Automatic Bayesian Inference of Reaction Networks via Guiding

Abstract

Jump process models based on chemical reaction networks are ubiquitous, especially in systems biology modeling. However, performing inference on the latent variables and parameters of such models is challenging, particularly when the observations of the system state are noisy and incomplete. This thesis presents CatalystFitting, a system for inferring the latent variables and parameters of stochastic reaction network models given observational data. CatalystFitting provides primitives for performing changes of measure on jump processes. Building on top of these primitives, CatalystFitting further provides a library of strategies for guiding a jump process to match an observation set. These strategies exploit the form of the underlying symbolic reaction network to automatically produce guides optimized to the particular reaction network structure of interest to the modeler, accelerating otherwise costly Bayesian inference procedures. We present inference results on a bistable switch system and a repressilator system.