S-PLM: Structure-Aware Protein Language Model via Contrastive Learning Between Sequence and Structure.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Proteins are crucial for biological and engineering applications.
• Protein Language Models (PLMs) accelerate protein function determination and design.
• Current PLMs often lack 3D structural information, limiting their predictive power.

Purpose of the Study:

• To develop a 3D structure-aware PLM (S-PLM) that integrates protein sequence and structure information.
• To enhance the prediction and design capabilities of PLMs for various biological tasks.
• To provide adaptable tools for diverse downstream protein prediction applications.

Main Methods:

• S-PLM employs multi-view contrastive learning to align protein sequence and 3D structure representations.
• Swin-Transformer is used to embed structural information from AlphaFold-predicted protein structures.
• Structural embeddings are fused with sequence embeddings from ESM2.

Main Results:

• S-PLM outperforms sequence-only PLMs on protein clustering and classification tasks.
• S-PLM achieves performance comparable to state-of-the-art methods using both sequence and structure.
• Lightweight tuning tools are provided for adapting S-PLM to specific prediction tasks.

Conclusions:

• S-PLM effectively integrates 3D structural information into protein language models.
• The model demonstrates superior performance in protein function prediction and classification.
• S-PLM offers a valuable tool for advancing protein research and engineering.