Molecular Dynamics-Powered Hierarchical Geometric Deep Learning Framework for Protein-Ligand Interaction.

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Accurate protein-ligand (PL) binding prediction is crucial for structure-based drug design.
• Existing equivariant graph neural network (EGNN) methods often overlook essential residue-level information in PL complexes.
• Understanding binding mechanisms requires incorporating both atom-level and residue-level structural and energetic data.

Purpose of the Study:

• To develop a novel SO(3)-equivariant hierarchical graph neural network (EHGNN) to capture biomolecular structure hierarchy.
• To propose Dynamics-PLI, a deep learning framework integrating molecular dynamics and energy guidance for enhanced PL interaction prediction.
• To improve the accuracy and interpretability of protein-ligand binding affinity and efficacy predictions.

Main Methods:

• Development of a SO(3)-EHGNN model to process hierarchical biomolecular data.
• Integration of molecular dynamics (MD) trajectories and energy information within a deep learning framework (Dynamics-PLI).
• Evaluation of the model's performance on binding affinity and ligand efficacy prediction tasks using established metrics.

Main Results:

• Dynamics-PLI achieved a 4.03% RMSE decrease in binding affinity prediction.
• The framework demonstrated an average increase of 3.95% in AUROC and AUPRC for ligand efficacy prediction.
• The SO(3)-EHGNN component showed strong performance without requiring pre-training, highlighting its inherent analytical capabilities.

Conclusions:

• The proposed Dynamics-PLI framework significantly outperforms state-of-the-art methods for protein-ligand interaction prediction.
• Incorporating residue-level information and MD data via EHGNNs enhances the understanding of binding mechanisms.
• The SO(3)-EHGNN offers a powerful and robust approach for analyzing complex biomolecular interactions in drug design.