Machine Learning-Enhanced Calculation of Quantum-Classical Binding Free Energies.

Area of Science:

• Computational chemistry
• Biophysics
• Drug discovery

Background:

• Accurate prediction of protein-drug binding free energies is crucial for drug discovery.
• Classical simulations struggle with metallodrugs, necessitating quantum chemical methods.
• Hybrid quantum mechanics/molecular mechanics (QM/MM) offers a potential solution but is computationally expensive.

Purpose of the Study:

• To develop an automated QM/MM-based workflow for efficient alchemical free energy simulations.
• To enable accurate binding free energy calculations for metallodrugs.
• To improve the efficiency and applicability of free energy simulations in computational drug design.

Main Methods:

• Hybrid quantum mechanics/molecular mechanics (QM/MM) calculations to sample the potential energy surface.
• Training a machine learning (ML) potential on QM/MM energies and forces.
• Developing an extended element-embracing atom-centered symmetry functions descriptor for QM/MM data.
• Incorporating electrostatic embedding and long-range electrostatics into the ML potential.

Main Results:

• Demonstrated a general and automated workflow for QM/MM-based free energy simulations.
• Successfully applied the workflow to protein-ligand complexes involving metallodrugs (NKP1339) and organic inhibitors (19G).
• The proposed ML descriptor efficiently represents systems with diverse chemical elements.

Conclusions:

• The developed workflow enables efficient and accurate alchemical free energy simulations for metallodrugs.
• This approach enhances the prediction of protein-ligand interactions, particularly for complex systems.
• The method holds significant promise for accelerating drug discovery and development.