Grappa - a machine learned molecular mechanics force field.

Area of Science:

• Computational chemistry
• Machine learning in science
• Molecular dynamics simulations

Background:

• Accurate and efficient force fields are crucial for simulating large molecular systems over extended timescales.
• Current E(3) equivariant neural networks improve accuracy but are computationally expensive compared to established molecular mechanics (MM) force fields.

Purpose of the Study:

• To develop a machine learning framework, Grappa, for predicting MM parameters from molecular graphs.
• To achieve the accuracy of advanced methods with the computational efficiency of traditional MM force fields.

Main Methods:

• Utilized a graph attentional neural network and a transformer with symmetry-preserving positional encoding.
• Developed Grappa, a machine learning framework to predict MM parameters directly from molecular graphs.
• Integrated Grappa with existing Molecular Dynamics (MD) engines like GROMACS and OpenMM.

Main Results:

• The Grappa force field demonstrated superior accuracy over tabulated and other machine-learned MM force fields at equivalent computational cost.
• Accurately predicted energies and forces for small molecules, peptides, and RNA, matching state-of-the-art MM accuracy.
• Successfully reproduced experimental J-couplings and showed transferability to large biomolecules, including a virus particle, in MD simulations.

Conclusions:

• Grappa offers a computationally efficient and accurate approach for molecular simulations, bridging the gap between traditional MM and advanced neural network force fields.
• The framework's data efficiency and simple input features facilitate extension to novel chemical spaces, exemplified by peptide radicals.
• Grappa paves the way for biomolecular simulations approaching chemical accuracy with the computational cost of established protein force fields.