Transferable Machine Learning Interatomic Potential for Bond Dissociation Energy Prediction of Drug-like Molecules.

Area of Science:

• Computational Chemistry
• Drug Metabolism Studies
• Machine Learning in Chemistry

Background:

• Developing accurate interatomic potentials is crucial for molecular simulations.
• Existing potentials often struggle with open-shell radical species common in drug metabolism.
• Cytochrome P450 (CYP) enzymes catalyze critical drug metabolism reactions involving hydrogen abstraction.

Purpose of the Study:

• To present a transferable MACE interatomic potential applicable to open- and closed-shell drug-like molecules.
• To extend the scope of applications to bond dissociation energy (BDE) prediction, particularly for CYP metabolism.
• To validate the MACE potential's accuracy and transferability against established datasets and methods.

Main Methods:

• Developed a MACE interatomic potential for molecules containing H, C, and O atoms.
• Validated the potential on the COMP6 dataset (closed-shell) and two CYP metabolism datasets (open- and closed-shell).
• Calculated aliphatic C-H BDEs and compared reaction energies for hydrogen abstraction.

Main Results:

• MACE potential demonstrated superior accuracy on the COMP6 dataset compared to ANI-2x.
• Achieved high accuracy on CYP metabolism datasets, with a BDE RMSE of 1.37 kcal/mol on the CYP 3A4 dataset.
• MACE showed better prediction of BDE ranks than AM1, GFN2-xTB, and ALFABET methods.
• Highlighted the potential's smoothness for bond elongation and its ability to find minimum energy reaction paths.

Conclusions:

• The transferable MACE potential accurately models open- and closed-shell molecules, including radical species relevant to drug metabolism.
• It provides a robust tool for calculating BDEs and reaction energies in CYP-mediated reactions.
• This work paves the way for broader applications in reaction modeling and exploring new chemical spaces.