Learning Regulation from the Ground Up: Combining Natural Selection, Thermodynamics and Data
Dr. William Cannon, Pacific Northwest National Laboratory & University of California, Riverside
Abstract: Modeling cells has many challenges: data is sparse, noisy, and measured over a population instead of over individuals or cell compartments. Moreover, parameters needed to build kinetic and thermodynamic models are extremely labor intensive to obtain. This makes model building a physics-based model a very hard problem. We address this challenge by taking advantage of the fact that natural selection selects for the most optimal individuals out of all solutions. We formulate fitness from a thermodynamic perspective to obtain the most likely model parameters, and then use data to constrain the solution space. Rate parameters that are statistically the most likely can be inferred in this way. Then we predict regulation of the cellular system using one of two approaches: Assuming that we have an optimal control problem and using control theory to infer regulation, or widely sample the solution space for regulation using reinforcement learning. The result is a model with reasonable parameters and predicts regulation for central metabolism that agrees with the literature.