A machine learning based intramolecular potential for a flexible organic molecule.

Area of Science:

• Computational Chemistry
• Machine Learning in Molecular Modeling
• Drug Discovery

Background:

• Quantum mechanical (QM) simulations are crucial for accurate molecular modeling but are computationally expensive for large systems and long timescales.
• Predictive modeling in chemistry and biology faces limitations due to the computational demands of QM methods.

Purpose of the Study:

• To develop a machine learning (ML) model that accurately reproduces the QM potential energy surface for a drug-like molecule.
• To enable efficient molecular simulations for condensed-phase systems and protein-ligand interactions.

Main Methods:

• Employed kernel regression and the Gaussian Approximation Potential (GAP) framework to create an analytical potential.
• Developed an iterative training protocol and a representation separating short and long-range interactions to handle high-dimensional configurational space.
• Integrated the ML potential with MCPRO for Monte Carlo simulations of a small molecule with proteins (p38 MAP kinase, leukotriene A4 hydrolase) and in water.

Main Results:

• Successfully constructed an ML-based analytical potential that accurately represents the QM potential energy surface.
• Demonstrated the transferability of the ML intramolecular model to condensed-phase simulations.
• Showed that accurate QM potential energy surface representation can refine protein-ligand binding free energies by up to 2 kcal mol-1.

Conclusions:

• Machine learning offers a viable approach to overcome the computational bottlenecks of QM simulations in chemistry and biology.
• The developed ML model provides a computationally efficient and accurate tool for molecular simulations, including protein-ligand binding.
• Accurate ML potentials are essential for reliable predictions of molecular interactions and binding affinities in drug discovery.