Optimal Prediction by Cellular Signaling Networks

TL;DR

This paper investigates how cellular signaling networks can be optimized to predict environmental signals, revealing the roles of network structure, noise suppression, and motifs like feedback in enhancing predictive accuracy.

Contribution

It introduces a theoretical framework for understanding optimal prediction in cellular networks, highlighting the importance of response kernels and network motifs for non-Markovian signal prediction.

Findings

Single-layer networks generate exponential kernels for Markovian signals.

Multilayer networks produce oscillatory kernels for non-Markovian signals.

E. coli chemotaxis can reliably predict concentration changes, balancing response speed and noise suppression.

Abstract

Living cells can enhance their fitness by anticipating environmental change. We study how accurately linear signaling networks in cells can predict future signals. We find that maximal predictive power results from a combination of input-noise suppression, linear extrapolation, and selective readout of correlated past signal values. Single-layer networks generate exponential response kernels, which suffice to predict Markovian signals optimally. Multilayer networks allow oscillatory kernels that can optimally predict non-Markovian signals. At low noise, these kernels exploit the signal derivative for extrapolation, while at high noise, they capitalize on signal values in the past that are strongly correlated with the future signal. We show how the common motifs of negative feedback and incoherent feed-forward can implement these optimal response functions. Simulations reveal that E. coli can reliably predict concentration changes for chemotaxis, and that the integration time of its response kernel arises from a trade-off between rapid response and noise suppression.