NPIF Parametrising gene expression models using data-efficient machine learning techniques

Biology is being transformed by the use of quantitative methods to analyse molecular processes as the system level and understand the underlying dynamics of the responses of living organisms to changes in their environment. For example, exposure of bacteria to antibiotics leads to modification of their gene expression profile at multiple time scales.

To understand these complex phenomena, mathematical models are essential but estimating parameters of these models from real experimental data is still very challenging because if the inherent noisiness of biological data.

Indeed, parameter estimation from noisy gene expression data is crucial to building predictive computational models of intracellular kinetics. The most commonly used approaches in the literature use the linear-noise approximation or moment-closure within a Bayesian framework [1,2]. These methods have been shown to give accurate estimates for simple gene regulatory models with large amounts of data however they are computationally inefficient because of the Markov Chain Monte Carlo (MCMC) optimisation step. It is also the case that large amounts of time-series data are not typically available from many experimental setups. In this project the aim is to design a new inference method based on recent advances in the estimation of parameters using data-efficient machine-learning techniques [3] which hold promise for accurate estimation from sparse data. The project will involve the extension of these techniques from partial differential equation models to master equation models and its application to infer the parameters of molecular processes at the heart of DNA and RNA dynamics. The methods will be applied to data produced in the ElKaroui lab to analyse the bacterial response to exposure of DNA damaging antibiotics at the single cell level using microfluidics and fluorescence microscopy.