Decoding the protein-ligand interactions using parallel graph neural networks.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Structural bioinformatics

Background:

• Protein-ligand interactions (PLIs) are vital for biological functions and drug development.
• Experimental methods for PLI characterization are accurate but time-consuming and labor-intensive.
• Existing computational methods often rely on limited 2D protein sequence data.

Purpose of the Study:

• To develop a novel parallel graph neural network (GNN) for accurate protein-ligand interaction prediction.
• To integrate 3D structural data and expert knowledge into deep learning models for PLI analysis.
• To create an interpretable AI tool for predicting binding activity, affinity, and compound potency.

Main Methods:

• Development of two distinct GNN architectures, one domain-aware and one knowledge-free.
• Utilizing 3D structural data for enhanced feature representation in GNNs.
• Adapting GNN models for both binary classification of interactions and regression of binding affinities.

Main Results:

• Achieved high test accuracy for predicting protein-ligand complex activity (0.979 for GNN1, 0.958 for GNN2).
• Demonstrated strong performance in predicting experimental binding affinities (Pearson correlation coefficient of 0.66 and 0.65).
• Outperformed 2D sequence-based models in predicting binding affinity and compound potency (pIC50).

Conclusions:

• The developed GNN models accurately predict protein-ligand interactions and biophysical properties using 3D structural information.
• The AI tool offers an interpretable approach for rational therapeutic design and lead optimization.
• Demonstrated utility in screening compounds against SARS-CoV-2 targets, validating its real-world applicability.