Parameter inference for stochastic biological models

Abstract

Parameter inference is the field concerned with estimating reliable model parameters from data. In recent years there has been a trend in the biology community toward single cell technologies such as fluorescent flow cytometry, transcriptomics and mass cytometry: providing a rich array of stochastic time series and temporal distribution data for analysis. Deterministically, there are a wide range of parameter inference and global optimisation techniques available. However, these do not always scale well to non-deterministic (i.e., stochastic) settings — whereby the temporal evolution of the system can be described by a chemical master equation for which the solution is nearly always intractable, and the dynamic behaviour of a system is hard to predict. For systems biology, the inference of stochastic parameters remains a bottleneck for accurate model simulation. This thesis is concerned with the parameter inference problem for stochastic chemical reaction networks. Stochastic chemical reaction networks are most frequently modelled as a continuous time discretestate Markov chain using Gillespie’s stochastic simulation algorithm. Firstly, I present a new parameter inference algorithm, SPICE, that combines Gillespie’s algorithm with the cross-entropy method. The cross-entropy method is a novel approach for global optimisation inspired from the field of rare-event probability estimation. I then present recent advances in utilising the generalised method of moments for inference, and seek to provide these approaches with a direct stochastic simulation based correction. Subsequently, I present a novel use of a recent multi-level tau-leaping approach for simulating population moments efficiently, and use this to provide a simulation based correction to the generalised method of moments. I also propose a new method for moment closures based on the use of Padé approximants. The presented algorithms are evaluated on a number of challenging case studies, including bistable systems — e.g., the Schlögl System and the Genetic Toggle Switch — and real experimental data. Experimental results are presented using each of the given algorithms. We also consider ‘realistic’ data — i.e., datasets missing model species, multiple datasets originating from experiment repetitions, and datasets containing arbitrary units (e.g., fluorescence values). The developed approaches are found to be viable alternatives to existing state-ofthe-art methods, and in certain cases are able to outperform other methods in terms of either speed, or accuracy