MGAPep: LLM-Augmented Multimodal Graph Attention for Protein-Peptide Binding Site Prediction and Cross-Domain...

Area of Science:

• Computational Biology
• Bioinformatics
• Artificial Intelligence in Biology

Background:

• Protein-peptide interactions are crucial for peptide therapeutics, precision design, and biomarker discovery.
• Existing prediction tools often fail to fully utilize complementary sequence and structural information.
• Large language models (LLMs) offer a promising avenue for improving multimodal approaches.

Purpose of the Study:

• To introduce MGAPep, a novel framework for predicting protein-biomolecule interactions.
• To leverage LLM embeddings and structural descriptors for enhanced prediction accuracy.
• To establish a generalizable paradigm for modeling biomolecular binding.

Main Methods:

• MGAPep fuses pre-trained LLM embeddings with protein sequence and structural descriptors.
• A residual graph attention backbone and multi-head dual-attention module capture fine-grained interface patterns.
• Self-supervised pre-training, transfer learning, and fine-tuning on large-scale interaction data are employed.

Main Results:

• MGAPep achieves state-of-the-art accuracy in protein-peptide binding site prediction.
• The model demonstrates robust generalization to unseen proteins and peptides.
• The framework shows effective cross-modal transferability, outperforming baselines in protein-nucleic acid binding site prediction.

Conclusions:

• MGAPep represents a significant advancement in predicting protein-biomolecule interactions.
• Graph-enhanced LLMs improve biomolecular binding modeling, establishing MGAPep as a general paradigm.
• The approach has broad applicability across various biomolecular interaction prediction tasks.