Unified protein-small molecule graph neural networks for binding site prediction

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Predicting small molecule binding sites on proteins is crucial for drug discovery but remains challenging.
• Current graph neural network methods often rely on limited local features and small datasets, hindering the modeling of long-range interactions.
• AlphaFold3 has advanced protein structure prediction, but identifying functional binding pockets is a separate, unresolved issue.

Purpose of the Study:

• To develop YuelPocket, a novel graph neural network (GNN) for accurate prediction of protein binding sites.
• To address limitations of existing GNNs by incorporating long-range interaction modeling and training on large-scale data.
• To provide a robust tool for identifying ligand binding pockets in both experimental and predicted protein structures.

Main Methods:

• Developed YuelPocket, a graph neural network operating in residue-level and coordinate-level prediction modes.
• Trained YuelPocket on the large-scale PLINDER dataset, which includes diverse protein-ligand interaction data.
• Evaluated YuelPocket using Distance to Closest Atom and Center-to-Center metrics against state-of-the-art methods.

Main Results:

• YuelPocket achieved higher success rates in binding site prediction compared to existing state-of-the-art methods.
• Demonstrated high accuracy and robustness on AlphaFold-predicted protein structures, even with deviations from experimental data.
• Successfully identified contact residues and pinpointed pocket centers with high precision.

Conclusions:

• YuelPocket offers a robust and accurate framework for identifying protein binding sites.
• The method enhances functional annotation and facilitates structure-guided drug discovery, particularly with predicted protein structures.
• YuelPocket's performance on large datasets and predicted structures represents a significant advancement in the field.