Recent advances in unlocking T cell signatures in cancer through deep learning

T-cell receptors (TCRs) recognize antigens, including cancer antigens, offering a promising avenue for immunological profiling and early detection. As cancer diagnostics increasingly shift towards less invasive methods, deep learning has emerged as a powerful tool in facilitating this transition. The identification of cancer-associated TCRs (caTCRs) from peripheral blood samples presents a compelling alternative to traditional biopsy-based approaches, enabling more scalable and non-invasive cancer screening. In this review, we examine twelve state-of-the-art caTCR detection frameworks, detailing their conceptual innovations and method workflows. We also perform a limited external benchmarking of a subset of publicly available pre-trained models on our own data and analyze their performance. We further discuss the biological and methodological constraints of caTCRs and their detection methods, and assess their ability to be adopted within clinical settings. By highlighting both recent advances and remaining limitations, this manuscript aims to provide a balanced perspective on the field and to inform future research towards more robust, interpretable, and clinically relevant caTCR-based diagnostic tools.