Likelihood-free nested sampling for biochemical reaction networks

Abstract

The development of mechanistic models of biological systems is a central part of Systems Biology. One major challenge in developing these models is the accurate inference of the model parameters. In the past years, nested sampling methods have gained an increasing amount of attention in the Systems Biology community. Some of the rather attractive features of these methods include that they are easily parallelizable and give an estimation of the variance of the final Bayesian evidence estimate from a single run. Still, the applicability of these methods is limited as they require the likelihood to be available and thus cannot be applied to stochastic systems with intractable likelihoods. In this paper, we present a likelihood-free nested sampling formulation that gives an unbiased estimator of the Bayesian evidence as well as samples from the posterior. Unlike most common nested sampling schemes we propose to use the information about the samples from the final prior volume to aid in the approximation of the Bayesian evidence and show how this allows us to formulate a lower bound on the variance of the obtained estimator. We proceed and use this lower bound to formulate a novel termination criterion for nested sampling approaches. We illustrate how our approach is applied to several realistically sized models with simulated data as well as recently published biological data. The presented method provides a viable alternative to other likelihood-free inference schemes such as Sequential Monte Carlo or Approximate Bayesian Computations methods. We also provide an intuitive and performative C++ implementation of our method.