Bayesian model selection in synthetic biology: factor levels and observation functions

Abstract

Data-driven inference of the most plausible mechanistic model within a set of candidates is a major hurdle in synthetic and systems biology. Probabilistic model selection is hampered by limitations in the quality and amount of biological data. Furthermore, the computational cost of discriminating between competing models often leads the user to skip model selection and subjectively choose a model.To challenge this practice, here we took a genetic toggle switch built in E. coli, considered three alternative models of it and used a Bayesian approach to rank these models based on the evidence from in vivo data. As the ranking depends on the information content of the data, we use Bayesian optimisation to design maximally informative inputs, i.e. chemical stimuli for the cells. We then explore how the optimality of such stimuli depended on the degrees of freedom in the optimisation (i.e. the number of segments in the input), showing a decrease of the attainable discriminatory power with the dynamic properties of the perturbation. We finally investigated the effect that the observable(s) selected in the optimisation exerts on the outcome of the latter. Our results suggest that Bayesian optimisation-based experimental design can be adopted as a means to discriminate between competing models of a gene regulatory network.