CSM-Potential: mapping protein interactions and biological ligands in 3D space using geometric deep learning

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in biochemistry

Background:

• Protein structure prediction has advanced significantly, enabling accurate 3D modeling of proteins.
• Understanding protein function requires knowledge of its interactions, which has been computationally challenging.
• Existing computational methods for predicting biological interactions are often limited in scope and accessibility.

Purpose of the Study:

• To develop a novel computational approach for identifying protein surface regions involved in interactions.
• To link protein 3D structure to biological function through interaction prediction.
• To provide an accessible tool for non-experts and bioinformatics pipelines.

Main Methods:

• Developed CSM-Potential, a geometric deep learning method.
• The approach identifies potential protein-protein and protein-ligand interaction sites on protein surfaces.
• Validated performance on independent blind tests.

Main Results:

• CSM-Potential demonstrated robust performance, outperforming existing methods.
• Achieved ROC AUC values up to 0.81 for identifying protein-protein binding sites.
• Reached up to 0.96 accuracy in biological ligand classification.

Conclusions:

• CSM-Potential effectively links 3D protein structure to biological function by predicting interactions.
• The method offers superior performance compared to existing approaches.
• A user-friendly web server and API are available for broad accessibility.