Data-driven parametrization of molecular mechanics force fields for expansive chemical space coverage

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Drug Discovery

Background:

• Force fields are essential for molecular dynamics (MD) simulations in computational drug discovery, balancing accuracy and efficiency.
• Traditional force field parameterization methods struggle with the expanding chemical space of drug-like molecules.
• Developing accurate and broadly applicable force fields remains a key challenge.

Purpose of the Study:

• To develop an Amber-compatible, data-driven force field (ByteFF) for drug-like molecules.
• To address the limitations of traditional parameterization methods in covering diverse chemical spaces.
• To enhance the accuracy and efficiency of molecular mechanics (MM) force fields for drug discovery.

Main Methods:

• Generated a large, diverse molecular dataset (2.4M geometries, 3.2M torsion profiles) using high-level theory (B3LYP-D3(BJ)/DZVP).
• Trained an edge-augmented, symmetry-preserving molecular graph neural network (GNN) on the generated dataset.
• Developed a robust training strategy for predicting MM force field parameters.

Main Results:

• ByteFF accurately predicts bonded and non-bonded MM parameters for drug-like molecules.
• The model demonstrates state-of-the-art performance on benchmark datasets for geometries, torsion profiles, and energies/forces.
• Achieved exceptional accuracy and broad chemical space coverage.

Conclusions:

• ByteFF offers a powerful data-driven solution for force field parameterization in computational drug discovery.
• Its accuracy and wide applicability make it suitable for various stages of the drug discovery pipeline.
• ByteFF represents a significant advancement in creating efficient and accurate molecular mechanics force fields.