BioText-SeqFormer: Integrating biomedical semantic representations and protein sequence embeddings for TCR-peptide interaction prediction.

The interaction between T cell receptors (TCRs) and antigenic peptides is central to adaptive immune recognition and is important for cancer immunotherapy, infectious disease research, and vaccine design. Although experimental screening methods are reliable, they are often costly and time-consuming, making it difficult to explore the vast combinatorial space of TCR-peptide pairs. Machine learning and deep learning methods provide efficient alternatives for TCR-peptide binding prediction, but most existing models mainly rely on amino acid sequence representations and make limited use of biomedical semantic information related to immune recognition. This limitation may restrict their ability to characterize complex binding patterns. To address this issue, we propose BioText-SeqFormer, a TCR-peptide interaction prediction framework that integrates task-adapted biomedical language semantic representations with protein sequence embeddings. The model constructs task-related biomedical textual representations and combines them with sequence embeddings extracted by a protein language model. Cross-modal Transformer encoding and the Complementary Fusion-Gating Aggregation (CFGA) module are used to model interactions among semantic, peptide, and TCR sequence features. Experiments on two public datasets, IEDB and McPAS-TCR, show that BioText-SeqFormer achieves stable and competitive predictive performance. Additional Top-k positive predictive value analysis, high-frequency peptide and TCR analysis, 9-mer peptide position-amino acid association analysis, and VDJdb-COVID transfer experiments further demonstrate its candidate screening ability, interpretability, and potential in few-shot transfer scenarios. Overall, this study provides a multimodal strategy for TCR-peptide interaction prediction by integrating language semantics and protein sequence representations.