Data-based optimal control of synthetic biology gene circuits

点击翻译按钮获取中文摘要

Synthetic Biology aims at the engineering of biological systems. Its most prominent application is the rational modification or (re-)design of living organisms, ideally in a way akin to the engineering of man-made devices, for their efficient use in sectors such as energy, biomedicine, drug production and food technology. The availability of control mechanisms that can ensure robust and optimal operation of engineered systems is one of the key factors behind the tremendous advances in engineering fields such as transportation, industrial production and energy. However, in the case of engineered biosystems, their accurate control must typically overcome two important hurdles: uncertainty and noise. Uncertainty arises from a high number of components that interact in a nonlinear (and often unknown) manner, and makes it often extremely hard to build accurate mathematical models of their behaviour. Noise, on the other hand, is ubiquitous in cellular systems since the environmental conditions in which they operate typically vary unpredictably and gene expression is inherently a stochastic process.In this research, we investigate the possibility of automatically learning to optimally control synthetic biology gene networks from input-output data collected from these gene networks, i.e. without using a mathematical model built a priori. In particular, we will develop algorithms that allow computer-based systems to autonomously learn how to vary the inputs of a given system so as to optimise its performance defined in terms of the time evolution of its measured outputs. The control strategies learned by our methods will take into account noise and uncertainties in the data and will be developed to be robust with respect to these. Such data-based strategies are analogous to, for example, the way we drive our cars: without any a priori mathematical model of the car behaviour on the road, we can effectively learn how and when to steer, accelerate and break (inputs) based on our observations of the car's position and velocity on the road (outputs) so as to, for example, minimise our lap time around an unknown track using appropriate input scheduling strategies.The algorithms we will develop will allow users to define the desired behaviour and performance objectives and will compute input-scheduling strategies that allow these objectives to be satisfied. The project will build on methods that I have developed and successfully applied to the optimal control of nonlinear systems in noisy environments, e.g., my work on data-based optimal drug-scheduling for HIV infected patients. The use of such purely data-based optimal control methods is particularly important in synthetic biology applications where the system to be controlled is typically poorly characterised and model uncertainties prevail, yet large amount of high-throughput input-output data are available or can be extracted. To showcase the potential of these computational techniques, we will develop data-based methods to optimally control two landmark synthetic biomodules: the light-inducible genetic toggle switch, and the light-inducible generalised repressilator, both of which are currently under implementation in my host Department.