ProtFormer-Site: Ultra-fast and Accurate Prediction of Protein-Protein Interaction Sites with Protein Language Model

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-protein interactions (PPIs) are vital for cellular processes.
• Experimental PPI site identification is resource-intensive.
• Existing computational methods struggle with accuracy-efficiency balance.

Purpose of the Study:

• To develop an accurate and computationally efficient framework for PPI site prediction.
• To integrate large protein language models (ESM2, SaProt) with parameter-efficient fine-tuning (LoRA).
• To introduce a specialized ProtFormer backbone with a recycling mechanism for feature refinement.

Main Methods:

• Developed ProtFormer-Site, a framework combining ESM2 and SaProt with LoRA.
• Implemented a ProtFormer backbone with iterative feature refinement.
• Created sequence-only and structure-enhanced model variants.

Main Results:

• Achieved significant improvements in Matthews correlation coefficient (MCC) (22.4%-61.5%) over state-of-the-art methods.
• Demonstrated exceptional scalability with substantially lower inference times.
• Outperformed existing methods on three benchmark datasets.

Conclusions:

• ProtFormer-Site provides a robust, accurate, and computationally efficient solution for PPI site prediction.
• The framework's efficiency is suitable for large-scale, high-throughput analyses.
• This approach advances the prediction of protein function, disease mechanisms, and drug design.