Multifidelity multilevel approximate Bayesian computation for stochastic biochemical reaction networks

Speaker: Dr David Warne
Affiliation: Queensland University of Technology

Abstract

Many biochemical processes, such as gene expression, are stochastic in nature. As a result, stochastic models of biochemical kinetics based on continuous-time Markov processes are widely utilised to study cellular regulatory functions. However, experimental data based on microscopy and fluorescent tagging represent incomplete observations of reality. This fact, along with the need to solve the chemical master equation, leads to the intractability of the likelihood function. As a result, it is common to apply likelihood-free approaches that replace likelihood evaluations with realisations of the model and observation process. However, likelihood-free techniques are computationally expensive for accurate inference as they may require millions of high-fidelity, expensive stochastic simulations.

To address this challenge, we develop a new method based on recent advances in the class of methods for estimation of expectations with respect to posterior distributions for parameter inference with partially observed Markov processes models.  Our novel approach combines the multilevel Monte Carlo telescoping summation, applied to a sequence of approximate Bayesian posterior targets, with a multifidelity rejection sampler that learns from low-fidelity, computationally inexpensive, model approximations to minimise the number of high-fidelity, computationally expensive, simulations required for accurate inference. Using several non-trivial examples of biochemical systems, we demonstrate improvements of more than two orders of magnitude over standard rejection sampling techniques.