This online talk will be given by Martin Packer from AstraZeneca.

Registration is free, but required. Use the link here to register.

Title: FEP, REST-MD and machine learning to drive molecular design

Abstract:

Free energy perturbation (FEP) models represent the state of the art for prediction of protein-ligand binding affinity. Readily available GPU hardware enables us to generate FEP results for thousands of virtual ligands during a drug design project, guiding selection of candidates for synthesis, but design teams will often generate very large virtual libraries which we would like to assess using FEP, without the associated expense of simulations. Active learning FEP uses machine learning algorithms to create FEP-based structure activity (SAR) models. We cycle between FEP and machine learning until we judge that models are predictive enough to spare further detailed FEP computation.

We have applied this approach to multiple drug design projects within AstraZeneca. We have used three distinct approaches to design large virtual libraries and seen positive impact across a diverse range of protein targets. We now generate more than 600,000 FEP-level data points per annum, helping to prioritise design ideas for synthesis. We also use the models to generate SAR maps for compound series, exemplified here using a published set for the kinase EphB4. In cases where further insight is required to explain specific FEP predictions, it is possible to use ligand-based REST-MD methods to explore the conformational dynamics of a ligand, providing further justification for a decision to proceed to synthesis, or to opt for an alternative virtual compound.

The potential high accuracy of a physics-based FEP/REST-MD approach, combined with machine learning, provides a drug design environment in which every virtual molecule can be assessed on its predicted affinity, enabling a focus on molecules most likely to meet multiple endpoints required for successful drug design.