LamNet: an alchemical-path-aware graph neural network to accelerate binding free energy calculations for drug...

Accurate prediction of protein-ligand binding free energies is critical yet computationally demanding in drug discovery. Alchemical free energy methods (AFEMs) offer high accuracy but suffer from significant computational costs and complex modeling setup, such as tuning the λ-schedule of alchemical transformation. While conventional deep learning (DL) models may instantly predict binding affinity, they often require a large training set and exhibit limited generalizability across chemical space. To address these challenges, we introduce LamNet, an alchemical-path-aware graph neural network. LamNet integrates endpoint molecular states and the bridging alchemical path (parametrized by λ) into a physics-informed representation learning framework, explicitly modeling free energy changes along a chosen thermodynamic transformation pathway. Trained on molecular-dynamics-simulated data along alchemical pathways and incorporating data reliability metrics, LamNet accurately predicts relative binding free energies and absolute binding free energies, and optimizes λ-schedules to improve traditional AFEM convergence. Evaluations on diverse datasets (463 ligands, 16 proteins) demonstrate that LamNet achieves superior or comparable performance to state-of-the-art methods, including traditional AFEM, but with up to 1000-fold acceleration. These findings establish LamNet as a generalizable, physics-grounded, and cost-effective tool that not only accelerates computations but also provides a novel framework for integrating rigorous computational physics into modern DL-driven drug discovery workflows.