Protein-peptide interactions drive peptide therapeutics, precision design, and biomarker discovery, yet most predictors underuse complementary sequence-structure information. LLM-augmented multimodal approaches offer a promising solution to these limitations. We introduce MGAPep, which fuses pre-trained large language model embeddings with protein sequence and structural descriptors via a residual graph attention backbone and a multi-head dual-attention module to capture fine-grained interface patterns. Leveraging large-scale corpora of protein fragment-peptide interaction data, MGAPep employs self-supervised pre-training, transfer learning, and task-specific fine-tuning to obtain rich, transferable representations. Extensive benchmarking shows consistent state-of-the-art accuracy for protein-peptide binding site prediction, with robust generalization to unseen proteins and peptides. The framework also transfers effectively across modalities, yielding superior performance to most baselines on protein-nucleic acid binding site prediction without architecture changes, underscoring broad applicability. Together with evidence that graph-enhanced LLMs improve biomolecular binding modeling, these results establish MGAPep as a general paradigm for protein-biomolecule interaction prediction.
Fu et al. (Thu,) studied this question.