2003) A role for dendritic protein synthesis in hippocampal late LTP. Eur J Neurosci 18(11):3150-3152 Brink DMvd, Brites P, Haasjes J, Wierzbicki AS, Mitchell J, Lambert-Hamill M, Belleroche Jd, Jansen GA, Waterham HR, Wanders RJ (2003) Identification of PEX7 as the second gene involved in Refsum disease. Am J Hum Genet 72(2):471-477 Brocher S, Artola A, Singer W (1992) Agonists of cholinergic and noradrenergic receptors facilitate synergistically the induction of long-term potentiation in slices of rat visual cortex. Brain Res 573:27-36 Brown A, Yates PA, Burrola P, Ortuno D, Ashish V, Jesselt TM, Pfaff SL, O'Leary DDM, Lemke G (2000) Topographic mapping from the retina to the midbrain is controlled by relative but not absolute levels of EphA receptor signaling. Cell 102:77-88 Brown C, Shreiber M, Chapman B, Jacobs G (2000) Information science and bioinformatics. In: N K (ed) Future directions of intelligent systems and information sciences. Springer, pp 251-287 Brown MPS, Grundy WN, Lin D, Cristianini N, Sugnet CW, Furey TS, Ares M, Jr., Haussler D (2000) Knowledge-based analysis of microarray gene expression data by using support vector machines. PNAS 97(1):262-267 Brown PO, Botstein D (1999) Exploring the new world of the genome with DNA microarrays. Nature 21:33-37
263 Brzustowicz LM, Hodgkinson KA, Chow EWC, Honer WG, Bassett AS (2000) Location of a Major Susceptibility Locus for Familial Schizophrenia on Chromosome 1q21-q22. Science 288(5466):678-682 Buiting K, Gross S, Lich C, Gillessen-Kaesbach G, e1-Maarri 0, Horsthemke B (2003) Epimutations in Prader-Willi and Angelman syndromes: a molecular study of 136 patients with an imprinting defect. Am J Hum Genet 72(3):571577 Bulik CM, Devlin B, Bacanu S-A, Thornton L, Klump KL, Fichter MM, Halmi KA, Kaplan AS, Strober M, Woodside DB, Bergen AW, Ganjei JK, Crow S, Mitchell J, Rotondo A, Mauri M, Cassano G, Keel P, Berrettini WH, Kaye WH (2003) Significant linkage on chromosome lOp in families with bulimia nervosa. Am J Hum Genet 72:200-207 Burnashev N, Rozov A (2000) Genomic control of receptor function. Cellular and Molecular Life Sciences 57: 1499-1507 Cacabelos R, Takeda M, Winblad B (1999) The glutamatergic system and neurodegeneration in dementia: preventive strategies in Alzheimer's disease. Int J Geriat Psychiatry 14:3-47 Cai L, Friedman N, Xie XS (2006) Stochastic protein expression in individual cells at the single molecule level. Nature 440:358-362 Cao Q, Martinez M, Zhang J, Sanders AR, Badner JA, Cravchik A, Markey CJ, Beshah E, Guroff JJ, Maxwell ME, Kazuba DM, Whiten R, Goldin LR, Gershon ES, Gejman PV (1997) Suggestive evidence for a schizophrenia susceptibility locus on chromosome 6q and a confirmation in an independent series of pedigrees. Genomics 43(1): 1-8 Carnevale NT, Hines ML (2006) The NEURON book. Cambridge University Press, Cambridge, UK Carpenter G, Grossberg S (1991) Pattern recognition by self-organizing neural networks. MIT Press, Cambridge, MA Carpenter G, Grossberg S, Markuzon N, Reynolds JH, Rosen DB (1991) Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analogue multi-dimensional maps. IEEE Trans on Neural Networks 3(5):698-713 Carrie A, Jun L, Bienvenu T, Vinet M-C, McDonell N, Couvert P, Zemni R, Cardona A, Buggenhout GV, Frints S, Hamel B, Moraine C, Ropers HH, Strom T, Howell GR, Whittaker A, Ross MT, Kahn A, Fryns J-P, Beldjord C, Marynen P, Chelly J (1999) A new member of the IL-1 receptor family highly expressed in hippocampus and involved in X-linked mental retardation. Nature Genetics 23:25-31 Carroll RC, Beattie EC, vonZastrow M, Malenka RC (2001) Role of AMPA receptor endocytosis in synaptic plasticity. Nature Rev Neurosci 2:315-324 Caspi A, Sugden K, Moffitt TE, Taylor A, Craig IW, Harrington H, McClay J, Mill J, Martin J, Braithwaite A, Poulton R (2003) Influence of life stress on depression: moderation by a polymorphism in the 5-HTT gene. Science 301(5631 ):386-389
264
References
Castellani GC, Quinlan EM, Cooper LN, Shouval HZ (2001) A biophysical model of bidirectional synaptic plasticity: dependence on AMPA and NMDA receptors. Proc Nat! Acad Sci USA 98(22):12772-12777 Cater MA, Forbes J, Fontaine SL, Cox D, Mercer JF (2004) Intracellular trafficking of the human Wilson protein: the role of the six N-terminal metal-binding sites. Biochem J 380(Pt 1):805-813 Cattaneo E, Rigamonti D, Zuccatto C, Squittieri F, Sipione S (2001) Loss of normal huntingtin function: new developments in Huntington's disease. Trends Neurosci 24:182-188 Cavalli-Sforza LL (2001) Genes, people and languages. Penguin Books, London Cavazos JE, Lum F (2005) Seizures and epilepsy: overview and classification. eMedicine.com, Inc., http://www.emedicine.comlneuro/topic415.htm Cerny V (1985) A thermodynamical approach to the travelling salesman problem: an efficient simulation algorithm. Journal of Optimization Theory and Applications 45:41-51 Chalmers DJ (1996) The conscious mind: in search of a fundamental theory. Oxford University Press, Oxford Chan ZSH, Kasabov N, Collins L (2006) A two-stage methodology for gene regulatory network extraction from time-course gene expression data. Expert Systems with Applications 30(1):59-63 Chen CK, Chen SL, Mill J, Huang YS, Lin SK, Curran S, Purcell S, Sham P, Asherson P (2003) The dopamine transporter gene is associated with attention deficit hyperactivity disorder in a Taiwanese sample. Mol Psychiatry 8(4):393-396 Chen L, Aihara K (2002) Stability analysis of genetic regulatory networks with time delay. IEEE Trans on Circuits and Systems - I: Fundamental Theory and Applications 49(5):602-608 Chenna R, Sugawara H, Koike T, Lopez R, Gibson TJ, Higgins DG, Thompson JD (2003) Multiple sequence alignment with the Clustal series of programs. Nucleic Acids Res 31(13):3497-3500 Chhabra J, Glezer M, Shkuro Y, Gittens SD, Reggia JA (1999) Effects of callosal lesions in a computational model of single-word reading. In: Reggia JA, Ruppin E, Glanzman DL (eds) Disorders of brain, behavior, and cognition: the neurocomputational perspective. Progress in brain research, vol 121. Springer, New York, pp 219-242 Chin HR, Moldin SO (eds) (2001) Methods in genomic neuroscience. Methods and new frontiers in neuroscience, CRC Press, Boca Raton Cho K, Aggleton JP, Brown MW, Bashir ZI (2001) An experimental test of the role of postsynaptic calcium levels in determining synaptic strength using perirhinal cortex of rat. J Physio1532(2):459-466 Chumakov I, Blumenfeld M, Guerassimenko 0, Cavarec L, Palicio M, Abderrahim H, Bougueleret L, Barry C, Tanaka H, Rosa PL (2002) Genetic and physiological data implicating the new human gene G72 and the gene for Damino acid oxidase in schizophrenia. Proc Nat! Acad Sci USA 99(1367513680)
265 Citron M (2004) Strategies for disease modification in Alzheimer's disease. Nature Rev Neurosci 5(9):677-685 Cloete I, Zurada J (eds) (2000) Knowledge-based neurocomputing, MIT Press, Cambridge, MA Cloninger CR (2002) The discovery of susceptibility genes for mental disorders. Proc Nat! Acad Sci USA 99(21):13365-13367 Clothiaux EE, Bear MF, Cooper LN (1991) Synaptic plasticity in visual cortex: comparison of theory with experiment. J Neurophysiol 66(5): 1785-1804 Cooper LN (1987) Cortical plasticity: theoretical analysis, experimental results. In: Rauschecker JP, Marler P (eds) Imprinting and cortical plasticity. John Wiley & Sons, New York, pp 117-191 Cooper LN, Intrator N, Blais B, Shouval HZ (2004) Theory of cortical plasticity. World Scientific, Singapore Corballis MC (2003) From mouth to hand: gesture, speech, and the evolution of right-handedness. Behav Brain Sci 26: 199-260 Crick F, Koch C (1995) Are we aware of neural activity in primary visual cortex? Nature 375:121-123 Crunelli V, Leresche N (2002) Childhood absence epilepsy: genes, channels, neurons and networks. Nature Rev Neurosci 3(5):371-382 Cybenko G (1989) Approximation by super-positions of sigmoidal function. Mathematics of Control, Signals and Systems 2:303-314 D'Haeseleer P, Wen X, Fuhrman S, Somogyi R (1999) Linear modeling ofmRNA expression levels during CNS development and injury. Proc. Pacific Symposium on Biocomputing, World Scientific, Singapore, pp 41-52 D'Haeseleer P, Liang S, Somogyi R (2000) Genetic network inference: from coexpression clustering to reverse engineering. Bioinformatics 16(8):707-726 Damasio AR (1994) Descartes' error. Putnam's Sons, New York Darwin C (1859) The origin of species by means of natural selection. John Murray, London de Jong H (2002) Modeling and simulation of genetic regulatory systems: a literature review. Journal of Computational Biology 9( I):67-102 DeFelipe J (1997) Types of neurons, synaptic connections and chemical characteristics of cells immunoreactive for calbindin-D28K, parvalbumin and calretinin in the neocortex. J Chern Neuroanat 14:1-19 Deisz RA (1999) GABAB receptor-mediated effects in human and rat neocortical neurones in vitro. Neuropharmacology 38:1755-1766 DelaTorre JC, Barrios M, Junque C (2005) Frontal lobe alterations in schizophrenia: neuroimaging and neuropsychological findings. Eur Arch Psychiatry Clin Neurosci 255(4):236-244 Delorme A, Gautrais J, vanRullen R, Thorpe S (1999) SpikeNET: a simulator for modeling large networks of integrate and fire neurons. Neurocomputing 2627:989-996 Delorme A, Thorpe S (200 I) Face identification using one spike per neuron: resistance to image degradation. Neural Networks 14:795-803 Dennett DC (1991) Consciousness explained. Penguin Books, New York
266
References
Destexhe A (1998) Spike-and-wave oscillations based on the properties of GABAB receptors. J Neurosci 18:9099-9111 Devlin JT, Gonnennan LM, Andersen ES, Seidenberg MS (1998) Categoryspecific semantic deficits in focal and widespread brain-damage: a computational account. J Cogn Neurosci 10(1):77-94 Diamond ME, Armstrong-James M, Ebner FF (1993) Experience-dependent plasticity in adult rat barrel cortex. Proc Nat! Acad Sci USA 90(5):2082-2086 Diamond ME, Petersen RS, Harris JA, Panzeri S (2003) Investigations into the organization ofinfonnation in sensory cortex. J Physiol Paris 97(4-6) :529-536 Dimitrov D, Sidorov I, Kasabov N (2004) Computational biology . In: Rieth M, Sommers W (eds) Handbook of theoretical and computational nanotechnology, vol 1. American Scientific, Los Angeles DiPellegrino G, Fadiga L, Fogassi L, Gallese V, Rizzolatti G (1992) Understanding motor events: a neurophysiological study. Experimental Brain Research 91:176-180 Drager LD, Layton W (1997) Initial value problems for nonlinear nonresonant delay differential equations with possibly infinite delay. Electronic Journal of Differential Equations (24) :1-20 Duch W, Adamczak R, Grabczewski K (1998) Extraction of logical rules from neural networks. Neural Proc Letters 7:211-219 Dudek SM, Bear MF (1993) Bidirectional long-term modification of synaptic effectiveness in the adult and immature hippocampus . J Neurosci 13(7):15181521 Eckert T, Eidelberg D (2005) Neuroimaging and therapeutics in movement disorders. NeuroRx 2(2):361-371 Edelman GM, Tononi G (2000) Consciousness. how matter becomes imagination. Penguin Books, London Egan MF, Goldberg TE, Kolachana BS, Callicott JH, Mazzanti CM, Straub RE, Goldman D, Weinberger DR (2001) Effect ofCOMT Vall 08/158 Met genotype on frontal lobe function and risk for schizophrenia. Proc Nat! Acad Sci USA 98(12):6917-6922 Elgersma Y, Silva AJ (1999) Molecular mechanisms of synaptic plasticity and memory. Current Opinion in Neurobiology 9(2):209-213 Elliott T, Shadbolt NR (1999) A neurotrophic model of the development of the retinogeniculocortical pathway induced by spontaneous retinal waves. J Neurosci 19(18):7951-7970 Enard W, Przeworski M, Fisher SE, Lai CSL, Wiebe V, Kitano T, Monaco AP, Paabo S (2002) Molecular evolution of FOXP2 , a gene involved in speech and language. Nature 418:869-872 Engel AK, Fries P, Konig P, Brecht M, Singer W (1999) Temporal binding, binocular rivarly, and consciousness. Consciousness and Cognition 8:128-151 Evans PD, Gilbert SL, Mekel-Bobrov N, Vallender EJ, Anderson JR, Vaez-Azizi LM, Tishkoff SA, Hudson RR, Lahn BT (2005) Microcephalin, a gene regulating brain size, continues to evolve adaptively in humans Science 309(5741): 1717-1720
267 Fahlman C, Lebiere C (1990) The cascade-correlation learning architecture . In: Touretzky DS (ed) Advances in neural information processing systems, vol 2. Morgan Kaufmann , San Francisco, CA Fairbanks LD, Jacomelli G, Micheli V, Slade T, Simmonds HA (2002) Severe pyridine nucleotide depletion in fibroblasts from Lesch-Nyhan patients . Biochern J 366(pt 1):265-272 Federici T, Boulis NM (2005) Gene-based treatment of motor neuron diseases, Muscle & Nerve, http://www3.interscience.wiley.comicgibin/fulltext/112l17863/HTMLSTART Fedor P, Benuskova L, Jakes H, Majernik V (1982) An electrophoretic coupling mechanism between efficiency modification of spine synapses and their stimulation. Studia Biophysica 92:141-146 Feng R, Rampon C, Tang Y-P, Shrom D, Jin J, Kyin M, Sopher B, Martin GM, Kim S-H, Langdon RB, Sisodia SS, Tsien JZ (2001) Deficient neurogenesis in forebrain-specific Presenilin-l knockout mice is associated with reduced clearance of hippocampal memory traces. Neuron 32:911-926 Fenton GW, Fenwick PBC, Dollimore J, Dunn TL, Hirsch SR (1980) EEG spectral analysis in schizophrenia. Brit J Psychiat 136:445-455 Fisher A, Walker MC, Bowery NG (2003) Mechanisms of action of anti-epileptic drugs. The National Society for Epilepsy, http://www.eepilepsy.org.uk/pages/articles/show_article.cfm?id= 111 Fogel D, Fogel L, Porto V (1990) Evolving neural networks. Bioi Cybernetics 63:487-493 Fogel DB (1995) Evolutionary computation - Toward a new philosophy of machine intelligence. IEEE Press, New York Fogel G, Corne D (2003) Evolutionary computation for bioinformatics. Morgan Kaufmann, San Francisco, CA Frank LM, Brown EN, Wilson MA (2001) A comparison of the firing properties of putative excitatory and inhibitory neurons from CAl and the entorhinal cortex. J Neurophysiol 86(4):2029-2049 Frank MJ (2005) Dynamic dopamine modulation in the basal ganglia : a neurocomputational account of cognitive deficits in medicated and nonmedicated Parkinsonism. J Cogn Neurosci 17:51-72 Fraser HB, Hirsh AE, Giaever G, Kurnn J, Eisen MB (2004) Noise minimization in eukaryotic gene expression. PLOS Biology 2(6):0834-0838 Freeman WJ (2000) Neurodynamics. An exploration in mesoscopic brain dynamics. Springer, London Freeman WJ (2003) Evidence from human scalp EEG of global chaotic itinerancy. Chaos 13(3):1-11 Fries P, Roelfsema PR, Engel AK, Konig P, Singer W (1997) Synchronization of oscillatory responses in visual cortex correlates with perception in interocular rivalry. Proc Nat! Acad Sci USA 94:12699-12704 Frith U (2001) Mind blindness and the brain in autism. Neuron 32:969-979 Fritzke B (1995) A growing neural gas network learns topologies . Advances in Neural Information Processing Systems 7:625-632
268
References
Froemke RC, Poo M-m, Dang Y (2005) Spike-timing-dependent synaptic plasticity depends on dendritic location. Nature 434:221-225 Funihashi K (1989) On the approximate realization of continuous mappings by neural networks. Neural Networks 2: 183-192 Furuhashi T, Nakaoka K, Uchikawa Y (1994) A new approach to genetic based machine learning and an efficient finding of fuzzy rules. Proc. Proc. WWW'94 Workshop, IEEE/Nagoya-University, Nagoya, pp 114-122 Gabanella F, Carissimi C, Usiello A, Pellizzoni L (2005) The activity of the spinal muscular atrophy protein is regulated during development and cellular differentiation. Hum Mol Genet 14(23):3629-3642 Ganesh S, Puri R, Singh S, Mittal S, Dubey D (2006) Recent advances in the molecular basis of Lafora's progressive myoclonus epilepsy. J Hum Genet 51(1):1-8 Gardenfors P (2000) Conceptual spaces. The geometry of thought. MIT Press, Cambridge, MA Gardiner RM (1999) Genetic basis of human epilepsies. Epilepsy Res 36:91-95 Gardiner RM (2003) Molecular genetics of the epilepsies. The National Society for Epilepsy, http://www.eepilepsy.org.uk/pages/articles/show_article.cfm?id=44 Geinisman Y, deToledo-Morrell L, Morrell F (1991) Induction of long-term potentiation is associated with an increase in the number of axospinous synapses with segmented postsynaptic densities. Brain Research 566:77-88 Genes and disease (2005), National Centre for Biotechnology Information (NCBI), The Nervous System, http://www.ncbi.nlm.nih.gov/books/bv.fcgi?rid=gnd.chapter. 75 George AL (2004) Inherited channelopathies associated with epilepsy. Epilepsy Currents 4(2):65-70 Gerstner W, Kistler WM (2002) Spiking neuron models. Cambridge Univ. Press, Cambridge, MA Glaze DG (2005) Neurophysiology of Rett syndrome. J Clin Neurol 20(9):740746 Gold JI, Bear MF (1994) A model of dendritic spine Ca2+ concentration exploring possible bases for a sliding synaptic modification threshold. Proc Natl Acad Sci USA 91:3941-3945 Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Addison-Wesley, Reading, MA Granato A, Santarelli M, Sbriccoli A, Minciacchi D (1995) Multifaceted alterations of the thalamo-cortico-thalamic loop in adult rats prenatally exposed to ethanol. Anat Embryol 191:11-23 Gray CM, Konig P, Engel AK, Singer W (1989) Oscillatory responses in cat visual cortex exhibit inter-columnar synchronization which reflects global stimulus properties. Nature 338:334-337 Greenbaum D, Colangelo C, Williams K, Gerstein M (2003) Comparing protein abundance and mRNA expression levels on a genomic scale. Genome Biology 4:117.111-117.118
269 Grice DE, Halmi KA, Fichter MM, Strober M, Woodside DB, Treasure JT, Kaplan AS, Magistretti PJ, Goldman D, Bulik CM, Kaye WH, Berrettini WH (2002) Evidence for a susceptibility gene for anorexia nervosa on chromosome 1. Am J Hum Genet 70:787-792 Grossberg S (1969) On learning and energy - entropy dependence in recurrent and nonrecurrent signed networks. J Stat Phys 1:319-350 Grossberg S (1982) Studies of mind and brain. Reidel, Boston Hakak Y, Walker JR, Li C, Wong WH, Davis KL, Buxbaum JD, Haroutunian V, Fienberg AA (2001) Genome-wide expression analysis reveals dysregulation of myelination-related genes in chronic schizophrenia. Proc Nat! Acad Sci USA 98(8):4746-4751 Hameroff S, Penrose R (1996) Orchestrated reduction of quantum coherence In brain microtubules: a model for consciousness? In: Hameroff SR, Kaszniak AW, Scott AC (eds) Toward a science of consciousness: the first Tucson discussions and debates. MIT Press, Cambridge, MA, pp 507-540 Hartemink AJ, Gifford DK, Jaakkola TS, Young RA (2001) Using graphical models and genomic expression data to statistically validate models of genetic regulatory networks. Proc. Pacific Symnposium on Biocomputing, vol. 6, pp 422-433 Hasselmo MF (1997) A computational model of the progression of Alzheimer's disease. MD Computing: Computers in Medical Practice 14(3):181-191 Hassibi B, Stork DG (1992) Second order derivatives for network prunning: optimal brain surgeon. In: Touretzky DS (ed) Advances in neural information processing systems, vol 4. Morgan Kaufmann, San Francisco, CA, pp 164-171 Hauptmann W, Heesche K (1995) A neural network topology for bidirectional fuzzy-neuro transformation. Proc. FUZZ-IEEE/IFES, IEEE Press, Yokohama, Japan, pp 1511-1518 Hayashi Y (1991) A neural expert system with automated extraction of fuzzy ifthen rules and its application to medical diagnosis. In: Lippman RP, Moody JE, Touretzky DS (eds) Advances in neural information processing systems, vol 3. Morgan Kaufmann, San Mateo, CA, pp 578-584 Haykin S (1994) Neural networks - A comprehensive foundation. Prentice Hall, Engelwood Cliffs, NJ Hebb D (1949) The Organization of Behavior. John Wiley and Sons, New York Heskes TM, Kappen B (1993) On-line learning processes in artificial neural networks) Mathematic foundations of neural networks. Elsevier, Amsterdam, pp 199-233 Hevroni D, Rattner A, Bundman M, Lederfein D, Gabarah A, Mangelus M, Silverman MA, Kedar H, Naor C, Komuc M, Hanoch T, Seger R, Theill LE, Nedivi E, Richter-Levin G, Citri Y (1998) Hippocampal plasticity involves extensive gene induction and multiple cellular mechanisms. J Mol Neurosci 10(2):75-98 Hinton GE (1989) Connectionist learning procedures. Artificial Intelligence 40:185-234 Hinton GE (1990) Preface to the special issue on connectionist symbol processing. Artificial Intelligence 46: 1-4
270
References
Hoffman RE, McGlashan TH (1999) Using a speech perception neural network simulation to explore normal neurodevelopment and hallucinated 'voices' in Shizophrenia. In: Reggia JA, Ruppin E, Glanzman DL (eds) Disorders of brain, behavior, and cognition: the neurocomputational perspective. Progress in brain research, vo112!. Springer, New York, pp 311-325 Holland JH (1975) Adaptation in natural and artificial systems. Univ Michigan Press, Ann Arbor, MI Holland JH (1998) Emergence. Oxford Univ Press, Oxford Holland LL, Wagner JJ (1998) Primed facilitation of homosynaptic long-term depression and depotentiation in rat hippocampus. J Neurosci 18(3):887-894 Honey GD, Bullmore ET, Soni W, Varathesaan M, Williams SC, Sharma T (1999) Differences in frontal activation by a working memory task after substitution of risperidone for typical antipsychotic drugs in patients with schizophrenia. Proc Nat! Acad Sci USA 96(23):13432-13437 Hong SJ, Li H, Becker KG, Dawson VL, Dawson TM (2004) Identification and analysis of plasticity-induced late-response genes. Proc Nat! Acad Sci USA 101(7):2145-2150 Hom D, Levy N, Ruppin E (1996) Neuronal-based synaptic compensation: a computational study in Alzheimer's disease. Neural Computation 8(6):12271243 Hornik K, Stinchcombe M, White H (1989) Multilayer feedforward networks are universal approximators. Neural Networks 2:359-366 Huber KM, Gallagher SM, Warren ST, Bear MF (2002) Altered synaptic plasticity in a mouse model of fragile X mental retardation. Proc Nat! Acad Sci USA 99:7746-7750 Hunter L (1994) Artificial intelligence and molecular biology. Canadian Artificial Intelligence 35:10-16 Impey S, Obrietan K, Wong ST, Poser S, Yano S, Wayman G, Deloume JC, Chan G, Storm DR (1998) Cross talk between ERK and PKA is required for Ca2+ stimulation of CREB-dependent transcription and ERK nuclear translocation. Neuron 21:869-883 Inoki K, Ouyang H, Li Y, Guan KL (2005) Signaling by target of rapamycin proteins in cell growth control. Microbiol Mol BioI Rev 69(1):79-100 Ishikawa M (1996) Structural learning with forgetting. Neural Networks 9:501521 Izhikevich EM (2003) Simple model of spiking neurons. IEEE Trans Neural Net 14(6): 1569-1572 Izhikevich EM, Desai NS (2003) Relating STDP to BCM. Neural Computation 15:1511-1523 Jang R (1993) ANFIS: adaptive network-based fuzzy inference system. IEEE Trans on Systems, Man, and Cybernetics 23(3):665-685 Jansen R, Greenbaum D, Gerstein M (2002) Relating whole-genome expression data with protein-protein interactions. Genome Research 12(1):37-46 Jedlicka P (2002) Synaptic plasticity, metaplasticity and the BCM theory. Bratislava Medical Letters 103(4-5):137-143
271 Jensen KF, Killackey HP (1987) Terminal arbors ofaxons projecting to the somatosensory cortex of adult rats. 1. The normal morphology of specific thalamocortical afferents. J Neurosci 7:3529 3543 Jensen 0 (2001) Information transfer between rhytmically coupled networks: reading the hippocampal phase code. Neural Computation 13:2743-2761 Jiang CH, Tsien JZ, Schultz PG, Hu Y (2001) The effects of aging on gene expression in the hypothalamus and cortex of mice. Proc Nat! Acad Sci USA 98(4): 1930-1934 Jouvet P, Rustin P, Taylor DL, Pocock JM, Felderhoff-Mueser U, Mazarakis ND, SarrafC, Joashi U, Kozma M, Greenwood K, Edwards AD, Mehmet H (2000) Branched chain amino acids induce apoptosis in neural cells without mitochondrial membrane depolarization or cytochrome c release: implications for neurological impairment associated with maple syrup urine disease. Mol BioI Cell 11(5):1919-1932 Kaas JH (1997) Topographic maps are fundamental to sensory processing. Brain Res Bull 44(2):107-112 Kandel ER, Schwartz JH, Jessell TM (2000) Principles of neural science, ed 4. McGraw-Hill, New York Kaplan BJ, Sadock VA (2000) Kaplan & Sadock's comprehensive textbook of psychiatry, ed 7. Lippincott Williams & Wilkins, New York Kasabov N (1996a) Foundations of neural networks, fuzzy systems and knowledge engineering. MIT Press, Cambridge, MA Kasabov N (1996b) Adaptable connectionist production systems. Neurocomputing 13(2-4):95-117 Kasabov N (1996c) Learning fuzzy rules and approximate reasoning in fuzzy neural networks and hybrid systems. Fuzzy Sets and Systems 82(2):2-20 Kasabov N (1998) Evolving fuzzy neural networks - Algorithms, applications and biological motivation. In: Yamakawa T, Matsumoto G (eds) Methodologies for the conception, design and application of soft computing. World Scientific, pp 271-274 Kasabov N (2001) Evolving fuzzy neural networks for supervised/unsupervised online knowledge-based learning. IEEE Trans Syst Man Cybern Part BCybern 31(6):902-918 Kasabov N (2002a) Evolving connectionist systems. Methods and applications in bioinformatics, brain study and intelligent machines. Springer, London Kasabov N (2002b) Evolving connectionist systems for adaptive learning and knowledge discovery: methods, tools, applications. Proc. First International IEEE Symposium on Intelligent Systems, pp 24-28 Kasabov N, Dimitrov D (2002) A method for gene regulatory network modelling with the use of evolving connectionist systems. Proc. ICONIP'2002 - International Conference on Neuro-Information Processing, IEEE Press, Singapore Kasabov N, Song Q (2002) DENFIS: Dynamic, evolving neural-fuzzy inference systems and its application for time-series prediction. IEEE Trans on Fuzzy Systems 10(2):144-154 Kasabov N (2003) Evolving connectionist systems. Methods and applications in bioinformatics, brain study and intelligent machines. Springer, London
272
References
Kasabov N, Benuskova L (2004) Computational neurogenetics. Journal of Computational and Theoretical Nanoscience 1(1):47-61 Kasabov N, Chan ZSH, Jain V, Sidorov I, Dimitrov D (2004) Gene regulatory network discovery from time-series gene expression data - a computational intelligence approach. In: Pal NR, Kasabov N, Mudi RK et al. (eds) Neural Information Processing - II th International Conference, ICONIP 2004 - Lecture Notes in Computer Science, vol 3316. Springer, Calcutta, India, pp 1344-1353 Kasabov N, Benuskova L (2005) Theoretical and computational models for neuro, genetic, and neuro-genetic information processing. In: Rieth M, Schommers W (eds) Handbook of computational and theoretical nanotechnology, vol X. American Scientific, Los Angeles, CA Kasabov N (2006) Evolving connectionist systems: the knowledge engineering approach. Springer, London Kasabov N, Bakardjian H, Zhang D, Song Q, Cichocki A, van Leeuwen C (2006) Evolving connectionist systems for adaptive learning, classification and transition rule discovery from EEG data: A case study using auditory and visual stimuli. Intl J Neural Systems (in press) Katok A, Hasselblat B (1995) Introduction to the modem theory of dynamical systems. Cambridge Univ Press, Cambridge, MA Kaufman S (1999) A model of human phenylalanine metabolism in normal subjects and in phenylketonuric patients. Proc Natl Acad Sci USA 96(6):31603164 Kecman V (200 I) Learning and soft computing: support vector machines, neural networks, and fuzzy logic models (complex adaptive systems). MIT Press, Cambridge, MA KEGG pathway database (2006), Kyoto Encyclopedia of Genes and Genomes, http://www.genome.jp/kegg/pathway.html Khan J, Simon R, Bitti1'er M, Chen Y, Leighton S, Pohida T, Smith P, Jiang Y, Gooden G, Trent J, Meltzer P (1998) Gene expression profiling of alveolar rhabdomyosarcoma with eDNA microarrays. Cancer Res 58(22):5009-5013 Kharazia VN, Wenthold RJ, Weinberg RJ (1996) GluRI-immunopositive interneurons in rat neocortex. J Comp Neurol 368:399-412 Kiddie G, McLean D, vanOojen A, Graham B (2005) Biologically plausible models of neurite outgrowth. In: Pelt Jv, Kamermans M, Levelt CN et al. (eds) Development, dynamics and pathology of neuronal networks: from molecules to functional circuits. Progress in brain research, vol 147. Elsevier, New York, pp67-79 Kim JJ, Foy MR, Thompson RF (1996) Behavioral stress modifies hippocampal plasticity through N-methyl-D-aspartate receptor activation. Proc Nat! Acad Sci USA 93(10):4750-4753 Kimura A, Pavlides C (2000) Long-term potentiation/depotentiation are accompanied by complex changes in spontaneous unit activity in the hippocampus. Journal of Neurophysiology: 1894-1906 Kirkwood A, Rioult MC, Bear MF (1996) Experience-dependent modification of synaptic plasticity in visual cortex. Nature 381(6582):526-528
273 Kirkwood A, Rozas C, Kirkwood J, Perez F, Bear MF (1999) Modulation oflongterm synaptic depression in visual cortex by acetylcholine and norepinephrine. J Neurosci 19(5):1599-1609 Kitano T, Schwarz C, Nickel B, Paabo S (2003) Gene diversity patterns at 10 X. chromosomal loci in humans and chimpanzees. Mol Biol Evol 20(8): 12811289 Kleppe IC, Robinson HPC (1999) Determining the activation time course of synaptic AMPA receptors from openings of colocalized NMDA receptors. Biophys J 77:1418-1427 Klinke R, Kral A, Heid S, Tillein J, Hartmann R (1999) Recruitment of the auditory cortex by long-term cochlear electrostimulation. Science 285: 1729-1733 Knerr I, Zschocke J, Schellmoser S, Topf HG, Weigel C, Dotsch J, Rascher W (2005) An exceptional Albanian family with seven children presenting with dysmorphic features and mental retardation: maternal phenylketonuria. BMC Pediatr 5(1):5 Ko DC, Binkley J, Sidow A, Scott MP (2003) The integrity of a cholesterolbinding pocket in Niemann-Pick C2 protein is necessary to control lysosome cholesterol levels. Proc Nat! Acad Sci USA 100(5):2518-2525 Ko DC, Milenkovic L, Beier SM, Manuel H, Buchanan J, Scott MP (2005) Cellautonomous death of cerebellar purkinje neurons with autophagy in NiemannPick type C disease. PLoS Genet 1(1):81-95 Koch C, Poggio T (1983) A theoretical analysis of electrical properties of spines. Proc Roy Soc Lond B 218:455-477 Koch C, Crick F (1994) Some further ideas regarding the neuronal basis of awareness. In: Koch C, Davis JL (eds) Large-scale neuronal theories of the brain. MIT Press, Cambridge, MA, pp 93-111 Koch C (1996) Towards the neuronal substrate of visual consciousness. In: Hameroff SR, Kaszniak AW, Scott AC (eds) Towards a science of consciousness: the first Tucson discussions and debates. MIT Press, Cambridge, MA, pp 247258 Koch C, Hepp K (2006) Quantum mechanics in the brain. Nature 440:611-612 Koester HJ, Sakmann B (1998) Calcium dynamics in single spines during coincident pre- and postsynaptic activity depend on relative timing of backpropagating action potentials and subthreshold excitatory postsynaptic potentials. Proc Nat! Acad Sci USA 95(16):9596-960 I Koetter R (2003) Neuroscience databases. Kluwer Academic, Norwell, MA Kohonen T (1984) Self-organization and associative memory. Springer, Berlin Kohonen T (1990) The self-organizing map. Proc IEEE 78: 1464-1497 Kohonen T (1997) Self-organizing maps, ed 2. Springer, Heidelberg Konig P, Engel AK, Singer W (1996) Integrator or coincidence detector? The role of the cortical neuron revisited. Trends Neurosci 19:130-137 Koza J (1992) Genetic Programming. MIT Press, Cambridge, MA Kudela P, Franaszcuk PJ, Bergey GK (2003) Changing excitation and inhibition in simulated neural networks: effects on induced bursting behavior. Biol Cybernetics 88(4):276-285
274
References
Kurkova V (1991) Kolmogorov's theorem is relevant. Neural Computation 3:617622 Langley K, Marshall L, Bree MVD , Thomas H, Owen M, O'Donovan M, Thapar A (2004) Association of the dopamine d(4) receptor gene 7-repeat allele with neuropsychological test performance of children with ADHD. Am J Psychiatry 161(1):133-138 Leblois A, Boraud T, Meissner W, Bergman H (2006) Competition between feedback loops underlies normal and pathological dynamics in the basal ganglia. J Neurosci 26:3567-3583 LeCun Y, Denker JS, Solla SA (1990) Brain damage. In: Touretzky DS (ed) Advances in neural information processing systems. Morgan Kaufmann, San Francisco, CA, pp 598-605 Lee C, Bae K, Edery I (1998) The Drosophila CLOCK protein undergoes daily rhythms in abundance, phosphorylation, and interactions with the PER-TIM complex. Neuron 21:857-867 Lee KS, Schottler F, Oliver M, Lynch G (1980) Brief bursts of high-frequency stimulation produce two types of structural change in rat hippocampus. J NeurophysioI44(2):247-258 Lee PS , Shaw LB, Choe LH, Mehra A, Hatzimanikatis V, KH KHL (2003) Insights into the relation between mRNA and protein expression patterns: II. Experimental observations in Escherichia coli . Biotechnology and Bioengineering 84(7):834-841 Leonard S, Gault J, Hopkins J, Logel J, Vianzon R, Short M, Drebing C, Berger R, Venn D, Sirota P, Zerbe G, Olincy A, Ross RG , Adler LE, Freedman R (2002) Association of promoter variants in the a7 nicotinic Acetylcholine receptor subunit gene with an inhibitory deficit found in schizophrenia. Arch Gen Psychiatry 59:10085-11 096 Leslie C, Eskin E, Noble WS (2002) The spectrum kernel : a string kernel for SVM protein classification. Proc. Pacific Symposium on Biocomputing, vol. 7, pp 566-575 Leutgeb JK, Frey JU, Behnisch YT (2005) Single celI analysis of activitydependent cyclic AMP-responsive element-binding protein phosphorylation during long-lasting long-term potentiation in area CA I of mature rat hippocampal-organotypic cultures. Neuroscience 131:601-610 Levy WB, Steward 0 (1983) Temporal contiguity requirements for long-term associative potentiation/depression in the hippocampus. Neuroscience 8(4):791797 Liao D, Hessler NA, Malinow R (1995) Activation ofpostsynapticalIy silent synapse s during pairing-induced LTP in CA I region of hippocampal slice. Nature 375: 400-404 Libet B (1985) Unconscious cerebral initiative and the role of conscious will in voluntary action. Behavioral and Brain Sciences 8:529-566 Libet B (1999) Do we have free will? Journal of Con sciou sness Studies 6(8-9):4757 Linden DJ (1999) The return of the spike: postsynaptic APs and the induction of LTP and LTD. Neuron 22(4):661-666
275 Liss B, Roeper J (2004) Correlating function and gene expression of individual basal ganglia neurons. Trends Neurosci 27(8):475-481 Livingstone M, Hubel D (1988) Segregation of form, color, movement, and depth: anatomy, physiology, and perception. Science 240:740-749 Lledo P-M, Zhang X, Sudhof TC, Malenka RC, Nicoll RA (1998) Postsynaptic membrane fusion and long-term potentiation. Science 279:399-403 Llinas RR, Ribary U (1994) Perception as an oneiric-like state modulated by senses. In: Koch C, Davis JL (eds) Large-scale neuronal theories of the brain. MIT Press, Cambridge, MA, pp 111-125 Lodish H, Berk A, Zipursky SL, Matsudaira P, Baltimore D, Darnell J (2000) Molecular cell biology, ed 4th. W.H. Freeman & Co., New York Lu T, Pan Y, Kao S-Y, Li C, Kohane I, Chan J, Yankner BA (2004) Gene regulation and DNA damage in the ageing human brain. Nature 429(24 June):883891 Lytton WW, Contreras D, Destexhe A, Steriade M (1997) Dynamic interactions determine partial thalamic quiescence in a computer network model of spikeand-wave seizures. J Neurophysiol 77(4):1679-1696 Maass W, Bishop CM (eds) (1999) Pulsed neural networks, MIT Press, Cambridge, MA MacBeath G, Schreiber S (2000) Printing proteins as microarrays for highthroughput function determination. Science 289(5485): 1760-1763 Mackay TFC (2000) Aging in the post-genomic era: simple or complex? Genome Biology 1(4) Magee JC, Johnston D (1997) A synaptically controlled associative signal for Hebbian plasticity in hippocampal neurons. Science 275:209-213 Maletic-Savatic M, Malinow R, Svoboda K (1999) Rapid dendritic morphogenesis in CAl hippocampal dendrites induced by synaptic activity. Science 283: 1923-1927 Mamdani E (1997) Application of fuzzy logic to approximate reasoning using linguistic synthesis. IEEE Trans on Computers 26(12): 1182-1191 Manchanda R, MalIa A, Harricharran R, Cortese L, Takhar J (2003) EEG abnormalities and outcome in first-eposode psychosis. Can J Psychiatry 48(11):722726 Maquet P (2001) The role of sleep in learning and memory. Science 294:10481052 Marcus G (2004a) The birth of the mind: how a tiny number of genes creates the complexity of the human mind. Basic Books, New York Marcus GF, Fisher SE (2003) FOXP2 in focus: what can genes tell us about speech and language? Trends in Cognitive Science 7(6):257-262 Marcus GF (2004b) Before the word. Nature 431:745 Marie H, Morishita W, Yu X, Calakos N, Malenka RC (2005) Generation of silent synapses by acute in vivo expression ofCaMKIV and CREB. Neuron 45:741752 Marini C, Harkin LA, Wallace RH, Mulley JC, Scheffer IE, Berkovic SF (2003) Childhood absence epilepsy and febrile seizures: a family with a GABAA receptor mutation. Brain 126:230-240
276
References
Markram H, Lubke J, Frotscher M, Sakmann B (1997) Regulation of Synaptic Efficacy by Coincidence of Postsynaptic APs and EPSPs. Science 275(5297):213-215 Marnellos G, Mjolsness ED (2003) Gene network models and neural development. In: vanOoyen A (ed) Modeling neural development. MIT Press, Cambridge, MA, pp 27-48 Martino S, Marconi P, Tancini B, Dolcetta D, Angelis MGD, Montanucci P, Bregola G, Sandhoff K, Bordignon C, Emiliani C, Manservigi R, Orlacchio A (2005) A direct gene transfer strategy via brain internal capsule reverses the biochemical defect in Tay-Sachs disease. Hum Mol Genet 14(15):2113-2123 Massimini M, Ferrarelli F, Huber R, Esser SK, Singh H, Tononi G (2005) Breakdown of cortical effective connectivity during sleep. Science 309:2228-2232 Maviel T, Durkin TP, Menzaghi F, Bontempi B (2004) Sites of neocortical reorganization critical for remote spatial memory. Science 305(5680):96-99 Mayeux R, Kandel ER (1991) Disorders oflanguage: the aphasias. In: Kandel ER, Schwartz JH, Jessell TM (eds) Principles of neural science, ed 3. Appleton & Lange, Norwalk, pp 839-851 Mayford M, Kandel ER (1999) Genetic approaches to memory storage. Trends Genet 15(11):463-470 McAdams HH, Arkin A (1998) Simulation of prokaryotic genetic circuits. Ann Rev Biophys Biomol Struct 27:199-224 McGuinness MC, Lu JF, Zhang HP, Dong GX, Heinzer AK, Watkins PA, Powers J, Smith KD (2003) Role of ALDP (ABCDI) and mitochondria in X-linked adrenoleukodystrophy. Mol Cell Bio123(2):744-753 McIntosh H (1998) Autism is likely to be linked to several genes. The APA Monitor online 29(11):http://www.apa.org/monitor/nov98/gene.html McNaughton BL, Barnes CA, Andersen P (1981) Synaptic efficacy and EPSP summation in granule cells of rat fascia dentata studied in vitro. J NeurophysioI46(5):952-966 Mehra A, Lee KH, Hatzimanikatis V (2003) Insight into the relation between mRNA and protein expression patterns: 1. Theoretical considerations. Biotechnology and Bioengineering 84(7):822-833 Meisler MH, Kearney J, Ottman R, Escayg A (2001) Identification of epilepsy genes in humans and mouse. Annu Rev Genetics 35:567-588 Mekel-Bobrov N, Gilbert SL, Vallender EJ, Anderson JR, Hudson RR, Tishkoff SA, Lahn BT (2005) Ongoing adaptive evolution of ASPM, a brain size determinant in Homo sapiens Science 309(5741):1720-1722 Melzack R (1999) Phantom limb. In: Wilson RA, Keil, F. (ed) The MIT Encyclopedia of the Cognitive Sciences. MIT Press, Cambridge, MA, pp 636-638 Mendel JM (2001) Uncertain rule-based fuzzy logic systems: introduction and new directions. Prentice Hall, New York Mieda M, Willie JT, Hara J, Sinton CM, Sakurai T, Yanagisawa M (2004) Orexin peptides prevent cataplexy and improve wakefulness in an orexin neuronablated model of narcolepsy in mice. Proc Nat! Acad Sci USA 101(13):46494654
277 Miller KD, MacKay DJC (1994) The role of constraints in Hebbian learning. Neural Computation 6(1):98-124 Miller MW, Dow-Edwards DL (1988) Structural and metabolic alterations in rat cerebral cortex induced by prenatal exposure to ethanol. Brain Res 474:316 326 Miltner WHR, Braun C, Arnold M, Witte H, Taub E (1999) Coherence of gammaband EEG activity as a basis for associative learning . Nature 397:434-436 Mitchell MT, Keller R, Kedar-Cabelli S (1997) Explanation-based generalization: a unified view. Mach Learn 1(1):47-80 Mitra S, Hayashi Y (2000) Neuro-fuzzy rule generation: survey in soft computing framework. IEEE Trans on Neural Networks 11(3):748-768 Mjolsness E, Sharp DH, Reinitz J (1991) A connectionist model of development. J Theor Bioi 152:429-453 Mockett B, Coussens C, Abraham WC (2002) NMDA receptor-mediated metaplasticity during the induction of long-term depression by low-frequency stimulation. Eur J Neurosci 15(11):1819-1826 Mogilner A, Grossman JAI, Ribary U, Joliot M, Volkmann J, Rapaport D, Beasley RW (1993) Somatosensory cortical plasticity in adult humans revealed by magneto encephalography. Proc Natl Acad Sci USA 90:3593-3597 Morales J, Hiesinger PR, Schroeder AJ, Kume K, Verstreken P, jackson FR, Nelson DL, Hassan BA (2002) Drosophila fragile X protein , DFXR, regulates neuronal morphology and function in the brain . Neuron 34:961-972 NeuCom (2006), Neuro-Computing Decision Support Environment, http://www.aut.ac.nz/research/research_instituteslkedri/research_centres/centre _for_novel_methods_oCcomputationaUntelligence/neucom.htm Ouyang Y, Kantor D, Harris KM, Schuman EM, Kennedy MB (1997) Visualization of the distribution of autopho sporylated calcium/calmodulin-dependent protein kinase II after tetanic stimulation in the CAl area of the hippocampus. J Neurosci 17(14):5416-5427 Pandey SC (2004) The gene transcription factor cyclic AMP-responsive element binding protein: role in positive and negative affective states of alcohol addiction. Pharmacol Ther 104(1):47-58 Pang S (2004) Data approximation for Bayesian network modelling. Inti J Computers, Systems and Signals 5(2) :36-43 Pang S, Kasabov N (2004) Inductive vs transductive inference, global vs local models: SVM, tSVM, and SVMT for gene expression classification problems. Proc. IntI. Joint Conf. Neural Net., IJCNN, IEEE Press, Budapest Pase L, Voskoboinik I, Greenough M, Camakaris J (2004) Copper stimulates trafficking of a distinct pool of the Menkes copper ATPase (ATP7A) to the plasma membrane and diverts it into a rapid recycling pool. Biochem J 378(Pt 3):1031-1037 Pastor P, Goate AM (2004) Molecular genetics of Alzheimer's disease. Curr Psychiatry Rep 6(2):125-133 Paulsen 0, Sejnowski TJ (2000) Natural patterns of activity and long-term synaptic plasticity. Current Opinion in Neurobiology 10(2):172-179
278
References
Penrose R (1994) Shadows of the Mind : A Search for the Missing Science of Consciousness. Oxford Univ. Press, Oxford Pevzner PA (2000) Computational molecular biology: An algorithmic approach. MIT Press, Cambridge, MA Philpot BD, Sekhar AK, Shouval HZ, Bear MF (2001) Visual experience and deprivation bidirectionally modify the composition and function ofNMDA receptors in visual cortex. Neuron 29(1):157-169 Ping TY, Shimizu E, Dube G, Rampon C, Kerchner G, Zhuo M, Guosong L, Tsien 1 (1999) Genetic enhancement of learning and memory in mice. Nature 401:63-69 Pollard KS, Salama SR, Lambert N, Lambot M-A, Coppens S, Pedersen JS, Katzman S, King B, Onodera C, Siepel A, Kern AD, Dehay C, Igel H, Manuel Ares J, Vanderhaeghen P, Haussler D (2006) An RNA gene expressed during cortical development evolved rapidly in humans. Nature Advance Online Publication (doi: 10.1038/nature051 13):1-6 Pomeroy SL, Tamayo P, Gaasenbeek M, Sturla LM, et aJ. (2002) Prediction of central nervous system embryonal tumour outcome based on gene expression. Nature 415(6870) :426 Poser S, Storm DR (2001) Role of Ca2+-stimu1ated adenylyl cyclase in LTP and memory formation . Int J Devl Neurosci 19:387-394 Protege (2006), http://protege.stanford.edul Rajavel KS, Neufeld EF (2001) Nonsense-mediated decay of human HEXA mRNA. Mol Cell Bioi 21(16) :5512-5519 Ralser M, NonhoffU, Albrecht M, Lengauer T, Wanker EE, Lehrach H, Krobitsch S (2005) Ataxin-2 and huntingtin interact with endophilin-A complexes to function in plastin-associated pathways. Hum Mol Genet 14(9):2893-2909 Ramaswamy S, Tamayo P, Rifkin R, Mukherjee S, et a1. (2001) Multiclass cancer diagnosis using tumor gene expression signatures. PNAS 98(26): 15149 Ranum LP, Day JW (2004) Myotonic dystrophy: RNA pathogenesis comes into focus. Am J Hum Genet 74(5):793-804 Raymond CR, Thompson VL, Tate WP, Abraham WC (2000) Metabotropic glutamate receptors trigger homosynaptic protein synthesis to prolong long-term potentiation. J Neurosci 20(3) :969-976 Reggia lA, Ruppin E, Glanzman DL (eds) (1999) Disorders of brain, behavior, and cognition: the neurocomputational perspective. Progress in brain research, Springer, New York Reid A, Willshaw D (1999) Modeling prefrontal cortex delay cells: the role of dopamine in schizophrenia. In: Reggia JA, Ruppin E, Glanzman DL (eds) Disorders of brain, behavior, and cognition: the neurocomputational perspective. Progress in brain research, vol 121. Springer, New York, pp 351-373 Reinitz J, Mjolsness E, Sharp DH (1995) Model for cooperative control of positional information in Drosophila by Bicoid and maternal Hunchback. 1 Exp ZooI271 :47-56 Rema V, Ebner FF (1999) Effect of enriched environment rearing on impairments in cortical excitability and plasticity after prenatal alcohol exposure. J Neurosci 19(24):10993-10006
279 Rhawn J (1996) Neuropsychiatry, neuropsychology, and clinical neuroscience: emotion, evolution, cognition, language, memory, brain damage, and abnormal behavior, ed 2. Lippincott Williams & Wilkins, Baltimore Ribary U, Ionnides K, Singh KD, Hasson R, Bolton JPR, Lado F, Mogilner A, Llinas R (1991) Magnetic field tomography of coherent thalamocortical 40Hz oscillations in humans. Proc Nat! Acad Sci USA 88:11037-11401 Rick JT, Milgram NW (1996) Frequency dependence of long-term potentiation and depression in the dentate gyrus of the freely moving rat. Hippocampus 6:118-124 Rieke F, Warland D, Steveninck RRdRv, Bialek W (1996) Spikes - Exploring the neural code. MIT Press, Cambridge, MA Rizzo1atti G, Fadiga L, Gallese V, Fogassi L (1996) Premotor cortex and the recognition of motor actions. Cognitive Brain Research 3:131-141 Rizzolatti G, Arbib MA (1998) Language within our grasp. Trends Neurosci 21: 188-194 Roberts AC, Robbins TW, Weikrantz L (1998) The prefrontal cortex. Oxford Univ. Press, Oxford Roberts S, Dybowski R, Husmeier D (2005) Probabilistic modeling in bioinformatics and medical informatics. Springer, London Robins A (1996) Consolidation in neural networks and the sleeping brain. Connection Science 8(2):259-275 Robinson PA, Rennie CJ, Rowe DL (2002) Dynamics of large-scale brain activity in normal arousal states and epileptic seizures. Phys Rev E 65(4):19-24 Rodriguez E, George N, Lachaux J-P, Martinerie J, Renault B, Varela FJ (1999) Perception's shadow: long-range synchronization of human brain activity. Nature 397:434-436 Roelfsema PR, Engel AK, Konig P, Singer W (1997) Visuomotor integration is associated with zero time-lag synchronization among cortical areas. Nature 385:157-161 Rolls ET, Treves A (1998) Neural networks and brain function. Oxford University Press, New York Rong R, Tang X, Gutmann DH, Ye K (2004) Neurofibromatosis 2 (NF2) tumor suppressor merlin inhibits phosphatidylinositol 3-kinase through binding to PIKE-L. Proc Nat! Acad Sci USA 101(52):18200-18205 Ropers H-H, Hoeltzenbein M, Kalscheuer V, Yntema H, Hamel B, Fryns J-P, Chelly J, Partington M, Gecz J, Moraine C (2003) Nonsyndromic X-linked mental retardation: where are the missing mutations? Trends Genet 19(6):316320 Rosenblatt F (1962) Principles of neurodynamics. Spartan Books, New York Rumelhart DE, Hinton GE, Williams RJ (1986) Learning internal representations by error propagation. In: Rumelhart DE, McClelland JL (eds) Parallel distributed processing: explorations in the microstructure of cognition, vol 1. MIT Press / Bradford Books, Cambridge, MA, pp 318-363 Rummery GA, Niranjan M (1994) On-line Q-learning using connectionist systems, Cambridge University Engineering Department
280
References
Saad D (ed) (1999) On-line learning in neural networks, Cambridge Univ. Press, Cambridge, MA Sabag AD, Dagan 0, Avraham KB (2005) Connexins in hearing loss: a comprehensive overview. J Basic Clin Physiol Pharmacol16(2-3):101-116 Sachdev RS, Lu SM, Wiley RG, Ebner FF (1998) The role of the basal forebrain cholinergic projection in somatosensory cortical plasticity. J Neurophysiol 79:3216-3228 Sahraie A, Weiskrantz L, Barbour JL, Simmons A, Williams SCR, Brammer MJ (1997) Pattern of neuronal activity associated with conscious and unconscious processing of visual signals. Proc Nat! Acad Sci USA 94(9406-9411) Salonen V, Kallinen S, Lopez-Picon FR, Korpi ER, Holopainen IE, Uusi-Oukari M (2006) AMPA/kainate receptor-mediated up-regulation of GABAA receptor d subunit mRNA expression in cultured rat cerebellar granule cells is dependent on NMDA receptor activation. Brain Res Salzberg SL (1990) Learning with nested generalized exemplars. Kluwer Academic, Boston, MA Sander JW (2003) The incidence and prevalence of epilepsy. The National Society for Epilepsy, http://www.eepilepsy.org.uk/pages/articles/show_article.cfm?id=26 Sankar A, Manmone RJ (1993) Growing and pruning neural tree networks. IEEE TransComput 42(3):291-299 Savage-Rumbaugh S, Lewin R (1994) Kanzi: the ape at the brink of the human mind. John Wiley & Sons, New York Schaal S, Atkeson C (1998) Constructive incremental learning from only local information. Neural Computation 10:2047-2084 Schnapp BJ, Reese TS (1986) New developments in understanding rapid axonal transport. Trends Neurosci 9:155-162 Schratt GM, Tuebing F, Nigh EA, Kane CG, Sabatini ME, Kiebler M, Greenberg ME (2006) A brain-specific micro RNA regulates dendritic spine development. Nature 439:283-289 Schule B, Albalwi M, Northrop E, Francis DI, Rowell M, Slater HR, Gardner RJ, Francke U (2005) Molecular breakpoint cloning and gene expression studies of a novel translocation t(4;15)(q27;q11.2) associated with Prader-Willi syndrome. BMC Med Genet 6(6):18 Schulz S, Siemer H, Krug M, Hollt V (1999) Direct evidence for biphasic cAMP responsive element-binding protein phosporylation during long-term potentiation in the rat dentate gyrus in vivo. J Neurosci 19(13):5683-5692 Schwaller B, Tetko IV, Tandon P, Silveira DC, Vreugdenhil M, Henzi T, Potier M-C, Celio MR, Villa AEP (2004) Parvalbumin deficiency affects network properties resulting in increased susceptibility to epileptic seizures. Mol Cell Neurosci 25:650-663 Searle J (2002) Consciousness and Language. Cambridge Univ. Press, Cambridge, MA Sebastian CS (2005) Mental retardation. eMedicine, Inc., http://www.emedicine.com/med/topic3095.htm
281 Segan S (2005) Absence seizures. eMedicine http://wwwemedicinecom/NEURO/topic3htm Seri B, Garcia-Verdugo JM, McEwen BS, Alvarez-Buylla A (2001) Astrocytes give rise to new neurons in the adult mammalian hippocampus. J Neurosci 21(18):7153-7160 Shadlen MN, Newsome WT (1998) The variable discharge of cortical neurons: implications for connectivity, computation, and information coding. J Neurosci 18:3870-3896 Sheng M, Lee SH (2001) AMPA receptor trafficking and the control of synaptic transmission. Cell 105:825-828 Shi SH, Hayashi Y, Petralia RS, Zaman SH, Wenthold RJ, Svoboda K, Malinow R (1999) Rapid spine delivery and redistribution of AMPA receptors after synaptic NMDA receptor activation. Science 284: 1811-1816 Shouval HZ, Bear MF, Cooper LN (2002) A unified model of NMDA receptordependent bidirectional synaptic plasticity. Proc Natl Acad Sci USA 99(16): 10831-10836 Shouval HZ, Castellani GC, Blais BS, Yeung LC, Cooper LN (2002) Converging evidence for a simplified biophysical model of synaptic plasticity. BioI Cybernetics 87:383-391 Siegel JM (2001) The REM sleep-memory consolidation hypothesis. Science 294: 10581063 Silvanto J, Cowey A, Lavie N, Walsh V (2005) Striate cortex (VI) activity gates awareness of motion. Nature Neurosci 8:143-144 Sinclair DA, Guarente L (2006) Unlocking the secrets of longevity genes. Scientific American 294(3):48-57 Singer W (1994) Putative function of temporal correlations in neocortical processing. In: Koch C, Davis JL (eds) Large-scale neuronal theories of the brain. MIT Press, Cambridge, MA, pp 201-239 Singer W (1999a) Neuronal synchrony: a versatile code for the definition ofrelations? Neuron 24:49-65 Singer W (1999b) The observer in the brain. In: Riegler A, Peschl M, Stein Av (eds) Understanding representation in the cognitive sciences. Kluwer Academic/Plenum, New York Sjostrom PJ, Turrigiano GG, Nelson SB (2001) Rate, timing, and cooperativity jointly determine cortical synaptic plasticity. Neuron 32:1149-1164 Smith LS, Hamilton A (eds) (1998) Neuromorphic systems: engineering silicon from neurobiology. Progress in Neural Processing, World Scientific, London Smolen P, Baxter DA, Byrne JH (2000) Mathematical modeling of gene networks. Neuron 26:567-580 Smolen P, Hardin PE, Lo BS, Baxter DA, Byrne JH (2004) Simulation of Drosophila circadian oscillations, mutations, and light responses by a model with VRI, PDP-I, and CLK. Biophys J 86(May):2786-2802 Somogyi R, Fuhrman S, Wen X (2001) Genetic network inference in computational models and applications to large-scale gene expression data. In: Bower JM, Bolouri H (eds) Computational modeling of genetic and biochemical networks. MIT Press, Cambridge, MA, pp 119-157
282
References
Song Q, Kasabov N (2006) TWNFI - Transductive weighted neuro-fuzzy inference system and applications for personalised modelling. Neural Networks Song S, Miller KD, Abbott LF (2000) Competitive Hebbian learning through spike-timing-dependent synaptic plasticity. Nature Neuroscience 3:919-926 Song S, Abbott LF (2001) Cortical development and remapping through spike timing-dependent plasticity. Neuron 32(2):339-350 Soosairajah J, Maiti S, Wiggan 0, Sarmiere P, Moussi N, Sarcevic B, Sampath R, (2005) Interplay between components of a novel Bamburg JR, Bernard LIM kinase-slingshot phosphatase complex regulates cofilin. EMBO ] 24(3):473-486 Spacek J, Harris KM (1997) Three-dimensional organization of smooth endoplasmatic reticulum in hippocampal CA 1 dendrites and dendritic spines of the immature and mature rat. J Neurosci 17:190-203 Spivak G (2004) The many faces of Cockayne syndrome. Proc Nat! Acad Sci USA 101(43):15273-15274 Stefansson H, Sigurdsson E, Steinthorsdottir V, Bjomsdottir S, Sigmundsson T, Ghosh S, Brynjolfsson J, Gunnarsdottir S, Ivarsson 0 , Chou TT (2002) Neuregulin 1 and susceptibility to schizophrenia. Am J Hum Genet 71(4):877-
°
892 Steinlein OK (2004) Genetic mechanisms that underlie epilepsy. Nature Rev Neurosci 5:400-408 Stevenson RE, Procopio-Allen AM , Schroer RJ, Collins JS (2003) Genetic syndromes among individuals with mental retardation. Am J Med Genet 123A:29-32 (1997) mRNA localization in neurons: a multipurpose mechanism? Steward Neuron 18:9-12 Stickgold R, Hobson JA, Fosse R, Fosse M (2001) Sleep , learning, and dreams: off-line memory reprocessing. Science 294: 1052-1 057 Storjohann R, Marcus GF (2005) NeuroGene: integrated simulation of gene regulation, neural activity and neurodevelopment. Proc . IntI. Joint. Conf. Neural Net., IJCNN 2005 , IEEE , Montreal, Canada, pp 428-433 Straub RE, Jiang Y, MacLean CJ, Ma Y, Webb BT, Myakishev MV, Harris-Kerr C, Wormley B, Sadek H, Kadambi B, Cesare AJ, Gibberman A, Wang X, O'Neill FA, Walsh D, Kendler KS (2002) Genetic variation in the 6p22.3 gene DTNBPI, the human ortholog of the mouse dysbindin gene, is associated with schizophrenia. Am J Hum Genet 71(2):337-348 Street VA, Goldy JD, Golden AS , Tempel BL, Bird TD, Chance PF (2002) Mapping of Charcot-Marie-Tooth disease type 1C to chromosome 16p identifies a novel locus for demyelinating neuropathies. Am J Hum Genet 70(1) :244-250 Stuart GJ, Sakmann B (1994) Active propagation of somatic action potentials into neocortical pyramidal cell dendrites. Nature 367(6458):69-72 Sudhof TC (1995) The synaptic vesicle cycle: a cascade of protein-protein interactions . Nature 375:645-653 Sugai T, Kawamura M, Iritani S, Araki K, Makifuchi T, Imai C, Nakamura R, Kakita A, Takahashi H, Nawa H (2004) Prefrontal abnormality of schizophre-
°
283 nia revealed by DNA microarray: impact on glial and neurotrophic gene expression. Ann N Y Acad Sci 1025(Oct):84-91 Suri V, Lanjuin A, Rosbash M (1999) TIMELESS-dependent positive and negative autoregulation in the Drosophila circadian clock. The EMBO Journal 18:675-686 Takagi T, Sugeno M (1985) Fuzzy identification of systems and its applications to modeling and control. IEEE Trans on Systems, Man, and Cybernetics 15:116132 Taylor JG (1999) The race for consciousness. MIT Press, Cambridge, MA Terman D, Rubin JE, Yew AC, Wilson CJ (2002) Activity patterns in a model for the subthalamopallidal network of the basal ganglia. J Neurosci 22:2963-2976 Theiler J (1995) On the evidence for low-dimensional chaos in an epileptic electroencephalogram. Phys Lett A 196:335-341 Thivierge J-P, Marcus GF (2006) Computational developmental neuroscience: exploring the interactions between genetics and neural activity. Proc. IntI. Joint Conf. Neural Net., IJCNN 2006, IEEE, Vancouver, Canada, pp 438-443 Thorpe S, Fize D, Marlot C (1996) Speed of processing in the human visual system. Nature 381 :520-522 Thorpe SJ, Fabre-Thorpe M (2001) Seeking categories in the brain. Science 291 :260-262 Tikovic P, Voros M, Durackova D (2001) Implementation of a learning synapse and a neuron for pulse-coupled neural networks. Journal of Electrical Engineering 52(3-4):68-73 Tononi G, Edelman GM (1998) Consciousness and complexity. Science 282: 1846-1851 Towell GG, Shawlik JW, Noordewier M (1990) Refinement of approximate domain theories by knowledge-based neural networks. Proc. Proc. 8th Natl. Conf. AI, AAAI PresslMIT Press, Boston, MA, pp 861-866 Towell GG, Shavlik JW (1993) Extracting refined rules from knowledge-based neural networks. Mach Learn 13(1):71-101 Towell GG, Shawlik JW (1994) Knowledge based artificial neural networks. Artificial Intelligence 70(4): 119-166 Toyoizumi T, Pfister J-P, Aihara K, Gerstner W (2005) Generalized BienenstockCooper-Munro rule for spiking neurons that maximizes information transmission. Proc Nat! Acad Sci USA 102(14):5239-5244 Traub RD, Miles R, Wong RK (1987) Models of synchronized hippocampal bursts in the presence of inhibition. I. Single population events. J NeurophysioI58(4):739-751 Traub RD, Whittington MA, Stanford 1M, Jefferys JGR (1996) A mechanism for generation of long-range synchronous fast oscillations in the cortex. Nature 383:621-624 Tsien JZ (2000) Linking Hebb's coincidence-detection to memory formation. Current Opinion in Neurobiology 10(2):266-273 Tsuda I (2001) Toward an interpretation of dynamic neural activity in terms of chaotic dynamicical systems. Behav Brain Sci 24:793-847
284
References
Turrigiano GG, Nelson SB (2000) Hebb and homeostasis in neuronal plasticity. Curr Opin Neurobiol10:358-364 Utlsh A, Siemon HP (1990) Kohonen's self-organizing feature maps for exploratory data analysis. Proc. IntI. Neural Networks Conf., INNC'90, Kluwer Academic, Paris, pp 305-308 van Ooyen A (ed) (2003) Modeling neural development, MIT Press, Cambridge, MA VanRossum MCW, Bi GQ, Turriggiano GG (2000) Stable Hebbian learning from spike timing-dependent plasticity. The Journal of Neuroscience 20(23):88128821 Vapnik V (1998) Statistical learning theory. John Wiley & Sons, New York Veenstra-Vanderweele J, Christian SL, E. H. Cook J (2004) Autism as a paradigmatic complex genetic disorder. Annu Rev Genomics Hum Genet 5:379-405 Villa AEP, Asai Y, Tetko IV, Pardo B, Celio MR, Schwaller B (2005) Crosschannel coupling of neuronal activity in parvalbumin-deficient mice susceptible to epileptic seizures. Epilepsia 46(Suppl. 6):359 Vreugdenhil M, Jefferys JGR, Celio MR, Schwaller B (2003) Parvalbumindeficiency facilitates repetitive IPSCs and related inhibition-based gamma oscillations in the hippocampus. J Neurophysiol 89: 1414-1423 Wang H, Wagner JJ (1999) Priming-induced shift in synaptic plasticity in the rat hippocampus. J Neurophysiol 82:2024-2028 Wang H, Fu Y, Sun R, He S, Zeng R, Gao W (2006) An SVM scorer for more sensitive and reliable peptide identification via tandem mass spectrometry. Proc. Pacific Symposium on Biocomputing, vol. 11, pp 303-314 Wang JC, Hinrichs AL, Stock H, Budde J, Allen R, Bertelsen S, Kwon JM, Wu W, Dick DM, Rice J, Jones K, Nurnberger J, Tischfield J, Porjesz B, Edenberg HJ, Hesse1brock V, Crowe R, Schuckit M, Begleiter H, Reich T, Goate AM, Bierut LJ (2004) Evidence of common and specific genetic effects: association of the muscarinic acetylcholine receptor M2 (CHRM2) gene with alcohol dependence and major depressive syndrome. Hum Mol Genet 13(17): 1903-1911 Wang NJ, Liu D, Parokonny AS, Schanen NC (2004) High-resolution molecular characterization of 15q 11-q 13 rearrangements by array comparative genomic hybridization (array CGH) with detection of gene dosage. Am J Hum Genet 75(2):267-281 Watts JA, Morley M, Burdick JT, Fiori JL, Ewens WJ, Spielman RS, Cheung VG (2002) Gene expression phenotype in heterozygous carriers of ataxia telangiectasia. Am J Hum Genet 71(4):791-800 Watts M, Kasabov N (1998) Genetic algorithms for the design of fuzzy neural networks. In: Usui S, Omori T (eds) Proc. 5th IntI. Conf. Neural Inf. Processing, vol 2. lOS Press, Kitakyushu, pp 793-796 Weaver DC, Workman CT, Stormo GD (1999) Modeling regulatory networks with weight matrices. Proc. Pacific Symposium on Biocomputing, World Scientific, pp 112-123 Weiler IJ, Irwin SA, Klintsova AY, Spencer CM, Brazelton AD, Miyashiro K, Comery TA, Patel B, Eberwine J, Greenough WT (1997) Fragile X mental re-
285 tardation protein is translated near synapses in response to neurotransmitter activation. Proc Nat! Acad Sci USA 94:5395-5400 Weisstein EW (1999-2006) Delay differential equations. Wolfram Research, MathWorld A Wolfram Web Resource http://mathworld.wolfram.com/DelayDifferentialEquation.html Wendling F, Bartolomei F, Bellanger H, Chauvel P (2002) Epileptic fast activity can be explained by a model of impaired GABAergic dendritic inhibition. Eur J Neurosci 15:1499-1508 Werbos P (1990) Backpropagation through time: what it does and how to do it. Proc IEEE 87: 10-15 Wessels LFA, van Someren EP, Reinders MJT (2001) A comparison of genetic network models. Proc. Pacific Symposium on Biocomputing, World Scientific, Singapore, Hawai, pp 508-519 White JA, Banks MI, Pearce RA, Kopell NJ (2000) Networks of interneurons with fast and slow g-aminobutyric acid type A (GABAA) kinetics provide substrate for mixed gamma-theta rhythm. Proc Nat! Acad Sci USA 97(14):81288133 Whitehead DJ, Skusa A, Kennedy PJ (2004) Evaluating an evolutionary approach for reconstructing gene regulatory networks. In: Pollack J, Bedau MA, Husbands P et al. (eds) Proc. 9th International Conference on the Simulation and Synthesis of Living Systems (ALIFE IX), MIT Press, Cambridge, MA, pp 427-432 Willshaw D, Price D (2003) Models for topographic map formation. In: vanOoyen A (ed) Modeling neural development. MIT Press, Cambridge, MA, pp 213244 Wittenberg GM, Sullivan MR, Tsien JZ (2002) Synaptic reentry reinforcement based network model for long-term memory consolidation. Hippocampus 12:637-647 Wittenberg GM, Tsien JZ (2002) An emerging molecular and cellular framework for memory processing by the hippocampus. Trends Neurosci 25(10):501-505 Wittner L, Eross L, Czirjak S, Halasz P, Freund TF, Magloczky Z (2005) Surviving CAl pyramidal cells receive intact perisomatic inhibitory input in the human epileptic hippocampus. Brain 128:138-152 Wu FX, Zhang WJ, Kusalik AJ (2004) Modeling gene expression from microarray expression data with state-space equations. Proc. Pacific Symposium on Biocomputing, World Scientific, Singapore, pp 581-592 Wu G-Y, Deisseroth K, Tsien RW (2001) Activity-dependent CREB phosphorylation: convergence of a fast, sensitive calmodulin kinase pathway and a slow, less sensitive mitogen-activated protein kinase activity. Proc Nat! Acad Sci USA 98(5):2808-2813 Wu L, Wells D, Tay J, Mendis D, Abbott M-A, Barnitt A, Quinlan E, Heynen A, Fallon JR, Richter JD (1998) CPEB-mediated cytoplasmic polyadenylation and the regulation of the experience-dependent translation of a-CaMKII mRNA at synapses. Neuron 21:1129-1139
286
References
Yamakawa T, Kusanagi H, Uchino E, Miki T (1993) A new effective algorithm for neo fuzzy neuron model. Proc . Fifth IFSA World Congress, IFSA, Seoul, Korea, pp 1017-1020 Yang JJ, Liao PJ, Su CC, Li SY (2005) Expression patterns of connexin 29 (GJE1) in mouse and rat cochlea. Biochem Biophys Res Commun 338(2):723-728 Yao X (1993) Evolutionary artificial neural networks. Inti J Neural Systems 4(3) :203-222 Zadeh L (1979) A theory of approximate reasoning. In: Hayes J, Michie D, Mikulich LI (eds) Machine intelligence, vol 9. Halstead Press, New York, pp 149194 Zadeh LA (1965) Fuzzy sets. Information and Control 8:338-353 Zador A, Koch C, Brown T (1990) Biophysical model of a Hebbian synapse. Proc Nat! Acad Sci USA 87:6718-6722 Zhukareva V, Sundarraj S, Mann D, Sjogren M, Blenow K, Clark CM, McKeel DW, Goate A, Lippa CF, Vonsattel JP, Growdon JH, Trojanowski JQ, Lee VM (2003) Selective reduction of soluble tau proteins in sporadic and familial frontotemporal dementias: an international follow-up study. Acta Neuropathol (Berl) 105(5):469-476 Zoghbi HY (2003) Postnatal neurodevelopmental disorders: meeting at the synapse? Science 302:826-830 Zoghbi HY (2005) MeCP2 dysfunction in humans and mice. J Child Neurol 20(9):736-740 Zubenko GS, Maher BS, Hughes HB, Zubenko WN, Stiffler JS, Kaplan BB, Marazita ML (2003 ) Genome-wide linkage survey for genetic loci that influence the development of depres sive disorders in families with recurrent, earlyonset, major depression. Am J Med Genet B Neuropsychiatr Genet 123(1):118 Zubenko GS, Maher BS, Hughes HB, Zubenko WN, Stiffler JS, Marazita ML (2004) Genome-wide linkage survey for genetic loci that affect the risk of suicide attempts in families with recurrent, early-onset, major depression. Am J Med Genet B Neuropsychiatr Genet 129(1):47-54 Zucker RS (1999) Calcium- and activity-dependent synaptic plasticity. Current Opinion in Neurobiology 9(3):305 -313
Index
action planning, 43 adult cortex, 66 aging, 224 Alzheimer's disease, 224 AMPA receptor, 55, 105 ANN, 81, 253 anosognosia, 25 artificial neural network, 81, 253 auditory cortex, 63 awareness, 46, 49 Bayesian methods, 247 BCM theory, 57, 68, 184 BGO,234 bifurcation analysis, 146 binding, 38, 41 binocular deprivation, 63 binocular rivalry, 39 blindsight, 43 Boolean methods, 251 brain, 23, 53 brain cancer, 97 brain diseases, 205 brain-gene ontology, 9, 234 Broca's aphasia, 32 CaMKJI, 61, 178 cAMP-responsive transcription factor, 186 cerebral cortex , 23, 56 chaos, 50 chromosome, 128 classification, 81, 124 clustering, 96, 122,248,249
CNGM, I, 155, 163, 169, 171, 174, 177, 196,203,205 coding , 78 codon, 141 coherence activity, 42, 43, 46, 47 computational intelligence, 247 computational neurogenetic model ing, I, 155, 163, 169, 171, 174,177,196,203,205 Computer Tomography, 20 conceptual spaces, 47 conduction aphasia , 32 connectionist, 84, 128 connectionist constructivism, 89 connectionist selectivism, 89 consciousness, 46, 49 cortical column, 74 CREB ,186 CREB phosphorylation, 189 crossover, 129 CT,20 Darwin, 136,256 dendrite, 54 dendritic tree, 54 DENFIS , 107, 116, 151 developmental plasticity, 62 dimensionality, 85 distance, 248 DNA , 137 dopamine, 211 dynamic core, 46 Dynamic Evolving Neural-Fuzzy Inference Systems , 107, 116, 151 dynamic synaptic modification threshold, 69
288
Index
dynamic systems , 169 ECOS,107 EEG, 20, 99 EFuNN, 108, 152 electroen cephalography, 20 epilepsy, 206 evolution, 128 evolutionary computation, 88, 127, 165 evolutionary processes, 127 evolv ing, 1, 109 evolv ing connectionist system s, 107 Evolving Fuzzy Neural Network, 108,152 excitatory, 55, 60, 102,214 experience-dependent, 61, 64, 79 expli cit memory, 27 firing threshold , 103 fitness, 129 fMRI,22 functional MRI, 22 fuzzy, 97, 109, 119, 249 fuzzy logic, 251 fuzzy set, 25 1 fuzzy variable, 251 GABA , 211 gamma oscillations, 41 gene , 128, 137 gene control, 156 gene expression, 97,142,155,162, 169, 188, 195 gene profile, 142 gene/protein regulatory network , 147,165,250 gene s and disease , 237 genetic algorithms, 128 genetic disorders , 237 Gestalt ,37 glutamate, 56, 2 13 GPRN, 147, 165, 250 gradient descent, 90
Hebbian synaptic plasticity, 180 hemiparalysis, 25 homo synaptic LTP, 200 immediate-early genes, 191 implicit or nondeclarative memory, 28 inhibito ry, 55, 102 innate factors, 62 input-output function, 82 ion channels, 55 knowledge, 91 knowledge-based, 254 language, 29, 35 language gene, 34 learning , 56, 84, 86, 120, 177,247 learning and memo ry, 25 lifelong learning, 87, 127 long-term memory , 27,191,223 long-term synaptic depression, 57, 178 long-term synaptic potentiation, 57, 178, 186 LTD, 57,178 LTP, 57,1 78,1 86 Magnetic Resonance Imaging , 21 magnetoencephalography, 20 MEG ,20 memory , 58, 61,177 mental retardation, 218 mentalization, 45 metaplasticity, 183, 193 microarray data, 150 micro array matrix , 142 mirror neurons, 34, 45 MLP, 98 monocular deprivation, 63 morphogenesis, 156 morphological changes, 61 motor, 75 MRI, 21
289 MSA,173 Multilayer Perceptron, 98 multiple sequence alignment, 173 mutation, 129 NeuCom, 235 neural code, 74 neural development, 156 Neural Gas, 89 neural representation, 36 neurogenesis, 28 neuro-information processing, 53 neuron, 53 neurotransmitter, 54 NMDA receptor, 55, 59, 73, 105, 178,213 NMDAR, 105, 187 non-coding, 61 non-REM sleep, 48 noradrenaline, 211 normalization, 122 ocular dominance, 62 ontology, 233 optimization, 84, 128, 155, 165,256 orientation selectivity, 62 oscillations, 38, 44 Parkinson disease, 229 peA,85 percept, 42 PET,21 phantom limb, 65 phase, 77 population, 129 Positron Emission Tomography, 21 postsynaptic potential, 55 prediction, 81 prefrontal cortex, 45 prenatal ethanol, 68 Principal Component Analysis, 85 protein, 141 PSP, 55 qualia,48
rate code, 77 receptors, 55, 59 reflective consciousness, 41 REM sleep, 48 representation, 65 reverse correlation, 77 ribosome, 140 RNA,137 robustness, 147 schizophrenia, 212 second messengers, 56 selection, 129 Self Organizing Map, 93 self-reflection, 45 sensory activity, 67 sensory awareness, 38,43 serotonin, 211 short-term memory, 26, 108, 191 similarity, 107 Single-Photon Emission Computed Tomography, 21 SNN, 102 SOM,93 somatosensory cortex, 64 SPECT,21 spike, 54 Spike Response Model, 102 spike timing, 77 spike timing-dependent plasticity, 180 Spiking Neural Network, 102 spiking neuron, 198 spine, 54, 59 SRM,102 STDP, 180 stochastic models, 250 subconscious, 47 subcortical structures, 23 subjective experience, 48 Support Vector Machine, 90, 249 SVM, 90, 249 synaptic modification threshold, 181 synaptic plasticity, 56, 58, 177, 199 synaptic strength, 53
290
Index
synaptic weight, 53, 68, 82 synchronization, 38, 40, 77 Takagi-Sugeno, 116, 151 thalamocortical noise, 72 thinking, 34 topographic, 161 topography, 64 topological map, 95 Transcranial Magnetic Stimulation, 19 transcription, 139
transductive inference, 121 transition matrix, 164 translation, 139 unsupervised learning, 83 vesicles, 54, 60 visual areas, 38 Wernicke's aphasia, 31 whiskers, 66