FEP Protocol Builder: Optimization of Free Energy Perturbation Protocols Using Active Learning

• DOI: 10.1021/acs.jcim.3c00681
• Publication Date: 2023-08-18T14:46:59Z
• Authors: César de Oliveira; Karl Leswing; Shulu Feng; René Kanters; Robert Abel; Sathesh Bhat
• Keywords: Overfitting; Time line

Significant improvements have been made in the past decade to methods that rapidly and accurately predict binding affinity through free energy perturbation (FEP) calculations. This has driven by recent advances small-molecule force fields sampling algorithms combined with availability of low-cost parallel computing. Predictive accuracies ∼1 kcal mol-1 regularly achieved, which are sufficient drive potency optimization modern drug discovery campaigns. Despite robustness these FEP approaches across multiple target classes, there invariably systems do not display expected performance default settings. Traditionally, required labor-intensive manual protocol development arrive at parameter settings produce a predictive model. Due (a) relatively large space be explored, (b) significant compute requirements, (c) limited understanding how combinations parameters can affect performance, take weeks months complete, often does involve rigorous train-test set splits, resulting potential overfitting. These timelines coincide tight project timelines, essentially preventing use calculations for systems. Here, we describe an automated workflow termed Protocol Builder (FEP-PB) generate accurate protocols perform well FEP-PB uses active-learning iteratively search develop protocols. To validate this approach, applied it pharmaceutically relevant where could models. We demonstrate previously challenging MCL1 system human intervention. also apply real-world setting p97 system. is able more than expert user, validating as amenable Additionally, workflow, gain insight into most important given results suggest robust tool aid developing increasing number targets technology.