Pacific Symposium on Biocomputing 3:18-29 (1998)

REVEAL, A GENERAL REVERSE ENGINEERING ALGORITHM

FOR INFERENCE OF GENETIC NETWORK ARCHITECTURES

SHOUDAN LIANG
SETI Institute, NASA Ames Research Center,
Moffett Field, CA 94035 (sliang@[Link])

STEFANIE FUHRMAN, ROLAND SOMOGYI

Molecular Physiology of CNS Development, LNP/NINDS/NIH, 36/2C02, Bethesda, MD
20892 ([Link]
sfuhrman@[Link]; rolands@[Link])

Given the immanent gene expression mapping covering whole genomes during
development, health and disease, we seek computational methods to maximize
functional inference from such large data sets. Is it possible, in principle, t o
completely infer a complex regulatory network architecture from input/output
patterns of its variables? We investigated this possibility using binary models
of genetic networks. Trajectories, or state transition tables of Boolean nets,
resemble time series of gene expression. By systematically analyzing the mutual
information between input states and output states, one is able to infer the sets
of input elements controlling each element or gene in the network. This process
is unequivocal and exact for complete state transition tables. We implemented
this REVerse Engineering ALgorithm (REVEAL) in a C program, and found the
problem to be tractable within the conditions tested so far. For n=50 (elements)
and k=3 (inputs per element), the analysis of incomplete state transition tables
(100 state transition pairs out of a possible 10 15) reliably produced the original
rule and wiring sets. While this study is limited to synchronous Boolean
networks, the algorithm is generalizable to include multi-state models,
essentially allowing direct application to realistic biological data sets. The
ability to adequately solve the inverse problem may enable in-depth analysis of
complex dynamic systems in biology and other fields.

Binary models of genetic networks

Virtually all molecular and cellular signaling processes involve several inputs and
outputs, forming a complex feedback network. The information for the construction
and maintenance of this signaling system is stored in the genome. The DNA

18
19 [Link]

sequence codes for the

a c input output structure and molecular
A B C A' B' C ' dynamics of RNA and
A B C 0 0 0 0 0 0 proteins, in turn determining
0 0 1 0 1 0
biochemical recognition or
0 1 0 1 0 0
signaling processes. The
0 1 1 1 1 1
regulatory molecules that
1 0 0 0 1 0
control the expression of
A’ B’ C’ 1 0 1 0 1 1
genes are themselves the
1 1 0 1 1 1
1 1 1 1 1 1 products of other genes.
b Effectively, genes turn each
other on and off within a
A' = B
proximal genetic network of
B' = A or C
C' = (A and B) or (B and C) or (A and C) transcriptional regulators
(Somogyi and Sniegoski,
1996). Furthermore,
Fig. 1 A simple Boolean network. a) Wiring diagram. b)
Logical (Boolean) rules. c) Complete state transition complex webs involving
table defining network. The input column corresponds to various intra- and
the state at time=t, the output column (elements marked extracellular signaling
by prime) corresponds to the state at time=t+1. systems on the one hand
depend on the expression of
the genes that encode them, and on the other hand control the expression of genes as
the signals terminate at transcriptional regulation. All in all, the information stored
in the DNA determines the dynamics of the extended genetic network, the state of
which at a particular time point should be reflected in gene expression patterns
(Somogyi and Sniegoski, 1996). We have developed the basic tools to measure
these gene expression patterns, and are now concerned with inferring the functional
network architectures from time series or state transition sets (Somogyi et al.,
1996).
A rational approach to designing genetic network analysis tools is based on
generating model systems on which the performance of the tools can be tested. The
simplest such model system is the Boolean network. Genes correspond to elements
in a Boolean net, the wiring of the elements to one another correspond to functional
links between genes, and the rules determine the result of a signaling interaction
given a set of input values. Genes are idealized as being either on or off, resulting in
binary elements interacting according to Boolean rules. Given a particular set of
elements, wiring, rules, a particular trajectory or the state transition table covering
all trajectories of a network can be calculated (Fig. 1). Such a trajectory must reach a
final repeating state cycle, for the simple reason that the network only has 2n states,
and each state transition is unequivocally determined (after maximally 2n iterations, a
repeating state must be found). An attractor may be a single state (point attractor,
corresponding to a “steady state”) or may comprise several states (dynamic attractor,
Pacific Symposium on Biocomputing 3:18-29 (1998) 20

corresponding to a “limit cycle”). Attractors may be envisioned as the “target area”

of an organism, e.g. cell types at the end of development, repaired tissue following a
response to injury, or even adaptation of metabolic gene expression following a
change in nutrient environment in bacteria (see Kauffman, 1993; Somogyi and
Sniegoski, 1996; Wuensche, 1992).
Testing of algorithms for extracting network architectures from state
transition measurements will require knowledge of the original network that was to
be inferred. This is not yet possible with living systems. Using Boolean networks,
we can facilely generate model state transition tables. Depending on the assumptions
we make about living genetic networks regarding size, connectivity, redundancy and
complexity, we can simulate these conditions in Boolean nets, and test how well
our reverse engineering procedure works against these different backdrops. Given
good results, our confidence in this approach should be warranted. Below we will
use systematic mutual information analysis of Boolean network state transition
tables to extract minimal network architectures.

Information theoretic principles of mutual information (M) analysis

Information theory provides us with a quantitative information measure, the

Shannon entropy, H. The Shannon entropy is defined in terms of the probability of
observing a particular symbol or event, p i, within a given sequence (Shannon &
Weaver, 1963),

H= - Σ pi log pi.

A few illustrations (Figs. 2 & 3) of a binary system shall help explain the
behavior of H. In a binary system, an element, X, may be in either of s=2 states,
say on or off. Over a particular sequence of events (Fig. 2a), the sum of the
probabilities of X being on, p(1) or off, p(0) must be equal to unity, therefore
p(1)=1-p(0), and H(X)=-p(0)*log[p(0)]-[1-p(0)] *log[1-p(0)]. H reaches its maximum
when the on and off states are equiprobable (Fig. 3a), i.e. the system is using each
information carrying state to its fullest possible extent. As one state becomes more
probable than the other, H decreases - the system is becoming biased. In the
limiting case, where one probability is unity (certainty) and the other(s) zero
(impossibility), H is zero (no uncertainty - no freedom of choice - no information).
The maximum entropy, Hmax, occurs when all states are equiprobable, i.e.
p(0)=p(1) =1/2. Accordingly,

Hmax=log(2).

Entropies are commonly measured in “bits” (binary digits), when using the
logarithm on base 2; e.g. Hmax=1 for a 2 state system.
21 [Link]

Our aim is to
compare different sequences a
using information measures X 0 1 1 1 1 1 1 0 0 0
to establish functional Y 0 0 0 1 1 0 0 1 1 1
relationships between
elements of a network. In a H(X) = -0.4log(0.4)-0.6log(0.6) = 0.97 (40% 0s and 60% 1s)
system of 2 binary elements, H(Y) = -0.5log(0.5)-0.5log(0.5) = 1.00 (50% 0s and 50% 1s)
X (index i) and Y (index j),
the individual and combined
Shannon entropies are defined b
essentially as above (Fig. 1 3 2
2b): Y

H(X)= - Σ pi log pi , 0 1 4

H(Y)= - Σ pj log pj , and 0 1

H(X, Y) = - Σ pi, j log pi, j X

H(X,Y) = -0.1log(0.1)-0.4log(0.4)-0.3log(0.3)-0.2log(0.2) = 1.85

There are 2 conditional
entropies which capture the Fig. 2 Determination of H. a) Single element.
relationship between the Probabilities are calculated from frequency of on/off
sequences of X and Y, values of X and Y. b) Distribution of value pairs. H i s
calculated from the probabilities of co-occurrence.
H(X|Y) and H(Y|X). These
are related as follows
(Shannon & Weaver, 1963):

H(X,Y) = H(Y|X) + H(X) = H(X|Y) + H(Y) . 1

In words, the uncertainty of X and the 0.8

remaining uncertainty of Y given knowledge
of X, H(Y|X), i.e. the information contained 0.6
H

in Y that is not shared with X, sum to the

0.4
entropy of the combination of X and Y.
We can now find an expression for 0.2
the shared or “mutual information”, M(X,Y),
also referred to as “rate of transmission” 0
between an input/output channel pair 0 0.25 0.5 0.75 1

(Shannon & Weaver, 1963): p(0)

M(X,Y) = H(Y) - H(Y|X) = H(X) - H(X|Y). Fig. 3 Shannon entropies for a 2-state
information source. Since the sum of
The shared information between X and Y the state probabilities must be unity,
corresponds to the remaining information of X p(1)=1-p(0) for 2 states.
Pacific Symposium on Biocomputing 3:18-29 (1998) 22

if we remove the information of X that is

not shared with Y. Using the above H(X) + H(Y)
equations, mutual information can be defined
directly in terms of the original entropies;
this formulation will be important for the
considerations below:

M(X,Y) = H(X) + H(Y) - H(X,Y).

H(X) H(Y)
The Venn diagrams of Fig. 4 illustrate the
relationships between these measures. We
will use these information principles to H(X,Y)
extract the critical connections between
network elements from binary network state
transition data. Similar analyses have been
previously explored in the classification of
Boolean rules (Somogyi & Fuhrman, 1997)
and genotype-phenotype mappings (Sawhill,
1995). H(X|Y) H(Y)
The core of REVEAL: systematic
M-analysis of state transition tables
M(X,Y)
The general strategy of our algorithm is to
use mutual information measures to extract
the wiring relationships from state transition
tables. These directly lead to the look-up
tables of the rules. We shall explain the
step-by-step workings of the algorithm (Fig.
5) in the analysis of the network example of
Fig. 4 Venn diagrams of information
Fig. 1.
relationships. In each case, add the
shaded portions of both squares t o
Step 1: Identification of k=1 links determine one of the following:
We begin by determining the pair-wise [H(X)+H(Y)], H(X,Y), and M(X,Y). The
mutual information matrix (Fig. 5) of all small corner rectangles represent
single input-output pairs (Fig. 1). A “prime” information that X and Y have i n
denotes the output state of an element, e.g. common. H(Y) is shown smaller than
A’. If M(A’,X)=H(A’), i.e. H(X) and with the corner rectangle o n
M(A’,X)/H(A’)=1, then X exactly determines the left instead of the right to indicate
A’. This is the case for B and A’ in Fig. 5. that X and Y are different, although they
have some mutual information.
Note: Since H(A’)=M(A’,X), then
23 [Link]

Input entropies

H(A) 1.00 H(X) = - ∑ p(x) log p(x)

H(B) 1.00 H(X,Y) = - ∑ p(x,y) log p(x,y)
H(C) 1.00
H(A,B) 2.00 M(X,Y) = H(X) + H(Y) - H(X,Y)
H(B,C) 2.00 M(X,[Y,Z]) = H(X) + H(Y,Z) - H(X,Y,Z)
H(A,C) 2.00
H(A,B,C) 3.00

Rule table for A

Determination of inputs for element A 2 rule no. 2
1 input output
H(A') 1.00 B A'
H(A',A) 2.00 M(A',A) 0.00 M(A',A) / H(A') 0.00 0 0
H(A',B) 1.00 M(A',B) 1.00 M(A',B) / H(A') 1 . 0 0 1 1
H(A',C) 2.00 M(A',C) 0.00 M(A',C) / H(A') 0.00

Rule table for B

Determination of inputs for element B 4 rule no. 14
3 input output
H(B') 0.81 A C B'
H(B',A) 1.50 M(B',A) 0.31 M(B',A) / H(B') 0.38 0 0 0
H(B',B) 1.81 M(B',B) 0.00 M(B',B) / H(B') 0.00 0 1 1
H(B',C) 1.50 M(B',C) 0.31 M(B',C) / H(B') 0.38 1 0 1
H(B',[A,B]) 2.50 M(B',[A,B]) 0.31 M(B',[A,B]) / H(B') 0.38 1 1 1
H(B',[B,C]) 2.50 M(B',[B,C]) 0.31 M(B',[B,C]) / H(B') 0.38
H(B',[A,C]) 2.00 M(B',[A,C]) 0.81 M(B',[A,C]) / H(B') 1.00
Rule table for C
6 rule no. 170
Determination of inputs for element C input output
5 A B C C'
H(C') 1.00 0 0 0 0
H(C',A) 1.81 M(C',A) 0.19 M(C',A) / H(C') 0.19 0 0 1 0
H(C',B) 1.81 M(C',B) 0.19 M(C',B) / H(C') 0.19 0 1 0 0
H(C',C) 1.81 M(C',C) 0.19 M(C',C) / H(C') 0.19 0 1 1 1
H(C',[A,B]) 2.50 M(C',[A,B]) 0.50 M(C',[A,B]) / H(C') 0.50 1 0 0 0
H(C',[B,C]) 2.50 M(C',[B,C]) 0.50 M(C',[B,C]) / H(C') 0.50 1 0 1 1
H(C',[A,C]) 2.50 M(C',[A,C]) 0.50 M(C',[A,C]) / H(C') 0.50 1 1 0 1
H(C',[A,B,C]) 3.00 M(C',[A,B,C]) 1.00 M(C',[A,B,C]) / H(C') 1.00 1 1 1 1

Fig. 5 Outline of progressive M-analysis underlying REVEAL for network example

shown in Fig. 1. Hs and Ms are calculated from the look-up tables according to the
definitions (shaded). The network wiring is extracted by M-analysis (left, odd steps).
Rule tables are then determined directly from the trajectory (right, even steps).

H(X)=H(A’,X), i.e. it is not even necessary to calculate M(A’,X) explicitly, making

the computation marginally faster. The measurement of M must be sufficiently
precise in the analysis of many state transition pairs, i.e. the determination of p (the
probability) must be able to distinguish a change of 1/T (T=number of state
transition pairs).
Pacific Symposium on Biocomputing 3:18-29 (1998) 24

Step 2: Determination of k=1 rule

The look-up table for the input/output pairs (Fig. 1) constitutes the rule. Redundant
input/output pair listings are eliminated in the proper rule table format (Fig. 5).
Step 3: Identification of k=2 links
If not all state transitions can be explained in terms of k=1, we will determine
entropies of input pair combinations with the remaining unsolved output elements.
If M(B’,[X,Y])=H(B’), or, more concisely, H(B’,X,Y)=H(X,Y), then the pair [X,Y]
completely determines B’. In Fig. 5, no single input can predict B’, but the pair [A,
C] accounts for all of the entropy of B’.
Step 4: Determination of k=2 rule (as above for k=1)
Step 5: Identification of k=3 links
If not all elements can be resolved in terms of k=1 and k=2, the next step is to
determine the entropies of input triplet combinations with the remaining unsolved
output elements. In our example (Fig. 5), since M(C’,[A,B,C])=H(C’), or, more
concisely, H(C’,A,B,C)=H(A,B,C), then [A,B,C] completely determines C’.
Step 4: Determination of k=3 rule (as above for k=1)
Step i: Identification of k=i links (i<=n, number of network elements).
This applies to networks of any size. If not all trajectories can be explained in terms
of k=i-1, i-2 . . . 1 inputs, the search pursues the entropies of i-let input
combinations with the remaining unsolved output elements. If M(Y, [X1, X 2, X 3,. .
. Xi])=H(Y), or H(Y,X1, X 2, X 3,. . . Xi)= H(X1, X 2, X 3,. . . Xi) then the i-let,
[X1,X2,X3,. . .Xi] completely determines Y (Y=output element). Naturally, the input
value combinations of the combined state transition tables covering Y and
[X1,X2,X3,. . .Xi] define the look-up table of the rule.

The advantage of this algorithm is that simple networks can be calculated very
quickly just by comparing Hs of state transition pairs. The algorithm will calculate
the Hs for higher k only as required. Of course, as k increases, the calculations of
the Hs will require progressively more time (see below). The goal is obviously to
minimize the number of computationally intensive operations. We are currently
exploring rational search optimization procedures (e.g. minimization of k
combination testing) based on probable rule restrictions. Moreover, REVEAL is
amenable to parallel computing, which we are planning to pursue in the future.

Implementation of the algorithm

A practical implementation of REVEAL will require careful considerations regarding

k
non-redundant and biologically feasible rule inference. Amongst 2( 2 ) input rules,
many do not depend on one or more of their inputs. These rules are equivalent to a
rule with a smaller keff (effective k; Somogyi & Fuhrman, 1997). Since we can only
detect rules that truly depend on all of their inputs, the best we can do is to infer the
25 [Link]

equivalent rule with minimum k eff. For this reason, the k-input rules we used in
constructing test networks are effective k-input rules, i.e. that they cannot be reduced
to a rule with smaller number of inputs.
There are two one-input (k=1) and ten two-input (k=2) rules that truly
depend on all their inputs. Two of the ten two-input rules, exclusive or and
equivalent, may be unlikely to occur biologically. They produce minimally
correlated behavior in networks (atypical for biological networks), and would be
difficult to encode in biomolecular interactions. One may consider eliminating such
rules from biologically feasible test networks.
For k=3 rules, there are 218 rules of keff=3, 30 of keff=2, 6 of keff=1 and 2
of keff=0. Of course, we limit the construction of model networks to rules of an
effective k. Moreover, k=3 rules may be further restricted according to biological
plausibility (as above for k=2).
In order to infer the rule for a particular gene, our strategy is to first test if
it is an effective one-input rule. Since the input for the gene could be anywhere in
the network, there are N possible inputs for each gene. Each one is tested in turn
using the mutual information analysis discussed earlier. For genes whose output is
not determined solely by any one input, the effective k for the rule of that gene is
larger than one. We next determine whether the gene is determined by a rule with
N ( N − 1)
two effective inputs. There are pairs of possible inputs for a two-input
2
rule. For each of the these input pairs, we use the M-analysis to determine whether
the input pair specifies the output value for the gene. In general we have
 N N!  N
 k = possible inputs for a k-input rule. All of the   input
  k!( N − k )  k
combinations are examined to find the correct input set.

Performance of Algorithm

In principle, the information theoretic approach requires computation of the

 N
probability over all 2N state transition pairs of the network for each of the  k
 
possible wirings of the k-input rule. For a network of moderate size (N=50), the
number of configurations becomes too large to compute. Fortunately, the criterion
for determining the causality relationship in the M-analysis is satisfied using any
finite input set provided that the same set is used in computing all quantities
involved. For example, the criterion (see above) for determining that C’ is the
output of A and B is H([A,B],C’)=H(A,B). A finite set of randomly selected input
patterns [A,B] can be used to construct a 4-bit histogram. From the histogram, we
Pacific Symposium on Biocomputing 3:18-29 (1998) 26

obtain the probabilities needed to compute H(A,B). H([A,B],C’) is computed from

an 8-bit histogram using the same set of input patterns combined with the output
C’. If [A,B] is the correct input for C’, then H([A,B],C’)=H(A,B) will be satisfied
for any input set.
However, if the number of input-output pairs is small, there will be a large
number of mis-identifications because of incidental degeneracy. In order to estimate
the size of the sample set needed to uniquely identify the right wiring for a gene, we
compute the probability of mis-identification as a function of increasing the sample
size, S, which is defined as the number of input-output pairs used in computing the
histogram. The probability is computed by counting the number of input wirings
 N
that satisfy the M-analysis criterion normalized to all possible  k  wirings for a
 
k-input rule. Fig. 6 shows that the probability declines exponentially with
increasing S. With a very small S, much smaller than the total number of all
possible state transition patterns 2N, we can already identify the correct wiring.

The networks used in testing REVEAL are constructed using a mixture of

1E+0
No. of mis-identified solutions

1E-1 k=1
1E-2 k=2
k=3
( normalized)

1E-3

1E-4
1E-5

1E-6
1E-7

1E-8
0 20 40 60 80 100
State transitions
Fig. 6 Reduction of mis-identified network wiring solutions. The number of
erroneous wirings identified by the M-analysis (normalized) versus the number
of state transition pairs used for effective k value k=1,2,3. The data was
obtained by averaging over 50 random wirings for a network with 50 elements.
Note that a correct solution is always found; this is subtracted from the plotted
number of solutions.
27 [Link]

1.0000
P (incorrect solution)

0.1000

0.0100

k=1
0.0010 k=2
k=3

0.0001
0 20 40 60 80 100
State transitions
Fig. 7 Convergence of solution in random network. The probability of not
finding the perfect solution, P, versus the number of state transition pairs used,
S, for effective k value k=1,2,3. Each data point is computed by averaging over
150 random wirings for a network with 50 elements. The network i s
constructed with one third each of one-input, two-input and three-input rules.
For each of the three cases, the rule selection is made at random amongst all the
effective k-input rules. As more transition pairs are used, the probabilities
decay exponentially at large S after a relatively flat plateau. Our data also
indicate that P becomes zero at S=100 for k=3; at S=60 for k=2; and at S=20
for k=1 (not graphically depictable on log scale).

one-input, two-input, and three-input rules with equal probability. When a gene is
assigned a k-input rule, one of k-input rules is selected for the gene at random from
all eligible rules. In the case of Fig. 7, all the rules that truly depend on all their
inputs are eligible. There are 2 such one-input rules, 10 such two-input rules and
218 such three-input rules.
Every rule assigned to genes has been correctly identified for all 150
networks used for Fig. 7. In the most difficult case of keff=3, all the rules have been
uniquely identified (perfect solution) when the number of state transition pairs
reaches 100. For one-input and two-input rules, the perfect solution is reached when
S=20 and S=60 respectively for all the genes in 150 networks. When S is smaller
than these limit values, some genes are allocated more than one set of inputs by the
M-analysis, i.e. there is more than one solution. The number of degenerate
solutions as a function of the number of state transition pairs was discussed in
Fig. 6.
Pacific Symposium on Biocomputing 3:18-29 (1998) 28

Outlook

We have shown that REVEAL performs well for networks of low k (number of
inputs per gene). For higher k, the algorithm should be accelerated through a)
parallelization, and b) increasing the search efficiency of solution space, e.g. by
taking maximal advantage of wiring and rule constraints. We are currently pursuing
these strategies.
Boolean networks are based on the notion that biological networks can be
represented by binary, synchronously updating switching networks. In real
biological systems, however variables change continuously in time. This behavior
can be approximated by asynchronous Boolean networks (reviewed in Thieffry &
Thomas, 1998), or continuous differential equations that capture the structure of
logical switching networks (Glass, 1975). The issue of determining the logical
structure of a continuous network based on knowledge of the transitions was
explicitly addressed in a previous work on oscillating neural networks (Glass and
Young, 1979). The point of REVEAL is to base causal inference on the most
fundamental and general correlation measure available, mutual information. While
we concentrated on idealized Boolean networks, mutual information measures can be
applied to multivalued discrete and also continuous data sets. Of course, once
multiple states are introduced, corresponding flexibility will also be found in the
timing. Since continuous behavior can be approximated by discrete systems given
sufficient resolution, REVEAL could be applied to appropriately discretized
continuous data sets. However, the introduction of multiple states will greatly
increase the number of theoretically possible state transitions; network and wiring
constraints must therefore be carefully considered when generalizing REVEAL to
multivalued networks. For example, integration of cluster analysis for the inference
of shared inputs (currently applied to continuous, large scale gene expression data
sets; see Michaels et al., 1998) could quickly identify wiring constraints and
simplify the overall inference process.
Finally, as REVEAL or potential successors become more refined, we need
to consider the data sets that must be generated to allow maximal depth of inference.
The algorithm relies on the analysis of state transitions or temporal responses of
gene expression patterns (or other relevant biological parameters!) to perturbations
or internal changes (e.g. development). What will be the proper time step across
which measurements need to be acquired and interpreted? How many perturbations
will be necessary to capture sufficient diversity? How many states (if more than
binary) need to be attributed to each biological parameter? The potential rewards of
fundamental insights into genetic and biological signaling networks should
encourage us to pursue these questions.
29 [Link]

Acknowledgments

S.L. acknowledges the support of NASA Cooperative Agreement NCC2-974. We

thank Ian McDonald, Leon Glass and Patrick D’haeseleer for helpful comments.

References

Glass, L. (1975). Classification of Biological Networks by Their Qualitative

Dynamics. J. Theor. Biol. 54:85-107.
Glass, L. and Young, R. (1979) Structure and Dynamics of Neural Network
Oscillators. Brain Res. 179:207-218.
Kauffman, S.A. (1993) The Origins of Order, Self-Organization and Selection in
Evolution. Oxford University Press.
Michaels G., Carr D.B., Wen X., Fuhrman S., Askenazi M. and Somogyi R.
(1998) Cluster Analysis and Data Visualization of Large-Scale Gene Expression
Data. Proc. Pacific Symposium on Biocomputing 1998. In press.
Sawhill, B.K. (1995) Genetic Function Analysis. Santa Fe Institute Working Paper
95-10-085.
Shannon, C.E. and Weaver, W. (1963) The Mathematical Theory of
Communication. University of Illinois Press.
Somogyi, R. and Fuhrman, S. (1997) Distributivity, a general information theoretic
network measure, or why the whole is more than the sum of its parts. Proc.
International Workshop on Information Processing in Cells and Tissues (IPCAT)
1997. In press.
Somogyi R. and Sniegoski C.A. (1996) Modeling the Complexity of Genetic
Networks: Understanding Multigenic and Pleiotropic Regulation. Complexity
1(6):45-63.
Somogyi, R., Fuhrman, S., Askenazi M., Wuensche A. (1996) The Gene
Expression Matrix: Towards the Extraction of Genetic Network Architectures.
Proc. Second World Congress of Nonlinear Analysts (WCNA96). Elsevier
Science.
Thieffry, D. and Thomas R. (1998) Qualitative Analysis of Gene Networks. Proc.
Pacific Symposium on Biocomputing 1998. In press.
Wuensche, A., Lesser, M.J. (1992) The Global Dynamics of Cellular Automata,
SFI Studies in the Sciences of Complexity, Volume 1. Addison Wesley.