Control of Gene Regulatory Networks Basin of Attractions with Batch Reinforcement Learning

Abstract

Basin of attraction contains biological functions and channels cell behavior, so when a gene network is in an unhealthy basin it may cause diseases. Control techniques can support the design of therapies that promote the transition of a biological system from diseased to healthier basins. Most control methods first infer a gene network and then derive a control strategy to avoid diseased states. However, this approach is limited to few genes and may cause other diseases, as the biological side of the problem is not considered. While changing between basins may change a diseased biological function for a healthier one, state avoidance can change functions in an unexpected way. We propose to extend a batch reinforcement learning method FQI-Sarsa, to change basin of attractions in a partial observable network. Using a batch reinforcement learning technique avoids the most time consuming phases that are the inference and control of the gene network. Results demonstrate that our method, BOAFQI-Sarsa, is more effective than previous studies that do not consider basins in their computations.