Integrating Machine Learning Interatomic Potentials with MMPBSA for Accurate Protein-Ligand Binding Free Energy...

End-point binding free energy (BFE) methods, such as molecular mechanics Poisson-Boltzmann surface area (MMPBSA), are widely used to estimate protein-ligand binding affinity due to their favorable balance between accuracy and computational efficiency. Their reliability, however, is fundamentally constrained by inherent statistical thermodynamic approximations and the limited accuracy of classical potential energy surfaces (PES). To overcome the PES bottleneck, we developed AIQM-PBSA, a novel hybrid framework integrating the ONIOM scheme with the PBSA model. Within this framework, the AIQM3 machine learning interatomic potential (MLIP)─an advanced Δ-learning quantum mechanical (QM) model─is employed to refine the molecular mechanics (MM) energy term, while solvation contributions are evaluated under the PBSA formalism. Extensive validation across diverse protein-ligand systems demonstrates that AIQM-PBSA substantially improves predictive accuracy, achieving Pearson R values of 0.84 and 0.82 on two primary benchmark data sets. On the rigorous Schrödinger JACS set, it yielded Pearson, Spearman, and Kendall correlations of 0.59, 0.58, and 0.42, respectively. By replacing classical force fields with MLIPs to describe gas-phase interaction energies, AIQM-PBSA significantly outperforms traditional MMPBSA and the classic ANI-2x. In summary, AIQM-PBSA offers a robust and generalizable framework that leverages advanced MLIPs to achieve QM-level accuracy while maintaining high computational efficiency, substantially improving the reliability of end-point free energy calculations in biomolecular recognition.