Computational prediction of ω-transaminase selectivity by deep learning analysis of molecular dynamics trajectories

Abstract

We previously presented a computational protocol to predict the enzymatic (enantio)selectivity of an ω-transaminase towards a set of ligands (Ramírez-Palacios et al. (2021) Journal of Chemical Information and Modeling 61(11), 5569–5580) by counting the number of binding poses present in molecular dynamics (MD) simulations that met a defined set of geometric criteria. The geometric criteria consisted of a hand-crafted set of distances, angles and dihedrals deemed to be important for the enzymatic reaction to take place. In this work, the MD trajectories are reanalysed using a deep-learning approach to predict the enantiopreference of the enzyme without the need for hand-crafted criteria. We show that a convolutional neural network is capable of classifying the trajectories as belonging to the ‘reactive’ or ‘non-reactive’ enantiomer (binary classification) with a good accuracy (>0.90). The new method reduces the computational cost of the methodology, because it does not necessitate the sampling approach from the previous work. We also show that analysing how neural networks reach specific decisions can aid hand-crafted approaches (e.g. definition of near-attack conformations, or binding poses).

Keywords

• Deep learning
• enantioselectivity predictions
• molecular dynamics trajectories
• ω-transaminases