Evaluation of Parallel Tempering to Accelerate Bayesian Parameter Estimation in Systems Biology

TL;DR

This paper evaluates the effectiveness of parallel tempering in accelerating Bayesian parameter estimation for systems biology models, demonstrating improved convergence over traditional methods especially in complex scenarios.

Contribution

The study compares parallel tempering with Metropolis-Hastings for Bayesian parameter estimation and provides a new MATLAB tool integrated with BioNetGen.

Findings

Parallel tempering accelerates convergence for simpler models.

PT often outperforms MH in complex models by escaping local minima.

A new MATLAB package PTempEst is introduced for Bayesian estimation.

Abstract

Models of biological systems often have many unknown parameters that must be determined in order for model behavior to match experimental observations. Commonly-used methods for parameter estimation that return point estimates of the best-fit parameters are insufficient when models are high dimensional and under-constrained. As a result, Bayesian methods, which treat model parameters as random variables and attempt to estimate their probability distributions given data, have become popular in systems biology. Bayesian parameter estimation often relies on Markov Chain Monte Carlo (MCMC) methods to sample model parameter distributions, but the slow convergence of MCMC sampling can be a major bottleneck. One approach to improving performance is parallel tempering (PT), a physics-based method that uses swapping between multiple Markov chains run in parallel at different temperatures to accelerate sampling. The temperature of a Markov chain determines the probability of accepting an unfavorable move, so swapping with higher temperatures chains enables the sampling chain to escape from local minima. In this work we compared the MCMC performance of PT and the commonly-used Metropolis-Hastings (MH) algorithm on six biological models of varying complexity. We found that for simpler models PT accelerated convergence and sampling, and that for more complex models, PT often converged in cases MH became trapped in non-optimal local minima. We also developed a freely-available MATLAB package for Bayesian parameter estimation called PTempEst (http://github.com/RuleWorld/ptempest), which is closely integrated with the popular BioNetGen software for rule-based modeling of biological systems.