Integrating Machine Learning into Free Energy Perturbation Workflows

Free energy perturbation (FEP) methods are among the most accurate tools in structure-based drug design for predicting protein–ligand binding affinities. However, their adoption remains limited due to high computational demands and complex setup procedures. This review explores how integrating machine learning (ML), especially active learning (AL) and deep learning (DL), can enhance the efficiency, accessibility, accuracy, and precision of FEP workflows. It examines three key areas where ML has been successfully applied: sampling strategies, protocol optimization, and force field development. AL algorithms can significantly reduce the number of FEP calculations needed during virtual screening by guiding the molecule selection. DL-based protein–ligand cofolding methods such as AlphaFold, NeuralPLexer, and DragonFold enable the automated generation of accurate complex structures for FEP, bypassing traditional docking and preparation steps. Additionally, ML-derived neural network potentials (NNPs), trained on quantum mechanical data, offer improved force field accuracy, although at the cost of higher computational expenses. This review emphasizes a hybrid approach combining human expertise with ML tools as the most promising strategy for accelerating and democratizing FEP-based drug discovery. Future developments in this interdisciplinary space are expected to expand the scope and impact of computer-aided drug design across pharmaceutical and materials science applications.