A physics-aware neural network for protein-ligand interactions with quantum chemical accuracy

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Quantum chemistry (QC) is vital for understanding molecular interactions like those in proteins and ligands.
• However, QC computations are often too computationally expensive for large systems.
• Machine-learned (ML) potentials offer a solution but struggle with long-range interactions.

Purpose of the Study:

• To develop a novel machine learning model for accurate and efficient prediction of intermolecular interactions.
• To address the limitations of existing ML potentials in capturing non-local interactions.
• To enable QC-quality energy predictions for large biomolecular systems.

Main Methods:

• Developed an atomic-pairwise neural network (AP-Net) incorporating physical constraints.
• Utilized a two-component equivariant message passing neural network architecture.
• Trained the model on a dataset of paired ligand and protein fragments, predicting monomer electron densities.

Main Results:

• AP-Net accurately predicts quantum chemistry-quality interaction energies for protein-ligand systems.
• Achieved orders of magnitude reduction in computational cost compared to traditional QC methods.
• Demonstrated potential applications in molecular crystal structure prediction.

Conclusions:

• AP-Net offers a computationally efficient and accurate method for modeling intermolecular interactions.
• The model shows promise for various applications in computational chemistry and drug discovery.
• Further research is needed to address limitations in modeling highly polarizable systems.