Simulating protein-ligand binding with neural network potentials.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning applications

Background:

• Drug molecules exhibit conformational flexibility impacting protein-ligand binding.
• Accurate prediction of drug binding affinities is crucial but challenging with existing computational models.
• Neural network potentials (NNPs) offer high accuracy and computational efficiency for molecular conformation stability.

Purpose of the Study:

• To investigate the utility of NNPs in molecular dynamics (MD) simulations for protein-bound drugs.
• To assess the ability of NNP-driven MD to predict drug binding poses and conformational energies.
• To compare NNP performance against traditional molecular mechanical models.

Main Methods:

• Integration of NNPs into MD simulations to model intramolecular forces of protein-bound drugs.
• Application of NNP-based MD to predict binding poses and Gibbs energy of binding for drug molecules.
• Comparative analysis with molecular mechanical models, specifically for erlotinib binding.

Main Results:

• NNP-driven MD simulations successfully predicted protein-ligand binding poses and conformational energy components.
• NNPs provided a more accurate conformational energy for erlotinib binding compared to overestimation by molecular mechanics.
• Reasonable binding poses were predicted for charged ligands, though not ideal for charged species in solution.

Conclusions:

• NNPs represent a powerful tool for modeling intramolecular forces in protein-bound drugs within MD simulations.
• This approach enhances the accuracy of predicting drug binding poses and conformational energies.
• NNPs show promise for computational drug discovery, offering significant advantages over conventional methods, with limitations for charged molecules in solution.