Speaker
Description
Approximate Bayesian Computation (ABC) is a common tool to tackle statistical inference problems for systems where the likelihood function is intractable, a feature common in biological settings due to the inherent complexity of the models under investigation. ABC replaces the likelihood with a comparison of experimental and simulated data, finding parameters which minimise any discrepancy. To improve the efficiency of ABC techniques such as Sequential Monte Carlo (SMC) are typically implemented where the tolerance used in the rejection sampling procedure is gradually reduced over multiple iterations. We have found that designing appropriate tolerance schedules is critical not only for efficiency but also reliability of ABC SMC, with commonly used techniques for schedule selection often leading to inferred parameter values associated with local minima in discrepancy space, rather than the global minima. This problem is particularly acute in stochastic systems. We propose a new procedure to overcome this issue underpinned by reliably choosing the value for the first tolerance to be just below any local minima, with subsequent iterations refining inference around the true parameter values. We have also found that inference of stochastic systems with limit-cycle dynamics is particularly challenging. However, we show identifiability is improved by incorporating reaction-event level stochastic statistics into the discrepancy metric used to compare observed and simulated data.
