Leveraging Alchemical Free Energy Calculations with Accurate Protein Structure Prediction

Small-molecule lead optimisation in early-stage drug discovery is broadly supported by computational chemistry approaches throughout industry. Over the last decade, Free Energy Perturbation (FEP) has grown into a mature physics-based tool that prospectively guides medicinal chemistry decision-making by accurately predicting ligand potencies at the level of precision that is required for the granular nature of the lead optimisation stage. Machine-learned ligand-protein co-folding models are at the forefront of accurate protein structure prediction and well-positioned to support early-stage drug discovery campaigns. This study investigates a hybrid framework that combines machine-learned ligand-protein co-folding models with FEP. By leveraging accurate pose and protein prediction, the method bypasses traditional, error-prone and time-consuming docking approaches, improving the reliability and scalability of FEP calculations. Benchmarking studies on a public kinase target (PFKFB3) and an internal target (target A) demonstrate that the hybrid framework achieves state-of-the-art accuracy while substantially lowering computational expense compared to more traditional FEP approaches. This approach integrates the machine learning and physical approach to affinity prediction, and represents a significant advancement in computational lead optimisation support, poising it to aid in accelerating the discovery of novel therapeutics.