PT - JOURNAL ARTICLE AU - Marçal Gabalda-Sagarra AU - Lucas Carey AU - Jordi Garcia-Ojalvo TI - State-dependent information processing in gene regulatory networks - Draft version AID - 10.1101/010124 DP - 2014 Jan 01 TA - bioRxiv PG - 010124 4099 - http://biorxiv.org/content/early/2014/10/08/010124.short 4100 - http://biorxiv.org/content/early/2014/10/08/010124.full AB - Single cells have the potential and the necessity to process the information they receive from their environment. In particular, they commonly need to process temporal information obtained simultaneously from multiple inputs. In addition, response to multiple temporally ordered inputs is evolvable under laboratory conditions, suggesting that genetic networks are constructed to enable organisms to integrate novel information over time. However, the logic used by cellular regulatory networks to perform such complex information processing tasks is not understood. Here we show that gene regulatory networks are consistent with a computation paradigm known as reservoir computing (RC), and that this network structure enables single cells to process temporal information. A core subnetwork of genes (the reservoir) encodes and classifies complex time-varying information in a state-dependent manner. Because the state of the reservoir can then be decoded by a single layer of readout genes, allowing cells to process temporal information and efficiently learn new complex environmental conditions. In support of claim, we analyzed transcription factor networks from a variety of organisms, and found that their topology is compatible with RC. We identified the reservoir cores of the regulatory networks, and tested them using the memory-demanding NARMA prediction task, used as a standard benchmark for RC systems in machine learning. Our results show that the gene regulatory networks perform, and significantly better than other, more constrained topologies reported to work as RC. Interestingly, we find that, in real biological subnetworks, the information processing capacity of of the subnetwork is not strongly dependent on on the number of genes that receive input from the environment. Therefore, reservoir computing is an efficient way to for cells to process information without needing to increase the number of genes or the structure of the network. This is in contrast to other network configurations.