Accurate Free Energy Calculation via Multiscale Simulations Driven by Hybrid Machine Learning and Molecular Mechanics Potentials

Our study focused on the implementation and testing of machine learning interatomic potentials (MLIPs) into the AMBER software suite. This implementation enables us to perform a novel type of molecular dynamics simulation utilizing the hybrid machine learning/molecular mechanics (ML/MM) potentials. To underpin the capabilities of ML/MM simulations, we first validated our implementation at a fundamental physical level by confirming energy and momentum conservation laws. The successful validation indicates that our implementation is able to produce adequate and physically interpretable samplings. Building upon this, for the first time to the best of our knowledge, we proposed an ML/MM-compatible thermodynamic integration (TI) protocol to tackle real-world challenges, such as solvation free energy calculation. Our results demonstrate that this computational protocol can predict hydration free energies with an accuracy of less than 1.00 kcal/mol compared to experimental data, paving the way for the use of ML/MM in multiscale simulations to addressing future drug design problems. Moreover, by applying ML/MM in molecular dynamics simulations of protein-ligand complexes, we demonstrated that the adequate samplings enable us to accurately reproduce experimental binding free energies. Thus, our implementation can offer new insights into biomolecular systems using the ML/MM "microscope". Last, we demonstrated that our implementation can achieve nanosecond timescale simulations daily after significant effort being put to improve the code performance. In a conclusion, we have successfully implemented ML/MM potential to AMBER software package after overcoming limitations in current multi-scale simulations including low computational efficiency. We have advanced TI theory allowing us to accurately predict free energies with ML/MM potentials.