State-dependent information processing in gene regulatory networks

Abstract

Cellular information processing is generally attributed to the complex networks of genes and proteins that regulate cell behavior. It is still unclear, however, what are the main features of those networks that allow a cell to encode and interpret its ever changing environment. Here we address this question by studying the computational capabilities of the transcriptional regulatory networks of five evolutionary distant organisms. We identify in all cases a cyclic recurrent structure, formed by a small core of genes, that is essential for dynamical encoding and information integration. The recent history of the cell is encoded by the transient dynamics of this recurrent reservoir of nodes, while the rest of the network forms a readout layer devoted to decode and interpret the high-dimensional dynamical state of the recurrent core. This separation of roles allows for the integration of temporal information, while facilitating the learning of new environmental conditions and preventing catastrophic interference between those new inputs and the previously stored information. This resembles the reservoir-computing paradigm recently proposed in computational neuroscience and machine learning. Our results reveal that gene regulatory networks act as echo-state networks that perform optimally in standard memory-demanding tasks, and confirms that most of their memory resides in the recurrent reservoir. We also show that the readout layer can learn to decode the information stored in the reservoir via standard evolutionary strategies. Our work thus suggests that recurrent dynamics is a key element for the processing of complex time-dependent information by cells.\n\nSummaryCells must monitor the dynamics of their environment continuously, in order to adapt to present conditions and anticipate future changes. But anticipation requires processing temporal information, which in turn requires memory. Here we propose that cells can perform such dynamical information processing via the reservoir computing paradigm. According to this concept, a structure with recurrent (cyclic) paths, known as the reservoir, stores in its dynamics a record of the cells recent history. A much simpler feedforward structure then reads and decodes that information. We show that the transcriptional gene regulatory networks of five evolutionary distant organisms are organized in this manner, allowing them to store complex time-dependent signals entering the cell in a biologically realistic manner.