LoFT-TCR: A LoRA-based Fine-tuning Framework for TCR-Antigen Binding Prediction

T cells recognize and eliminate diseased cells by binding their T cell receptors (TCRs) to short endogenous peptides (antigens) presented on the cell surface. While central to adaptive immunity, the identification of TCR-antigen binding pairs remains experimentally labor-intensive and constrained by limited reagents. To address this, we propose LoFT-TCR, a low-rank adaptation (LoRA)-based fine-tuning framework for predicting TCR-antigen binding. Our approach begins with fine-tuning the protein language model ESM-2 on antigen-specific TCR datasets via LoRA to capture discriminative sequence representations. We then construct a heterogeneous interaction graph in which nodes encode sequence features and edges denote interaction relationships. By integrating sequence and structural information through a graph learning framework, LoFT-TCR enhances predictive accuracy. Systematic experiments show that ESM-2 fine-tuning significantly improves the extraction of informative TCR sequence features, which are essential for accurate specificity prediction. LoFT-TCR consistently outperforms state-of-the-art methods in TCR-antigen binding prediction and TCR specificity discrimination, demonstrating substantial gains in predictive accuracy and potential for advancing personalized T cell-based immunotherapy.