CLIFF: A component-based, machine-learned, intermolecular force field

Area of Science:

• Computational chemistry
• Drug discovery
• Molecular modeling

Background:

• Accurate computation of intermolecular interactions is crucial but challenging in drug discovery.
• High-accuracy ab initio methods are computationally prohibitive for drug-protein systems.
• Classical force fields offer feasibility but require laborious parameterization for new molecules.

Purpose of the Study:

• To introduce the component-based machine-learned intermolecular force field (CLIFF) for automated and accurate molecular interaction calculations.
• To combine physics-based equations with machine learning for efficient parameterization.
• To enable routine application of accurate interaction energy calculations in drug discovery.

Main Methods:

• Developed CLIFF using functional forms for electrostatic, exchange-repulsion, induction/polarization, and London dispersion components based on Symmetry Adapted Perturbation Theory (SAPT).
• Fit molecule-independent parameters using SAPT2+(3)δMP2/aug-cc-pVTZ.
• Obtained molecule-dependent atomic parameters (widths, multipoles, Hirshfeld ratios) via machine learning models for common atoms (C, N, O, H, S, F, Cl, Br).

Main Results:

• CLIFF achieved mean absolute errors (MAEs) of ≤0.70 kcal mol⁻¹ for total and component energies on a diverse dimer set.
• On protein fragment interactions, CLIFF yielded an MAE of 0.27 kcal mol⁻¹, outperforming other methods.
• In drug-protein models, CLIFF accurately ranked ligand binding strengths with <10% error compared to SAPT.

Conclusions:

• CLIFF provides an accurate and computationally efficient approach for calculating intermolecular interactions.
• The automated parameterization of CLIFF addresses a key bottleneck in applying accurate force fields.
• CLIFF shows promise for improving drug discovery pipelines through reliable prediction of binding strengths.