Physica A 342 (2004) 207 213

www.elsevier.com/locate/physa

Random walk networks

de Matem atica Aplicada y Estadstica, Escuela Superior de Ingenieros Aeron auticos, Universidad Polit ecnica de Madrid, Plaza Cardenal Cisneros 3, 28040 Madrid, Spain b Observatorio Astron omico, Universidad de Valencia, Edicio Institutos de Investigaci on, Pol. La Coma, 46980 Paterna, Valencia, Spain Received 7 November 2003; received in revised form 21 January 2004 Available online 18 May 2004

Abstract Random Boolean networks are among the best-known systems used to model genetic networks. They show an ono dynamics and it is easy to obtain analytical results with them. Unfortunately very few genes are strictly ono switched. On the other hand, continuous methods are in principle more suitable to capture the real behavior of the genome, but have di culties when trying to obtain analytical results. In this work, we introduce a new model of random discrete network: random walk networks, where the state of each gene is changed by small discrete variations, being thus a natural bridge between discrete and continuous models. c 2004 Elsevier B.V. All rights reserved.
PACS: 05.40.Fb; 02.50.r; 02.70.Uu Keywords: Random networks; Stochastic processes; Random walks; Monte Carlo methods

1. Introduction Random Boolean networks (RBN) are a classical discrete approach in complex systems to gene networks [1,2]. In RBN, each gene have only two states (on/o ) with a dynamic based in Boolean functions. Nevertheless, very few genes are strictly ono switched, rather they have continuous behavior [3]. In order to handle this

Corresponding author. E-mail addresses: [email protected] (B. Luque), [email protected] (F.J. Ballesteros).

0378-4371/$ - see front matter c 2004 Elsevier B.V. All rights reserved. doi:10.1016/j.physa.2004.04.080

208

B. Luque, F.J. Ballesteros / Physica A 342 (2004) 207 213

continuous behavior, models of many-genes activity based on di erential equations have been developed, although they have evident computational and analytical problems. In Ref. [4], it is proposed a continuous-discrete hybrid set of piece-wise linear di erential equations in order to get round this problem. Unfortunately, it is di cult to obtain analytical results in such continuous or semi-continuous models. In this paper, we introduce random walk networks (RWN). These networks work with small variations in the gene states, being in this way a discrete approach to differential equations. RWN are formed by N automata, each one connected with another K automata in a random manner. Each automaton can be in a number s of possible states ranging from 1 to s and the changes among states occur in discrete steps of {+1; 1}. The dynamics of each automaton is constrained by two re ecting states: 1 and s. These extreme states represent null activity and saturation activity respectively, and they act as barriers for the automaton. A function rule table of K variables, generated in a random manner, is associated with each automaton. This function has as input the state values of the K input automata. In contrast with RBN, the output of this function does not dene the new value of the automaton, but a variation +1 or 1 in its value, and hence its link with di erential equations. These variations modify the state of a given automaton producing (in the chaotic case) a random walk-like behavior of its value, which gives the name to these networks. For example, let us suppose K = 2 and s = 4 and the automaton i with state in time t , xit = 2. Let us also suppose that its two automata input have states xit1 = 1 and xit2 = 3 in time t . We search the input combination (1,3) in the rule table for the automaton i and nd +1 as output. Then the value xit = 2 is updated as xit +1 = 2 + 1 = 3 in time t + 1. In the case xit = 4 (saturation state) the update would be xit +1 = 4. Similarly, if the automaton has a value of xit = 1 (null activity) and the output is 1 the update will be xit +1 = 1. That is, the extreme values act as barriers. As in RBN, the evolution of the system is updated synchronously and we can dene a new parameter p, the bias [5], but here it is the probability of having +1 as the output value in the rule tables (and having 1 with probability (1 p)) when we generate in a random manner all the rule tables at the beginning of the denition of the network. Note that connections and rule tables are generated in a random manner at the beginning of the denition of each network but afterwards are maintained (quenched) in time. RWN present complex dynamical behavior and similar orderdisorder transition as RBN [5], random threshold networks [6] or asymmetric neural networks [7], for example. For a xed number s of states, two well-dened phases appear in RWN, separated by a critical line in the space p K : (1) an ordered phase when the value of the bias p is far away from 0.5 and/or the connectivity K is low, in which the networks freeze in a pattern after a short transient. In this phase, almost all of the automata remain in a completely frozen state, and (2) a disordered phase, in the contrary case for p and K , where all patterns are lost and the automata appear to be in a completely disordered state, switching from one state to another seemingly at random.

B. Luque, F.J. Ballesteros / Physica A 342 (2004) 207 213

209

100
-

80

automaton

60 40 20

20

40

60

80

100

120

140

t
10 9 8 7 6 5 4 3 2 1 0 20 40 60 80

state

autom. 33 autom. 53 autom. 88

100

120

140

t
Fig. 1. Up: dynamics of a RWN with N = 100 automata, K = 2, p = 0:79 and s = 10. The horizontal axis represents time steps and the vertical axis shows the 100 automata. Di erent gray intensities indicate di erent states for each automaton, ranging from 1 (white) to 10 (black). Down: evolution of the automata states (automata values) in time for the same RWN. For simplicity we only show automata number 33 (broken line), 53 (dotted line) and 88 (solid line). The positions of these automata are marked in the vertical axis of the upper gure.

We can observe that RWN reduces to RBN when s = 2. In this sense, RWN are a generalization of RBN. Fig. 1 shows the time evolution of an example RWN with N = 100, K = 2, p = 0:79 and s = 10. 2. Annealed and quenched model Following Ref. [8], we can obtain the boundaries separating order and disorder for the RWN annealed model. For the annealed model, in contrast with quenched RWN, we dene the automata inputs and the functions of each automaton each time step in order to avoid correlations or system memory. This is strictly true for other systems, as RBN, but for annealed RWN the memory dependence is not totally avoided. Here, the state values xit have in uence in the values xit +1 . Nevertheless, the RWN annealed system shows an ordereddisordered transition and we can compute it as a mean eld of the quenched RWN. The behavior of each automaton is constrained by the re ecting states 1 and s. For a given network automata, we dene Pn as the probability of being in state

210

B. Luque, F.J. Ballesteros / Physica A 342 (2004) 207 213

n {1; 2; : : : ; s}, that is, having value n. If we assume absence of correlations, we can write the evolution of the probabilities as P1 (t + 1) = (1 p)(P1 (t ) + P2 (t )) . . . Pi (t + 1) = (1 p)Pi+1 (t ) + pPi1 (t ) . . . Ps (t + 1) = p(Ps1 (t ) + Ps (t ))

(1)

with i = 2; 3; : : : ; s 1. Following the annealed model [8], we need to compute the overlap between two annealed replicas, with initial state conditions randomly chosen. The overlap in time t : a(t ) [0; 1] is dened as the normalized number Na(t ) of elements with common states in two annealed replicas. The time evolution of the overlap in the annealed model is given by a(t + 1) = a(t )[aK (t ) + (1 aK (t ))(p2 + (1 p)2 )] + 2(1 a(t )) aK (t )(pPs1 Ps + (1 p)P1 P2 ) s 2 1 n=1 Pn

+ (1 aK (t )) (p2 Ps1 Ps + (1 p)2 P1 P2

+p(1 p)
n=2

Pn1 Pn+1

(2)

This equation has as xed point, the value a = 1 and a stability analysis leads to the following critical surface separating the ordered and disordered phases for annealed RWN:
2 1 n=1 Pn =1 ; K 2 + 1= (1 p)P 2 1=pPs 1 s

(3)

where now Pn are the stationary values of Pn (t ) calculated from Eq. (1) in the limit when t : Pn = 1 z p 1p
n1

(4)

being z =

1p= (1p)

p 1( 1p )s

the normalization constant.

B. Luque, F.J. Ballesteros / Physica A 342 (2004) 207 213

211

s=2
0.8

s=2

0.6

Annealed N=10.000 K=2 p=0.65

0.8

s=3

0.6

a(t)

a(t)

s=3
0.4

s=5 s=9

s=5 s=9

0.4

0.2

0.2

Quenched N=10.000 K=2 p=0.65

0 0

10

20

30

40

50

Fig. 2. Evolution of the overlap a(t ) as a function of the number of states s. Left: Evolution of the overlap a(t ) for annealed RWN for K = 2, p = 0:65 and di erent values of s. In all of them we have taken as initial conditions a(0) = 1=s. Each point represents the average of 100 di erent networks, with N = 10000 automata each one. Solid lines represent the theoretical evolution using Eqs. (1) and (2). Right: Evolution of the overlap a(t ) for quenched RWN for the same parameter values and initial conditions than left. Again, each point represents the average of 100 di erent networks, with N = 10000 automata each one.

As expected, using s = 2 in the critical surface equation (Eq. (3)) this value leads to the well-known critical curve for RBN [5]: K 2p(1 p) = 1 : 3. Numerical simulations Fig. 2 shows the results in the evolution of the overlap a(t ) between two given congurations of the same system. Solid lines in Fig. 2 left are the theoretical predictions of the evolution for the annealed RWN from Eqs. (1) and (2), meanwhile points (circles, squares, diamonds and triangles) are the simulation results for the annealed RWN. K and p have a xed value and it shows the e ect of the variations in s. As can be seen, the agreement between theory (continuous line) from Eqs. (1) and (2) and simulation is complete. Compare it with Fig. 2 right, corresponding to the quenched RWN. It is clear how the values of the overlap a(t ) in the annealed RWN (left) are always smaller than the values in the quenched RWN (right). Thus, due to the strong correlations of the quenched system, the annealed model underestimates the quenched one and the critical surface obtained for the annealed model is not applicable to the quenched model (unlike RBN). Fig. 3 shows the dependence with p and s of the stationary value of the overlap, a , i.e.: a(t ) when t tends to innity. As can be seen in these gures, 1 a acts as order parameter, having order when 1 a = 0 and disorder otherwise. Fig. 3 left corresponds to the theoretical annealed RWN from Eqs. (1) and (2) when t , and Fig. 3 right to the quenched model. As can be seen, both cases show a clear transition from order to disorder, although the annealed model predicts a critical p bigger than the quenched one. (5)

212
1

B. Luque, F.J. Ballesteros / Physica A 342 (2004) 207 213

K=2
0.8

K=2
0.8

Annealed

Quenched N=10.000

0.6

0.6

1-a*

0.4

1-a*
0.4 0.2 0.2 0 0.5 0.6 0.7 0.8 0.9 1 0 0.5

0.6

0.7

0.8

0.9

Fig. 3. Behavior of the stationary value of the overlap, a , as a function of the bias p. We plot 1 a , which acts as an order parameter, versus p. Left: Theoretical representation of 1 a as a function of p for the annealed RWN. Each curve corresponds from the bottom to the top to s = 3; 4; 5; 6; 7; 8; 9. Right: Numerical simulations of 1 a as a function of p for the quenched system. Each point in the curves has been obtained by averaging 100 simulations. Each curve corresponds from the bottom to the top to s = 3; 4; 5; 6; 7; 8; 9, for quenched RWN with N = 10000 and K = 2.

4. Summary The RBN was one of the rst ways proposed to statistically study genome global properties. Unfortunately, this is a model of allnothing behavior where the state of a given gene is completely determined by the input genes and where smooth variations of state are not allowed, in contrast with the real genome. On the other hand, systems of di erential equations, which in principle should give a better understanding of such continuous behavior, are di cult to solve and it is hard to obtain analytical results. In this paper, we have proposed RWN, a new simple random network model which has great resemblances with real genetic networks. RWN are a natural link between discrete allnothing models (like RBN) and continuous systems of di erential equations, as RWN allow small, almost continuous variations in the behavior of the genome. RWN have a clear orderdisorder transition and allow analytical treatment. In this paper, we have deduced the critical frontier for the annealed model, although due to the memory of the system, the annealed solution does not coincide with the quenched critical frontier. Acknowledgements We would like to thank Alberto Fernandez for his valuable opinions. BL has been supported by CICYT BFM2002-01812. References
[1] S.A. Kau man, The Origins of Order, Oxford University Press, New York, 1993.

B. Luque, F.J. Ballesteros / Physica A 342 (2004) 207 213

213

[2] M. Aldana, S. Coppersmith, L.P. Kadano , in: E. Kaplan, J.E. Marsden, K.R. Sreenivasan (Eds.), Perspectives and Problems in Nonlinear Science, Applied Mathematical Sciences Series, Springer, Berlin, 2003. [3] T.A. Brown, Genomes, Wiley, New York, 1999. [4] K. Kappler, R. Edwards, L. Glass, Signal Processing 83 (2003) 789798. [5] B. Luque, R.V. Sol e, Phys. Rev. E 55 (1997) 257260. [6] T. Rohlf, S. Bornholdt, Physica A 310 (2002) 245259. [7] J.A. de Sales, M.L. Martins, D.A. Stariolo, Phys. Rev. E 55 (1997) 3262. [8] B. Derrida, Y. Pomeau, Europhys. Lett. 1 (1986) 45.