Computational Neurogenetic Modeling (Topics in Biomedical Engineering. International Book Series)

(J

respec-

• Building a Takagi-Sugeno fuzzy inference engine dynamically The Takagi-Sugeno fuzzy inference engine is used in both on-line and off-line modes of DENFIS. The difference between them is that for forming a dynamic inference engine, only first-order Takagi-Sugeno fuzzy rules are employed in DENFIS on-line mode and both first-order TakagiSugeno fuzzy rules and expanded high-order Takagi-Sugeno fuzzy rules are used in DENFIS off-line modes. To build such a fuzzy inference engine, several fuzzy rules are dynamically chosen from the existing fuzzy rule set depending on the position of current input vector in the input space. • Dynamic creation and updating of fuzzy rules All fuzzy rules in the DENFIS on-line mode are created and updated during a 'one-pass' training process by applying the Evolving Clustering Method (ECM) and the Weighted Recursive Least Square Estimator with Forgetting Factors (WRLSE). • Local generalization Similar to EFuNNs, DENFIS model has local generalization to speed up the training procedure and to decrease the number of fuzzy rules in the system. • Fast training speed

118

5 Evolving Connectionist Systems (ECOS)

In the DENFIS on-line mode, the training is a 'one-pass' procedure and in the off-line modes, WLSE and small-scale MLPs are applied, which lead DENFIS to have the training speed for complex tasks faster than some common neural networks or hybrid systems such as multi-layer perceptron with backpropagation algorithm (MLP-BP) and Adaptive Neural-Fuzzy Inference System (ANFIS), both of which adopt global generalization. • Satisfactory accuracy Using DENFIS off-line modes, we can achieve a high accuracy especially in non-linear system identification and prediction. 5.4.1 Dynamic Takagi-Sugeno Fuzzy Inference Engine

The Takagi-Sugeno fuzzy inference engine (Takagi and Sugeno 1985) utilized in DENFIS is a dynamic inference model. In addition to dynamically creating and updating fuzzy rules in the DENFIS on-line mode, the major differences between such inference engine and the general Takagi-Sugeno fuzzy inference engine are described as follows: - First, depending on the position of the current input vector in the input space, different fuzzy rules are chosen from the fuzzy rule set, which has been estimated during the training procedure, for constructing an inference engine. If there are two input vectors very close to each other, especially in DENFIS off-line modes, two identical fuzzy inference engines are established and they may be exactly the same. In the on-line mode, however, although sometimes two inputs are exactly same, their corresponding inference engines are probably different. This is because these two inputs corne into the system from the data stream at different moments and the fuzzy rules probably have been updated during this interval. - Second, also depending on the position of current input vector in the input space, the antecedents of fuzzy rules, which have been chosen from the fuzzy rule set for forming an inference engine, may be different. An example is illustrated in Fig. 5.4, where two fuzzy rule groups, FG I and FG2, are estimated depending on two input vectors XI and X2 respectively in a 2-D input space. We can know from this example that, for instance, the region C represents a linguistic meaning 'large' in FG I on the XI axis but it represents a linguistic meaning 'small' on that in FG2 • Also, the region C is presents as different membership functions respectively in FG r and FG2 .

5.4 DENFIS

119

a

H

C A

I

I

I

I

I

D
I

I

A

H

I

C

b

E C D

I

I

I

I

I

~ I

I

I

I

C E D

Fig. 5.4. Two fuzzy rule groups corresponding with input X I and.e, in a 2-D space

5.4.2 Fuzzy Rule Set, Rule Insertion and Rule Extraction

Fuzzy rules in a DENFIS are created during a training procedure, or come from rule insertion. In the on-line mode, the fuzzy rules in the rule set can also be updated as new training data appear in the system (Kasabov and Song 2002). As the DENFIS uses a Takagi-Sugeno fuzzy inference engine the fuzzy rules inserted to or extracted from the system are Takagi-Sugeno type fuzzy rules. These rules can be inserted into the rule set before or during the training procedure and they can also be exacted from the rule set during or after the training procedure.

120

5 Evolving Connectionist Systems (ECOS)

The inserted fuzzy rules can be the rules that are extracted from a fuzzy rule set created in previous training of DENFIS, or they can also be general Takagi-Sugeno type fuzzy rules. In the latter, the corresponding nodes of the general Takagi-Sugeno fuzzy rules have to be found and located in the input space. For an on-line learning mode, their corresponding radiuses should also be defined. The region can be obtained from the antecedent of a fuzzy rule and the centre of this region is taken as the node corresponding with the fuzzy rule. A value of (0.5 ~ I )Dthr can be taken as the corresponding radius.

5.5 Transductive Reasoning for Personalized Modeling Most of learning models and systems in artificial intelligence developed and implemented so far are based on inductive methods, where a model (a function) is derived from data representing the problem space and this model is further applied on new data. The model is usually created without taking into account any information about a particular new data vector (test data). An error is measured to estimate how well the new data fits into the model. The inductive learning and inference approach is useful when a global model ("the big picture") of the problem is needed even in its very approximate form. In contrast to the inductive learning and inference methods, transductive inference methods estimate the value of a potential model (function) only in a single point of the space (the new data vector) utilizing additional information related to this point (Vapnik 1998). This approach seems to be more appropriate for clinical and medical applications of learning systems, where the focus is not on the model, but on the individual patient. Each individual data vector (e.g.: a patient in the medical area; a future time moment for predicting a time series; or a target day for predicting a stock index) may need an individual, local model that best fits the new data, rather then a global model, in which the new data is matched without taking into account any specific information about this data. An individual model AI; is trained for every new input vector Xi with data use of samples D, selected from a data set D, and data samples DO,i generated from an existing model (formula) M (if such a model is existing). Data samples in both D, and DO,i are similar to the new vector Xi according to defined similarity criteria.

5.5 Transductive Reasoning for Personalized Modeling

121

Data D, selected from D in the vicinity of the input vector Xi

M, the

Outputy,

Data Do,! selected from D in the vicinity of the input vectorXi

Fig. 5.5. A block diagram of a transductive reasoning system

Transductive inference is concerned with the estimation of a function in single point of the space only. For every new input vector Xi that needs to be processed for a prognostic task, the N, nearest neighbors, which form a sub-data set D, are derived from an existing data set D. If necessary, some similar vectors to vector Xi and their outputs can also be generated from an existing model M. A new model M, is dynamically created from these samples to approximate the function in the point Xi - Fig. 5.6. The system is then used to calculate the output value Yi for this input vector Xi .

0

0

0 0

0

0

L>.

0

•

XI

0

0

0 0

L>.

L>.

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 000<;:;>00000

€>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 <, "

Solution space

.

........... Goal solution

Fig. 6.1. A block diagram of a genetic algorithm - evolution of individual models

in time is presentedas a trace in space where severalgenerations of populations of individuals are traced

While Fig. 6.1 shows graphically how a GA searches for the best solution in the solution space, the outline of GA is given as: 1. Generate initial population of individuals - each individual defined as a chromosome containing parameters - genes. 2. Evaluate the fitness of each individual using a fitness function. 3. Select a subset of individuals based on their fitness. 4. Apply a crossover procedure on the selected individuals to create a new generation of a population. 5. Apply mutation.

130

6 Evolutionary Computation for Model and Feature Optimization

6. Continue with the previous procedure 2-6 until a desired solution (with a desired fitness) is obtained, or the run time is over. When using the GA method for a complex multi-optional optimization problem, there is no need for in-depth problem knowledge, neither a need for many data examples stored beforehand. What is needed here is merely a "fitness" or "goodness" criterion for the selection of the most promising individuals (they are partial solutions to the problem). This criterion may require a mutation as well, which is a heuristic approach of a "trial-error" type. This implies keeping (recording) the best solutions at each of the stages. A class of simple genetic algorithms introduced by John Holland, is characterized by: • Simple, binary genes. The genes take values of 0 and I only. • Simple, fixed single-point crossover operation. The crossover operation is done by choosing a point where a chromosome is divided into two parts swapped with the two parts taken from another individual. • Fixed-length encoding. The chromosomes had fixed length of 6 genes.

Many complex optimization problems find their way to a solution through genetic algorithms. Such problems are, for example, the Traveling Salesman Problem (TSP) - finding the cheapest way to visit n towns without visiting a town twice; the Min Cut problem - cutting a graph with minimum links between the cut parts; adaptive control; applied physics problems; optimization of the parameters of complex computational models; optimization of neural network architectures; finding fuzzy rules and membership functions (Furuhashi et al. 1994), etc. The main issues in using genetic algorithms relate to the choice of genetic operations (crossover, selection, mutation). In the case of The Traveling Salesman the crossover operation can be merging different parts of two possible roads ('mother' and 'father' road) until new usable roads are obtained. The criterion for the choice of the most prospective ones is minimum length (or cost). Genetic algorithms comprise a great deal of parallelism. Thus, each of the branches of the search tree for best individuals can be utilized in parallel with the others. This allows for an easy realization of the genetic algorithms on parallel architectures. Genetic algorithms are search heuristics for the "best" instance in the space of all possible instances. The following issues are important for any genetic algorithm:

6.3 Genetic Algorithms

131

• Encoding scheme. How to encode the problem in terms of genetic algorithms, what variables to choose as genes, how to construct the chromosomes, etc. • Population size. How many possible solutions should be kept for further development. • Crossover operations. How to combine old individuals and produce new, more prospective ones. • Mutation heuristic. When and how to apply mutation.

In short, the major characteristics of the genetic algorithms are the following: • They are heuristic methods for search and optimization. As opposed to the exhaustive search algorithms, the genetic algorithms do not produce all variants in order to select the best one. Therefore, they may not lead to the perfect solution but to one which is closest to it taking into account the time limits. But nature itself is imperfect too, (partly due to the fact that the criteria for perfection keep changing), and what seems to be close to perfection according to one "goodness" criterion, may be far from it according to another. • They are adaptable, which means that they have the ability to learn; to accumulate facts and knowledge without having any previous knowledge. They begin only with a "fitness" criterion for selecting and storing individuals (partial solutions), which are "good", and dismissing those which are "not good". Genetic algorithms can be incorporated in learning modules as part of an expert system or of other information processing systems. Other EC techniques are: • Evolutionary strategies. These techniques use only one chromosome and a mutation operation, along with a fitness criterion, to navigate in the solution (chromosomal) space. • Evolutionary programming. These are EC techniques applied to the automated creation and optimization of sequence of commands (operators) that constitute a program (or an algorithm) to solve a given problem (Koza 1992). The theory of GA and the other EC techniques includes different methods for selection of individuals from a population, different cross over techniques, different mutation techniques. Selection is based on fitness. A common approach is proportional fitness, i.e. 'if I am twice as fit as you, I have twice the probability of being

132

6 Evolutionary Computation for Model and Feature Optimization

selected'. Roulette wheel selection gives chances to individuals according to their fitness evaluation (see Table 6.1). Table 6.1. Selection of individuals from a population based on fitness: each individual is assigned an evaluated fitness number and then either the top few are selected for crossover or every one is given a probability (chance) - the roulette strategy Individual

2

3

4

5

6

7

8

9

10

1.8

1.6

1.4

1.2

1.0

0.8

0.6

0.4

0.2

Fitness

2.0

Selection

0.180.160.150.130.110.090.070.060.030.02

Other selection techniques include tournament selection (every time of selection, the roulette wheel is turned twice, and the individual with the highest fitness is selected), rank ordering, and so on (Fogel et al. 1990). Important feature of the selection procedure is that fitter individuals are more likely to be selected. The selection procedure can involve also keeping the best individuals from the previous generation (if this principle was used by the Nature, Michelangelo would have been alive nowadays, as he is one of the greatest artists ever, with the best genes in this respect). This operation is called elitism. After the best individuals are selected from a population, a cross over operation is applied between these individuals. Different cross-over operations can be used: • One-point cross-over; • Three-point cross over or more (Fig. 6.2), etc.

-

Parents

...

Offspring

Fig. 6.2. Three point cross over operation, where the two individual cross over through exchanging their genes in 4 sections, base on usually randomly selected 3 points

Mutation can be performed in the following ways:

6.4 EC for Model and Parameter Optimization

133

• For a binary string, just randomly 'flip' a bit. • For a more complex structure, randomly select a site, delete the structure associated with this site, and randomly create a new sub-structure. Some EC methods just use mutation (no crossover, e.g. evolutionary strategies). Normally, however, mutation is used to search in a "local search space", by allowing small changes in the genotype (and therefore hopefully in the phenotype).

6.4 EC for Model and Parameter Optimization EC has been widely used in bioinformatics for optimization purposes (see (de Jong 2002, Fogel and Corne 2003). Here, a general procedure for a GA optimization of a feature set and a model is presented and illustrated in Figs.6.3 and 6.4. 1. Starting with an initial feature set Gm and a model M, create a population of K chromosomes (models), each having different feature subsets from Gm and slightly different parameter values from the parameter values of the model M. The chromosome contains a binary part, where a feature is present (1) or not present (0) in a model, and a part of continuous values - the parameters of the model. If, for example, the model is ECF ECOS (see Chap. 5), the parameters are: R max , R min, Number of fuzzy membership functions, Number of iterations of training the ECF model. 2. FOR J= 1 to P generations DO a. Select randomly from the data set S a subset Stst for testing and the rest Str for training. b. Train all K models on Str and test them on Stst. c. Select the best models (e.g. maximum accuracy). d. Apply cross validation and mutation to the chromosomes to create the next generation ofK new models. END (FOR) 3. Select the final best model (the model with the best accuracy) that has an optimized feature set and optimized model parameter values. 6.4.1 Example Fig. 6.3 shows a GA optimization of the parameters and the set of input variables (features) of an ECOS model for classification of samples into

134

6 Evolutionary Computation for Model and Feature Optimization

two classes - class of control and a class of cancer for CNS data (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, 12, and variables 2, 5, 6, 8 and 10 are not used. Optimal values of the ECF parameters (R max , R min , m of n, epochs) are shown in the figure. ~L::::

) SUrw.re Genetic Altoflthm For Ofillne (CF OplimlWtion

..:J

S"'llio no ICHSl2va-

r "'- ,in

'1:01_""._ rro-

,...-----E_ -

-

100

----,

I': ~ reldcraet_

,-

(; Iotnflllldcrsetas

, -

ro E_ a.. ..

,-

90 ~

femdnOlsetat

,-

-=-

i ~

J;> 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........ Le<¢

28

'.-010.....

N~ofF&atu'e1

27

2

4

6

8

10 12

I 25

1~5

. - - - - - Staa NoNetwot t.o.ded

Scang

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 (Pomeroyet 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, 12,(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. See Color Plate 6

After the optimal set of features and model parameters is selected (optimized) using GA - see Fig. 6.3, now the final model is created using the optimal parameters and only 7 features out of 12 - see Fig. 6.4. Fig. 6.5 presents the clusters (profiles) of the expression of the 7 genes for class 1

6.4 EC for Model and Parameter Optimization

135

(that is patients not responding to treatment) and Fig. 6.6 for class 2 (eNS cancer surviving children).

.......

} .lflW... l(f

......

r:.~

- !XJ

..

....... I .....

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. See Color Plate 7

) ct.. ,

.

10

Gono_

.:..

,r:-.--.. .--,-II J

Fig. 6.5. Individual cluster profiles for class I (not responding to treatment) obtained using 7 genes selected through GA optimization as shown in Fig. 6.3. See Color Plate 7

136

6 Evolutionary Computation for Modeland Feature Optimization

.

0-_

..:.J

r

I

Fig. 6.6. Individual cluster profiles for class 2 (cancer survivors) obtainedusing 7 genes selectedthroughGA optimization as shownin Fig. 6.3. See Color Plate 8

6.5 Summary This chapter introduces main principles of EC and GA and shows how these methods can be applied to optimize parameters of models and input features. Parameter estimation is a very difficult task in inferring GRN models, mainly because of the lack of observation data relative to the number of genes involved. In this respect, evolutionary computation (EC) that is a robust, global optimization method becomes an important tool to accurately infer and optimize GRNs. EC methods are inspired by the Darwin theory of evolution and have been used for parameter estimation or optimization in many engineering applications. Unlike classical derivativebased (like Newton) optimization methods, EC is more robust against noise and high dimensionality in the search space. In addition, EC does not require the derivative information of the objective function and is thus applicable to complex, black box problems. These characteristics make EC highly suitable for identifying parameters ofGRN, in particular because derivative information of the underlying model is usually not available and data is scarce and noisy, thus requiring a robust, global optimization algorithm that is not easily misled by these drawbacks. And finally, qualitative inference of parameters is difficult with small number of observations compared to the large number of genes involved. For these reasons, GA and EC are used in Chap. 7 and Chap. 8 for optimizing GRNs and CNGMs.

7 Gene/Protein Interactions - Modeling Gene Regulatory Networks (GRN)

This chapter presents background knowledge from Bioinfonnatics on gene and protein information processing in a biological cell with the emphasis on their dynamic interaction. In a cell and in a neuron in particular, DNA, RNA and proteins interact continuously, affecting the functioning of the whole cell and the phenotype of an organism. Here, the main principles of these interactions are presented. Some of these interactions are subject to modeling in a CNGM when related to the output signals/functions of a cell/neuron.

7.1 The Central Dogma of Molecular Biology With the completion of the first draft of the human genome and the genomes of some other species the task is now to be able to process this vast amount of ever growing dynamic information and to create intelligent systems for prediction and knowledge discoveries at different levels of life, from cell to whole organisms and species. The DNA (Deoxyribonucleic Acid) is a chemical chain, present in the nucleus of each cell of an organism, and it consists of ordered in a double helix pairs of small chemical molecules (bases) which are: Adenine (A), Cytosine (C), Guanine (G), and Thymine (T), linked together by deoxyribose sugar phosphate nucleic acid backbone. The RNA (ribonucleic acid) has a similar structure as the DNA, but here Thymine (T) is substituted by Uracil (U) (see Fig. 7.1). The central dogma ofthe molecular biology (see Fig. 1.2) states that the DNA is transcribed into RNA, which is translated into proteins. This process is shown in a bit more details in Fig. 7.2. The DNA contains millions of base pairs, but only 5% or so is used for the production of proteins, and these are the segments from the DNA that contain genes. Each gene is a sequence of base pairs that is used in the cell to produce proteins. Genes have length of hundreds to thousands of bases.

138

7 GenelProtein Interactions - Modeling Gene Regulatory Networks (GRN)

RNA

DNA

Ribonucleic acid

Deoxyribonucleicacid

Fig. 7.1. Structure of DNA and RNA

I

-

I

I

protein

mRNA

DNA \

•••••.•••••

tr¥7scription

/

/fanslation..

(

'_%i_~_

I I I

I I

" \

I \

Final destination

\

,~

Nucleus

'. -..

• ..-

......

)

-~( (

.

\

....../

Goigi Complex

Ribosome Endoplasmic Reticulum

Fig. 7.2. The central dogma of molecular biology - a schematic representation

Each gene consists of two types of segments - exons, that are segments translated into proteins, and introns - segments that are considered redundant and do not take part in the protein production. Removing the introns and ordering only the exon parts of the genes in a sequence is called splic-

7.1 The Central Dogma of Molecular Biology

139

ing and this process results in the production of a messenger RNA (or mRNA) sequences. mRNAs are directly translated into proteins. Each protein consists of a sequence of amino-acids, each of them defined by a RNA base triplet, called a codon. From one DNA sequence there are many copies of mRNA produced, the presence of certain gene in all of them defines the level of the gene expression in the cell and can indicate what and how much of the corresponding protein will be produced in the cell. The above description of the central dogma of the molecular biology is very much a simplified, but that would help to understand the rationale behind using connectionist and other information models in bioinformatics (Brown, Shreiber et al. 2000). Genes are complex chemical structures and they cause dynamic transformation of one substance into another during the whole life of an individual, as well as the life of the human population over many generations. When genes are "in action", the dynamics of the processes in which a single gene is involved are very complex, as this gene interacts with many other genes, proteins, and is influenced by many environmental and developmental factors. Modeling these interactions, learning about them and extracting knowledge, is a major goal for Bioinformatics. Bioinformatics is concerned with the application of the methods of information sciences for the analysis, modeling and knowledge discovery of biological processes in living organisms (Brown, Grundy et al. 2000, Baldi and Brunak 2001). The whole process of DNA transcription, gene translation, and protein production is continuous and it evolves over time. Proteins have 3D structures that unfold over time governed by physical and chemical laws. Proteins make some genes to express and may suppress the expression of other genes. The genes in an individual may mutate, change slightly their code, and may therefore express differently at a next time. So, genes may change, mutate, and evolve in a life time of a living organism. Modeling these processes is extremely complex task. The more new information is made available about DNA, gene expression, protein creation, metabolism, the more accurate the information models will become. They should adapt to the new information in a continuous way. The process of biological knowledge discovery is also evolving in terms of data and information being created continuously. Through evolutionary processes (evolution) genes are slowly modified through many generations of populations of individuals and selection processes (e.g. natural selection). Proteins provide the majority of the structural and functional components of a cell. The area of molecular biology that deals with all aspects of

140

7 GenelProtein Interactions - Modeling Gene Regulatory Networks (GRN)

proteins is called proteomics. So, far about 30,000 proteins have been identified and labeled, but this is considered to be a small part of the total set of proteins that keep our cells alive. The mRNA is translated by ribosomes into proteins. A protein is a sequence of amino-acids, each of them defined by a group of 3 nucleotides (codons). There are 20 amino acids all together, denoted by letters (A, CH, I, K-N, P-T, V, W, Y). Chemical formulae of each of the amino acids are shown in Fig. 7.3. Amino acids with hydropho bic side groups COO-

COO'

COO'

COO'

+H"N-b-H

+H"N-b-H

+H"N-b-H

I

I

I

I

C'" I

CH H"C/' 'CH"

+H"N-b-H

H-C-CH"

y'" yH"

I CH"

CH H"c' 'CH"

I CH"

I

CH"

Leucine

Isoleucine

(val)

(leu)

(lle)

I

e5

S

Valine

COO' +H"N-b-H

Methionine (met)

Phenylalanine (phe)

Amino acids with hydrophilic side groups COO-

COO' +H"N-t-H

+H"N-b-H

t",

(asn)

coo-

COO' +H"N-b-H

c'"

C'" I

b-tf

c'"

yH"

coo",N

Glu1amic acid (glu)

)",

I

C'" I

C'" I

CH" I

c'" I

c'" I

" >H

HC-N H

"'0

Glu1amine (gin)

COO+H"N-b-H

I

I

I

C'" I

H"N' "0 Asparagine

+H"N-b-H

I

I

C'"

COO' +H"N-b-H

COO-

y'"

c'" I

NH

b=NH,,+

NH,,+

Histidine (his)

I

N",

Lysine (lys)

+H"N-b-H

I

Arginine (arg)

C'" I

COO'

Aspartic acid (asp)

Amino acids that are in betwe en COO'

COO-

+H"N-6-H

+H"N-b-H

I

I

H-C-OH

H

Glycine (gly)

Alanine (ala)

~

I

y'" SH

Cysteine (cys)

CH"

'I'hIe onine (thr)

(ser) COO'

COO+H"N-b-H

H-t-OH

I

Serine

+"'N--b-H

I

COO'

COO'

+H"N-t-H

+H"N-J-H

COO+H"N-t-H

I

"'C'c/C'"

'"

Proline (pro)

Fig. 7.3. The chemical structure of the aminoacids

I

Q OH

Tyrosine (tyr)

I

c'"

J=:D H

Tryptophan (trp)

7.2 Gene and Protein Expression DataAnalysis and Modeling

141

The length of a protein in number of amino-acids is from tens to several thousands. Each protein is characterized by some characteristics, for example (Brown and Botstein 1999): • • • • • •

Structure Function Charge Acidity Hydrophilicity Molecular weight

An initiation codon, that is a particular triplet of bases, defines the start position of a gene in an mRNA where the translation of the mRNA into protein begins. A stop codon defines the end position. Proteins with a high similarity are called homologous. Homologues that have identical functions are called orthologues. Similar proteins that have different functions are called paralogues. Proteins have complex structures that include: • Primary structure (a linear sequence of the amino-acids); • Secondary structure (3D, defining functionality); • Tertiary structure (high level folding and energy minimization packing of the protein). Polypeptides with a tertiary level of structure are usually referred to as globular proteins, since their shape is irregular and globular in form; • Quaternary structure (interaction between two or more protein molecules).

7.2 Gene and Protein Expression Data Analysis and Modeling The recent advent of cDNA microarray and gene-chip technologies means that it is now possible to simultaneously interrogate thousands of genes. The potential applications of this technology are numerous and include identifying markers for classification, diagnosis, disease outcome prediction, therapeutic responsiveness, and target identification. Microarray analysis might not identify unique markers (e.g. a single gene) of clinical utility for a disease because of the heterogeneity of the disease. Prediction of the biological state/disease is likely to be more accurate by identifying clusters of gene expressions (gene expression profiles).

142

7 Gene/Protein Interactions- Modeling Gene RegulatoryNetworks (GRN)

Each point (pixel, cell) in a microarray matrix represents the level of expression of a single gene. Five principal steps in the microarray technologyare shown in Fig. 7.4. They are: tissue collection; RNA extraction; microarray gene expression evaluation; scanning and image processing; data analysis. Tissuesamples

a

RNA

Microarray ........

Data analysis

~ ~§~ ........... e••

:1:' -, I~ ,~ ~

C . ... ..

•••• ••••••• ••••••• .. ••••"

. ;' .

f: ~. :

.:..

, . ~'.-

.

RNA extractionand Reverse transcription Scanning and Normalization RNA amplification tocDNA and labeling imageprocessing and preprocessing

b

.,

.

I'

I

~ \

I

II H

, I

I:

' :. ;

,,

" III

"

..

I

,

, "

•ItI

1 $ I

I '

"

.I·

.1

f ~ d

II.

!'

Fig. 7.4. (a) Steps in microarray gene expression data collection, analysis and modeling; (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. See Color Plate 8

One of the contemporary directions while searching for efficient drugs for many terminal illnesses, such as cancer or HIV, is the creation of gene profiles of these diseases and subsequently finding targets for treatment through gene expression regulation. A gene profile is a pattern of expres-

7.2 Gene and ProteinExpression Data Analysis and Modeling

143

sion of a number of genes that is typical for all, or for some of the known samples of a particular disease. Having such profiles for a particular disease makes it possible to set an early diagnostic test, so a sample can be taken from a patient, the data related to the sample processed, and a profile related to the sample - obtained. This profile can be matched against existing gene profiles. Based on similarity, it can be predicted with certain probability if the patient is in an early phase of a disease or he/she is at risk of developing the disease in the future with certain probability. A usual way of processing gene (and also protein) expression data S of genes G consists of the following steps: 1. FOR i=l to N (in particular, it could be leave-one-out method, where N is the number of available samples) DO a. Select a data sub-set Sv,; from S for validation. b. For the rest of the data So, select the most discriminative subset of genes G;; Card(G;)«Card(G). c. Create a model AI; based on the data So and the gene set G; as input variables. d. Validate AI; On the test set Sv,I and calculate the error E;. END (FOR) 2. Calculate the average error from E;. 3. Define a set Gm of the most frequently selected genes in all N iterations. 4. Create a final model M based on the whole data S and the set of genes Gm • After the above steps, a model M is created using the most frequently appearing genes in all N iterations - a gene set Gm • The gene set Gm though may not be optimal in terms of representing a minimum set of genes that through their interaction "cover" all clusters of the problem space. The parameters of the model M derived above were not optimized and may not be optimal either. A GA optimization algorithm for the optimization of both the gene set and the model parameter values as an additional procedure to the procedure above is given in Chap. 6. 7.2.1 Example An example is shown in Fig. 7.5, where class profiles of 14 types of cancer are extracted from a trained EFuNN (see Chap. 5) on previously extracted 399 gene expression variables and 14 output classes using publicly available data from (Ramaswamy et al. 2001) http://wwwgenome.wi.mit.edu/MPRlGCM.html. The profiles of each class can be

144

7 Gene/Protein Interactions- Modeling Gene RegulatoryNetworks (GRN)

modified through a threshold tuned for each individual class that defines the membership degree above which a gene should be either overexpressed or under-expressed in all rules of this class in order for this gene to appear in the profile. The last profile is of the CNS cancer. The profiles show that different genes are expressed highly in different types of cancer. The analysis is performed with the use of a proprietary methodology and software system SIFTWARE (www.peblnz.com) (Kasabov 2001). The profiles show that different genes are expressed highly in different types of cancer.

PR LU

co H

LV 8L

;

u

12

... UT

LE PE PA

eN

.. E

eNS lID

!ll

'!ll

;m Gtn. Nurnbff

'-----'"",1-'.1

- . - - f ' R I I-'.1"

I ...,jfO" r j ...,j[QT r

...... --1IJ11-'.1 J e-.·--ICO,I-'.1 j '_LV1!-'.1 j ,.........-""-lJ)L'I-'.1 j

M_

'1 -'.1"

~l"O&5r

...,j[QTr ...,j("'iii""r ...,jfO"r j ...,j[QTr

:.'!ll

IU1'W ..... -'-' 1-'.1

U_ _ _

R.... """"_~II -'.1

•....-_ -.... 11-'.1"" o-__ IOVII.:..J

- ..-,,,,,1-'.1 e-..-_ItHSI !-'.1

zn

l5/J

j~107'r

j ...,j[QTr I ...,jj""OW r J ...,j("'iii""r j ...,jl"O&5r j ...,jlO7' r j ...,jf"Oi'"r

399

uo- -'.1"

j

~ [QT

v_'__ 1 s...r;."

I

........'·-1

Fig. 7.5. Class profiles of 14 types of cancer extracted from a trained 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. See Color Plate 9

Among the CNS cancer group there are 3 clusters that have different gene expression profiles, as detected by an EFuNN ECOS trained system in Fig. 7.5 and shown in Fig. 7.6. The analysis is again performed with the software system SIFTWARE (www.peblnz.com). The system allows accessing a selected gene in the GenBank. For example, the highly expressed

7.3 Modeling Gene/Protein Regulatory Networks (GPRN)

145

gene (a darker line) in cluster 1 is gene with an accession number AA363338, which is not expressed in clusters 2 and 3 of the same eNS cancer data. (g)

.) unl,.1 neryous,,,,Iem (CNS) PEBl Genee.presslOt! Class ~ules

~

I

~td =

hlgn, Green = low)

-.J V;

........

I

~

s_.. G....

'O""i5" ,

I

Fig. 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. See Color Plate 9

7.3 Modeling Gene/Protein Regulatory Networks (GPRN) The aim of computational system biology is to understand complex biological objects in their entirety, i.e. at a system level. It involves the integration of different approaches and tools: computer modeling, large-scale data analysis, and biological experimentation. One of the major challenges of the systems biology is the identification of the logic and dynamics of gene-regulatory and biochemical networks. The most feasible application of systems biology is to create a detailed model of a cell regulation to provide system-level insights into mechanism-based drug discovery. System-level understanding is a recurrent theme in biology and has a long history. The term "system -level understanding" is a shift of focus in understanding a system's structure and dynamics in a whole, rather than

146

7 Gene/Protein Interactions - Modeling Gene Regulatory Networks (GRN)

the particular objects and their interactions. System-level understanding of a biological system can be derived from insight into four key properties (Dimitrov et al. 2004). 1. System structures. These include the gene regulatory network (GRN) and biochemical pathways. They can also include the mechanisms of modulation the physical properties of intracellular and multi cellular structures by interactions. 2. System dynamics. System behavior over time under various conditions can be understood by identifying essential mechanisms underlying specific behaviors and through various approaches depending on the systems nature: metabolic analysis (finding a basis of elementary flux modes that describe the dominant reaction pathways within the network), sensitivity analysis (the study of how the variation in the output of a model can be apportioned, qualitatively or quantitatively, to different sources of variation.), dynamic analysis methods such as phase portrait (geometry of the trajectories of the system in state space) and bifurcation analysis (bifurcation analysis traces time-varying change(s) in the state of the system in a multidimensional space where each dimension represents a particular system parameter (concentration of the biochemical factor involved, rate of reactions/interactions, etc). As parameters varied, changes may occur in the qualitative structure of the solutions for certain parameter values. These changes are called bifurcations and the parameter values are called bifurcation values). 3. The control method. Mechanisms that systematically control the state of the cell can be modulated to change system behavior and optimize potential therapeutic effect targets of the treatment. 4. The design method. Strategies to modify and construct biological systems having desired properties can be devised based on definite design principles and simulations, instead of blind trial-and-error. As it was mentioned above, in reality analysis of system dynamics and understanding the system structure are overlapping processes. In some cases analysis of the system dynamics can give useful predictions in system structure (new interactions, additional member of system). Different methods can be used to study the dynamical properties of the system: • Analysis of steady-states allows finding the systems states when there are no dynamical changes in system components. • Stability and sensitivity analyses provide insights into how systems behavior changes when stimuli and rate constants are modified to reflect dynamic behavior.

7.3 Modeling Gene/Protein Regulatory Networks (GPRN)

147

• Bifurcation analysis, in which a dynamic simulator is coupled with analysis tools, can provide a detailed illustration of dynamic behavior. The choice of the analytical methods depends on availability of the data that can be incorporated in into the model and the nature of the model. It is important to know the main properties of the complex system under investigation, such as robustness. Robustness is a central issue in all complex systems and it is very essential for understanding of the biological object functioning at the system level. Robust systems exhibit the following phenomenological properties: • Adaptation, which denotes the ability to cope with environmental changes. • Parameter insensitivity, which indicates a system's relative insensitivity (to a certain extent) to specific kinetic parameters. • Graceful degradation, which reflects the characteristic slow degradation of a system's functions after damage, rather than catastrophic failure.

All the above features are present in many of the CI methods and techniques and make them very suitable to modeling complex biological systems. Revealing all these characteristics of a complex living system helps choosing an appropriate method for their modeling, and also constitutes an inspiration for the development of new CI methods that posses these features. Modeling living cells in silico (in a computer) has many implications; one of them is testing new drugs through simulation rather than on patients. According to recent statistics, human trials fail for 70-75% of the drugs that enter them. Modeling gene regulatory networks (GRN) is the task of creating a dynamic interaction network between genes that defines the next time expression of genes based on their previous levels of expression. A simple GRN of 4 genes is shown in Fig. 7.7. Each node from Fig. 7.7 represents either a single gene/protein or a cluster of genes that have a similar expression over time, as illustrated in Fig. 7.8. A detailed discussion of the methods for GRN modeling can be found in (Dimitrov et al. 2004). Models of GRN, derived from gene expression RNA data, have been developed with the use of different mathematical and computational methods, such as: statistical correlation techniques, evolutionary computation, ANN, differential equations (both ordinary and partial), Boolean models; kinetic models; state-based models and others.

148

7 Gene/Protein Interactions - Modeling Gene Regulatory Networks (GRN)

CJ Fig. 7.7. A simplified gene regulatory network where each node represents a gene/protein (or a group of them) and the arcs represent the connection between them - either excitatory (+) or inhibitory (-)

1.5

0 .5

c:

o

·en

...~ - 0 .5

Co >C

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 -
•••••••

.+

+•

..,

. To other nodes

Link to the neuron's

~~ Fig. 8.2. Schematic illustration of the gene/protein regulatory network (GPRN). Description of symbols is in the text. Protein properties and concentrations are linked to neuronal parameters, like for instance magnitude of excitation and inhibition, etc. GPRN illustrationwas inspired by Fig.l in (Chen and Aihara 2002)

We will consider the dynamics of genes and proteins to be continuous that is, we can describe their changes in a continuous time. We will represent the mRNA levels of all the relevant genes with the vector m = (mr, mz, ... , m n ) and the corresponding protein levels with the vector p = (pr, P2, ···,Pn)' Components ofmRNA vector change as: dm ' dt

= Am;

(n

K

am, LWijPj(t-r p ) + LV;kXk(t-r FI

)

X)

+ bm, -Am;m;(t)

(8.10)

k=l

where:

mlt) is the overall level of mRNA for the

t h gene at time t;

am, is a nonlinear sigmoid regulation-expression (activation) function for the t h gene; th Am, is the amplitude of the activation function for the i gene;

8.3 Continuous Model of Gene-Protein Dynamics

167

are the interaction coefficients between the t h andj" gene (while the interaction itself is mediated by proteins activating corresponding transcription factors); - Pi is the level of the /h protein or protein subunit;

-

wij

T Pi is the delay, with which the /h protein influences the transcription of the t h gene; - Vik is the influence of the J(h external factor upon the gene i (influence of hormone, drug, etc.); - x, is the concentration of the k 1h external factor; 1h - 'x, is the delay, with which the k external factor influences the tran1h scription of the i gene; h - bm. is the bias, i.e., the basal expression level of the t gene group; - Am. is the degradation rate of the mRNA of the i'h gene group. -

Analogically, protein levels change as:

:' = Api O'p, mi(t-T m)+ ~UikYk(t-Ty)+bp, -Ap,Pi(t)

d. .

(

K'

)

(8.11)

where:

- plt) is the level of a protein (or protein subunit) coded for by the t h gene

-

-

-

-

-

-

(by the level of protein we mean the concentration of a fully functional protein); h 0' p, is a nonlinear sigmoid synthesis function for the t protein (note, we consider that one protein is coded for by only one gene); h A p is the amplitude of the synthesis function for the t protein; 1h m, is the overall level ofmRNA for the i gene; 'm, is the delay from initiation of transcription of the i1h gene family till the end of synthesis of the i 1h protein (on the order of tens of minutes); 1h Uik is the influence of the k external factor upon the protein (hormone, drug, but also a concentration of free ribosomes, etc.); Yk is the concentration of the k' h external factor; 'v, is the delay, with which the k'h external factor influences the protein level; 1h b p, is the bias, i.e., the basal level of the i protein; 1h Ap , is the degradation rate of the i protein;

r

The linear difference form of Eq. 8.10 without decay was used by (D'Haeseleer et al. 1999) to model mRNA levels during the rat eNS development and injury. Similar equation as Eq. 8.11 for the protein (gene

168

8 CNGMas Integration of GPRN,ANN and Evolving Processes

product) dynamics was used by (Marnellos and Mjolsness 2003), to model early neurogenesis in Drosophila. Eq. 8.11 is a more general form of the commonly used mathematical descriptions of the protein mass balance (McAdams and Arkin 1998). However it is less complex than the detailed model of Mehra et al., which takes into account: ratio of total mRNA to total ribosomes, product of average ribosomal binding affinity and total mRNA concentration, percentage of free ribosomes, distribution of active ribosomes over the polysomes, ratio of irreversible initiation or elongation rate constant to the termination rate constant for each mRNA, and ratio of protein degradation rate constant to translation termination rate constant (Mehra et al. 2003). This model was developed to provide insight into a reported lack of correspondence between mRNA levels and protein levels (Lee et al. 2003). There are presumably at least three reasons for the poor correlation generally reported in the literature between the level of mRNA and the level of protein: (1) complicated and varied post-translational processes, (2) variable degradation of proteins, and (3) error and noise in both protein and mRNA measurements (Greenbaum et al. 2003). However, this lack of correlation does not hold for all classes of proteins/mRNAs. In fact protein complexes that have clearly defined interactions between their subunits have highly correlated levels with mRNA expression levels (Jansen et al. 2002, Greenbaum et al. 2003, Fraser et al. 2004). 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 provide a special case as they have to be available in stoichiometric amounts of proteins for the complexes to functions (Greenbaum et al. 2003). And this is exactly the case of proteins in our model, which are receptors and ion channels, comprised of exact ratios of subunits. Thus, the use of Eq. 8.2 is justified in our model. Eq. 8.10 and Eq. 8.11 are the so-called delay differential equations. Delay differential equations are similar to ordinary differential equations, but their evolution involves past values of the state variable. The solution of delay differential equations therefore requires knowledge of not only the current state, but also of the state a certain time previously (Weisstein 1999-2006). For DDEs we must provide not just the value of the solution at the initial point, but also the history, that is the solution at times prior to the initial point (Drager and Layton 1997). As of MATLAB 6.5 (Release 13), the DDEs solver dde23 is part of the official MATLAB release. The next step is to link parameter values to protein concentrations. Let P, denotes the j'h parameter of a model neuron. If Pj E (0, 1) is the normal-

8.3 Continuous Model of Gene-Protein Dynamics

169

ized level of protein concentration, then the value of parameter P, is directly proportional to the concentration of protein Pt- in such a way that P/t)

=

P j (t)(Pt

ax

-

Pt

n

) - Pj

ffiin

(8.12)

where: Pj max and P,min are maximal and minimal values of the lh parameter, respectively. Other relations between proteins and parameters are also possible. The system of introduced equation allows for investigation of how deleted or mutated genes can alter the activity of a model neural network. Fig. 8.3 summarizes temporal evolution of activity of three system levels, i.e. three dynamic systems that we want to integrate into one integrated dynamic system of CNGM. The first dynamic system is the ANN with its measurable output, for instance level of electrical activity. Below that we have changes in protein levels. Changes in protein concentrations lag behind the changes in gene expression levels by the order of tens of minutes, even hours, due to the process of transcription, translation, and posttranslational modification (Lodish et al. 2000). Protein levels are directly related to parameters of neuronal signaling, like fast and slow excitation (inhibition), etc. Gene expression levels (expressed as mRNA levels) change slowly - on the order of hours and so do protein levels and parameter values. Ideally we would know all the parameters in Eqs. 8.9 and 8.10, that is, ideally we would know all the delays, biases, shapes of activation functions, amplitudes, degradation rates, and last but not least, the interaction matrix W = {wij}' Then we could simulate the integrated CNGM and obtain Fig. 8.3 for the real time. Unfortunately, the situation is far from ideal. In reality, we would have to proceed in the following steps: 1. Obtain gene expression data (mRNA levels) from the relevant neural system for discrete sampling intervals let us say for every hour. 2. Then use the extrapolation technique (for instance the extended Kalman filter (Kasabov et al. 2004)) to extrapolate missing values for mRNA levels, to obtain the curves for all m;(t)'s. In such a way, we can obtain the values of m, for, let us say, every minute. 3. Then we need to calculate the levels of proteins according to Eq. 8.10 to estimate the values of parameters according to the Eq. 8.11. 4. For predictions on gene interactions we need to infer the values of wij interaction coefficients between the pairs lh and lh gene by means of reverse engineering (Kasabov et al. 2004).

170

8 CNGM as Integration of GPRN, ANN and Evolving Processes

I Mea surabl e output: level of some

activity

)

Time (hrs)

P3

I

102040 min

c

Time (hrs)

Samp ling rate : every hour

mRNA

Gl G2

G3 2

3

4

5

7

Ti me (hrs)

Fig. 8.3. Summary of three dynamic systems within one integrated dynamic system - computational neurogenetic model. (a) abstract ANN activity like average spiking frequency, level of glucose consumption, etc., (b) protein dynamics, and (c) gene dynamics

The most important is the first step, i.e. the extrapolation , which will enable us to work with values of mRNA s and consequently parameter values in minute intervals. Different methods can be used , for instance the Kalman filter (Kasabov et al. 2004 ), evolutionary optimization (Whitehead et al. 2004), state space equations (Wu et al. 2004), etc. Then the biggest challenge is to estimate the delays from initiation of trans cription of the gene families till the end of synthesis of proteins. Alternatively , protein level s can be directly determined by mass spectrometry (MacBeath and Schreiber 2000). We need temporal courses of protein concentrations for updating neuronal parameters to simulate ANN. Proper updating of neuronal parameters is crucial for explaining changes in the ANN output why they occur and when . We can make rough qualified estimates of protein-synthesis delays usin g the information of the length of the relevant proteins and the time , which is needed for their gene s transcription and the

8.4 Towards the Integration of CNGM and Bioinfonnatics

171

subsequent protein translation and posttranslational modifications (Lodish et al. 2000). Alternatively, instead of simulating the ANN for the whole time of evolving gene-protein dynamics, we would like to suggest simulations of ANN only at particular interesting intervals of the gene-protein dynamics, as it is illustrated in Fig. 8.4. ANN output behavior: for instance the level of activity

B~EJ i \ r

Protein level

EJ

.:

P1

P3

\

t

Time (hrs)

Sampling rate: at every interesting interval of gene-protein dynamics

Fig. 8.4. Sampling of the ANN output at interesting time intervals, based on some heuristics. The average level of activity can be measured as the number of spikes, level of glucoseconsumption, etc.

Interesting intervals for sampling the ANN output can be based on some kind of heuristics - for instance based on the knowledge that at that particular time instant something has happened - for instance, an animal fell asleep. Otherwise this sampling can occur at intervals where the parameters have their extreme values, at intersections of values, etc.

8.4 Towards the Integration of CNGM and Bioinformatics Is there a genetic basis for the interaction matrix W between many genes, which is at the heart of our abstract CNGM? In this section we would like to provide an argument that such a genetic basis indeed exist. We have investigated through sequence analysis, the excitatory and fast inhibitory receptors (i.e., AMPAR, NMDAR and GABRA) gene groups (Benuskova et al. 2006).

172

8 CNGM as Integration of GPRN, ANN and Evolving Processes

Table 8.1. List of subunit proteins for AMPAR, GABRA and NMDAR. Glu means glutamate, GABA is the y-aminobutyric acid

gilnumber] gi11169959[ gil23831146[ gi[1169961[ gil1346142[ gil114969711 gi[142856031 gi[145481621 gil24926291 gi118201966[ gi138327554[ gil13460781 gi[45576031 gil1346079! gil3995191 gil238311281 gi1455946[ gi11207731 gil278201211 gi[38788155[ gi113959689[

Human Protein Glu AMPAR 1; (GluR-l) (GluR-A) (GluR-Kl) Glu AMPAR 2; (GluR-2) (GluR-B) (GluR-K2) Glu AMPAR 3; (GluR-3) (GluR-C) (GluR-K3) Glu AMPAR 4; (GluR-4) (GluR4) (GluR-D) Glu NMDAR 1 isoform NRl-l precursor Glu NMDAR subunit epsilon 1 precursor; (NR2A) (NMDAR2A) Glu NMDAR subunit epsilon 2 precursor; (NR2B) (NMDAR2B) Glu NMDAR subunit epsilon 3 precursor; (NR2C) (NMDAR2C) Glu NMDAR subunit epsilon 4 precursor; (NR2D) (NMDAR2D) GABA A receptor, alpha 1 precursor GABAA receptor alpha-2 subunit precursor GABA A receptor, alpha-3 precursor GABA A receptor alpha-4 subunit precursor GABA A receptor alpha-5 subunit precursor GABAA receptor beta-I subunit precursor GABAA receptor beta 2 subunit GABAA receptor beta-3 subunit precursor GABAA receptor gamma-l subunit precursor GABA A receptor, gamma 2 isoform 1 precursor GABA A receptor gamma-3 subunit precursor

Gene

Sequence length

GRIAI

906 aa

GRIA2

883 aa

GRIA3

894aa

GRIA4

902 aa

GRIN 1

885 aa

GRIN2

1464 aa

GRIN3

1484 aa

GRIN4

1233 aa

GRIN5

1336 aa

GABRAI

456 aa

GABRA2

451 aa

GABRA3

492 aa

GABRA4

554 aa

GABRA5

462 aa

GABRBI

474 aa

GABRB2

474 aa

GABRB3

473 aa

GABRGI

465 aa

GABRG2

475 aa

GABRG3

467 aa

We mainly focused on them because they playa major role in most of the mental disorders through their direct or indirect interactions with several other genes/proteins and through specific parameter functions (e.g.,

8.4 Towards the Integration ofCNGM and Bioinforrnatics

173

excitation and inhibition). Each of these proteins is comprised of several subunits, and each subunit is coded by a separate gene (see Table 8.1). Particular subunit content of these receptor proteins can vary depending on the brain region (Burnashev and Rozov 2000) therefore we perform the bioinformatics analysis on all subunit genes. Initially, the information related to expression of these subunit genes, mutations, etc. was collected through relevant literature survey followed by sequences retrieval from NCBI database. We did preliminary analysis for searching the common motifs (patterns that occurs repeatedly in a group of related protein or DNA sequences) among these subunits. Through this initial analysis we observed that all detected motifs were belonging to similar protein families, like ligand-gated ion channel, receptor family ligand binding region and neurotransmitter-gated ion channel ligand binding domain. This observation motivated us for performing the detailed residual inspection. For our investigation, we employed the standard operating bioinformatics procedure, i.e. comparative analysis from similarity measures that is widely accepted approach by most of neuroscientists. The method is familiarly known as multiple sequence alignments (MSA). Scope of this bioinformatics technique is twofold: (a) gains understanding to identify the shared regions of homology; (b) determines the consensus sequence of several aligned sequences. For such strategies, there are however many applications available free to academic users, but here in our case we used the CLUSTALW package (http://align.genome.jp/) developed by Koichi Ohkubo, (Genome Net), for details see Chenna (Chenna et al. 2003). A specific reason for picking this software was that it is mostly cited and secondly its output is a much readable representation than others in the field. Sequence similarity between all subunits was visualized, then we later used Box-Shade program V3.21 (http://www.ch.embnet.org/software/ BOXjorm.html) in order to format our multiple alignment results for explanation purpose. Fig. 8.5 highlights the most important observation and in general indicates that comparison of sequences of the 20 subunit proteins has revealed a number of conserved residues. As a biological fact we know the number of structurally conserved residues increases with the binding site contact size, thus we expect similar function for these amino acids in this case also. Here we have observed the extent to which these conserved residues are preserved, which is generally the case in most channel proteins and neuroreceptors, but the most interesting investigation is the consistent conservation of phenylalanine (F at position 269) and leucine (L at position 353) in all 20 proteins with no mutations. We expect these residues to play the roles as binding centers for interaction of these proteins with other

174

8 CNGM as Integrationof GPRN, ANN and Evolving Processes

genes/proteins that have a regulatory effect upon these receptors. Therefore, we expect that the gene interaction observations, which we infer from our abstract CNGM, may be justified because of such truly conserved residues. However we must say that all such hypotheses remain unproven and these predictions need to be tested through laboratory experimentation. For these neuronal proteins we also obtained a dendogram (Fig. 8.6) that represents the evolutionary related closeness between these subunit proteins. More closely related pairs of neuronal protein sequences have aligned most readily to each other than more divergent pairs. Gene

GABRA3 GABRAS GABRAI GABRA2 GABRA4 GABRGI GABRG2 GABRG3 GABRB2 GABRB3 GABRBI GRIA2 GRIA3 GRIM GRIAI GRIN2 GRIN3 GRIN4 GRINS

'"

Position

260

o

IGEYVVMII IGBYIIMI IGEYVVMII IGBYIVMI IGBYIVMI

LKRKIGYFV LKRKIGYFV LKRKIGYFV LKRKIGYFV LRRKMGYFM

SGDYVIMII

LSRRMGYFI SVCFIFV'F LSRRMGYFI SVCFIFVPS LSRRMGYFI IVCFLFVF LKRNIGYFI MGCFV LKRNIGYFI MGCFVFVFL LKRNIGYFI MGCFVFVF DPLAYEIWM LIIISSYI DPLAYEIWM ..• LIIISSYI DPLAYEIWM LIIISSYI DPLAYEIWM LIIISSYI EPFSASVWV VIFLASYI EPFSADVWV VIFLASYI EPYSPAVWV VIFLASYI BPYSPAVWV VIFLASYI

'" SGDYVVMS . . . AGDYVVJfII

IGSYPRLSL IGAYPRLSL IGAYPRLSL PQKSKPGVF

POKSKPGVF PQKSKPGVF PQKSKPGVF .SNGIVSPS .SNGIVSPS .SNGIVSPS .SNGIVSPS

350 AVCYAFVFS AVCYAFWFS AYCYAFVPS AVCYAFVFS AVCFAFVFS

360 BFAIVNYPI EFAIVNYPI EFAIVNYFI EFAIVNYFI EFAAVNYFI YGILHYFI YGILHYFV YAILNYYS EYALVNYIF EYAFVNYIF EYAFVNYIF FLIVERMV

. . . . . . . . . . . .

FLIVERNV ••• LIVERMV . FLIVERMV . FMIQEEFV . FHIQEBYV ••.

FMIQBQYI ••• FMIQBBYV •••

Fig. 8.5. Multiple alignments of all 20 subunits of 3 neuronal informationprocessing proteins showing consistent conservation of phenylalanine (F) at position 269 and leucine (L) at 353

r-c.

GRIN2 GRIN3 GRIN5 GRIN4

I

rlGABRB3

GABRB2 GRIA2

GABRGI r-

GABRG2 GABRG3

~GABRAI

GABRA4 GABRAS GABRAJ GABRA2

GRINI

rI GA!lRBl '1

GRIA3 GRIA4

GRIAI

Fig. 8.6. Phylogenetic tree representing 20 subunit proteins of AMPA, GABAA and NMDA receptors

8.5 Summary

175

By means of bioinformatics analysis of subunit proteins of AMPA, GABA A and NMDA neuroreceptors new information about preserved protein segments has been extracted. In particular, the evolutionary linkage of protein families indicates that most of the neuronal subunits are very closely related with each other, which further support the hypothesis that these neuroreceptor subunits are all required for proper and complete functioning of that receptor protein and thus their expression should be coordinated within one gene group. Based on the result of analysis that clearly showed us the extent of conservation of many amino-acid residues among all investigated receptor subunits, we assume that these regions can be the basis for mutual interactions not only between the subunit genes of one receptor but also between different neuroreceptors. In fact, mutual interaction between different receptors (i.e. AMPAR, NMDAR and GABRA) has been recently confirmed experimentally to occur on the genetic level (Salonen et al. 2006).

8.5 Summary Complex interactions between genes and proteins in neurons affect the dynamics of neural networks in the brain. Gene expression values may change due to internal dynamics of a GPRN or due to external factors, like hormones, electrical activity, etc. We can expect that different initial gene expression values, and even different gene interactions can lead to the same outcome in terms of neuronal activity. However, in the diseased brain, either altered initial expression values, mutated genes and/or altered interactions within GPRN lead to abnormalities in network activity. Biological sequence analysis and representation of biological properties like discovered conserved residues and motifs, location of gene on a chromosome, mutation etc. of these interacting molecules enhances the chances of better tuning of the gene interaction network used in the CNGM experiments and also in assignment of initial gene/protein expression values through which different states of the neural network operation can be achieved. Protein function usually imposes tight constraints on the evolution of specific regions of protein structure. Thus, proteins with a common evolutionary history and function can frequently be identified from the occurrence of clusters of conserved residues in their amino acid sequences. Residues directly or indirectly involved in function are often clustered in a short sequence motif (signature, pattern, or fingerprint) that is conserved across the different proteins sharing that function. Therefore molecular se-

176

8 CNGM as Integration of GPRN, ANN and Evolving Processes

quence analysis through bioinformatics approach can strongly enhance the biological plausibility of neurogenetic models. Here in our case study, sequence analysis was done on each specific subunit protein of AMPA, GABA A and NMDA receptors. Each subunit is encoded by individual gene. Particular subunit content of these receptor proteins can vary depending on the brain region (Burnashev and Rozov 2000) therefore we performed the bioinformatics analysis on all subunit genes. Based on the bioinformatics analysis we hypothesize that the conserved regions can be the basis for mutual interactions between these subunits both at the level of their composite receptor and between different receptors. In real neural networks neuronal parameters that define the functioning of a neural network depend on genes and proteins in a complex way. Gene expression values change due to internal dynamics of the gene/protein regulatory network, initial conditions of the genes and external conditions. All this may affect gradually or quickly the functioning of the neural network as a whole. Realistic models of gene networks within neural networks should account for these processes. Future research will concentrate on animal data and around the following issues: 1. "What-if analysis". What happens if one or few particular genes are erased or mutated (i.e. data are collected from knock-out gene technology)? What happens if interactions within the GRN change? What happens if external factors are included? In such a way our approach can serve as a noninvasive test system. 2. Introduction of learning rule(s) into an ANN model and corresponding genes into the GRN. First steps are presented in the next chapter of this book. 3. Exploration of possibilities of modeling genetically caused brain disorders such as epilepsy, Parkinson's disease, etc. The goal is to make predictions about gene interactions to aid experimental research on gene interactions in various states and conditions of the brain. Basic genetic data are presented and some ideas are conceived in the last chapter of this book.

9 Application of CNGM to Learning and Memory

Before we introduce a tentative CNGM of learning and memory at a cellular level, we need to consider the relevant rules of synaptic plasticity, which represent knowledge about the basic mechanisms of cellular and molecular memory storage. Then we need to enhance these rules with connections to gene/protein regulatory network (GPRN) of genes and proteins that are relevant to learning and memory. Finally, a provisional CNGM of learning and memory will be constructed. In what follows, we will discuss only the activity-dependent plasticity of excitatory synaptic connections of the brain, because they are thought to mediate the information storage within biological neural networks (Kandel et al. 2000).

9.1 Rules of Synaptic Plasticity and Metaplasticity At present, it is widely accepted that origins and targets of brain synapses are determined genetically as well as the properties of neurotransmission. However, the efficacy of signal transfer at synapses and even the number of synapses can change throughout life as a consequence of learning (Kandel et al. 2000). At present, bidirectional changes in the strength of excitatory synaptic weights are thought to be fundamental to information storage within neuronal networks. To unravel the cellular and molecular mechanisms of synaptic plasticity, which are the basis of learning and memory, is an ambitious goal of neurobiology and neurogenetics. Donald Hebb predicted a rule of synaptic plasticity driven by correlation of preand postsynaptic activity. He postulated that repeated firing of one neuron by another one, across a particular synapse, increases its strength (Hebb 1949). But more than 50 years later, there are still many open questions as to how this happens as well as how the weakening of synapses occurs. Synaptic plasticity is a process in which synapses change their efficacy as a consequence of their previous activity. Sometimes it is also called activity- or experience-dependent synaptic plasticity. Synaptic efficacy (synaptic weight, synaptic strength) is directly proportional to the magnitude of the postsynaptic potential (PSP) on the postsynaptic membrane, which

178

9 Application of CNGM to Learning and Memory

arises as a consequence of the defined unit stimulation of the presynaptic terminal of that synapse (Benuskova 1988). Synaptic efficacy is a measure of the synapse's contribution to the summary somatic postsynaptic potential, which determines the time and frequency of the postsynaptic spike train generated after exceeding the excitation threshold of the neuron. Synaptic weight (PSP after unit stimulation of the synapse) depends on two groups of factors: • Presynaptic factors: released amount of neurotransmitter. • Postsynaptic factors: number of postsynaptic receptors, receptor types and properties, and input electric impedance (depends on the geometry of dendritic input surface and its electric properties). Change of these synaptic properties leads to the change of synaptic strength. This change can be short- or long-lasting, negative or positive. In many regions of the brain, long-term synaptic potentiation (LTP), a longlasting increase in synaptic efficacy, is produced by high-frequency stimulation (HFS) of presynaptic afferents (Bliss and Lomo 1973) or by pairing presynaptic stimulation with postsynaptic depolarization (Markram et al. 1997). Long-term synaptic depression (LTD), a long-lasting decrease in the strength of synaptic transmission, is produced by low-frequency stimulation (LFS) of presynaptic afferents. The majority of synapses in many brain regions and in many species that express LTP also express LTD. Thus, the regulation of synaptic strength by activity is bidirectional (Bear 1995, Castellani et al. 2001). Molecules of glutamate neurotransmitter released during afferent activity bind to AMPA, NMDA and metabotropic glutamate (mGlu) receptors to produce postsynaptic response. The intracellular Ca2+ level varies as Ca2+ enters the neuron via the voltage-gated calcium channels (VGCCs) and NMDA receptor-channel complex, and as Ca2+ is released from internal storage sites as a result of the mGlu receptormediated G-protein cascade (Abraham and Tate 1997, Zucker 1999). When the postsynaptic membrane is depolarized by the actions of the nonNMDA (AMPA) receptor channel (as occurs during high-frequency stimulation), the depolarization relieves the Mg2+ blockage of the NMDA channel. High-frequency afferent activity results in high levels of intracellular Ca2+, which preferentially activate protein kinases (CaMKII, PKC, tyrosine kinase fyn); low-frequency afferent activity results in low levels of Ca2+, which preferentially activate protein phosphatases (calcineurin, PP 1). The induction of LTP and LTD appears to depend on the relative activity of kinases and phosphatases (Elgersma and Silva 1999, Mayford and Kandel 1999). With both enzymes present, predominant kinase activity leads to LTP (via phosphorylation of various substrates) while predominant phosphatase activity leads to LTD (via de-phosphorylation of various sub-

9.1 Rules of Synaptic Plasticity and Metaplasticity

179

strates). The intracellular calcium concentration [Ca2+]i is the principal trigger for the induction of LTD/LTP (Artola and Singer 1993, Zucker 1999, Shouval, Bear et al. 2002). Calcium influx through NMDARs plays a crucial role in the induction of LTD/LTP. NMDAR is unique because its activation requires the presynaptic release of glutamate to occur within certain time window as postsynaptic depolarization (NMDAR is both receptor- and voltage-gated channel). Thus, NMDAR serves as a molecular coincidence detector for detecting the two simultaneous presynaptic a postsynaptic events, thus implementing Hebb's rule at synapses (Paulsen and Sejnowski 2000, Tsien 2000). Conventional view for induction of LTP is that strong synaptic input (i.e. many synchronously active afferent fibers) produces a local depolarization that unblocks NMDARs, and concurrently released glutamate provides the necessary Ca2+ signal through NMDARs. Alternative model is based on backpropagating action potentials (Stuart and Sakmann 1994). Backpropagated APs (triggered by large EPSPs or by other inputs) can provide additional postsynaptic depolarization on the top of the AMPAmediated depolarization, for the voltage-dependent relief of Mg2+ block of NMDARs, when they bind glutamate, so that more Ca2+ pures in (Magee and Johnston 1997, Koester and Sakmann 1998, Linden 1999). This important role of backpropagating APs in switching between LTD and LTP is also supported by computational modeling studies (Shouval, Bear et al. 2002, Shouval, Castellani et al. 2002) (Fig. 9.1).

Postsynaptic spike occurs too early \0 coincide with EPSP

Postsynaptic spike occu rs in prope r time window \0 coincide with EPSP

Fig. 9.1. Illustration of the spike timing-dependent plasticity. Ions of sodium and calcium enter the postsynaptic spine through the NMDA-receptor gated ion channels. When postsynaptic neuron fires, postsynaptic spikes backpropagate to spines. When their timing coincides with EPSP, more calcium enters spines, and thus there is a bigger chance to achieve LTP

180

9 Application of CNGM to Learning and Memory

Experimental evidence indicates that long-term modification of synaptic efficacy indeed depends on the timing of pre- and postsynaptic APs (spike timing-dependent plasticity, STOP) (Song and Abbott 2001). It has been shown that the temporal order of the synaptic input and the postsynaptic spike determines whether LTP or LTO is elicited. Repeated pairing of postsynaptic spiking after presynaptic spikes results in larger Ca2+ influx and LTP (EPSP coincides with the backpropagating AP), whereas postsynaptic spiking before presynaptic spike (EPSP follows the AP) leads to small Ca2+ transient and LTO (Markram et al. 1997, Bi and Poo 1998, Froemke et al. 2005). Such a rule holds for cortical pyramidal neurons and for excitatory neurons in hippocampus (Abbott and Nelson 2000). This temporally asymmetric Hebbian synaptic plasticity supports sequence learning because it tends to wire together neurons that form causal chains (Paulsen and Sejnowski 2000).

a

A W [%]

b

ij

0.4

0.2

A t[m.] -40

- 20 --0.2

--0.4

ep 0

V

'j

c

.~

-" 0

E

., '"

~ '0 °

Q.

ftl

ce-, til

:0

0 Time between presynapticand postsynaptic spikes

' i

Fig. 9.2. STDP rule of synaptic plasticity. Presynaptic spikes that precede (follow) postsynaptic spikes within a certain time window produce long-term strengthening (weakening) of synapses, respectively. (a) Illustration of quant itative relationships for cortical neurons (Song et al. 2000); (b) illustration of experimentally measured points versus theoretical curves

Thus, the synaptic change, ~w(~f) , where ~f = fpost - f pre , due to a single pre- and postsynaptic pair is determined by this equation (Song et al. 2000, Izhikevich and Desai 2003) (see Fig. 9.2): ~w(M)=

{

~W+ = A+ exp(-~t /,+ )if M > 0 ~w_ =A_e xp( -M / ,_) i f M

< 0

(9.1)

9.1 Rules of Synaptic Plasticity and Metaplasticity

181

Parameters A+ and A_ determine the amplitude of synaptic change, which occurs when I1t is close to zero, while T+ and L determine the time windows over which synaptic changes occur. There are different implementations of the STDP rule, i.e. considering all the presynapticpostsynaptic spike pairs or just nearest neighbors, before and/or after a given presynaptic spike (Izhikevich and Desai 2003). We employ the nearest-neighbor additive implementation of STDP, that is, for each presynaptic spike, only two postsynaptic spikes are considered: the one that occurs before and the one that occurs after the presynaptic spike, i.e.: w(t + b't) = w(t)(l + I1w+ -l1wJ

(9.2)

Where Jt is a time step of weight calculation. This implementation as well as the nearest spike implementation were shown to faithfully reproduce the STDP experimental data from hippocampal and cortical preparations (VanRossum et al. 2000, Sjostrom et al. 2001). Moreover, nearest neighbor and nearest spike implementations of STDP have been shown to lead to the Bienenstock, Cooper and Munro (BCM) rule of synaptic plasticity (Izhikevich and Desai 2003). Besides precise timing of pre- and postsynaptic spikes, another factor determining the sign and magnitude of synaptic change is frequency of presynaptic firing (Bear 1995, Abraham and Bear 1996). Experimental data from the developing visual cortex had led to the formulation of a synaptic modification rule, known as the Bienenstock-Cooper-Munro (BCM) rule (Bienenstock et al. 1982). The model has two main features: First, it postulates that a neuron possesses a synaptic modification threshold (LTP/LTD threshold or 8M), which dictates whether the neuron's activity at any given instant will lead to strengthening or weakening of its input synapses. Thus, the modification threshold, Brv!, determines the direction of synaptic efficacy change. Synaptic modification varies as a nonlinear (parabolic) function (rf) of postsynaptic activity (c), i.e. ¢Xc) = c(t) (c(t) - Brv!(t)) (see Fig. 9.3). In the original BCM formulation, activity means either average or instantaneous firing rate rather than individual spikes. Although the firing rate of a neuron c(t) depends in a nonlinear fashion on the postsynaptic potentials, BCM theory considers that the region between the excitation threshold and saturation may be reasonably approximated by a linear input-output relationship of the model neuron (Benuskova, Rema et al. 200 I), thus c(t) is defined as the product between presynaptic activity vector (x) and synaptic efficacy vector (w), i.e. c(t) = w(t) . x(t). The function ¢Xc) changes sign at a particular value of c, that is equal to the value of modification threshold Brv!. Brv! is the point of crossover from LTD to LTP. If postsynaptic activity is below 8M (c < Brv!), but above baseline, ¢Xc) < 0, and synaptic efficacies are weakened. Con-

182

9 Application of CNGM to Learningand Memory

vcrs ely, if c exceeds B:vi, modification function ¢i..c) > 0, and active synapses potentiate. Synaptic weight vector (w) changes according to Hebb 's learning rule, which requires correlated pre- and postsynaptic activity at the synapse, i.e. (9.3)

LlW=rytP X where

n> 0 is the learning rate constant. ¢ > 0: potentiation

c

,,/ (L

Postsynaptic activity c'

c ¢ < 0: depression Fig. 9.3. Sliding modification LTP/LTD threshold Bt..1 in the BCM theory. Synap-

tic strengthening/weakening depends on the current value of Bt..1, which is not constant but instead changes proportionally to the average of the past postsynaptic activity c. In ABS theory, B_and B+. mean thresholds for LTD and LTP, respectively

The second important feature of the BCM rule is that the value of B:vi is not fixed but instead changes according to a nonlinear function of the average output of the cell (Cooper et al. 2004). B:vi varies according to a (running) time average of prior postsynaptic activity, i.e. 8M is a sliding modification threshold. The current value of 8M changes proportionally to the square of the neuron's activity averaged over some recent past "eM, i.e. (9.4) ~ slides as a function of the prior history of the postsynaptic cell, and its effective position depends upon the parameter a and the function qi.,t), which in turn depend on levels of neuromodulators like acetylcholine, dopamine, etc, and on the overall number of NMDARs (Cooper 1987, Benuskova, Rema et al. 2001) . Sliding LTP/LTD threshold is a homeostatic mechanism, which keeps the modifiable synapses within a useful dynamic

9.1 Rules of Synaptic Plasticity and Metaplasticity

183

range (Abraham et al. 2001). It acts against Hebbian positive-feedback process, in which effective synapses are strengthened, making them even more effective, and ineffective synapses are weakened, making them less so. This tends to destabilize postsynaptic firing rates increasing or decreasing them excessively. BCM stabilizes Hebbian plasticity by negative feedback, i.e. the LTP/LTD threshold increases (elevates, slides rightward) if the postsynaptic neuron is highly active, making LTP more difficult and LTD easier to induce. The opposite process occurs when the activity of the postsynaptic neuron is overly reduced, like in the cases of sensory deprivation (Cooper et al. 2004). Thus sliding of ~ guarantees synaptic stability without imposing subtractive or multiplicative controls on synaptic weights (Miller and MacKay 1994). Thus, the modification threshold ~, which regulates the ability to undergo LTP/LTD is itself regulated. The term "metaplasticity" has been introduced to describe the changes in the neuron's ability to undergo LTP and LTD, because it has become apparent from experimental data that the synaptic plasticity can indeed be altered by previous synaptic activity, or by neuromodulators (Abraham and Bear 1996). Metaplasticity is the plasticity of synaptic plasticity (plasticity at a higher level). For example, previous stimuli may make it easier or harder to induce LTP, like with the sliding BCM modification threshold. Metaplasticity mechanisms do not change synaptic efficacy but instead affect synaptic physiology in such a way that subsequent attempts to induce synaptic plasticity will be modified (Bear 1995, Abraham and Tate 1997). Artola, Brocher and Singer later formulated a similar synaptic plasticity rule (called the ABS rule), which defines two modification thresholds, (1for LTD induction and ()+ for LTP induction (Artola et al. 1990). According to the ABS rule, direction of the synaptic gain change again depends on the membrane potential of the postsynaptic cell or on the concentration of [Ca2+Ji. If the first threshold ()_ is reached, a mechanism is activated that leads to LTD. If the second threshold ()+ is reached, another process is triggered that leads to LTP. Experimental evidence supports the existence of the LTD threshold (Mockett et al. 2002) so the original BCM theory has been slightly extended to incorporate this and another findings to result in the so called calcium control hypothesis (Shouval, Bear et al. 2002, Shouval, Castellani et al. 2002). According to the calcium control hypothesis, the temporal derivative of the lh neuron's synaptic weight depends on the calcium levels so that: (9.5)

184

9 Application of CNGMto Learningand Memory

where [Ca 2+1 is the calcium concentration at synapse j, learning rate '7 is supposed to be calcium-dependent and increase monotonically (sigmoidally) with calcium levels, and the Q-function is defined in the following way: when 0< [Ca 2+1 < B_, the synaptic weight does not change, when B_ < 2 [Ca 2+1 < B+, the synaptic weight is depressed (LTD), and when [Ca +lJ > B+, the synaptic weight is potentiated (LTP). Both modification thresholds are the whole cell property and can change as a function of average past cell activity. Let us go back to the modification LTP/LTD threshold 8M. The BCM theory makes several explicit assumptions about how synapses modify, and there appears to be a good correspondence to experimental results. First, 8M exists: there is evidence in both hippocampus and visual cortex that the change in frequency of afferent activity generates a bidirectional pattern of LTD and LTP induction very similar to that predicted by the BCM model (Dudek and Bear 1993, Rick and Milgram 1996, Cho et al. 2001). Second, 8M slides: the threshold 8M does indeed appear to be adjustable, depending on the prior history of postsynaptic activity (Kirkwood et al. 1996, Rick and Milgram 1996, Abraham and Tate 1997). Third, 8M is a whole cell property and a function of previous postsynaptic cell firing (Holland and Wagner 1998, Wang and Wagner 1999, Abraham et al. 2001). Thus, it seems that the BCM theory indeed captures a real feature of a biological learning rule. Recently (Toyoizumi et al. 2005) asked what would be the optimal synaptic update rule so as to guarantee that a spiking neuron transmits as much information as possible. Under the (realistic) assumption of Poisson firing statistics, the synaptic rule exhibits all of the features of the BCM rule, in particular, regimes of synaptic potentiation and depression separated by a sliding threshold. The learning rule is found by maximizing the mutual information between presynaptic and postsynaptic spike trains under the constraint that the postsynaptic firing rate stays close to some target firing rate. Thus, the BCM rule has been extended to the spiking neuron, but still formulated in terms of instantaneous rate and not spikes (see equation 9.3). Another way of merging the BCM rule with the spiking nature of neurons is to proceed along the lines suggested by (Izhikevich and Desai 2003). They started with the STDP rule expressed by Eq. 9.1 and Eq. 9.2, and derived the expected magnitude of synaptic modification per one presynaptic spike, when the postsynaptic spike train is a Poisson process with firing rate x, i.e.: Liw(x) = x(

_~++x + _~-+x )

T+

T_

(9.6)

9.2 Towarda GPRNof Synaptic Plasticity

185

with the meaning of symbols as in Eq. 9.1. Derivation of the LTD/LTP threshold (the zero crossing of ~w(~) ) leads to (9.7)

The latter equation can be used to relate the sliding of ~ to biophysical and biochemical processes underlying LTD and LTP, either through dynamic changes of time constants for LTP and LTD as suggested in (Izhikevich and Desai 2003), and/or through the dynamic changes of their amplitudes. For instance, the amplitudes of LTD and LTP can change in the following way (Benuskova and Abraham 2006):

A+ (t) =

~:~~~

and

A~ (t) = A_ (O)BM (t)

(9.8)

where A(O) are initial (constant) values and ~ is the modification threshold. The time average of postsynaptic activity for ~ in Eq. 9.4 can be calculated as in (Benuskova, Rema et al. 200 I), that is by numeric integration of the following integral: (9.9)

If the postsynaptic activity c(t) was defined as c(t) = 1, if there is a postsynaptic spike at time t, and c(t) = 0, otherwise, then of course there is no need for the second power. In the following sections, we will relate the synaptic plasticity rules, i.e. STDP and BCM sliding modification threshold, to relevant proteins and genes, and the gene/protein regulatory network (GPRN), in order to propose a CNGM of learning and memory.

9.2 Toward a GPRN of Synaptic Plasticity Information revealed in this section will serve as a basis for the construction of links between synaptic plasticity variables from theoretical equations and particular proteins and genes within neurons that are involved in learning and memory formation. Then we will attempt to build a GPRN relevant to learning and memory molecular mechanisms. The induction of associative LTP/LTD depends on Ca 2+-dependent phosphorylation and dephosphorylation of various proteins, respectively (Mayford and Kandel 1999), and upon insertion and removal of AMPARs into and from the postsynaptic membrane (Carroll et al. 2001, Sheng and

186

9 Application of CNGM to Learning and Memory

Lee 2001). Long-term potentiation (LTP) in the mouse can be either short(1-3 hours) or long-lasting (> 24 hours), depending on the number of highfrequency stimulation trains administered to presynaptic inputs. However, in both cases, the induction of LTP requires Ca2+ influx through the Nmethyl-D-aspartate (NMDA) glutamate receptors. The short-lasting form of LTP or early E-LTP requires the participation of a Ca2+/calmodulindepenendent protein kinase II (CaMKII). CaMKII in the basal state is completely dependent upon Ca2+/calmodulin for its activity, but upon activation can rapidly convert to a Ca2+-independent kinase by autophosphorylation, which is necessary for the induction ofE-LTP. A molecular basis of E-LTP is the insertion of new AMPARs into the postsynaptic membrane (Shi et al. 1999). AMPARs mediate most of excitatory postsynaptic response in glutamatergic synapses; therefore changing their number (and/or properties) is a powerful way to control the strength of synaptic transmission. Insertion of new AMPAR occurs through exocytosis, i.e. fusion of storage vesicles with new AMPARs with the postsynaptic membrane (Lledo et al. 1998). Ca2+ and CaMKII play a key role in this process (Sudhof 1995). Localization of excitatory synapses on spines in turn provides for induction of strong intra-spine electric field that can electrophoretically drive vesicles for fusion with the postsynaptic membrane (Benuskova 2000). The long-lasting form of LTP or late L-LTP requires the activation of the cAMP signaling pathway through PKA (cAMP-dependent protein kinase, protein kinase A), ERKIMAPK (extracellular signal-regulated protein kinase/ mitogen-activated protein kinase), and RSK2 (ribosomal S6 kinase 2). Activation of gene transcription via CREB (cAMP-responsive transcription factor) pathway follows (Mayford and Kandel 1999). Constitutively active CaMKII activates a related kinase, CaMKIV that contributes to early Ca2+ -stimulated CREB phosphorylation (Wu et al. 200 I). CAMP is activated through Ca2+/calmodulin-dependent adenylyl cyclase. (Alternatively, a neurotransmitter or neuromodulator like dopamine binds to G-protein-coupled receptors. G-protein activates adenylyl cyclase, which catalyzes production of cAMP). CAMP then activates PKA. Ca2+ also activates MAPK and ERK. Facilitated by cAMP, both CaMKII and CaMKIV translocate to the cell nucleus along with PKA and ERKIMAPK to activate gene transcription via phosphorylation of CREE. Translocation ofERK to the nucleus requires activation ofPKA (Poser and Storm 2001). Activation of the ERKlMAPK pathway may trigger nuclear translocation of RSK2 and thus phosporylation and transactivation of CREB (Impey et al. 1998). Transition from E-LTP to L-LTP also requires phosphorylation by PKA of inhibitor-I (1-1), which dephosphorylates and turns off protein phos-

9.2 Towarda GPRNof Synaptic Plasticity

187

phatase 1, PPI (Mayford and Kandel 1999). The action on inhibitor-1 is opposed by ca1cineurin, which activates the phosphatase cascade and leads to LTD, which is the mirror opposite ofLTP. In fact, PKA and ca1cineurin phosphorylate and dephosphorylate the same residue on inhibitor-I. Calcineurin has this role in LTD because it has a high affinity for Ca 2+, even higher than that of CaMKII. At low-frequency stimulation, the amount of Ca2+ coming into the cell through the NMDARs is small. This will activate ca1cineurin but not CaMKII. Activation of ca1cineurin leads to a removal of AMPARs from the postsynaptic membrane through the endocytosis of membranous vesicles (Beattie et al. 2000). In addition, calcineurin dephosphorylates protein phosphatase I-I, which in tum phosphorylates and activates PPI that also participates in mechanisms of LTD. Processes leading to LTD and LTP described above are summarized in Fig. 9.4. NMDAR stimulation

D

::::==- CaMKil

2

[Ca +]i

D[hi9h]~ Calmodulin

D

Calcineurin

~

Adenylyl cyclase

cAMP

.Pt1'-1p~ i:o':';.1 (l

AMPAR removal

LTD

D D

_p +P

~~"'t1'.1

PKA

~

~

+P ERK I MAPK

I

RSK'

pCRES Gene expression

D

L-L TP

pCaMKl1

~«~

C,MKIV

~

~ <;:::"

AM PAR insertion

E-LTP

Fig. 9.4. Illustration of postsynaptic molecular events leading to induction of LTD or LTP after NMDAR stimulation. +P means phosphorylation; -P means dephosphorylation of the target protein, respectively. Meanings of all other symbols are explained in the text

188

9 Application ofCNGM to Learning and Memory

Some researchers argue that sometimes there may be an intermediate phase between E-LTP and L-LTP, which is dependent upon a local protein synthesis in spines based upon the pre-existing mRNA (Bortolotto and Collingridge 1998). In particular the activation of metabotropic glutamate receptors (mGluRs) before or during tetanus can trigger transcriptionindependent local protein synthesis and thus prolong the duration of LTP (Raymond et al. 2000). There is a local machinery for protein synthesis at spines and dendrites consisting of polyribosomes, tRNAs, initiation factors and mRNAs for glutamate receptors, structural proteins and kinases like CaMKII (Steward 1997). 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. Now we want to tum our attention to the transition from E-LTP to LLTP, especially the process of initiation of gene transcription, nuclear protein synthesis and its consequences. Neuronal nucleus has many different types of input (multiple converging signal-transduction cascades) and generates many different types of output (multiple genes that can be expressed in different patterns and at different levels). Some of these gene expressions are in response to inputs derived from synaptic stimuli. Extensive studies of stimulus-dependent gene expression in CNS neurons have shown that nuclear signaling pathways are able to discriminate between features of the electrical stimuli, such as their frequency, intensity, duration or pattern of repetition (Bito et al. 1997). As we have seen, such discrimination may well involve Ca 2 + signaling mechanisms. In principle, Ca 2+ influx could link up to Ca 2+ targets with different Ca 2+ sensitivity, thereby providing discrimination between Ca 2+ signals of varying amplitudes, like we have seen it happened for calcineurin and calmodulin (Fig. 9.4). In construction of our learning and memory GPRN we have to address several questions: 1. What are the various molecular pathways that link patterns of synaptic activity with gene expression? 2. What are the specific transcription factors involved in stimulustranscription coupling? 3. What are the specific downstream genes initiated by synaptic activity and how is their expression modified? 4. How do these genes give rise to activity-induced changes of neuronal synapses and what is the nature of these synaptic changes? 5. How is the expression of transcription factors themselves modified? The answer to the first question is summarized in Fig. 9.4. Upon repeated activation of NMDAR (and/or G-protein-coupled receptor) adenylyl cyclase catalyzes production of cAMP. In tum, cAMP activates

9.2 Toward a GPRN of Synaptic Plasticity

189

PKA. Prolonged Ca2+ increase activates PKA and ErkiMAPK. CaMKII activates CaMKIV, and all these kinases translocate to nucleus to activate gene transcription via CREB, which is the main transcription factor linking synaptic stimulation with gene expression (Mayford and Kandel 1999). CREB is activated by phosphorylation of Serl33. Phospho-CREB (pCREB) induces specific immediate-response genes via CRE (cAMPresponsive element) in their promoter regions. Negative regulation of CREB in hippocampal neurons has been found to occur through calcineurin-dependent regulation of nuclear PP1. Sustained but not transient elevation of nuclear CREB phosphorylation is required for efficient stimulus-transcription coupling. Prolongation of the synaptic input on the time scale of minutes, and the activity-induced inactivation of calcineurin pathway, greatly extends the period over which pCREB levels are elevated, thus affecting induction of downstream genes (Bito et al. 1997). The correlation between the duration of LTP and CREB phosphorylation was tested by quantifying changes in CREB phosphorylation throughout the induction and maintenance of L-LTP in pyramidal neurons in mature hippocampalentorhinal cortex slices (Leutgeb et al. 2005). High-frequency stimulation of Schaffer collaterals consisted of 2 stimulus trains of 100 pulses at 100 Hz with a 10 min intertrain interval. LTP was recorded after increasing time intervals, and lasted for 4 hours. At the same time intervals, measuring the extent of CREB phosphorylation at a single cell resolution, using confocal microscopy, revealed that the pCREB/CREB ratio was significantly increased already at 30 min after the LTP induction, and continued to increase until the end of measurement, i.e. after 4 hours. In another study, done in freely moving rats, LTP was evoked in hippocampal dentate gyrus and phosphorylation of CREB was measured by immunocytochemistry (Schulz et al. 1999). CREB phosphorylation occurred in a biphasic manner, with a first short-lasting peak at 30 min, almost a zero minimum at 1 hr, and a second long-lasting peak beginning 2 hr after tetanic stimulation and lasting for at least 24 hr. In (Schulz et al. 1999), only synaptic stimulation that generated nondecremental L-LTP promoted a sustained hyper-phosphorylation of CREB but not stimulation that produced deeremental E-LTP. Strong stimulation consisted of 20 trains of 15 impulses of the frequency 200 Hz with the inertrain interval of 5 s. Weak stimulation consisted of only 2 trains. Animals that received a weak stimulation showed (if at all) a small and transient rise in CREB phosphorylation within 2 hr after stimulation, and after 6 hr, none of the animals showed any significant CREB phosphorylation. This study is in accordance with the discovery that different protein kinases that phosphorylate CREB, do so with different temporal courses (Wu et al. 2001). In hippocampal neurons, fast CaMK-dependent pathway is followed by a slower MAPK path-

190

9 Application of CNGMto Learning and Memory

way . In vitro pCREB formation in the culture of dissociated CA3 /CAI pyramidal neurons was overwhelmingly dominated by the CaMK phosphorylation during the first 0-10 min, with the peak at around 5 min, and then decreasing over the next hour. MAPK-driven CREB phosphorylation slowly increased with the peak at 60 min (end of measurement). At 30 min poststimulus, both kinases contributed about equally to the CREB phosphorylation. Thus, several waves of CREB phosporylation and subsequent gene induction can follow after synaptic stimulation. Now, we would like to know , what are the downstream genes regulated by pCREB or in other words which proteins are synthesized due to the action of CREE. In Aplysia, result of CREB-induced immediate-response genes (with CRE in their promoter regions) transcription are two proteins (Mayford and Kandel 1999): (I) ubiquitin-hydrolase recruits ubiquitin proteosome, which cleaves the inhibitory subunit of PKA and thus causes its persistent activity and prolongation of facilitation up to 12-24 hrs, (2) transcription factor clEBP , which binds to the CAAT region of DNA and initiates transcription of proteins that are coded for by late genes and are responsible for growth of new synapses and a prolonged stabilization of synaptic transmission. Long-term plasticity of the central nervous system (CNS) in mammals involves induction of a whole set of genes whose identity and purpose is not completely characterized. Attempts to identify candidate plasticityrelated genes (CPGs) in the hippocampus dentate gyrus (DG) yielded 362 upregulated CPGs and 41 downregulated transcripts (dCPGs) (Hevroni et al. 1998). Of these , 66 CPGs and 5 dCPGs are known genes that encode for a variety of signal transduction proteins, transcription factors, trophic factors, various enzymes (kinases, phosphatases, etc) and structural proteins. To test relevance to neural plasticity, 66 CPGs were tested for induction by stimuli producing long-term potentiation (LTP). Approximately one-fourth of the genes examined were upregulated by LTP. These results indicate that an extensive genetic response is induced in mammalian brain after glutamate receptor activation, and imply that a significant proportion of this activity is coinduced by LTP. Based on the identified CPGs, it is conceivable that in fact multiple cellular mechanisms underlie long-term plasticity of the nervous system . In a more recent study, a powerful geneidentification method, differential analysis of primary cDNA library expression (DAzLE), and cDNA microarray from primary cortical neurons were used to identify plasticity-induced genes (PLINGs) (Hong et al. 2004) . 661 out of 1,152 arrayed genes are found to be consistently upregulated between I and 24 hr with differential expression patterns after NMDA receptor activation. This indicates that there is a broad and dynamic range of long-lasting neuronal responses that occur through NMDA

9.2 Toward a GPRN of Synaptic Plasticity

191

receptor activation. Based on a different temporal course of expression after NMDARs activation, 6 groups of late plasticity-induced genes were distinguished. Based on their functional categorization, 20 gene categories were identified, i.e.: neurotransmitter related, cellular receptors, cell survival, proteolysis, microtubule function related, proteosome components, RNA processing/stability, ion pumps/transporters, protein folding /chaperones, cell adhesion/matrix, actin cytoskeleton regulation, protein/vesicle trafficking, energy metabolism, transcription regulators, cellular oxidation-reduction, intracellular signaling, and finally genes with an unknown function. Next step is to use gene knockout mice to learn which of these genes are crucial for learning and memory formation. Mutant mice lacking a-and 8isoforms of CREB display intact short-term memory but deficient longterm memory in three independent learning tasks concomitantly, while LLTP in hippocampal CA1 is also impaired. In contrast to the striking phenotype of the CREB-deficient mice, knockouts of immediate early transcription factor zif/268 (also known as NGFI-A), NFAT (nuclear factors of activated T cell) and CAMP-response element modulator (CREM), all of which are induced due to NMDARs stimulation, have not yet yielded remarkable behavioral changes. Deletions of c-fos and fosB cause behavioral abnormalities, but their neurobiological basis is not understood (Bito et aI. 1997). These results are surprising since immediate-early genes for transcription factors like c-fos.fosis, zif268, CREM were among the first to be believed to bring about long-term plastic changes. More work is needed to discover which genes and proteins related directly to the long-term changes in synaptic weights are being expressed due to the action of CREE. However, recently it has been discovered that 24 hours of activation of CaMKIV and CREB results in the generation of new synapses that contain both new NMDARs and new AMPARs in CAl pyramidal cells (Marie et aI. 2005). CREB is probably involved in generation of the socalled silent synapse, i.e. synapse that contain only NMDARs. CaMKIV is probably involved in AMPARs insertion. Silent synapses are unsilenced during LTP by insertion of new AMPARs into the postsynaptic membrane. The process of new spine growth requires about 20 min (Maletic-Savatic et aI. 1999), thus the rate limiting processes should be related to initiation of gene transcription, subsequent gene translation and protein synthesis. To complete the plasticity GPRN, we must also consider how expression of CREB itself is regulated (Mayford and Kandel 1999, Smolen et al. 2000). CREB activator is called CREB-l a, and it gets into active state also by phosphorylation. CREB2 and CREB-l b are endogenous repressors of CREB-la. CREB also induces ICER (cAMP-responsive early repressor). ICER then binds to CREs and represses its own transcription as well as

192

9 Application ofCNGM to Learning and Memory

that of CREB. Also, a positive feedback-loop appears to exist, with CREB binding to its own gene CREs and activating its own transcription. PhospoCREB binds with the CREB-binding protein (CBP) that has been shown to be a co-activator of transcription not only for CREB, but also for a number of other transcription factors, including (but not necessarily restricted to) activator protein 1 (API), early region 1A (ElA) protein, nuclear hormone receptors (NHRs), and STATs (signal transducers and activators of transcription) (Bito et al. 1997). The convergence of signals onto CBP is all the more remarkable in light of evidence suggesting an additional role for CBP as a histone acetyltransferase critical for transcriptional initiation. Taken together, the multiplicity of signaling mechanisms acting on CREB and CBP provide a rich array of possibilities for input-specific patterns of gene expression. Figure 9.5 summarizes the core gene/protein regulatory network around CREB, and the long-term synaptic changes resulting from CREB activation. CREB2

CREB-1 b

-\ 1-

AM PARs NMDARs (new synapses)

+

CREB-1a

\+

+

+

_r(;;ER

STATs AP1

E1A

Fig. 9.5. Gene regulatory network for CREB, the major transcription factor of learningand memory, whichwhen activated by signal-transduction pathways from stimulated synapses leads to new protein synthesis and new synapse formation. This network is by no means exhaustive

9.3 Putative Molecular Mechanisms of Metaplasticity

193

9.3 Putative Molecular Mechanisms of Metaplasticity Next we have to consider molecules involved in metaplasticity, plasticity of synaptic plasticity itself (Abraham and Bear 1996). In theoretical equations, processes of LTP and LTD are metaplastically modulated by the variable(s) called modification threshold(s) that depend on the previous history of cell activity and upon other inborn and age/state related factors. Although NMDARs mediate LTD and LTP, they are regulated separately from AMPARs (Turrigiano and Nelson 2000, Sheng and Lee 2001). Therefore changes in numbers/properties of NMDARs rather belong to several candidate mechanisms underlying metaplasticity as we will see below. In general, any molecular action, which affects the characteristics of temporal changes in [Ca2+]i during and after NMDAR stimulation, in tum influences both plasticity and metaplasticity (Abraham and Tate 1997). We will mention these main metaplasticity candidates while not claiming the list is exhaustive: 1. NMDARs, since they playa critical role in postsynaptic calcium entry, which in tum plays a fundamental role in bidirectional synaptic modification. As such, changes in NMDAR number/function will dramatically alter the properties of experience-dependent synaptic plasticity and metaplasticity. NMDARs are heteromeric ion channels, composed of NR 1 and NR2 subunits. Each of four subtypes of the NR2 subunit (2A-2D) confers distinct functional properties to the receptor. NMDARs composed ofNRI and NR2B subunits mediate long duration currents (i.e. larger [Ca2+]i), whereas inclusion of the NR2A subunit results in NMDARs with faster kinetics (i.e. lower [Ca2+] J NMDARs composed ofNRl and NR2B are observed in the neocortex at birth, and over the course of development there is an increase in the ratio NR2A/ NR2B (Bliss 1999). Thus the age can set the effective position of ~ expressed by the constant a (Eq. 9.4). In fact, in previous computer modeling studies, a was introduced as a constant and for the immature cortex it was about ten times smaller than a for the mature cortex (Clothiaux et al. 1991, Benuskova et al. 1994). That means, in computer simulations of cortical plasticity, effective position of the modification threshold was higher for the adult cortex than for the immature cortex. Another critical factor for the effective position of ~ is the number of NMDARs. When the NMDAR number is reduced to one half, like in the neocortex treated by alcohol during prenatal development, value of a in the model that successfully simulates impaired synaptic plasticity increases twice compared to the model of the normal neocortex

194

9 Application of CNGMto Learning and Memory

(Benuskova, Rema et aI. 2001). Thus, the value of the constant a (Eq. 9.4) can depend on the inborn and/or genetic factors. In addition, NMDARs can also contribute to the changes in the temporal average of activity expressed by (C(t)zM. In the immature cortex, the composition and thus the function of synaptic NMDARs can be acutely and bidirectionally modified by severely altered input sensory activity on the time scale of hours (sensory activation) and days (sensory depression) (Philpot et aI. 2001). Thus ~ changed over hours in the biophysical model of the BCM rule (Castellani et aI. 2001). Thus changes in the NMDAR subunit content can contribute to a slow change of ~ (t). 2. Neuromodulators. It appears that ~ effective position expressed by the function cAt) may be affected by various modulatory factors like various neurotransmitters, neuromodulators, hormones, etc. For instance, binding of corticosteroids to its glucocorticoid receptor results in a shift of the LTP modification threshold to the right, and a shift of the LTD modification threshold to the left (Kim et al. 1996). Stress and glucocorticoids appear to exert a metaplastic effect through the modulation of Ca 2+ levels. These data indicate that B_, postulated by Artola, Brocher and Singer (Artola et al. 1990), may be modifiable as well as ~ (B+) and that a single stimulus such as glucocorticoid receptor activation may shift these two thresholds in opposite directions. These effects can be either temporary or permanent, depending on whether the alterations in levels of neuromodulators are temporary or permanent. Actually, it seems that certain levels of neuromodulatory substances like acetylcholine (ACh), and maybe dopamine (DA) and norepinephrine (NE) as well, are necessary for synaptic plasticity to occur at all. For instance when the cholinergic basal forebrain fibers are completely destroyed, synaptic plasticity in the somatosensory cortex is completely abolished as well (Sachdev et aI. 1998). It has been shown that application of Ach and or NE or their agonists facilitates the expression of both LTD and LTP in cortical preparations, as if the thresholds B_ and ~ (B+) were both shifted to the left (Brocher et aI. 1992, Kirkwood et al. 1999). Thus, different neuromodulators can have different effects on the position of B_ or ~ ( B+) or both. 3. CREB/CRE induced gene expression. Stimulation ofNMDARs leads to the CREB/CRE induced gene expression that further facilitates maintenance ofLTP (Schulz et al. 1999, Wu et al. 2001, Leutgeb et aI. 2005). In the next section we will demonstrate how to incorporate the dynamics of CREB phosporylation into the modification threshold calculation. 4. Most promising candidate mechanism for relatively fast postsynaptic activity-dependent sliding of ~ (expressed by (C(t)zM in Eq. 9.4) is the

9.3 Putative Molecular Mechanisms of Metaplasticity

195

dynamics of CaMKII. The temporal 8M integration window for neuron's memory on past activity was equated to 22 min in the modeling study of visual cortex (Clothiaux et al. 1991), from seconds to minutes in the modeling studies of somatosensory cortex (Benuskova et al. 1994, Benuskova, Rema et al. 2001). Increased intracellular Ca2+ results in autophosphorylation of CaMKII, thereby converting the enzyme to a Ca2+ independent (autonomous) form (Abraham and Tate 1997). It has been proposed, that the level of Ca2+-independent CaMKII sets the current value of 8M (Bear 1995). It appears that higher level of autonomous CaMKII shifts the LTP threshold to the right and so the subsequent incoming signal may be less effective in inducing LTP. Phosphorylated CaMKII binds Ca2+ and calmodulin more tightly and limits the availability of Ca2+ and calmodulin for other enzymes essential for LTP to be established. This may result in a metaplastic state such that a second burst of synaptic activity might lead to such elevation in Ca2+ concentration that could be now in the range for selective activation of specific phosphatases. In this way the induction of LTD may be facilitated (Abraham and Tate 1997). It has been shown that the effect of tetanus on the kinase and its activity is confined to both synapses and neuronal soma. In addition, in dendrites the increase of phosphorylated CaMKII is accompanied by the increase of nonphosphorylated CaMKII (Ouyang et al. 1997), probably due to the local synthesis (Wu et al. 1998). To sum up, it seems that 8M is comprised of some constant portion dependent upon genetically based or inborn factors (like the number of NMDARs, overall levels of neuromodulators, and so on) that sets the effective position of 8M on the x-axis. Into this constant component, agerelated factors can fall too, because they can be considered constant over the time of measurement. Then there is a slowly changing component over hours or days that can be for instance dependent upon the composition of NMDARs, other activity-related slow influences like motivation or attention related chemicals, and CREB/CRE-induced gene expression. Last but not least, there is a fast component that adjusts 8M in the matter of minutes. This latter component can be dependent for instance on the phosphorylation state of CaMKII and perhaps also on other fast mechanisms and molecular changes. Perhaps the choice of factors entering the calculation of 8M depends on the experimental situation, which we want to model. Many other influences will be hidden in the proportionality constants and scaling factors. In the next section we step towards the construction of CNGM of synaptic plasticity that underlies learning and memory formation.

196

9 Application ofCNGM to Learningand Memory

9.4 A Simple One Protein-One Neuronal Function CNGM We will formulate a set of rules and equations that can be used in any of neuron spiking models (Maass and Bishop 1999). For better clarity, we will repeat those equations, which were introduced earlier, and which are relevant to our CNGM of learning and memory. The main difference against the previous interpretations will be that some of the parameters will be governed by the dynamics of CREB/CRE-induced gene expression, which has been proven to be critical for the long-term memory formation (Poser and Storm 2001). Thus, we will assume that each cortical excitatory synaptic weight changes according to the additive nearest neighbor STDP rule (former Eq. 9.2), i.e.: wet + Sf) = w(t)(l + ~w+ - ~wJ

(9.10)

where & is the time step of weight updating, let us say 1 ms. Synaptic change is comprised of contributions from only two nearest spikes. That is for each presynaptic spike, only two postsynaptic spikes are considered: the one that occurs right before and the one that occurs right after the given presynaptic spike, respectively, so that L'lw+ = A+ exp(-M / rJ for M > 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

::


..

0 .01 0 -0 .0 1 -0 .02 -0 .0 3 -0 .04 -0 .0 5 - 1 0 0 -8 0

-6 0

-40

-2 0

0

20

40

60

80

100

T im e difference (ms)

Fig. 9.8. Positive and negative weight changes (above and below zero on the yaxis, respectively) recorded during the course of the model GC simulation as a function of the temporal difference between the postsynaptic and presynaptic spike timing

Changeable windows for the induction of LTD and LTP are crucial factors for the successful reproduction of experimental results. Whether these windows are changeable and whether this dynamics indeed depends on the average past activity of the postsynaptic cell, remains to be proven. So far, it is one of the predictions of our model. Another prediction of our model

9.6 Sununary and Discussion

203

is that the width of temporal windows for the induction of LTDILTP also depends on particular long-term processes like the time course of concentration of biochemical factors that are necessary for the induction and maintenance of LTP. In our particular experimental study of hippocampal L-LTP we have employed the time course of the elevation of concentration of phosphorylated CREB, the main transcription factor involved in the induction and maintenance of the late LTP known nowadays. It is to be expected that there are many factors influencing the LTDILTP windows . In fact, we have made a preliminary list of them in section 9.3 on putative mechanisms of metaplasticity. It may be a worthy exercise to incorporate these other factors into our learning rule to model different kinds of experimental situations. In this chapter, we have given the neurogenetic basis and have introduced the construction of CNGM of learning and memory at a cellular level. It should be mentioned that it is not the only way how to construct CNGM of learning and memory. Instead, this theory and methodology should serve as a recipe of how to proceed in the set up of such a model and how to verify it on the real data. As an experimental study we have modeled synaptic plasticity phenomena in hippocampal dentate gyrus (Schulz et al. 1999). This brain structure is often studied in relation to mechanism of learning and memory. Our CNGM is of course based on the present knowledge of molecular mechanisms of learning and memory. These can be much more elaborated in future years, although hopefully some of present knowledge will be incorporated. At the same time, the methodology of current CNGM can be used with any particular knowledge of molecular mechanisms of learning and memory to bring more insight into the basis ofleaming and memory in the brain.

10 Applications of CNGM and Future Development

This chapter presents information on neurogenetic causes of brain diseases and points to future directions of building CNGM for these diseases. With the advancement of molecular research technologies more and more data and information is available about the genetic basis of neuronal functions and diseases (see Table A.I in Appendix 1). This information can be utilized to create models of brain functions and diseases that include models of gene and protein interactions, i.e. computational neurogenetic models (CNGM) (Kasabov and Benuskova 2004,2005). This research has many open questions, some of them listed below we will attempt to address in this chapter: 1. Which real neuronal parameters are to be included in an ANN model of a particular brain condition and how to link them to activities of genes/proteins? 2. In turn, which genes/proteins are to be included in the model and how to represent gene interaction over time within each neuron? 3. How to integrate in time the activity of genes, proteins and neurons in an ANN model? 4. How to integrate internal and external variables in a CNGM (e.g., genes and neuronal parameters with external signals acting on the brain) and how to treat various perturbations? 5. How to create and validate a CNG model in the presence of scarce data? 6. How to measure brain activity and the CNGM activity in order to validate the model? 7. What is the effect of connectivity? How it interacts with the effects of GPRN? 8. What useful information can be derived from CNGM? These and probably many more questions remain to be addressed in the future. We will suggest answers to some of them, or at least directions in which the answers may lie. In the present chapter we will suggest the construction of CNGM for these diseases and processes: epilepsy, schizophrenia, mental retardation, brain aging and Alzheimer disease, and Parkinson

206

10Applications ofCNGM and FutureDevelopment

disease. We conclude the final chapter with the introduction of the braingene ontology project.

10.1 CNGM of Epilepsy Epilepsy is a disorder characterized by the occurrence of at least two unprovoked seizures (Cavazos and Lum 2005). Seizures are the manifestation of abnormal hypersynchronous discharges of neurons in the cerebral cortex. The clinical signs or symptoms of seizures depend on the location and extent of the propagation of the discharging cortical neurons. The prevalence of active epilepsy is about 1%, which however means about 50 millions people worldwide are affected (Cavazos and Lum 2005). People at any age can develop epilepsy. However, it has been noted that around half of people developing epilepsy do so before the age of 15 years. Recent epidemiological evidence suggests, however, that increasing numbers of patients are developing epilepsy in old age: this is partly because of demographic changes in the population (with an increasing proportion of the population in the elderly age range) and also due to an increasing incidence of degenerative cerebrovascular disease in old age. Most, but not all, studies have found a slight male preponderance (Sander 2003). Seizures are often a common, nonspecific manifestation of neurologic injury and disease, which should not be surprising because the main function of neurons is the transmission of electrical impulses. A genetic contribution to a etiology has been estimated to be present in about 40% of patients with epilepsy (Gardiner 2003). Pure Mendelian epilepsies, in which a single major locus can account for segregation of the disease trait are considered to be rare, and probably account for no more than 1% of patients (Gardiner 2003). The common familial epilepsies tend to display complex inheritance, in which the pattern of familial clustering can be accounted for by the interaction of several loci together with environmental factors. 10.1.1 Genetically Caused Epilepsies The following Table 10.1 is devoted to different types of genetically caused epilepsies, associated brain pathologies, symptoms, and putative mutated genes to be included in future CNGM. This table is by far not complete, since every day new data on genes involved in the so-called idiopathic epilepsies (i.e. those with unknown cause) emerge.

10.1 CNGM of Epilepsy

207

Table 10.1. Epilepsies caused genetically, putative mutated genes, and affected functions of brain neurons in humans

Epilepsy Mutated genes / Brain abnormality chromosome location (if

Symptoms

References

Autosomal dominant nocturnal frontal lobe epilepsy (ADNFL)

a4 subunit of the nicotinic AchR (CHRNA4) / 20q 1. ~2 subunit of same receptor (CHRNB2) / lp

Reduced nAchR channel opening time and reduced conductance leading to hyperexcitability.

Partial seizures during night that may generalize, arising from the frontal lobes, motor, tonic, postural type.

(Gardiner 1999, Meisler et al. 2001, Stein le in 2004)

Benign familial neonatal convul-

EBN1 (K+ chan- Alteration of the gatnel gene KCQ2) ing properties of the /20q K+ channel leading to poor control of repetiEBN2 (K+ chan- tive firing. nel gene KCNQ3)/8q

Generalized epilepsy of the newborns, seizures is frequent and brief, episodes resolve within a few days.

(Gardiner 1999, Meisler et al. 2001, George 2004, Steinlein 2004)

SIOns

(BFNC1 and BFNC2) Childhood absence epilepsy (CAE)

y2 subunit gene for the GABA A receptor gene GABRG2 / 5q gene CLCN2 / 3q

Fast and part of slow GABAergic inhibition is reduced, voltage-gated Cl" channel function is impaired.

Absence seizures (consciousness impaired) up to 200 times a day, bilateral 2-4 Hz spike and slowwave EEG.

(Crunelli and Leresche 2002, Marini et al. 2003, Steinlein 2004, Segan 2005)

Generalized epilepsyand febrile seizures plus (GEFS+)

~1 subunit of the Na+ channel gene SCN1B / 19q a1 and a2 subunits, gene SCN1A and gene SCN2A/2q GABRG2/5q

Normal inactivation kinetics of the Na+ channel is reduced causing persistent Na+ influx and hyperexcitability, reduced function of the GABAAR.

Childhood onset offebrile seizures, with febrile and afebrile generalized seizures continuing beyond 6 yrs of age.

(Gardiner 1999, Meisler et al. 2001, Steinlein 2004)

208

10 Applications of CNGM and Future Development

Epilepsy Mutated genes / Brain abnormality chromosome location (if known)

Symptoms

Intractable childhood epilepsy

al subunit of the Na+ channel, gene SCNIA / 2q

Rapid recovery of the Na+ channel from inactivation or very slow inactivation.

Frequent intracta- (George 2004) ble generalized tonic-clonic seizures.

Juvenile absence epilepsy (JAE)

al /5q, a5 /I5q, y2 /5q subunit genes for the GABA A receptor gene (CLCN2) / 3q

Fast and part of slow GABAergic inhibition is reduced, voltchannel age-gated function is impaired.

Similar like CAE (George but the seizures 2004, Sestart after year 10, gan 2005) seizures may be less frequent and last longer than few seconds.

Juvenile myoclonic epilepsy (JME)

a7 subunit of the nicotinic AchR (CHRNA7) / 15q gene CLCN2 / 3q, 134 subunit of Ca 2+ channel, (CACNB4) / 19p

Reduced function of the nicotinic AChR, voltage-gated Clchannel and voltage-gated Ca2+ channel have reduced conductance.

Myoclonic jerks or seizures shortly after awakening, generalized tonicclonic seizures, and sometimes absence seizures.

(Gardiner 1999, Cavazos and Lum 2005)

Dravet syndrome, severe myoclonic epilepsy of infancy (SMEI)

a I subunit of Complete loss of acthe Na+ channel, tivity of the Na+ gene SCN1A / channel. 2q

Both generalized and localized seizures, clonic and myoclonic seizure types.

(George 2004, Steinlein 2004)

Lafora disease (progressive myoclonus epilepsy)

Laforin gene EPM2A / 6q24 malin gene EPM2B /6p22.3

cr

Presence of Lafora bodies (granules of accumulated carbohydrates).

References

Myoclonic jerk- (Ganesh et ing, ataxia, mental al. 2006) deterioration leading to dementia.

10.1 CNGM of Epilepsy

209

10.1.2 Discussion and Future Developments

There are many computational models of temporal lobe-like epileptic seizures based on neural network dynamics (Traub et al. 1987, Biswal and Dasgupta 2002, Wendling et al. 2002, Kudela et al. 2003). We distinguish these latter models from models of thalamocortical slow spike-and-wave seizures like in CAE (Lytton et al. 1997, Destexhe 1998, Robinson et al. 2002). However none of existing models take into account the influence of genes and their internal networks upon neural dynamics, which is the goal of CNGM. Using the information about genetic causes of epilepsies listed in Table 10.1, existing neural network or dynamical models can be enhanced to incorporate this information and thus lead to future CNGM. There is quite high percentage of pharmacoresistant epilepsies, i.e. 30-40% (Fisher et al. 2003), therefore, it is important to seek new ways of modeling epilepsy to gain more insight into the mechanisms of seizure spread and maintenance in seizure-prone cortex. By the development of new computational models the aim is to better understand the mechanisms of drug resistance. In humans with intractable temporal lobe epilepsy (TLE), many of surviving inhibitory interneurons lose their PV content or PV immunoreactivity (Wittner et al. 2005). It has been proposed that efficient Ca2+ buffering by PV and its high concentration in PV-expressing inhibitory cells is a prerequisite for the proficient inhibition of cortical networks (DeFelipe 1997). To investigate this hypothesis, Schwaller and coworkers used mice lacking PV (PV-/-), which had previously been produced by homologous recombination. These mice show no obvious abnormalities and do not have epilepsy (Schwaller et al. 2004). However, the severity of generalized tonic-clonic seizures induced by Pentylenetetrazole (PTZ) was significantly greater in PV-/- than in PV+/+ animals. Extracellular single-unit activity recorded from over 1000 neurons in vivo in the temporal cortex revealed an increase of units firing regularly and a decrease of cells firing in bursts. In addition, control animals showed a lesser degree of synchronicity and mainly high frequency components above 65 Hz in the LFP spectrum compared to PV-/- mice. On the other hand, PV-/- mice were characterized by increased synchronicity and by abnormally high proportion of frequencies below 40 Hz (Villa et al. 2005). In the hippocampus, PV deficiency facilitated the GABAAergic current reversal induced by high-frequency stimulation, a mechanism implied in the generation of epileptic activity (Vreugdenhil et al. 2003). Through an increase in inhibition, the absence of PV facilitates hypersynchrony through the depolarizing action of GABA (Schwaller et al. 2004). In setting up the future CNGM, for modeling the generation of LFP we can use the SNN model of cerebral cortex introduced in the end of Chapter

210

10Applications of CNGM and Future Development

4. Or we can use more sophisticated simulation environments like GENESIS (Bower and Beeman 1998) and NEURON (Carnevale and Hines 2006). Behind each of the neuron's parameters there is a particular protein be it the ion channel or postsynaptic receptor or enzyme. Proteins in neurons like receptors and ion channels are complex proteins comprised of several subunits each of them is coded for by a separate gene. These genes are expressed in a coordinated manner so we can treat them as a one gene group G, with the overall normalized expression level or we can treat each subunit individually. During the time interval of LFP measurement, we can assume that the gene expression level is constant but at the same time depends on expression levels of all genes in the selected set of genes such that (10.1) where (J is the sigmoid function between 0 and 1, gk(t) is the expression level of gene k at time t and Wjk E (-1, I) is the coefficient of an abstract gene interaction matrix W. Neuron's parameter value Pit) is proportional to gene group expression level git) such that Pit)

=

P(O) git)

(10.2)

where Pj(t) is the value of parameter j at time t, P(O) is the initial value of that parameter and gj(t) E (0, I) is the normalized level of expression of thej" gene in the model GRN. Genes and proteins involved in GRN should be comprised of ion channels, excitatory and inhibitory receptors with fast and slow kinetics, calcium buffering proteins, etc. Then we can use optimization procedures as introduced in Chapter 8 to find the coefficients of the gene interaction matrix. The objective function for optimization will be mouse LFP with particular spectral characteristics. Once we have obtained the gene interaction matrix W, we can simulate gene deletion and mutation and drug effects upon LFP characteristics, spiking activity of neurons, synchronicity, oscillations, etc. Predictions from CNGM can aid experimental research of causes of idiopathic epilepsies and drug resistance for various drugs.

10.2 CNGM of Schizophrenia It is estimated that there are about 50 different neurotransmitters acting in the human brain. The three major categories of substances that act as neurotransmitters are (I) amino acids (primarily glutamate, GABA, aspartic acid and glycine), (2) peptides (vasopressin, somatostatin, neurotensin,

10.2 CNGM of Schizophrenia

211

etc.) and (3) monoamines (norepinephrine, dopamine and serotonin) plus acetylcholine. There are also other categories like opioids, tachykinins, and so on. A vast majority of neurotransmitters is being produced in evolutionary older subcortical nuclei. It can be said, that almost every subcortical nucleus has its own neurotransmitter(s) (they may synthesize more than just one). These different subcortical groups of neurons and their axons which are sent and spread all over the brain including the cortex constitute different neurotransmitter systems. Serotonin (5-HT, 5hydroxytryptophan), dopamine (DA), and noradrenaline (norepinephrin) are produced in three subcortical nuclei (Fig. 10.1).

Prefrontal associat ion cortex

Occipital cortex

Limbic asociation cortex

Fig. 10.1. Schematic illustration of diffuse projections that originate in the noradrenergic (NA), serotonergic (5-HT) and dopaminergic (DA) nuclei in the brain stem, and spread all over the cerebral cortex (projections into other brain parts are not illustrated)

On the other hand, major excitatory neurotransmitter of cortical excitatory neurons is glutamate. Major inhibitory neurotransmitter in cortex is GABA (y-amino-butyric acid). However in effect, many neurotransmitters act upon every neuron within a cortex. Action of different neurotransmitters causes influx of different ions and activations of different second messengers. Thus, the information-processing properties of neurons are subject to complex short- and long-term influences. Different neurotransmitter systems and thus different subcortical and cortical brain parts are related to different mental processes and functions. Extraordinary low or high activity of neurotransmitter systems leads to disorders of mental functions

212

10Applications ofCNGM and Future Development

(Kaplan and Sadock 2000). Thus, neurotransmitters themselves are endogenous psychoactive drugs. Since levels of neurotransmitter production in every nucleus are pre-programmed genetically, a gene error or gene mutation can lead to a permanently altered production of a particular neurotransmitter, its receptor(s), or any molecule in the second messenger systems. Neuronal genetic program itself is subject to such influences as hormones, stress, aging, and as we now know, also to activity of neurons themselves. The so-called psychoactive substances, that is drugs and medicines used to treat mental disorders, act in synapses upon different types of receptors. They are psychoactive because their chemical structure resembles that of endogenous neurotransmitters. Thus, they can chemically bind to their receptors. Neurons react as if there was a natural neurotransmitter present. Biochemical and neuroscientific research finds systematic and reproducible changes in the brain that are typical for particular mental disorders like depression and schizophrenia (Kaplan and Sadock 2000). There are dozens of studies, which demonstrate that schizophrenia has the so-called polygenetic type of heredity. Individuals carrying a specific genetic predisposition (i.e. complex tiny genetic mutations) mayor may not develop a disorder. It is said that these individuals are vulnerable against a given disease. Individuals carrying these mutations are vulnerable in a sense that various life stress events may trigger or switch on the disease process. If the endogenous genetic predisposition is too strong, the process of disease may start spontaneously, that is without a triggering external event. 10.2.1 Neurotransmitter Systems Affected in Schizophrenia

Schizophrenic symptoms can be very bizarre and vary a lot from person to person. Statistically, about 1% of population suffers schizophrenia regardless of the race, gender, education, family and social status, etc. There are several subtypes of schizophrenia with typical symptoms, different severity of symptoms and different courses of the disease. The most characteristic symptoms are delusions, hallucinations, and various thinking and perceptual disorders. Schizophrenic withdrawal from reality can manifest itself in many peculiar ways. Disorder is accompanied by serious deterioration from previous level of functioning in such areas as work, social relations, and self-care. Neuropathological research on the brains of schizophrenics has shown that there are specific alterations on some types of neurons within the frontal and temporal cortices (Kaplan and Sadock 2000). These cortical neurons seem to be less mature, with altered and retarded signs of differentiation. These subtle neurodevelopmental morpho-

10.2 CNGM of Schizophrenia

213

logical abnormalities may be the consequence of defects in major neurotransmitter systems that affect the developing cortex. Major altered neurotransmitter systems seem to be the dopaminergic (DA), glutamatergic (Glu) and serotonergic (5-HT) systems. There are four relatively separated dopaminergic systems in the brain. The first one acts on hypothalamus and affects the neurohormonal secretion. The second one, acting on basal ganglia, plays a crucial role in executing muscular, particularly involuntary, movements. The third DA system, the so-called mesolimbic system, originates in ventral tegmentum in the brain stem and innervates the limbic system and the limbic association cortex. It regulates expression of emotions, feelings of satisfaction, reward and pleasure. The fourth DA system, the so-called mesocortical system, has its dopaminergic neurons also in the ventral tegmentum. These neurons send their axons to the frontal and prefrontal cortices. Release of DA in these cortices regulates motivation, concentration and goal-directed planned behavior, which requires a complex organization of thoughts. Based on experimental data, it is assumed that a disbalance ofDA levels in the brain includes its lack in the frontal and prefrontal cortices, and its excess in the limbic and subcortical areas. Reduction of DA levels in the frontal cortex results in the so-called hypofrontality, that is unusually reduced levels of activity in the frontal cortex, revealed by means of brain imaging techniques (DelaTorre et al. 2005). Reduced frontal activity is hypothesized to result in such cognitive deficits as a specific "emptiness", an absence of cognitive and emotional contents in thoughts and motivation. On the other hand, increased levels ofDA in the emotional and subcortical centers may lead to an impaired filtration and discrimination of stimuli, thus in tum leading to delusions in thinking and hallucinations in perception. For instance, a long-term abuse of amphetamines, e.g. drugs which increase levels of DA in the brain, results in development of paranoid delusions. The second major system altered in schizophrenia is a glutamatergic system. Glutamate is being released from axonal terminals of cortical excitatory neurons that make synapses with other cortical neurons and with subcortical neurons in the brain stem nuclei to which they relay their feedback influence. Neurochemical research has revealed a decrease in the number of glutamate receptors and glutamate itself in the frontal cortices of schizophrenics. Decreased glutamatergic transmission in frontal cortex can also contribute to the hypofrontality phenomenon. Importance of the role played in schizophrenia by the reduced action of glutamate is related to the fact that it is a neurotransmitter of learning. Long-term changes of synaptic weights occur in glutamate synapses through NMDA receptors. Thus, reduced levels of glutamate and NMDA receptors in the frontal cor-

214

10 Applications ofCNGM and Future Development

tex may lead to altered synaptic plasticity during learning (adaptation). The drug PCP (phencyclidin, "angel dust"), that blocks postsynaptic glutamate receptors and increases levels of dopamine and serotonin, causes similar psychic experiences like in schizophrenia, e.g. hallucinations, disorders of thinking and cognition, emptiness. Neurons that produce serotonin are located in the brain stem in the nucleus raphe. They send their axons all over the cerebral cortex, limbic system, basal ganglia, thalamus and hypothalamus. Target neurons possess different types of serotonin receptors, thus somewhere the serotonin acts as an excitatory NT and elsewhere as an inhibitory NT. Serotonin regulates sleep, appetite, and libido. It also suppresses aggressive behavior - it can be said that it is a neurotransmitter of "mental comfort". Antipsychotic medicines (previously called neuroleptics) used in psychiatry to alleviate schizophrenic symptoms, block postsynaptic receptors for DA. Differences between effects of different antipsychotics depend on which subtypes of DA receptors they affect. Recently, it has been found that antipsychotics that reduce not only the DA synaptic transmission but also a serotonergic synaptic transmission, have more desirable effects upon schizophrenic symptoms and can restore the activity in frontal areas of the brain (Honey et al. 1999). The involvement of 5-HT system is not surprising since it has been known for a long time, that LSD (lysergic acid dyethylamid) enhances serotonergic transmission and produces gross perceptual and thinking alterations. Treatment of schizophrenia is usually a lifelong process. Since medicines still do not cure the original biochemical defect, they can only help the shattered neurotransmitter systems to regain more or less stable balance.

10.2.2 Gene Mutations in Schizophrenia Twin and adoption studies demonstrated that susceptibility to schizophrenia is strongly heritable even if children are reared apart from their biological parents. When one twin has schizophrenia, the risk of schizophrenia in the co-twin is greater in monozygotic twins (45%) than in dizygotic twins (15%). However, 40% of the monozygotic co-twins of a person with schizophrenia are clinically normal (Cloninger 2002). Furthermore, the risk of illness decreases with degree of genetic relationship more rapidly than can be explained by a single gene or the sum of effects of several such genes. Thus, the inheritance pattern of schizophrenia suggests that multiple genes, each of small effect, interact nonlinearly with one another and with environmental factors to influence susceptibility to schizophrenia (Cloninger 2002). This prediction has been confirmed by more than 20 ge-

10.2 CNGM of Schizophrenia

215

nomewide linkage scans in more than 1,200 families of schizophrenics. These studies found evidence for several genes of small effect; that is, genes that modify susceptibility but are neither necessary nor sufficient to cause the disorder. However, no evidence was found for any genes with a large individual effect, such as a Mendelian subtype of schizophrenia. Linkage studies for susceptibility genes produced the list of regions of interest to include target regions on Iq21-q22 (Brzustowicz et al. 2000), 6q21-q22.3 (Cao et al. 1997), and 13q34. Chumakov and colleagues (Chumakov et al. 2002) describe a new human gene, G72, on chromosome 13q34 that interacts with the gene for D-amino acid oxidase (DAAO) on 12q24 to regulate glutaminergic signaling through the N-methyl-Daspartate (NMDA) receptor pathway. Using traditional positional cloning techniques of linkage and linkage disequilibrium, they show that both of these genes are associated with increased susceptibility to schizophrenia. Therefore, this is the first discovery ofa specific gene that also provides a pathogenic molecular mechanism that can account for the major symptoms of a psychiatric disorder. Similarly, two other groups reported that the gene dysbindin on 6p22.3 (Straub et al. 2002) and the gene neuregulin I on 8p (Stefansson et al. 2002) also influence susceptibility to schizophrenia and may operate via the same NMDA mechanism. Each of these gene discoveries came from association analysis targeting chromosomal regions first identified by linkage analysis. The linkage of schizophrenia to the 15q14 locus of the a-7 nicotinic receptor has also been replicated (Leonard et al. 2002). A gene called COMT (catecho-o-methytransferase) had long been suspected of being involved because it codes for an enzyme that breaks down dopamine after it is secreted into the synapse. COMT gene chromosome location is 22q 11.2. Egan et al. have linked a gene variant that reduces dopamine activity in the prefrontal cortex to poorer performance and inefficient functioning of the prefrontal cortex during working memory tasks, and to slightly increased risk for schizophrenia (Egan et al. 2001). The finding emerged from an ongoing study of people with schizophrenia and their siblings. Their brain imaging studies had revealed that both well siblings and patients falter on tasks of working memory and show reduced activation of the prefrontal cortex, which is required for this function. Studies have shown that the chemical messenger dopamine plays a pivotal role in tuning the activity of the prefrontal cortex during working memory tasks. People inherit two copies of COMT (one from each parent), each in either of two forms. The first variant, Val, reduces prefrontal dopamine activity, while a second form, Met, increases it. Egan et al. found that those who had inherited 2 copies of Val, on average, performed worse than those with only one copy. Those with two copies of Met performed best in the

216

10 Applications of CNGM and Future Development

working memory task. At the same time, people with two copies of Val had the lowest prefrontal activity. Brain activity of those with one copy of each variant was more efficient, while activity of siblings who inherited two copies of Met showed the highest brain efficiency, on average. Among 104 pairs of parents studied, the investigators discovered that the Val form of COMT was transmitted to offspring who eventually developed schizophrenia more often than would be expected by chance: 75 times for Val, compared to 51 times for Met. Inheriting two copies of the Val form accounts for a 1.5-fold increase risk for schizophrenia. It is not yet known exactly how the COMT Val variant impairs prefrontal efficiency. The researchers suspect that COMT's effect, while modest, may be amplified through interaction with other susceptibility genes and environmental factors. It seems that the COMT Val allele is certainly not a necessary or sufficient causative factor for schizophrenia. However, its biological effect on prefrontal function and the relevance of prefrontal function for schizophrenia implicate a mechanism by which it increases liability for the disorder. Although some insights into the etiology of schizophrenia have been developed from these studies, an understanding of the disease on the molecular level remains elusive. Efforts to identify molecular aberrations associated with the disease may be confounded by the subtle structural and cellular changes that occur and the polygenic nature ofschizophrenia. Neuropathological and brain imaging studies suggest that schizophrenia may result from neurodevelopmental defects. Cytoarchitectural studies indicate cellular abnormalities suggestive of a disruption in neuronal connectivity in schizophrenia, particularly in the dorsolateral prefrontal cortex (Arnold and Trojanowski 1996). Cellular aberrations, such as decreased neuronal size, increased cellular packing density, and distortions in neuronal orientation, have been observed in immunocytochemical and ultrastructural studies. Yet, the molecular mechanisms underlying these findings remain unclear. To identify molecular substrates associated with schizophrenia, DNA microarray analysis was used to assay gene expression levels in post-mortem dorsolateral prefrontal cortex of schizophrenic and control patients (Hakak et al. 2001). DNA microarray analysis is one such technique that allows for the quantitative measurement of the transcriptional expression of several thousand genes simultaneously. Genes determined to have altered expression levels in schizophrenics relative to controls are involved in a number of biological processes, including synaptic plasticity, neuronal development, neurotransmission, and signal transduction. Most notable was the differential expression of myelination-related genes suggesting a disruption in oligodendrocyte function in schizophrenia. Oligodendrocytes increase neuronal conduction velocity through their insulating properties and pro-

10.2 CNGM of Schizophrenia

217

vide extrinsic trophic factors that promote neuronal maturation and axonal survival. Myelination ofthe prefrontal cortex has been observed to occur in late adolescence and early adulthood, which is typically the age of onset of schizophrenia (Benes 1989). The consequence of abnormal myelination can also be related to abnormal timing in signal transfer between the two brain hemispheres as reported for instance in (Barnett et al. 2005). Genes involved in schizophrenia that can be used in the future CNGM are listed in Table 10.2. Table 10.2. Genes involved in GPRN of schizophrenia(SCH) for future CNGM Location / Gene sym- Presumed function bol lq2l-q22 Myelination related genes. Abnormal myelination leads to abnormal signal transfer between neurons Splicing and expression of myelin-relatedgenes 6q21-q22.3 / QKI G72 gene interacts with the gene for DAAO to regu13q34/ G72 late glutamatergictransmissionthrough NMDARs D-amino acid oxidase (DAAO) l2q24 / DAAO Involved in NMDAR mechanism 6p22.3 / dysbindin 8p / neuregulin Involved in NMDAR mechanism l5q 14 / a7 subunit of Acetylcholine(ACh) is involved in such cognitive nicotinic ACh receptor functions as learning and memory COMT is an enzyme that breaks down dopamine after 22q11.2/ COMT it has been released in the synapse. SCH gene variant seems to reduce activity in the prefrontal cortex

10.2.3 Discussion and Future Developments Computational models of neural networks have been applied to modeling neurological diseases and mental illnesses including schizophrenia (Reggia et al. 1999). In the model of aberrant prefrontal cortex (PC) function, neural network approach helps to elucidate how multiple pathologies can lead to PC dysfunction, be it low or high levels of dopamine, reduction or increase in D 1 receptors function, loss or increase in GAB A global inhibition (Reid and Willshaw 1999). The authors developed a spiking network model of PC and simulated the situation when neurotransmitter/receptor levels are altered. The reason for alteration can be either a gene mutation and subsequent altered protein biosynthesis or a psychoactive drug. Changes in neurotransmitter/receptor action led to the disruption of coherence of firing patterns of neurons in the model Pc. Another model has targeted neurodevelopmental impairment in schizophrenia and its consequences upon neural networks functioning (Hoffman and McGlashan

218

10 Applications of CNGM and Future Development

1999). The modeling goal in this work was to simulate that pruning of corticocortical synaptic connections in the speech perception system can lead to hallucinated voices, which is a frequent schizophrenic symptom. The effect of myelination impairment upon the function of corpus callosum and interhemispheric information transfer was computationally investigated in (Chhabra et al. 1999). The model shows that an impaired performance of the brain can be a consequence of the imbalance in the interhemispheric excitation/inhibition and also due to loss of transmission of specific information between the left and right hemispheres via the corpus callosum. Such computational and neuroengineering models as mentioned above provide a conceptual bridge between molecular/cellular pathology and cognitive performance. On the other hand, computational neurogenetic models as introduced in this book can offer insights into the gene/protein interactions within neurons that underlie a particular brain activity. For instance genetic mutations of genes coding for proteins influencing the function of DA, NMDARs and myelination can be incorporated into the neural GRN together with other genes/proteins they can influence. Then a hierarchical model of LFP/EEG generation can be used to optimize the coefficients of GRN interaction matrix W. The target signal can be for instance EEG recorded from a schizophrenic patient. There are several reports that the spectral characteristics of EEG in schizophrenic patients have abnormalities (Fenton et al. 1980, Manchanda et al. 2003). The CNGM can explain these abnormalities by means of mutations in certain genes and/or by means of altered interactions within GRN. In addition, the CNGM can simulate the effects of antipsychotic drugs upon GRN, neural network dynamics and the resulting EEG. Thus, the knowledge extracted from CNGM of schizophrenia will be qualitatively new and can be used for improving the treatment of this disease on the individual basis by simulating the processes occurring between the genes/proteins and neural dynamics.

10.3 CNGM of Mental Retardation Mental retardation (MR) is a developmental deficit, beginning in childhood, which results in significant limitation of intellect and cognition and poor adaptation to the demands of everyday life (Sebastian 2005). MR is defined as an overall intelligence quotient (IQ) lower than 70 associated with functional deficits in adaptive behavior (such as daily living skills, social skills and communication), with an onset before 18 years. MR affects approximately 2-3% of population. Moderate to severe MR (IQ < 50) is estimated to affect 0.4-0.8% of the population and the mild MR (50

10.3 CNGM of Mental Retardation

219

< IQ < 70) about 2%. Intellectual disability is the developmental consequence of some pathogenic process. The underlying causes of MR are very heterogeneous. They include non-genetic factors that cause brain injury and act prenatally (like maternal drug consumption or malnutrition) or during early infancy (like hypoxia, infections, etc.), as well as established genetic causes. Yet the underlying cause is established in only about one half of cases (Stevenson et al. 2003).

10.3.1 Genetic Causes of Mental Retardation

The Down syndrome is the best-known example of prenatal genetic MR (Sebastian 2005). Although it can have several forms, basically it is caused by trisomy 21, in which the fetus acquires an extra chromosome 21 in all cells, for a total of 47 chromosomes. Another example is cri-du-chat syndrome, which is characterized by a high-pitched voice and is caused by a deletion in chromosome 5p3. Similar microscopic deletions (microdeletions) of DNA have been reported in chromosome 15qll-12 in the PraderWilli and Angelman syndromes. The Prader-Willi syndrome results when the microdeletion is in the chromosome of paternal origin and the Angelman syndrome results when it is of maternal origin. Tuberous sclerosis is another example of the disorders in this group, which might be associated with mental retardation. It is caused by a mutation in a gene affecting the formation of the ectodermal layer of the embryo. Phenylketonuria is the best known and most common of the metabolically caused MR. The enzymatic defect is diminished activity uf phenylalanine hydroxylase, caused by single mutated gene, which leads to a high serum phenylalanine level, affecting, among other things, myelination of the CNS. Seizures and tremors are common, as are eczema and psychotic manifestations. There is a big group of chromosome X-linked forms of MR (XLMR). Approximately a third of XLMR patients have syndromic forms of XLMR, where MR is associated with recognizable clinical signs such as somatic abnormalities or dysmorphic facial features. MR with fragile X (FRAXA) is the most common form of syndromic XLMR known to date. XLMR are categorized as syndromic if associated with characteristic clinical features like in fragile X syndrome (FRAXA), or non-specific (non-syndromic) if MR is the only symptom in affected individuals. Finding the molecular causes of nonsyndromic (NS) XLMR has turned out to be much more difficult because of genetic heterogeneity, which precludes pooling of linkage data from different families. Therefore, mapping intervals have remained comparatively wide. Until 1998, only a single gene, FMR2, could be isolated, because of its association with another fragile site, FRAXE. Analo-

220

10Applications of CNGM and Future Development

gous to the mechanism leading to inactivation of FMRI in patients with FRAXA, transcriptional silencing of FMR2 was found to be caused by the expansion and subsequent hypermethylation of a CCG trinucleotide repeat in the 50 noncoding region of the gene. FRAXA or fragile X syndrome is the most common inherited form of syndromic XLMR and, after Down syndrome, its most common genetic form. It has dominant inheritance with lower penetrance in females because of the haploid status in males for most genes on the X chromosome. Because of a constriction at the location of Xq27.3, it appears as if the chromosome is fragile and a part of it is breaking off (Genes and disease 2005). Prepubertal boys with this syndrome look quite normal. They often are restless and hyperactive and have a short attention span. Their developmental milestones, especially speech development, are delayed. After puberty, the characteristic phenotypical features may appear. They include an oblong face, prominent ears and jaw, and macroorchidism (enlarged testicles). Fragile X syndrome results from a repeat-expansion mutation in the FMRI gene, which encodes the protein named FMRP. In normal individuals, the FMRI gene is transmitted stably from parent to child. However, in Fragile X individuals, there is a mutation in one end of the gene (the 5' untranslated region), consisting of an amplification of a CGG repeat. Patients with fragile X syndrome have 200 or more copies of the CGG motif. The huge expansion of this repeat means that the FMRI gene is not expressed, so no FMRI protein is made (Genes and disease 2005). Although the exact function of FMRI protein in the cell is unclear, it is known that it is translated near synapses in response to neurotransmitter activation and plays a role in synaptic plasticity and possibly also maturation (Weiler et al. 1997). Recently, the search for NS-XLMR genes has gained considerable momentum by identifying 8 of the 13 genes presently known to playa role in NS-XLMR. Table 10.3 lists genes involved in NS-XLMR together with their location on chromosome X, coded protein and presumed function (Carrie et al. 1999, Bienvenu et al. 2002, Kitano et al. 2003, Ropers et al. 2003). These genes can form the basis of the future CNGM of mental retardation. As we have envisaged in previous chapters, changes in synaptic connectivity are likely to be a key mechanism by which nervous system organization is permanently changed by experience. Local translation of some proteins in dendrites is increasingly considered to be important for changes in synaptic plasticity (Raymond et al. 2000, Bradshaw et al. 2003). Certain mRNAs are known to be targeted to dendrites, and are observed in or near dendritic spines, more frequently at newly forming synapses. Dendrites have been shown to be equipped with components necessary for protein

10.3 CNGM of Mental Retardation

221

synthesis such as polyribosomal aggregates. Protein translation induced by metabotropic receptor stimulation has previously been proposed to playa role in LTP (Bradshaw et al. 2003). Protein translation near synapses may playa role in activity-dependent modification of synapses and also in synaptic maturation and stabilization during development. Table 10.3. Genes involved in X-linked mental retardation for future CNGM Location / Gene symbol Protein and presumed function XpilA / TM4SF2 Tetraspanin protein, interacts with integrins, control of neurite outgrowth? XqI2 /0PHNI RhoGAP involved in regulating actin cytoskeleton dynamics and neuronal morphogenesis Xq22 /PAK3 Rac/Cdc42 effector involved in regulating actin cytoskeleton dynamics and neuronal morphogenesis Xp22.1 / ARX Transcription factor Xp22.1-p21.3 / Unknown ILlRAPLl Xp22.3-22.1 / Serine/threonine kinase RPS6KA3/RSK2 Xq22.3 / FACL4 Fatty acid coenzyme A ligase, involved in vesicle transport, membrane fusion and gene expression Xq24 / AGTR2 Angiotensin II receptor, type 2, unknown function Xq26 / ARHGEF6 Rho guanine nucleotide exchange factor 6, role in integrin-mediated signaling, regulates Rho-G'TPases Rac l/Cdc42 Xq28 / FMR2/FRAXE Transcription factor? Xq28 / GDlI Rab GDP-dissociation inhibitor involved in synaptic vesicle fusion and neuronal morphogenesis Xq28 / MECP2 Methyl CpG-binding protein, involved in Chromatin remodeling and gene silencing Xq28 / SLC6A8 Solute carrier family 6, member 8, Creatine transporter Weiler with co-workers report that the mRNA for fragile X mental retardation protein (FMRP) rapidly associates with synaptic polyribosomal complexes in synaptoneurosomes after stimulation by a specific mGluRl agonist DHPG (3,5-dihydroxyphenylglycine) (Weiler et al. 1997). This association happens within 1 min after stimulation of mGluRl. Moreover, immunostaining of the synaptosomal proteins at short intervals after stimulation shows increased FMRP expression relative to unstimulated samples, indicating rapid synthesis of FMRP in response to synaptic activation during the first 5 min after stimulation ofmGluRl. Huber with colleagues investigated whether mice with a null mutation in the FMR 1 gene showed any changes in synaptic plasticity that could ac-

222

10Applications of CNGM and Future Development

count for the effects of the human mutation on the brain (Huber et al. 2002). Previous work had shown that the protein-synthesis-dependent phase of long-term potentiation (LTP) in the hippocampus of the knockout mouse was normal, but Huber with colleagues decided to look at another form of hippocampal plasticity - long-term depression (LTD) - following the demonstration that LTD also requires local protein synthesis. Surprisingly, they found that LTD in the knockout mice was enhanced, so that a train of stimulation produced greater depression of synaptic function than in control mice. LTD can also be induced by directly stimulating metabotropic glutamate receptors using the agonist DHPG (3, 5dihydroxyphenylglycine). The knockout mice also showed stronger LTD in response to DHPG application than did wild-type mice. However, it should be noted that a different type of LTD that is mediated by NMDA (N-methyl-D-aspartate) receptors rather than by metabotropic glutamate receptors was not affected. Further insights into the function of FMRP have been obtained using the drosophila FMRI homolog, DFXR (also called DFMR1) (Morales et al. 2002). The protein product, DFXR, is an RNA binding protein capable of interacting with itself and the mammalian counterparts, thus implying some degree of functional conservation. The study has found that DFXR regulates neurite branching and extension in the developing brain. In wildtype flies, neurites branch in a regular array, forming a stereotypical gridlike structure. In the mutants, small ectopic branches were observed. The gain of DFXR function also affects neurite morphology. High levels of DFXR cause a complete failure of axon extension. This is qualitatively very similar to the loss-of-function mutant, but more severe. These findings suggest a dose-dependent role of DFXR in neurite growth. Also studies in mammalian brains suggest that FMRP may be involved in dendritic spine maturation, as both MR patients and FMRI knockout mice exhibit an immature spine morphology (Weiler et al. 1997). Autopsy results indicate that forebrain synapses in fragile X patients exhibit a thin, elongated neurite morphology and a reduced synaptic contact size in electron microscopy. This suggests that synaptically regulated synthesis of FMRP may be involved in dendritic spine maturation, and it is not unreasonable to postulate that this process may be impaired in cases of fragile X syndrome. Although the specific function of FMRP that leads to spine and neurite abnormalities has not been determined, FMRP has been shown to bind to ribosomes. Polyribosomal aggregates are seen in dendritic spines, particularly during development and synaptogenesis, and protein translation occurs in developing synapses in response to neurotransmitter stimulation. It has been also shown that a rat homologue of FMRP is produced at synapses and its translation is accelerated upon mGluR1 stimulation

10.3 CNGM of Mental Retardation

223

(Weiler et al. 1997). Thus FMRP may be involved in regulating translation of proteins at the synapse, and its absence might impede this synthesis and consequently impair synaptic maturation. 10.3.2 Discussion and Future Developments It may not be surprising that in genetically caused MR, mutated genes code

for proteins involved in structural synaptic plasticity and neural morphogenesis. These processes are considered to be the material basis of learning and long-term memory as well as behavior of animals and humans. Thus the CNGM of MR can be developed along the lines of synaptic plasticity models described in chapters 2 and 11 of this book. Both models, one for the sigmoid (linear) neurons and the other one for a spiking neuron, work with the notion of a movable synaptic modification LTD/LTP threshold. In fact the theory and model of cortical synaptic plasticity described in chapter 2 and in (Benuskova, Rema et al. 2001) accounts for an impaired synaptic plasticity in the cortex of the rat model of mental retardation. The explanatory variable is the synaptic modification LTD/LTP threshold that is set to high values in the model of the impaired cortex. The same mathematical operation can work for an explanation of an increase in LTD observed in the FMRI-KO mice (Huber et al. 2002). Albeit causes and underlying molecular processes are different (prenatal alcohol versus FMRI gene deletion) manifestations at the synaptic plasticity levels are the same, i.e. enhancement of LTD. More detailed dynamical models in relation to the internal dynamics of the underlying GRN can concentrate on the design of the complex functional relationship for the fast and slow component of the movable synaptic modification LTD/LTP threshold. Another CNGM approach to modeling MR in future can be using biological models of neurite outgrowth that are based on dynamic concentrations of relevant proteins as developed by (Kiddie et al. 2005). Protein concentrations in this model can be made functions of gene expressions of the underlying GRN. In such a way the CNGM can lead to explanation of aberrant neurite extension and branching by means of altered gene/protein dynamics. Similar strategy for modeling impaired neural development can be applied to modify the neurogenetic model of normal neural development in Drosophila (Marnellos and Mjolsness 2003). That is the normally functioning model GRN can be modified to include gene deletions and/or mutations to see what the consequences upon neural development would be.

224

10Applications of CNGM and Future Development

All these models then can be used to seek corrections at the genetic and molecular level that should be applied to alleviate damage to the neural system in MR.

10.4 CNGM of Brain Aging and Alzheimer Disease The aging of the human brain is the major risk for Alzheimer's disease and a cause of cognitive decline in the elderly (Lu et al. 2004). As the world population grows older, understanding the genetic and environmental factors that contribute to aging is becoming very important. Evolutionary biologists have sought to formulate theories as to why organisms age. Current views on aging range from the very complicated genetic picture to those who seek one or few major mechanisms that will explain why we age. Cellular aging is accompanied by a decline in metabolic activity, decreased resistance to stress, gene dysregulation, decreased rates of protein synthesis and degradation. Attractive hypotheses for a unitary basis for cellular aging include telomere shortening, accumulation of somatic mutations in mitochondria, oxidative stress and cellular damage by free radicals, errors in DNA replication and repair, and altered rates and accuracy of protein synthesis (Mackay 2000). Further, the one simple factor that leads to increased longevity in mammals is dietary restriction (Sinclair and Guarente 2006). Changes in gene expression with advancing age in brain tissue are consistent with inflammatory and oxidative stress responses and are attenuated by caloric restriction. In skeletal muscle, aging results in a gene expression pattern indicative of a marked stress response and lower expression of metabolic and biosynthetic genes. Most of these changes in gene expression are prevented or ameliorated by caloric restriction (Sinclair and Guarente 2006). Thus it is not resolved yet whether there are numerous mechanisms of aging, which are modulated by hundreds of gene effects, depending upon gene-gene and gene-environmental interactions or whether there is only a single major mechanism of aging, which is modulated by only a few major gene effects (Mackay 2000). Perhaps the most likely answer is that many viewpoints are to some extent correct, and the real question becomes one of determining all of the potential genetic mechanisms and the extent to which each contributes to genetic variation in aging. Understanding the molecular effects of aging in the brain will reveal processes that lead to age-related brain dysfunction, like Alzheimer's disease. More than 12 million individuals worldwide have Alzheimer's disease (AD), and it accounts for most cases of dementia that are diagnosed

lOA CNGM of BrainAging and Alzheimer Disease

225

after the age of 60. AD is clinically characterized by a global decline of cognitive function that progresses slowly and leaves end-stage patients bedridden, incontinent and dependent on custodial care. Death occurs, on average, 9 years after diagnosis. All of the currently used drugs are of limited benefit, because they have only modest symptomatic effects. Other drugs are used to manage mood disorder, agitation and psychosis in later stages of the disease, but no treatment with a strong disease-modifying effect is currently available (Citron 2004). A large body of evidence suggests that oxidative stress results in DNA damage that subsequently leads to changes in gene expression and aging of tissues. The time in life when brain ageing begins is undefined. Transcriptional profiling of 11,000 genes from the human frontal cortex from individuals ranging from 26 to l06 years of age defined a set of genes with reduced expression after age 40 (Lu et al. 2004). These genes play central roles in synaptic plasticity, vesicular transport and mitochondrial function. This is followed by induction of stress response, antioxidant and DNA repair genes. DNA damage is markedly increased in the promoters of genes with reduced expression in the aged cortex. DNA damage may reduce the expression of genes involved in learning, memory and neuronal survival, thus initiating a program of brain aging that starts early in adult life. It was interesting to note that middle age group ranging in age from 45-71 exhibited much greater heterogeneity in gene expression than other age groups, with some cases resembling the young group and others resembling the aged group. These results suggest that a genetic signature of human cortical aging may be defined starting in young adult life, and that the rate of age-related change may be heterogeneous among middle age individuals. Genes that playa role in synaptic function and the plasticity that underlies learning and memory were among those most significantly affected in the ageing human cortex (Lu et al. 2004). Several neurotransmitter receptors that are centrally involved in synaptic plasticity showed significantly reduced expression after age 40, including the GluRl AMPA receptor subunit, the NMDA R2A receptor subunit, and subunits of the GABA A receptor. Moreover, the expression of genes that mediate synaptic vesicle release and recycling was significantly reduced notably VAMPl/synaptobrevin, synapsin II, RAB3A and SNAPs. Members of the major signal transduction systems that mediate long-term potentiation (LTP) and memory storage were age-downregulated, notably the synaptic calcium signaling system, with reduced expression of calmodulin 1 and CAM kinase IIa. The major calcium-binding proteins calbindins land 2, the calcium pump ATP2B2, and the calcium-activated transcription factor MEF2C that promotes neuronal survival, were also significantly reduced. Furthermore, multiple members of the protein kinase C (PKC) and Ras-

226

10Applications of CNGM and Future Development

MAP kinase signaling pathways showed decreased expression. In addition, genes involved in vesicular/protein transport showed reduced expression in the aged cortex, including multiple RAB GTPases, sortilin, dynein, and clathrin light chain. Moreover, microtubule-associated proteins (MAPIB, MAP2, tau and kinesin IB) that stabilize microtubules and promote axonal transport were consistently and robustly reduced. The p35 activator of cyclindependent kinase- 5 (cdk5), which regulates intraneuronal protein trafficking and synaptic function, was also significantly reduced. A number of genes involved in protein turnover also showed reduced expression in aged cortex, including ubiquitin conjugating enzymes, the lysosomal proton pump, and the enzymes D-aspartate O-methyltransferase and methionine adenosyltransferase II, which repair damaged proteins. The aging of the human frontal cortex was also associated with increased expression of genes that mediate stress responses and repair. These included genes involved in protein folding (heat shock protein 70 and a crystallin), antioxidant defense (nonselenium glutathione peroxidase, paraoxonase and selenoprotein P) and metal ion homeostasis (metallothioneins IB, IG and 2A). Genes involved in inflammatory or immune responses, such as tumor-necrosis factor (TNF)-a, were also increased. Increased expression of the base-excision repair enzymes 8-oxoguanine DNA glycosylase and uracil DNA glycosylase is consistent with increased oxidative DNA damage in the aged cortex. Analogical changes were observed in the brain of aging mice (Jiang et al. 2001). Interestingly, the genes playing the role in learning and memory whose expression was affected by aging was oppositely affected by exposure of mice to an enriched environment. Brain tissue of AD patients is characterized by two main findings: (1) senile plaques made of fragmented brain cells surrounded by amyloid family proteins, and (2) tangles of cytoskeleton filaments within cells (Pastor and Goate 2004). Two main disease mechanism are based on the involvement of two proteins - amyloid-B (AB) and tau - in AD pathology. A13 is the main constituent of senile plaques - one of the key pathological characteristics of AD. Tau is the main component of neurofibrillary tangles, the other hallmark lesion of AD. Genetic and pathological evidence strongly supports the amyloid cascade hypothesis of AD, which states that amyloid1342 (A1342), a proteolytic derivative of the large transmembrane protein amyloid precursor protein (APP), has an early and vital role in all cases of AD (Citron 2004). A1342 forms aggregates that are thought to initiate the pathogenic cascade, leading ultimately to neuronal loss and dementia. Although A1342 is constitutively produced, for a long time it was assumed that only A13 that is deposited in compact neuritic plaques is neurotoxic. However, more recent data indicate that soluble oligomeric species also

lOA

CNGM of BrainAging and Alzheimer Disease

227

have toxic properties. Therefore it is now thought that the formation of toxic aggregates is the first truly pathological step in the disease. The amyloid cascade is initiated by the generation of AB42. In familial early onset AD, AB42 is overproduced owing to pathogenic mutations. In sporadic AD, various factors can contribute to an increased load of AB42 oligomers and aggregates. Amyloid-B oligomers might directly injure the synapses and neurites of brain neurons, in addition to activating microglia and astrocytes. Tau pathology, which contributes substantially to the disease process through cytoskeletal tangles, is triggered by AB42. AB is generated proteolytic ally from a large precursor molecule, APP, by the sequential action of two proteases, B-secretase and y-secretase. A third protease, a-secretase, which competes with B-secretase for the APP substrate, can preclude the production of AB by cleaving the peptide in two. This outline immediately points to three strategies to reduce AB: inhibition of B-secretase, inhibition of y-secretase and stimulation of a-secretase. However, these enzymes have other substrate proteins as well, thus the potential side effects of this approach are not known (Citron 2004). After brain aging, family history is the second-greatest risk factor for AD. AD appears to follow an age-related dichotomy: early-onset familial AD (EOFAD) and late-onset AD (LOAD) without obvious familial segregation. Rare and highly penetrant early-onset EOFAD mutations in different genes are transmitted in an autosomal dominant fashion, while LOAD is thought to be explained by the common disease/common variant (CD/CV) hypothesis. According to this theory common disorders are also governed by common DNA variants (such as single nucleotide polymorphisms). These variants significantly increase disease risk but are insufficient to actually cause a specific disorder. Current empirical and theoretical data support this hypothesis, although there remains great uncertainty as to the number of the underlying risk factors and their specific effect sizes. EOFAD represents only a small fraction of all AD cases (-5%) and typically presents with onset ages younger than 65 years, showing autosomal dominant transmission within affected families. To date, more than 160 mutations in 3 genes have been reported to cause EOFAD. These include the AB precursor protein (APP) on chromosome 21, presenilin 1 (PSEN1) on chromosome 14, and presenilin 2 (PSEN2) on chromosome 1 (Bertram and Tanzi 2005). The most frequently mutated gene, PSEN1, accounts for the majority of AD cases with onset prior to age 50. While these AD-causing mutations occur in 3 different genes located on 3 different chromosomes, they all share a common biochemical pathway, i.e., the altered production of AB leading to a relative overabundance of the AB42 species, which eventually results in neuronal cell death and dementia. An

228

10Applications of CNGM and Future Development

up-to-date overview of disease-causing mutations in these genes can be found at the Alzheimer Disease & Frontotemporal Dementia Mutation Database (Alzheimer disease & frontotemporal dementia mutation database 2006). LOAD, on the other hand, is classically defined as AD with onset at age 65 years or older and represents the vast majority of all AD cases. While segregation and twin studies conclusively suggest a major role of genetic factors in this form of AD, to date, only 1 such factor has been established, the £4 allele of the apolipoprotein E gene on chromosome 19q13 (APOE) (Bertram and Tanzi 2005). The risk effect of APOE-£4 has been consistently replicated in a large number of studies across many ethnic groups. In addition to the increased risk exerted by the s-l-allele, a weak, albeit significant, protective effect for the minor allele, £2, has also been reported in several studies. Unlike the mutations in the known EOFAD genes, APOE£4 is neither necessary nor sufficient to cause AD but instead operates as a genetic risk modifier by decreasing the age of onset in a dose-dependent manner. Despite its long-known and well-established genetic association, the biochemical consequences of APOE-£4 in AD pathogenesis are not yet fully understood but likely encompass AB-aggregation/clearance and/or cholesterol homeostasis. Numerous additional LOAD loci and probably also EOFAD loci remain to be identified, since the 4 known genes together account for probably less than 50% of the genetic variance of AD (Table lOA). It is currently unclear how many of these loci will prove to be risk factors as opposed to causative factors. As candidates for the former, more than three dozen genes have been significantly associated with AD (Cacabelos et al. 1999). However, no single gene has been shown to be a risk factor with even nearly the same degree of replication or consistency as has APOE-£4. An up-to-date overview of the status of these and other potential AD candidate genes, including metaanalyses across published association studies, can be found at the Alzheimer Research Forum genetic database (Alzheimer research forum genetic database of candidate genes 2006). The table can also include genes involved in learning and memory, synaptic function, intraneuronal protein trafficking, protein turnover, genes that mediate stress responses and repair, genes involved in inflammatory and immune responses, etc. (Cacabelos et al. 1999, Jiang et al. 2001, Citron 2004, Bertram and Tanzi 2005). In the future CNGM, each gene and protein can be linked to the corresponding molecular pathway and cellular function using the Kyoto Encyclopedia of Genes and Genomes (KEGG) (KEGG pathway database 2006). The next step will be to link protein levels to neuronal parameters in the neural network model. One brain region that is severely affected by AD processes is hippocampus. There are models of

10.5 CNGM of Parkinson Disease

229

hippocampal functioning (Rolls and Treves 1998) that can be utilized to include internal gene networks. Several computational models of AD have been proposed, e.g. (Hom et al. 1996, Hasselmo 1997, Devlin et al. 1998), which can be also exploited for linking mutated proteins to neuronal parameters and the investigations of altered dynamics of gene expressions and the consequences upon deterioration of neural functions like learning and memory formation in AD. Table 10.4. Maingenes involved in Alzheimer'sdisease (AD) for future CNGM Location / Gene symbol Protein and presumed function

21q21 / APP l4q24 / PSEN1 1q31 / PSEN2 19q13 / APOE

A~ precursor proteinis mutated, whichresults in alteredproduction and aggregation of A~ Altered A~ production Altered A~ production Apoliprotein E. It is a risk factor, the role of which is unknown. Alterations in AB-aggregationiclearance and/orcholesterol homeostasis are likely.

10.5 CNGM of Parkinson Disease Parkinson disease (PD) is the second most common neurodegenerative disease of adult onset. Histopathologically, it is characterized by a severe loss of dopaminergic neurons in the substantia nigra and cytoplasmic inclusions consisting of insoluble protein aggregates (Lewy bodies). These malformations lead to a progressive movement disorder including the classic triad of tremor, bradykinesia, and rigidity, with an average onset age between 50 and 60 years. Genetics has played a major role in elucidating the causes of nigrostriatal neuronal loss across a wide spectrum of clinically and histopathologically heterogeneous PD cases. As in AD, there appears to be an age-dependent dichotomy: the majority of individuals with an early or even juvenile onset show typical Mendelian inheritance. However, unlike in AD, these cases show a predominantly autosomal-recessive mode of inheritance, and there is an ongoing debate as to whether genetic factors play any substantial role in contributing to disease risk in cases with onset beyond approximately 50 years. Notwithstanding these uncertainties, mutations in at least five genes have now been shown to cause familial early-onset parkinsonism (Bertram and Tanzi 2005). These genes include genes for a-synuclein (SNCA or PARKl), parkin (PRKN or PARK2), DJ-l or PARK7, PTEN-induced putative kinase I (PINKI or PARK6), and leucine-rich repeat kinase 2 or dardarin (LRRK2 or

230

10Applications of CNGM and Future Development

PARK8), with several other linkage regions pending characterization and/or replication. The locus PARKI on chromosome 4q21 involves the protein that is the major constituent of one of the classic neuropathological hallmarks of the disease, i.e., a-synuclein, which can be found at the core of Lewy bodies. While the exact mechanisms underlying a-synuclein toxicity currently remain only incompletely understood, recent evidence suggests that some SNCA mutations may change normal protein function quantitatively rather than qualitatively, via duplication or triplication of the a-synuclein gene. Another mutation occurs at LRKK2 locus on chromosome 12q12. While the functional consequences of LRRK2 mutations are still unknown, it was suggested they could interfere with the protein's kinase activity. While changes in SNCA and LRRK2 are the leading causes of autosomaldominant forms of PD, the majority of affected pedigrees actually show a recessive mode of inheritance. The most frequently involved gene in recessive Parkinsonism is parkin (PRKN) on chromosome 6q25, which causes nearly half of all early-onset PD cases. Parkin is an ubiquitin ligase that is involved in the ubiquitination of proteins targeted for degradation by the proteasomal system. The spectrum of parkin mutations ranges from amino acid-changing single base mutations to complex genomic rearrangements and exon deletions, which probably result in a loss of protein function. It has been speculated that this may trigger cell death by rendering neurons more vulnerable to cytotoxic insults, e.g., the accumulation of glycosylated a-synuclein. In addition to parkin mutations, genetic analyses on other pedigrees revealed two independent, homozygous mutations in DJl on chromosome Ip36. Both mutations result in a loss of function of DJ-I, a protein that is suggested to be involved in oxidative stress response. While several studies have independently confirmed the presence of DJ-l mutations in other PD cases, the frequency of disease-causing variants in this gene is estimated to be low (~l %). Additional PD-causing mutations were discovered in PINKI. This enzyme is expressed with particularly high levels in brain, and the first two identified mutations were predicted to lead to a loss of function that may render neurons more vulnerable to cellular stress, similar to the effects of parkin mutations. While Lewy bodies are typically not found in brains of patients bearing parkin mutations, it is currently unclear whether these are present in PD cases with mutations in DJl and PINKl. At least 6 additional candidate PD loci have been described, including putative disease-causing mutations in the ubiquitin carboxyterminal hydrolase Ll (UCHLl) on chromosome 4p14, and in a nuclear receptor of subfamily 4 (NR4A2 or NURR1) located on 2q22 (Bertram and Tanzi 2005). These genes may actually be susceptibility factors rather

10.5 CNGM of Parkinson Disease

231

than causal PD genes. Table 10.4 lists major genes involved in PD for future constructions of CNGM. Table 10.5. Main genes involved in Parkinson disease (PD) for future CNGM Location / Gene symbol Protein and presumed function 4q21 / SNCA Protein u-synuclein leads to neurotoxicity by aggregation 12q12 / LRRK2 Enzyme Leucine-rich repeatkinase 2, function unknown 6q25/PRKN Parkin mutation leads to impaired proteindegradation via proteasome Ip36 / DJ1 Protein DJ-l, mutation leads to impaired oxidative stress response PTEN-induced putativekinase 1, mutation may lead to Ip36 / PINKI mitochondrial dysfunction While a number of whole-genome screens across several late-onset PD family samples have been performed, only a few overlapping genomic intervals have been identified (Bertram and Tanzi 2005). One of the more extensively studied regions is 17q21, near the gene encoding the microtubule-associated protein tau (MAPT). Previously, it had been shown that rare missense mutations in MAPT lead to a syndrome of frontotemporal dementia with Parkinsonism (FTD with Parkinsonism linked to chromosome 17), but to date no mutations have been identified as causing Parkinsonism without frontotemporal degeneration. Despite the lack of evidence for genetic linkage to chromosome 19q13, variants in APOE have also been tested for a role in PD and related syndromes. Across the nearly 3 dozen different studies available to date, some authors report a significant risk effect of APOE-£4 for PD, while others only see association with certain PD phenotypes or even a risk effect of the £2 allele. In addition to the findings in autosomal-dominant familial PD, there is also some support for a potential role of SNCA variants on the risk for late-onset PD (Bertram and Tanzi 2005). In the future CNGM, each gene and protein can be linked to the corresponding molecular pathway and cellular function using molecular pathways in Kyoto Encyclopedia of Genes and Genomes (KEGG pathwaydatabase 2006). In addition, new ways of combining electrophysiological and molecular single neuron analysis with gene-expression profiles in the same neuron are being developed and bring first results (Liss and Roeper 2004). After constructing the gene and protein regulatory network for PD, the next step will be the model of affected neural circuits and functions (Bear et al. 2001). The main brain region affected in PD is substantia nigra (i.e.

232

10Applications of CNGM and Future Development

black substance). Substantia nigra (SN) is an elongated nucleus located medial to the basis pedunculi throughout the rostrocaudal extent of the midbrain. Although a component of the brainstem, SN is considered along with the basal ganglia because it has reciprocal connections with basal ganglia and is functionally related to them. The SN is divided into two components that have different connections and distinct neurotransmitters. A more ventral part is called the substantia nigra pars reticulata (SNr) and a dorsal part is called the substantia nigra pars compacta (SNc). Neurons of the SNc use dopamine (DA) as a neurotransmitter and project primarily to the neostriatum. They also contain the black pigment neuromelanin; a product like dopamine produced from tyrosine, hence the name black substance. Neurons in the SNr project principally to the thalamus (ventral anterior, ventral lateral, and dorsomedial nuclei) but also to brain stem nuclei (superior colliculus, pedunculopontine nucleus) and use GABA as the neurotransmitter. The SNr also receives striatal input that uses GABA (and substance P) as transmitters and is inhibitory. The SNr is a subnucleus that synthesizes GABA. The basal ganglia together with cerebellum modify movements generated by motor cortex on a minute-to-minute basis. Motor cortex sends information to both, and both structures send information right back to cortex via the thalamus. The output of the cerebellum is excitatory, while the basal ganglia are inhibitory. The balance between these two systems allows for smooth, coordinated movements, and a disturbance in either system will show up as a movement disorder, like for instance the PD triad. There are several computational models of the basal ganglia system that can be exploited for future CNGM development, i.e. (Terman et al. 2002, Frank 2005, Leblois et al. 2006).

10.6 Brain-Gene Ontology The explosion of biomedical data and the growing number of disparate data sources are exposing researchers to a new challenge - how to acquire, represent, maintain and share knowledge from large and distributed databases in the context of rapidly evolving research. In the biomedical domain, for instance, the problem of knowledge discovery from biomedical data and making biomedical knowledge and concepts sharable over applications and reusable for several purposes is both complex and crucial. It is central to support the decision in the medical practice as well as to enable comprehensive knowledge-acquisition by medical research communities and molecular biologists involved in biomedical discovery.

10.6 Brain-Gene Ontology

233

Ontology is defined in the artificial intelligence literature as a specification of a conceptualization. Ontology specifies at a higher level the classes of concepts that are relevant to the domain and the relations that exist between these classes. Ontology captures the intrinsic conceptual structure of a domain. For any given domain, its ontology forms the heart of the knowledge representation. Cell functions through the use of its genes to produce proteins and although each cell within an organism will usually contains the same set of genes there are significant differences in how the genes are utilized between them. Research from structural genomics towards functional genomics is calling for availability of new biological knowledge. The development of reliable automated sequencing techniques made possible the growth of genomics as a commonplace element of modern biology. The genome sequencing projects have provided tremendous information about the central dogma of molecular biology and helps in understanding the proper functioning of organism in much better way. Microarray technology makes use of the sequence resources created by the genome projects and other sequencing efforts to answer the questions which genes are expressed in a particular cell type of an organism, at a particular time, under particular conditions. For instance, they allow comparison of gene expression between normal and diseased (e.g. epileptic, schizophrenic, etc) cells. DNA microarray consists of an orderly arrangement of DNA fragments representing the genes of a particular organism and assessment transcription profiles at the genomic scale can be achieved using this technology. Microarray analysis permits scientists to detect thousands of genes in a small sample simultaneously and to analyze the expression of those genes. Integrating the domain of discovery outputs from different experimentation techniques with extracted appropriate biological and medical knowledge, there is a scope to discover new metabolic pathways and modeling of the metabolic and regulatory networks in living organisms, and ultimately to understand the pathogenesis of various diseases. In this context the biggest problem today we are faced with is an overwhelming array of nomenclature for genes, proteins, drugs and even diseases. And thus biomedical community is suffering from communication problem and the ability to use resources to search the vast sources of information more effectively, i.e. to extract appropriate meanings. Many researchers and databases use (at least partially) idiosyncratic terms and concepts for representing biological information. It has been observed that often, terms and definitions differ between groups, with different groups not infrequently using identical terms with different meanings. The concept 'gene', for example, is used with different semantics by the major international genomic databases.

234

10Applications ofCNGM and Future Development

Ontology can provide conceptual framework and factual knowledge that is necessary to deal critically with the rapidly changing science of biology. Ontology is one way to provide a semantic repository of systematically ordered relevant concepts in molecular biology. Such repositories can be used, for example, to bridge the different notions in various databases by explicitly specifying the meaning of and relation between fundamental concepts. Thus in general, ontology permits researchers to define and share domain-specific vocabularies. Gene Ontology (GO), for example, has been used to "produce a controlled vocabulary that can be applied to all organisms even if knowledge of genes and proteins is changing". GO is the basis for systems that address the problem of linking biology knowledge and literature facts. However, in addition to research-based literature the amount of data produced daily by medical information systems and medical decision support systems is growing at a staggering rate. We must consider that scientific biomedical information can include information stored in the genetic code, but also can include experimental results from various experiments and databases, including patient statistics and clinical data. Large amounts of information and knowledge are available in medicine. Making medical knowledge and medical concepts shared over applications and reusable for different purposes is crucial. Biomedical ontologies should provide conceptuallinks between data from seemingly disparate fields. This might include, for example, the information collected in clinical patient data for clinical trial design, geographical and demographic data, epidemiological data, drugs, and therapeutic data, as well as from different perspectives as those collected by nurses, doctors, laboratory experts, research experiments and so on. At the same time the framework should reuse and integrate as many as possible different ontologies. The ontologies should integrate terminologies, as well as domain specific ontologies, such as disease ontologies and GO, in order to support the knowledge discovery process. The project Brain-Gene Ontology (BGO) that runs at KEDRI AUT in New Zealand aims to build multi-dimensional biomedical informatics ontology, able to share knowledge from different sources and experiments undertaken across aligned research communities in order to connect areas of science seemingly unrelated to the area of immediate interest. BGO is a frame-based ontology developed using Protege environment (Protege 2006) (see Fig. 10.2). Protege includes knowledge acquisition tools that allow domain expert and ontology engineers to built and refine the knowledge representation at the same time that populates instances in the knowledge base. BGO is primarily focused on the gene-disease relationship. We are representing graphically, the relationships in a way that enables visualization and creation of new relationships. We are planning to integrate the

10.7Summary

235

data mining techniques available in the NeuCom (NeuCom 2006) that was also developed at KEDRI AUT. NeuCom is a complete software package for generic data processing with emphasis on evolving connectionist modeling techniques. It contains a set of tools for data visualization, normalization, analysis, feature extraction, clustering and modeling with cross validation and genetic algorithm with a consistent and easy to use interface. It is designed to help the user to extract knowledge from any given dataset and, if the user desires, develop prediction or classification models on it. The NeuCom modules can adapt to new incoming data in an on-line incremental, lifelong learning mode, and can extract meaningful rules that would help people discover new knowledge in their respective fields.

...... . -_. ·-

, ~X

~

·.-...-- , ....... .. . ..............-.....-......

.. .

o-.~ ..eo-

•

.. . 0.. ..,.

.. • e-........... ..-..; - .

"

._--. . . . . .o-.

• __..-=-....-. ,1 ~

• e a.-..JIo-I'o .. • ~J~

. •c:ao-.......• "'-S--

., ~

~

..

Fig. 10.2. Snapshot from the BGO developed at KEDRI (www.kedri.info). See ColorPlate 11

10.7 Summary The book introduces a novel computational approach to brain neural network modeling that integrates dynamic gene networks with a neural network model. The bridging system is the protein network, in which individual proteins that are coded for by genes are related to neuronal parameters.

236

10Applications ofCNGM and Future Development

Interaction of genes in neurons affects the dynamics of the whole neural network 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 ANN parameters, particular target states of the neural network operation can be achieved. A particular ANN model and its observable output should be chosen based on the concrete neural system and a particular problem, which one wants to model. Principles of building such models can be found for instance in (Rolls and Treves 1998). Although we speak about brain diseases and functions, having in mind mammals and higher vertebrates, the approach suggested in this book is applicable also to simpler organisms such as Drosophila to aid the explanation of genetic basis of their behavior. Another system that can be used to validate our approach could be the brain slices or cell cultures taken from normal and genetically modified animal brain.

Appendix 1

A.1 Table of Genes and Related Brain Functions and Diseases Most of the genetic disorders of the human neural system featured in this Appendix are the direct result of a mutation in one gene. Even in those cases, the exact molecular processes leading to disease symptoms may not be fully understood yet, as is the case of the Huntington's disease (Cattaneo et al. 2001). However, one of the most difficult problems ahead is to find out how genes contribute to diseases that have a complex pattern of inheritance, such as in the cases of mental diseases like schizophrenia, depression, Parkinson's disease, etc. In all these cases, no one gene has the decisive power to say whether a person has a disease or not. It is likely that more than one gene mutation is required before the disease is manifest, and a number of genes may each make a subtle contribution to a person's susceptibility to a disease; genes may also affect how a person reacts to environmental factors like life stress (Caspi et al. 2003). The following Table A.I.I contains the list of human brain diseases, mutated genes / location on chromosomes, affected brain cell functions and resulting symptoms. The main source for Table A.I.l was the public domain chapter The Nervous System in Genes and Disease (Genes and disease 2005). From the site of Genes and Disease you can follow the links into many online related resources with free and full access including the human genome sites to see the location of the genes implicated in each disorder. You can also find related gene sequences in different organisms. For the very latest information, you can search for complete research articles via PubMed, and look into other books in the NCBI electronic bookshelf.

238

Appendix I

Table A.l.l. Genes and related brain functions and diseases in humans Disease

Mutations of genes identified so far / chromosome location if known (protein role if known)

Brain abnor- Symptoms mality

Adrenoleukodystrophy (ALD)

ADL gene (affects the function of the fatty acid enzyme) / unknown chromosome

Myelin sheath Progressive neu- (McGuinon nerve fibers rological disabil- ness et al. in the brain is ity and death. 2003) lost, and the adrenal cortex degenerates.

References

Alcoholism CREB gene (role in synaptic plasticity associated with drug addictive behaviors) / unknown chromosome

Compromised CREB signaling in nucleus accumbens and amygdala.

Compulsive and (Pandey uncontrolled pat- 2004) tern of alcohol drinking in spite of the adverse consequences of its abuse.

AlcoholCHRM2 gene / 7 ismdepression

Attention, learning, memory and cognition are affected.

People are at risk (Wang, for developing Hinrichs et both diseases, al. 2004) alcoholism linked to depression.

Alzheimer PSEN2 (early-onset AD)/lq31 disease PSEN I (early onset (AD) AD) /14q24 APP gene (earlyonset AD) / 21q21 APOE-e4 (late-onset AD)/19qI3

Plaques made of fragmented brain cells surrounded by amyloidfamily proteins, tangles of cytoskeleton filaments.

Progressive inability to remember facts and events and, later, to recognize friends and family.

(Citron 2004, Pastor and Goate 2004, Bertram and Tanzi 2005)

A.l Table of Genes and Related Brain Functions and Diseases

239

Disease

Mutations of genes identified so far / chromosome location if known (protein role if known)

Brain abnor- Symptoms mality

References

Amyotrophic lateral sclerosis (ALS)

SODI / 2lq22 (prevents superoxide radicals attack from the cell inside)

Progressive degeneration of motor neuron cells in the spinal cord and brain.

Loss of motor control which ultimately results in paralysis and death.

(Bertram and Tanzi 2005, Federici and Boulis 2005)

Mutations in UBE3A disrupt protein break down during brain development.

Mental retardation, abnormal gait, speech impairment, seizures, frequent laughing, smiling and excitability.

(Buiting et al. 2003, Wang, Liu et al. 2004)

Angelman UBE3A / deletion of syndrome segment l5q ll-q13 (AS) on maternally derived chromosome 15 (protein degradation)

Anorexia nervosa (AN)

Susceptibility locus / Severely reObsessive fear lp stricted eating of weight gain, extremely low body weight

(Grice et al. 2002)

Ataxia te- ATM gene / 11 (aflangiecta- fects DNA damage sia (A-T) response)

Cerebellar de- Lack of balance (Watts et al. generation and slurred 2002) speech, immunodeficiency, radiosensitivity, predisposition to cancer.

Attention DATI /5p15.3 deficit hyperactivity DRD4/ llp15.5 disorder (ADHD)

Brain dopaminergic neurotransmission is compromised.

Trouble keeping attention on tasks or play activities, easily distracted and forgetful, excessive senseless aggressive hyperactivity.

(Chen et al. 2003, Langley et al. 2004)

240

Appendix 1

Disease

Mutations of genes identified so far / chromosome location if known (protein role if known)

Brain abnor- Symptoms mality

Autism

Genes related to GABAergic and serotonergic neurotransmission, early developmental genes affecting embryonic development of CNS / unknown chromosome(s)

Brain inhibi- Communication (McIntosh tory and sero- problems, inabil- 1998, Veenstratonergic trans- ity to read the mind of others, Vanderweele mISSIOn IS et al. 2004) compromised; social impairment, and unthe immune usual or repetisystem is weakened to tive behaviors. allow viral infections to damage the brain during embryonic development.

Bulimia nervosa (BN)

Susceptibility locus on chromosome lOp

Severe distur- Binge-eating, bances in eat- self-induced ing behavior. vomiting.

References

(Bulik et al. 2003)

Charcot- Type IA PMP22 /11, MarieTooth dis- Type 1B Cx32 / X, ease (CMT) ge- Type 1C EMP2 / 16 netically heterogeneous group of diseases

Peripheral neuropathy resulting from insufficient myelin coating of peripheral nerves.

Slowly progres- (Street et al. 2002) sive degeneration of the musc1es in the foot, lower leg, hand, forearm and a mild loss of sensation

Cockayne CSA/5 syndrome CSB / unknown chromosome (components of the transcriptional machinery and DNA repair)

Loss of transcriptioncoupled repair of DNA in all cells in the body.

Premature aging (Spivak of the brain and 2004) the whole body, short stature, sensitive to sunlight.

A.l Table of Genes and Related Brain Functions and Diseases Disease

Mutations of genes Brain abnor- Symptoms identified so far mality / chromosome location if known (protein role if known)

241

References

Creutzfeld Gene encoding the -Jacob dis- prion protein PrP ease (CJD) (PRNP) / 20p13

Neurodegeneration with spongiosis and amyloid plaques consisting ofPrP.

(Bertram Rapidly proand Tanzi gressing neurodegeneration 2005) leading to death.

Deafness

Cx26 / 13

Congenital deafness.

Deafness is not (Sabag et al. accompanied by 2005, Yang other symptoms. et al. 2005)

Depression (linked to 19 different genetic regions)

1-2 copies of the short allele of the 5HTT promoter / unknown chromosome

Serotonergic Stressful life (5-HT) neuro- events (loss, transmission in threat, humiliathe brain is tion, defeat) lead compromised. to clinical depression.

(Caspi et al. 2003, Zubenko et al. 2003, Zubenko et al. 2004)

Epilepsy (at least 8 forms of epilepsy possessing some genetic basis)

Many genes coding for ion channels, neurotransmitter receptors, transcription factors, etc.

Abnormal cell Recurring seifiring in the zures brain originating in different parts of the brain.

(Gardiner 1999, Meisler et al. 2001, George 2004, Steinlein 2004)

Fragile X expansion of the Defects in syndrome CGG motif in FMRI neurite density / X (function unand morpholknown, FMRI protein ogy and synbinds RNA) aptic plasticity.

The most common inherited form of mental retardation.

(Weiler et al. 1997, Huber et al. 2002, Morales et al. 2002)

242

Appendix 1

Disease

Mutations of genes identified so far / chromosome location if known (protein role if known)

Frontotemporal dementia (FTD)

Microtubule associ- Protein accuated protein tau gene mulates in neurofibrillary (MAPT) / l7q2l tangles within neurons to cause malfunction.

Brain abnor- Symptoms mality

First personality and behavioral changes, later forgetfulness, disorientation, confusion, like in AD.

Hunting- Expansion of CAG in Dilatation of Degenerative ton disease HD gene for protein ventricles and neurological disease that leads to atrophy of (HD) Huntigtin /4p16 caudate nudementia and death. cleus.

References

(Zhukareva et al. 2003, Bertram and Tanzi 2005)

(Cattaneo et al. 2001, Bertram and Tanzi 2005, RaIser et al. 2005)

HPRTl gene / X (af- Accumulation Painful deposits (Fairbanks et Leschal. 2002) fects recycling of of uric acid of uric acid in Nyhan the skin, kidney syndrome purines) and bladder, (LNS) self-mutilation, mental retardation, muscle weakness. Lewy body Gene for Leucine-rich dementia repeat kinase 2 (LBD) (LRRK2) /12q12 a-synuclein gene (SNCA) / 4q2l

Maple Syrup Urine Disease (MSUD)

Six gene loci encoding for the BCKDH enzyme / unknown chromosome

Presence of cortical and subcortical Lewy bodies, amyloid plaques and neurofibrillary tangles

Progressive cog- (Bertram nitive impairand Tanzi ment, recurrent 2005) visual hallucinations, and Parkinsonism.

Disruption of the metabolism of certain amino acids.

Progressive neu- (Jouvet et al. rodegeneration 2000) leading to death within the first months of life.

A.l Table of Genes and Related Brain Functions and Diseases Disease

Mutations of genes Brain abnor- Symptoms identified so far maUty / chromosome location if known (protein role if known)

243

References

Severe cerebral (Pase et al. degeneration and 2004) arterial changes, resulting in death in infancy.

Menkes X-linked (can not syndrome transport copper)

Disruption of the cells' ability to absorb copper.

Myotonic Gene, found on dystrophy chromosome 19

Muscles con- Mental defi(Ranum and tract but have ciency, hair loss Day 2004) decreasing and cataracts. power to relax.

Narcolepsy Unknown

Low levels of hypocretins (orexins), suggesting the loss of the brain cells that secrete hypocretin.

Falling into a (Mieda et al. deep sleep at any 2004) time, drowsiness, sudden weakness of the muscles that leads to collapse.

Neurofi- NF2 gene / 22 (tumor Impairment of brake on cell bromato- suppressing gene) sis, type 2 growth and division. (NF-2)

Development of (Rong et al. malignant brain 2004) tumors and benign tumors on both auditory nerves.

Niemann- Gene on chromosome Pick type 18 (plays a role in C (NP-C) cholesterol homeostasis

Brain and nerv- (Ko et al. ous system im- 2003, Ko et pairment leading al. 2005) to death.

Excessive build-up of cholesterol inside cells, causing ill processmg.

244

Appendix 1 Brain abnor- Symptoms mality

References

Parkinson a-synuclein gene disease (SNCA) /4q21 (PD) parkin gene (PRKN) / 6q25 DJl / 1p36 gene for PTENinduced putative kinase 1 (PINK1) / 1p36 gene for dardarin (LRRK2) /12q12

Neurotoxicity by aggregation of asynuclein, impaired protein degradation via parkin, impaired oxidative stress response via DJ1, mitochondrial dysfunction.

Presence of an inclusion bodies, called the Lewy bodies, in basal ganglia resulting in tremor, muscular stiffness and difficulty with balance and walking.

(Bertram and Tanzi 2005, Eckert and Eidelberg 2005)

Phenylke- Both alleles of the gene for phenylatonuria lanine hydroxylase (PKU) (PAH) /12 (converts phenylalanine to tyrosine)

Concentration of phenylalanine in the body builds up to toxic levels.

Mental retarda- (Kaufman tion, organ dam- 1999, Knerr age, unusual et al. 2005) posture.

Disease

Mutations of genes identified so far / chromosome location if known (protein role if known)

PraderWilli syndrome (PWS)

Absence of segment 11-13 on the long arm of the paternally derived chromosome 15 (mRNA processing)

Impaired in- Mental retardatermediate step tion, decreased between DNA muscle tone, transcription short stature, and protein emotionallabilformation. ity and an insatiable appetite.

Refsum disease

PAHX gene / 10 (required for the metabolism of phytanic acid)

Impaired PAHX -phytanic acid hydrolase

(Wang, Liu et al. 2004, Schule et al. 2005)

(Brink et al. Degenerative 2003) nerve disease (peripheral neuropathy), ataxia, retinitis pigmentosa (vision disorder), bone and skin changes

A.l Table of Genes and Related Brain Functions and Diseases

245

Disease

Mutations of genes identified so far / chromosome location if known (protein role if known)

Brain abnor- Symptoms mality

References

Rett syndrome (RTT)

MeCP2 gene / X (controls gene expression)

It is not clear

Loss of purposewhat the ful use of hands mechanism is. and speech, reduced muscle tone, wringing hand movements, autisticlike behavior and seizures.

(Zoghbi 2003, Glaze 2005, Zoghbi 2005)

Schizophrenia (SCH) (Many genes may be involved)

COMT Val allele (decreases prefrontal dopamine), myelination and growth-related genes, G72/ l3q34 (interacts with the gene for D-amino acid oxidase (DAAO) on l2q24 (NMDAR signaling )

Impairment of prefrontal cortex function, disruption in oligodendrocyte function, disorders of neural development.

Personality changes, disorder of thoughts (delusions), perceptory phenomena (hallucinations), degradation of ability to perform daily functions.

(Egan et al. 2001, Cloninger 2002, Sugai et al. 2004, Barnett et al. 2005)

Spinal Muscular Atrophy (SMA)

Mutation of both alleles of the survival motor neuron gene (SMNl) /5

Death of spinal motor neurons and subsequent muscle paralysis.

Neuromuscular disorder that may result in death from respiratory failure (type I), reduced life expectancy (type II), inability to walk (types II and III).

(Federici and Boulis 2005, Gabanella et al. 2005)

Degeneration of the spinal cord and the cerebellum.

Loss of muscle coordination, spasticity.

(RaIser et al. 2005)

Spinocere- Expansion of a CAG triplet in the SCA1 bellar gene /6 (function unataxia known)

246

Appendix 1

Disease

Mutations of genes Brain abnor- Symptoms identified so far mality / chromosome location if known (protein role if known)

References

Tay-Sachs Mutations in both alleles of a gene HEXA disease / 15 (helps to degrade a lipid called GM2 ganglioside)

Excessive accumulation of the GM2 ganglioside in neurons.

Paralysis, dementia, blindness and early death. Neuron dysfunction and psychosis in chronic adult form.

Tuberous sclerosis

Growth of cells proceeds in an unregulated fashion, resulting in tumor formation.

Benign, tumor- (Inoki et al. like nodules of 2005) the brain and/or retinas, skin lesions, seizures and/or mental retardation

TSCI /9 TSC2/16 (tumor suppressor functions)

Williams Deletion of genes for Deletion of LIM kinase syndrome LIM kinase and leads to imelastin / 7 pairment of visuospatial cognition.

Wilson's disease

ATP7B /13 (metabo- Copper acculism of copper) mulation and toxicity to the liver and brain.

High competence in language, mUSIC and interpersonal relations, with low IQ, heart and vein problems.

(Rajavel and Neufeld 2001, Martino et al. 2005)

(Basalyga et aI. 2004, Soosairajah et al. 2005)

Liver and neuro- (Cater et al. logical disease, 2004) cornea of the eye can also be affected.

Appendix 2

A.2 A Brief Overview of Computational Intelligence Methods Computational Intelligence (CI) is the area of developing generic intelligent information processing methods and systems with wider applications. CI methods, in its majority, are inspired by the human intelligence. They are characterized by learning, generalization, adaptation, pattern recognition, rule extraction, knowledge representation, which are characteristics of the living systems too. The methods of CI include: • Probabilistic and statistical learning methods (e.g. Bayesian classifiers, data clustering, Markov Models, Support Vector Machines (SVM), transductive SVM and SVM trees). • Rule-based systems (propositional logic dated back to Aristotle) and fuzzy systems (introduced by Zadeh (Zadeh 1965, Zadeh 1979)). • Artificial neural networks. • Evolutionary computation, genetic algorithms, particle swarm intelligence and other artificial life approaches. • Quantum computation and nanotechnology. • Hybrid systems (e.g. knowledge-based neural networks; neuro-fuzzy systems; neuro-fuzzy-genetic systems; evolving connectionist systems). A.2.1 Probabilistic and Statistical Methods

These methods are based on probability of event estimation and their statistical analysis. Bayesian methods are among the most popular (Pang 2004). They are based on the Bayesian probability that represents the conditional probability between two events C and A (Bradley and Louis 1996):

248

Appendix 2

p(A IC)= p(A IC)p(A) p(C)

(A.2.l)

Sometimes, using the Bayesian formula involves difficulties, mainly concerning the evaluation of the prior probabilities peA), p( C), p( C IA). Bayesian networks are used to describe relationships between genes in a genetic regulatory network (GRN) (Baldi and Long 2001, Hartemink et al. 2001). Unlike other approaches such as clustering, a Bayesian network can describe arbitrary combinatorial control of gene expression and thus it is not limited to pair-wise interactions between genes. Due to their probabilistic nature, Bayesian networks are robust in the face of both imperfect data and imperfect models. Most importantly, the models are biologically interpretable and can be scored rigorously against observational data. A very popular statistical technique for discovering patterns in data is clustering. Based on a measured distance between instances (objects, points, vectors) from the problem space, groups of close instances can be defined. These groups are called clusters. They are defined by their cluster centers and the membership of the data points to them. A centre c, of a cluster C, is defined as an instance, the mean of the distances to which from each instance in the cluster, is less than its distance to another cluster centre. Let us have a set X of p data items represented in an n-dimensional space. A clustering procedure results in defining k disjoint subsets (clusters), such that every data item (n-dimensional vector) belongs to one only cluster. A cluster membership function M is defined for each of the clusters Cr, Cz,...,

c.

M, : X ---+ {O,l}

(A.2.2)

1' XE C, M.(x) = { O,xrteC, ,

where x is a data instance (vector) from X. A significant characteristic of clustering is how distance between vectors is measured. The distance between two data points in an ndimensional geometrical space can be measured in several ways, e.g.: • Hamming distance: (A.2.3) • Euclidean distance: (A.2.4)

A.2 A Brief Overview of Computational Intelligence Methods

249

A special type of clustering is called fuzzy clustering in which clusters may overlap, so that each of the data instances may belong to each of the clusters to a certain degree (Bezdek 1981). The procedure aims at finding the cluster centers Vi (i = 1,2, ..., c) and the cluster membership functions jJi, which define to what degree each of the n examples belong to the /h cluster. The number of clusters c is either defined a priori (supervised type of clustering), or chosen by the clustering procedure (unsupervised type of clustering). The result of a clustering procedure can be represented as a fuzzy relation jJi,h such that:

1.

L>u =

2.

all clusters equals 1); >0, for each i = 1,2, ..., c (there are no empty clusters).

1,

for each k = 1,2, ..., n; (the total membership of an instance to

Lfl"

Probabilistic and statistical methods are used widely in Bioinformatics tools and systems for biological data analysis and modeling. Regression and discrimination analysis, along with Support Vector Machines (SVM) (Vapnik 1998) are popular techniques to build a classifier function. The idea of SVM is to transform original data into higher dimensional feature space via a kernel computation, and to construct a separating hyper plane with maximum margin between the class samples (see Fig. A.2.1 as an example). These kernel functions could be polynomial, radial basis, linear, etc.

Fig. A.2.t. Support vector machines define vectors (called support vectors) at the border betweenthe samples of two classes

SVM are widely used for gene expression and protein expression data classification and profiling (Leslie et al. 2002, Pang and Kasabov 2004, Wang et al. 2006).

250

Appendix2

Stochastic Models Stochastic models deal with the dynamic history of each object of the model. In other word, for each object the next state must be calculated using a set of probabilistic rules. Each rule shows the probability for object to be changed in particular interval of time, and probability to come to each state. So, the change of state in this type of model is probabilistic, not deterministic. Let us assume that object x in the system has a finite state space with L states (like in a kinetic logic model ): {Xi, Xl, , Xr} . For each time step tHI there is a transition probability P(xHdxo, ,Xk) and a chain xo, ..., X k represents the history of the system. Variables X k form Markov chain if and only if for any k: (A.2.5) In other words, a future state depends only on the current state . All probability values P(xilxj) that represent the probability for the system to move from the i h to the state form a transition matrix. Suppose that a system can move to state Xi at time t, with a transition rate ~. The probability of the system to move to the lh state at the time tk

l

IS: L

(A.2.6)

i A-"A - L.J k j =!

The formula for calculating a next time point depends on the distribution of the move s tk+ i - tk, and, for instance, in the case of an exponential process, the next time point is the following: tH i = t k - In(r) /2, where r, is a random value uniformly distributed in (0, l). Stochastic models are used for modeling gene regulatory networks (GRN) (Baldi and Hatfield 2002 , Roberts et al. 2005).

A.2.2 Boolean and Fuzzy Logic Models

Boolean Models k

Consider the set of N objects at time h {t/ , t/ , ..., tn } and each object can be in only two different states: on/off, 1/0, False/True, etc. For simplicity, let us assume:

A.2 A Brief Overview of Computational Intelligence Methods

x;

E

{O,l}

i = 1,...,N

251

(A.2.7)

The state of the system at a time moment can be described as the states of all objects in this set. The state of a given object at the next time step tHI can be determined by a Boolean logic function (returning only two values: 0 or 1) that takes as an input the current state of the system: (A.2.8)

A Boolean function B = {B I , B2, ••• , BN}can be represented as a truth taN ble which consists of all possible system states (2 ) . This function represents relations between all system's states and can be represented as a diagram. Examples of Boolean functions are given below: (A.2.9)

where: I, logical OR; and ~, logical NOT. Boolean methods are used for gene regulatory network modeling to represent the expression of a gene (1 expressed; or 0 not expressed) and the connection between the genes (1- excitatory, and -1 - inhibitory) (Somogyi et al. 2001). Boolean models are very limited in terms of representing the "grayness" in biological systems and the "smoothness" in the interaction of its elements.

Fuzzy Logic Models Fuzzy logic is a logic system that is based on fuzzy relations and fuzzy propositions, the latter being defined on the basis of fuzzy sets (Zadeh 1965). A fuzzy set is a set defined by a membership function to which each domain value can belong to any degree of membership between 0 and 1 and not just (1 (belong) and 0 (does not belong) as in ordinary sets. A variable that can take as values symbolic concepts, such as Small, Medium, High, each defined by their fuzzy membership function, is called fuzzy variable. Fuzzy propositions are propositions which contain fuzzy variables with their fuzzy values. The truth value of a fuzzy proposition " X is A " is given by the membership function JiA of the fuzzy value A. Fuzzy relations link two fuzzy sets in a predefined manner. Fuzzy relations make it possible to represent ambiguous relationships, like: "the expressions of the genes in cluster 2 and cluster 3 are similar", or "model A performed more or less better than model B", or "the more the gene X is expressed, the higher the risk of cancer".

252

Appendix 2

Fuzzy logic allows for representation of "grayness", "smoothness", inexactness, flexibility, mobility of relations and concepts, which is important when dealing with biological data and concepts. Examples are: "high/low activity of a brain segment", "high/low gene expression values", "strong/week binding between proteins", "more or less similar structures", "fast growing culture", and many more. Short

Medium

Tall

0.7

OL..--"'-----"';:O".......::;.---L..--.,;:----......... 30 170 250 Height (em)

Fig. A.2.2. Fuzzy concepts 'short', 'medium' and 'tall' defined by membership functions

Afuzzy model is represented usually as a set of fuzzy rules and an inference algorithm. An example of a set of two general fuzzy rules, where each rule has two fuzzy input variables xl and x2 and one fuzzy output variable is: Rule r.: IF xl is Short (DIll) and x2 is Short (DI2I)

(A.2.l0)

THEN Output is Short (CF1) Rule r{ IF xl is Tall (DIl) and x2 is Tall (DI2})

(A.2.II)

THEN Output is Tall (CF})

'Short' and 'tall' are fuzzy concepts defined by their respective fuzzy membership functions and DI and CF are degree of importance (membership) and certainty factors respectively (see Fig. A.2.2). These rules are facilitated in the ANN structure shown in Fig. A.2.3 and explained later in the next section.

A.2 A Brief Overview of Computational Intelligence Methods

-,

253

FuzzyOLtputs ..........

-,

-.

!

-, \.

Fig. A.2.3. Fuzzy rules 1.10 and 1.11 implemented in EFuNN architecture

A.2.3 Artificial Neural Networks

Artificial neural networks (A1\TN) (connectionist systems) are computational models that mimic vaguely the nervous system in its main functions of adaptive learning and generalization (Amari 1990). They are universal computational models so that any algorithm or function can be realized as an ANN model. Moreover, A1\TN can learn functions from data without specifying the type of the function. They are called model-free estimators . ANNs provide a model of computation that is different from traditional algorithms. Typically, they are not explicitly programmed to perform a given task; rather, they learn to do the task from examples of desired input/output behavior. The networks automatically generalize their processing knowledge into previously unseen situations, and they perform well for the noisy, incomplete or inaccurate input data. In general view, artificial neural network is the model consisting of interconnected units evolving in time. A connection between units i and j is usually characterized by a weight denoted as w ij ' There are three important architectures ANN based on the connectivity: • Recurrent (contains direct loops from output units (nodes) back to input nodes). • Feedforward (contains no direct loops). • Layered (units organized into layers and connections are between layers).

254

Appendix 2

The behavior of each unit in time can be described by a time-dependent function, or a stochastic process, or a Bayesian formula, etc. So, i h unit receives total input x from the units connected to it and generates a response based on an activation function. When this function is a threshold function 1,X> 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