Optimal approach to the design of experiments for the automatic characterisation of biosystems

Abstract

The Design-Build-Test-Learn cycle is the main approach of synthetic biology to re-design and create new biological parts and systems, targeting the solution for complex and challenging paramount problems. The applications of the novel designs range from biosensing and bioremediation of water pollutants (e.g. heavy metals) to drug discovery and delivery (e.g. cancer treatment) or biofuel production (e.g. butanol and ethanol), amongst others. Standardisation, predictability and automation are crucial elements for synthetic biology to efficiently attain these objectives. Mathematical modelling is a powerful tool that allows us to understand, predict, and control these systems, as shown in many other disciplines such as particle physics, chemical engineering, epidemiology and economics. Yet, the inherent difficulties of using mathematical models substantially slowed their adoption by the synthetic biology community.

Researchers might develop different competing model alternatives in absence of in-depth knowledge of a system, consequently being left with the burden of with having to find the best one. Models also come with unknown and difficult to measure parameters that need to be inferred from experimental data. Moreover, the varying informative content of different experiments hampers the solution of these model selection and parameter identification problems, adding to the scarcity and noisiness of laborious-to-obtain data. The difficulty to solve these non-linear optimisation problems limited the widespread use of advantageous mathematical models in synthetic biology, broadening the gap between computational and experimental scientists. In this work, I present the solutions to the problems of parameter identification, model selection and experimental design, validating them with in vivo data. First, I use Bayesian inference to estimate model parameters, relaxing the traditional noise assumptions associated with this problem. I also apply information-theoretic approaches to evaluate the amount of information extracted from experiments (entropy gain). Next, I define methodologies to quantify the informative content of tentative experiments planned for model selection (distance between predictions of competing models) and parameter inference (model prediction uncertainty). Then, I use the two methods to define efficient platforms for optimal experimental design and use a synthetic gene circuit (the genetic toggle switch) to substantiate the results, computationally and experimentally. I also expand strategies to optimally design experiments for parameter identification to update parameter information and input designs during the execution of these (on-line optimal experimental design) using microfluidics. Finally, I developed an open-source and easy-to-use Julia package, BOMBs.jl, automating all the above functionalities to facilitate their dissemination and use amongst the synthetic biology community.