Design and statistical properties of robust functional networks: A model study of biological signal transduction Pablo Kaluza, 1 Mads Ipsen, 2 Martin Vingron, 3 and Alexander S. Mikhailov 1 1 Department of Physical Chemistry, Fritz Haber Institute of the Max Planck Society, Faradayweg 4-6, D-14195 Berlin, Germany 2 Department of Chemistry, H. C. Ørsted Institute, University of Copenhagen, Universitetsparken 5, Copenhagen Ø2100, Denmark 3 Department of Computational Molecular Biology, Max Planck Institute for Molecular Genetics, Ihnestrasse 63-73, D-14195 Berlin, Germany Received 10 March 2006; published 19 January 2007 A simple flow network model of biological signal transduction is investigated. Networks with prescribed signal processing functions, robust against random node or link removals, are designed through an evolutionary optimization process. Statistical properties of large ensembles of such networks, including their characteristic motif distributions, are determined. Our analysis suggests that robustness against link removals plays the principal role in the architecture of real signal transduction networks and developmental genetic transcription networks. DOI: 10.1103/PhysRevE.75.015101 PACS numbers: 89.75.Hc, 89.20.a, 89.75.Fb Cells in a biological organism function in a stable, precise way despite noise and destructive mutations 1. If the prin- ciples of biological robustness were understood, they could be further applied for the design of complex industrial pro- duction and transportation systems, or for understanding so- cial processes 2,3. The cells operate as dynamical networks and their robustness is, to a large extent, determined by a special network architecture. This has been demonstrated for various biological functions, including chemotaxis 4, me- tabolism 5, signal transduction 6, and the cell cycle 7. It was shown that genetic networks with robust expression pat- terns may spontaneously develop through biological evolu- tion 8. The problems of robustness have also been dis- cussed in an abstract context for large random networks, aimed at developing optimal defense strategies for the Inter- net and the WWW 911. The networks of a living cell are not random. They are selected by biological evolution to execute certain functions. In particular, a cell should activate a fixed group of genes in response to each arriving stimulus. These networks should maintain their prescribed, specific functions, although possibly exposed to random damage or parameter variations. The structure of such networks reflects their functions. Is it possible, by adjustment of the network structure, to develop systems with prescribed functions that are, furthermore, robust against damage? How strongly would the requirements of robustness against a particular kind of damage affect their architecture? In this paper, we study a toy flow model of biological signal transduction. The network transports signals, applied to input nodes, through a number of middle redistribution nodes to a set of output nodes. In a cell, the analogy would be to a particular set of genes that are turned on upon arrival of a certain stimulus at the cell surface. This mapping be- tween input stimulusand output gene activityis mediated by a network of interactions among proteins in the cell. These proteins are modeled as nodes in our networks, while interactions between them are reflected in the existence of links. Physically, a signal from the cell surface is passed on through processes like protein phosphorylation or dephos- phorylation, translocation, structural change, etc. In a gross oversimplification of the real processes, we model this signal transduction process by an abstract network flow. Proteins undergo mutations which in turn can affect the links among them or completely delete some nodes and introduce new ones. Thus the network topology is subject to random local changes. By running an optimization process with structural muta- tions and subsequent selection, we show that networks with predefined output patterns can be constructed. Then, we ex- tend the optimization criterion and design networks which, while approximately retaining a fixed output pattern, become robust against removal of randomly chosen links or nodes. Statistical properties of robust functional networks, for an ensemble of different optimization trajectories starting with various initial conditions, are considered and distributions of structural motifs in two kinds of networks, robust against link or node removals, are then determined. A considered network of size N = N in + M + N out consists of N in input nodes, M middle nodes, and N out output nodes. Its architecture is specified by a directed graph of connections between the nodes with the adjacency matrix A ij we have A ij =1, if there is a link from node j to node i, and A ij =0 otherwise. An input node can be connected only with the middle nodes, a middle node can be connected with other middle nodes and with the output nodes see Fig. 1. Each link j i carries some signal flux u ij . The sum of all incom- ing fluxes for any node is equal to the sum of all outgoing fluxes. For any node, all outgoing fluxes are equal in inten- sity and are obtained by splitting the total incoming signal flux in equal parts between the outgoing connections. Thus we have u ik =  l A lk -1 j A kj u kj for any node k. Introducing the total fluxes x i = j A ji u ji passing through nodes i, this re- distribution law can also be written as x i = j A ij x j  k A kj -1 for i =1,2,... N. External fluxes can be applied to the input nodes and sinks are attached to the output nodes. An external unit flux x =1, applied to an input node =1,2,..., N in , becomes distributed after passing through the network and fractions x = Q  of the applied flux reach different output nodes =1,2,..., N out . The matrix Q with the elements Q  represents the output pattern of a given network. Note that Q  = 1. The performance FGof a given network G is its output pattern Q, i.e., FG= Q. The ideal performance of a PHYSICAL REVIEW E 75, 015101R2007 RAPID COMMUNICATIONS 1539-3755/2007/751/0151014©2007 The American Physical Society 015101-1