Beyond the cutoff: Hybrid ML/MM electrostatics for neural network potentials

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Biophysics

Background:

• Atomistic Neural Network Potentials (NNPs) offer scalable, low-cost prediction of molecular properties but struggle with long-range interactions.
• Existing NNPs are local, limiting their reliability for condensed-phase systems crucial in biology, engineering, and pharmaceuticals.
• Previous work integrated long-range electrostatic information into NNPs (ANI/MM), drawing parallels to quantum mechanics/molecular mechanics (QM/MM).

Purpose of the Study:

• To develop and train an ANI/MM neural network for predicting binding energies in protein-ligand complexes.
• To evaluate the accuracy of the ANI/MM network in predicting forces and its potential for molecular dynamics simulations.
• To compare the performance of the developed ANI/MM NNP against established classical force fields.

Main Methods:

• Retraining the ANI NNP to incorporate electrostatic potentials from the molecular environment, creating an embedded ANI/MM model.
• Training the ANI/MM network specifically on two protein-ligand complexes to predict binding energies.
• Assessing force prediction accuracy and comparing performance against the ab initio-fitted classical force field Q-Force.

Main Results:

• The ANI/MM network accurately predicts forces with errors under 1 kcal/mol/Å, enabling geometry optimizations and molecular dynamics.
• The developed ANI/MM NNP surpasses the performance of the Q-Force field for the studied complexes.
• The model demonstrates good transferability to new solutes when training data includes relevant structural fragments.

Conclusions:

• Hybrid ML/MM neural architectures provide a promising pathway for chemically accurate and scalable modeling of intricate molecular systems.
• The ANI/MM NNP offers a significant advancement for simulating biological, engineering, and pharmaceutical applications.
• This work highlights the potential of integrating long-range interaction information into neural network potentials for enhanced molecular modeling.