SOPE-MsL: Synergy-Optimized Protein Language Model Embeddings with Multiscale Learning for Interpretable...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Drug Discovery

Background:

• Protein-small-molecule interactions are crucial for cellular functions and drug development.
• Identifying binding residues is key for understanding molecular recognition and rational drug design.
• Current experimental methods are costly and limited, while computational approaches lack comprehensive residue representation.

Purpose of the Study:

• To develop an accurate and interpretable computational method for predicting protein binding sites.
• To leverage synergistic effects of multiple Protein Language Model (PLM) embeddings.
• To improve upon existing methods for structure-aware sequence analysis.

Main Methods:

• SOPE-MsL integrates PLM embedding fusion with multiscale learning for binding site prediction.
• Systematic evaluation identified ProstT5 and Ankh embeddings as an effective pair.
• A network combining multiscale convolutional operations and attention mechanisms was employed, along with weighted focal loss to handle class imbalance.

Main Results:

• SOPE-MsL achieves competitive performance in binding site prediction across multiple datasets.
• Synergistic embedding fusion demonstrated advantages over single-model representations via t-SNE and SHAP analyses.
• The method provides residue-level interpretability.

Conclusions:

• SOPE-MsL offers a robust and interpretable tool for identifying protein-small-molecule interaction sites.
• The approach advances structure-aware sequence analysis for drug discovery.
• PLM embedding fusion with multiscale learning is a promising strategy for binding site prediction.