Identifying Internal Control in Biological Systems Through Machine Learning

Homeostasis, the biological phenomenon of maintaining a stable internal environment in the face of externally changing conditions, is at the focus of research in the intersecting fields of Control Theory, Biology, and more recently, Machine Learning. Failure of homeostatic control is associated with a variety of diseases and it is therefore vital to identify and characterise the homeostatic mechanisms. Related data-driven methodologies that given observations of a system infer the control mechanisms include Inverse Optimal Control, Inverse Reinforcement Learning and Symbolic Regression. However, lacking the plant-controller separation that characterizes human engineered systems, these methodologies are not suitable for analyzing biological systems and different dedicated algorithmic tools are required. In this paper we elaborate on the inability of the engineered inspired identification tools to correctly infer biological control mechanisms and on a dedicated algorithm, Identifying Regulation with Adversarial Surrogates, a datadriven Machine Learning algorithm. This algorithm automates classical hypothesis checking via surrogate data methods and finds the control objective as the solution of a min-max style optimization problem. Finally we conduct a case study of identifying the control mechanism in a realistic biologically inspired model of cell-division, demonstrating the success of the datadriven algorithm.