PSMa: Learning Protein Surface Representations with Physicochemical Masked Autoencoders.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Traditional protein modeling relies heavily on sequence data, neglecting crucial geometric and physicochemical information from molecular surfaces.
• This underutilization limits the depth and accuracy of current protein representation learning methods.

Purpose of the Study:

• To introduce PSMa (Protein Surface Masked autoencoder), a novel self-supervised framework for learning protein representations directly from molecular surface data.
• To evaluate PSMa's effectiveness in downstream tasks like protein-protein interface binding-site identification and thermal stability prediction.

Main Methods:

• Developed PSMa, a self-supervised framework operating on point clouds of molecular surfaces with geometric and physicochemical attributes.
• Trained PSMa on approximately 220,000 Protein Data Bank (PDB) structures by reconstructing masked surface patches.
• Assessed performance on held-out proteins for protein-protein interface binding-site identification and thermal stability prediction.

Main Results:

• PSMa significantly outperformed sequence-only and structure-only baselines on both downstream tasks.
• Achieved state-of-the-art performance in AUROC and AUPRC for binding-site identification and stability prediction.
• Demonstrated improved threshold-dependent classification performance compared to existing methods.

Conclusions:

• Surface-centric pretraining using PSMa provides a powerful and generalizable foundation for structure-aware protein machine learning.
• The learned representations are transferable across different protein lengths and structure types (experimental and predicted).
• PSMa highlights the importance of incorporating molecular surface geometry and physicochemical properties for advanced protein modeling.