ANI Neural Networks Meet Electrostatics: A ML/MM Implementation in Amber

We present a novel integration of the ANI neural networks into the Amber software suite, offering a sophisticated machine learning/molecular mechanics (ML/MM) framework. The implementation is designed as a general-purpose tool for the simulation of neutral organic molecules, requiring no additional training for its use beyond the initial setup. The framework leverages a new ANI potential that accurately predicts geometry-dependent atomic partial charges at the Minimal Basis Iterative Stockholder (MBIS) level, enhancing the modeling of electrostatic interactions within ML/MM systems. Additionally, we incorporate a polarization correction to address the distortion effects on the ML subsystem from MM point charges. Our approach is validated through simulations of solvation profiles, vibrational spectra, and torsion free energy profiles of small molecules in aqueous environments, as well as protein-ligand interactions. Our findings demonstrate that this ML/MM framework can approximate QM/MM electrostatic embedding with significantly reduced computational demands, paving the way for more efficient and accurate simulations in computational chemistry.