A hybrid molecular simulation/machine-learning framework for rapid and accurate computation of absolute binding free energies of lead-like molecules

Free Energy Perturbation (FEP) methods are increasingly used to guide in silico potency optimisation of preclinical candidate compounds. FEP is most commonly used in the context of Relative Binding Free Energy (RBFE) calculations that are well suited to hit-to-lead or lead optimisation stages of a drug discovery campaign. However, RBFE methods are limited to the calculation of differences in binding affinity between structurally related molecules. It is highly desirable to develop methodologies that achieve accuracy comparable to RBFE but are applicable to a broader class of drug design problems. Examples of high value problems outside the scope of RBFE include: prediction of binding modes; ranking of diverse chemotypes; prediction of binding selectivity profiles. Such problems can in principle be tackled using Absolute Binding Free Energy (ABFE) calculation methods. However ABFE are currently considered too computationally intensive and unreliable to be widely used.
This project will leverage preliminary results from the Michel lab to substantially increase the efficiency of ABFE calculations. In collaboration with the Cole lab, new simulation protocols that combine GPU-accelerated molecular dynamics simulations with machine learning of forcefields and sampling algorithms will be devised. The protocols will be benchmarked on diverse protein-ligand datasets of interest to AstraZeneca. The overall aim is to make ABFE calculations sufficiently rapid and accurate to enable routine use in industrial R&D. This is an exciting opportunity to develop next-generation computer-aided drug design software and methodologies.