Integrating Machine Learning into Free Energy Perturbation Workflows.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in science

Background:

• Free energy perturbation (FEP) is a highly accurate method for predicting protein-ligand binding affinities in structure-based drug design.
• However, FEP's widespread adoption is hindered by significant computational costs and complex implementation requirements.

Purpose of the Study:

• This review explores the integration of machine learning (ML), including active learning (AL) and deep learning (DL), to enhance FEP workflows.
• The goal is to improve the efficiency, accessibility, accuracy, and precision of FEP applications in drug discovery.

Main Methods:

• The review examines ML applications in three key FEP areas: sampling strategies, protocol optimization, and force field development.
• Active learning (AL) guides molecule selection to reduce FEP calculations in virtual screening.
• Deep learning (DL) models like AlphaFold automate accurate protein-ligand complex structure generation for FEP.

Main Results:

• ML integration significantly reduces the computational burden and complexity associated with FEP calculations.
• Deep learning (DL) methods streamline the generation of accurate protein-ligand complex structures, bypassing traditional docking.
• ML-derived neural network potentials (NNPs) offer enhanced force field accuracy, albeit with increased computational demands.

Conclusions:

• A hybrid approach combining human expertise with ML tools is the most effective strategy for accelerating FEP-based drug discovery.
• Future interdisciplinary developments in ML and FEP will broaden the impact of computer-aided drug design in pharmaceuticals and materials science.