Unscented approaches to inference and design for systems and synthetic biology

Daniel Silk, Michael Stumpf, Biotechnology And Biological Sciences Research Council (Great Britain)
Progress in systems and synthetic biology is driven by mathematical and experimental collaboration. Mathematical models can be used for hypothesis generation, design and knowledge summary, while experiments allow predictions to be tested and provide data for model creation and refinement. While there have been many high profile successes, modelling biological systems remains very challenging. In particular, the high dimensionality and complexity of biological systems along with significant
more » ... imental noise, means that tasks such as parameter fitting and model selection often require far more data than is practical. Here we approach these issues on three levels - data choice, model selection and parameter estimation. Firstly, we show how qualitative information can be utilized independently or in addition to experimental data to both reverse engineer and design biological systems exhibiting desired dynamical behaviours, such as oscillations, chaos and hyperchaos. Next we provide a framework for the rational design of experiments, with the purpose of predicting and maximizing the model discriminatory power of the generated data. Within the parameter inference context we are concerned with sequential approximate Bayesian computation (ABC) algorithms - an increasingly important class of likelihood-free inference techniques. Here we propose a method for choosing the ABC threshold parameter and avoiding local minima, based upon efficient prediction of the threshold-acceptance rate curve. We demonstrate the approaches within different contexts, including the design of synthetic oscillators, epidemiological models, the Hes1 innate immune signalling system, various models of the JAK-STAT signalling pathway and the identification of crosstalk connections between signalling pathways. By exploiting the efficiency of the unscented transform for propagating probability distributions through non-linear functions, the tools developed here allow us to pose and investigate some fundamental questions about the modelling of such s [...]
doi:10.25560/18393 fatcat:suorgof7mbgypef2ia32hcdcom