Orienteering: Constructing protein fitness landscapes via epistasis-aware machine learning methods

Our main goal is to develop more efficient protein engineering (PE) methods by combining state-of-the-art experimental techniques with novel machine learning (ML) methods that incorporate nonlinear epistatic effects. The domain expert has pioneered high-throughput methods for enzyme engineering, whereas the methodological expert has developed cutting-edge methods in computer science based on Bayesian optimization, a ML algorithm forsolving optimization problems in low-data regimes. This collaboration aims to extend the previous work of both experts by focusing on the challenges of multiparametric optimisation and non-linear epistatic effects that are challenging in PE. As a model system, we will use an NADPH-dependent enzyme called formate dehydrogenase (FDH) that interconverts carbon dioxide (CO2) and formic acid, thus exhibiting potential for recycling biological cofactors in industrial biocatalysis (and in the reverse direction CO2 “capture”). Our computational methods are target agnostic and will be thus transferable to any protein or enzyme to support biotechnnologists optimizing complex systems in which non-linear, epistatic and trade-off effects between multiple properties are wide-spread and difficult to handle.