|
Abstract:
|
Parameter and structural estimation in stochastic reaction networks is a notoriously hard problem. These systems arise frequently in the reverse engineering of gene regulatory networks and epidemic models. The main obstacles associated with inference in such systems are largely computational, owing to the fact that the exact likelihood is often intractable, or impossible to specify. Since vast amounts of prior biological information have been produced through years of experimentation, the Bayesian paradigm is often desirable. Thus, elaborate strategies for likelihood evaluation, like for instance particle filtering, or approximate posterior sampling; i.e., approximate Bayesian computation, are required for analysis. We develop a novel Markov chain Monte-Carlo algorithm, related to the idea of a synthetic likelihood, for inference in such systems. The synthetic likelihood is based on statistics, which are projections of the original time series data into Euclidean space. Their form leads to highly efficient posterior sampling and better inferential properties in some systems, as is demonstrated empirically via simulation and on real biological data.
|
