Improving mechanistic modelling using machine learning algorithms

 

Professor Jan Hasenauer

University of Bonn

 

Mechanistic modelling is an essential tool for systems biology and has revolutionised the understanding of complex biological processes. Nevertheless, the computational requirements are high and certain processes are still difficult to capture with fully mechanistic descriptions.

In this talk, I will give a brief overview of the various points of contact between mechanistic modelling and machine learning, ranging from semi-automatic model construction to parameter estimation. Following this, I will first present our recent work on tailored mini-batch optimisation techniques for large mechanistic models. This approach allows for a significant speed-up compared to previous methods, thus scaling mechanistic modelling to large datasets. Secondly, I will present a marginalisation approach that enables further speed-up in the presence of semi-quantitative data. This approach can contribute to enabling Bayesian uncertainty quantification for large-scale problems.

Overall, I think that machine learning technique will contribute to lifting mechanistic modelling on the next level. In turn, ideas from mechanistic modelling can help to tailor machine learning methods and facilitate their interpretation.