Self-rewiring signaling networks:
Organization as distributed control

In the near term I will focus on the structure, function and evolution of distributed combinatorial control networks. The networks I have in mind are primarily molecular, in particular intra- and intercellular signal transduction networks⁶.

Is there a network paradigm underlying molecular signaling networks? I believe so and decided to focus in particular on one putative aspects of signaling and their consequences: self-rewiring.

Some signaling components are able to redefine with whom another component of the same network is communicating with. This means that the architecture or topology of the network is dynamic and (at least in part) controlled by the network state (the concentration vector) itself. A signaling component (such as a kinase) is here defined in terms of a repertoire of possible downstream targets (within the same network) that it can interact with. Yet, which of these interactions actually do occur at a given time t, is controlled by the concentration of other signaling components. This generates a feedback loop between the network architecture and the dynamics induced by it: a particular network topology determines the concentration changes of network components which, in turn, rewire the network topology. In a system of this kind, a fixed pool of components represents a variety of possible control networks from which a particular one is induced in response to an external signal.

The obvious resemblance of this scheme to neural networks⁷ suggests to investigate the learning capabilities of self-rewiring control networks. Learning is, perhaps, the construction of memory or at least a history-dependent long-term modification of parameters governing a dynamical system. If self-rewiring signaling systems do learn, they should respond in a characteristic fashion to the re-presentation of a signal to which they have been exposed before (memory) or they should trigger long-term architectural changes in response to it. The latter may simply require the addition of a gene expression layer to the model. If genes coding for the components of a network are among the targets controlled by that same network, long-term endogenous architectural changes, such as the removal of nodes (``knock-outs'') or the addition of nodes (``knock-ins''), become possible.

In a first approach, the network dynamics is modeled in terms of mass-action kinetics. This may be appropriate for some portions of a signaling network, but hardly for the whole network. The overall behavior of many control networks depends on single events, such as checkpoints. These are not meaningfully described by traditional kinetics, but emphasize logical aspects of behavior. Building a unified framework for kinetics and logic is a significant challenge that must be met for understanding cellular control networks.

Walter Fontana, Santa Fe Institute