PT - JOURNAL ARTICLE
AU - Warne, David J.
AU - Baker, Ruth E.
AU - Simpson, Matthew J.
TI - Accelerating computational Bayesian inference for stochastic biochemical reaction network models using multilevel Monte Carlo sampling
AID - 10.1101/064170
DP - 2016 Jan 01
TA - bioRxiv
PG - 064170
4099 - http://biorxiv.org/content/early/2016/07/15/064170.short
4100 - http://biorxiv.org/content/early/2016/07/15/064170.full
AB - Investigating the behavior of stochastic models of biochemical reactionnetworks generally relies upon numerical stochastic simulation methods to generate many realizations of the model. For many practical applications, such numerical simulation can be computationally expensive. The statistical inference of reaction rate parameters based on observed data is, however, a significantly greater computational challenge; often relying upon likelihood-free methods such as approximate Bayesian computation, that requirethe generation of millions of individual stochastic realizations. In this study, we investigate a new approach to computational inference, based on multilevel Monte Carlo sampling: we approximate the posterior cumulative distribution function through a combination of model samples taken over a range of acceptance thresholds. We demonstrate this approach using a variety of discrete-state, continuous-time Markov models of biochemical reactionnetworks. Results show that a computational gain over standard rejection schemes of up to an order of magnitude is achievable without significant loss in estimator accuracy.Author Summary We develop a new method to infer the reaction rate parameters for stochastic models of biochemical reaction networks. Standard computational approaches, based on numerical simulations, are often used to estimate parameters. These computational approaches, however, are extremely expensive, potentially requiring millions of simulations. To alleviate this issue, we apply a different method of sampling allowing us to find an optimal trade-off between performance and accuracy. Our approach is approximately one order of magnitude faster than standard methods, without significant loss in accuracy.