CELL INTELLIGENCE · 肿瘤细胞治疗研究

Adoptive T cell therapies, including tumor-infiltrating lymphocyte (TIL) transfer and engineered chimeric antigen receptor (CAR) or T cell receptor (TCR) therapies, have transformed the immuno-oncology landscape but remain limited by a fundamental variation between antigenic variety (breadth) and target precision. TIL therapies capture diverse antigen recognition but often fail to enrich tumor-reactive clones with sustained proliferative potential. On the other hand, CAR-T therapies achieve potent, antigen-specific cytotoxicity, yet are constrained by tumor heterogeneity and the need for pre-identified targets. Recent advances in machine learning (ML) promise to bridge this gap. Platforms such as PredicTCR and TRTpred, trained on paired TCR sequences and single-cell transcriptomics, can predict tumor-reactive clones with > 90% accuracy from a single tumor biopsy, enabling high-variety, high-precision selection within days. Complementary deep learning frameworks, including MATE-Pred and BertTCR, extend predictive capacity across diverse epitopes and HLA backgrounds. These innovations foreshadow a paradigm in which ML-driven algorithms guide the rapid design of personalized TCR-engineered products, potentially reducing manufacturing timelines from months to weeks. However, challenges remain in validating model generalizability across tumor types, minimizing false predictions, and integrating safety profiling into computational selection pipelines. By converging computational intelligence with cellular immunotherapy, ML-enhanced adoptive T cell therapies may overcome current limitations and realize the long-sought goal of individualized, tumor-specific immunotherapy for solid tumors. Overall, the study aims to highlight the use of ML platforms to unify variety and precision in adoptive T-cell therapies, enabling rapid, personalized selection of tumor-reactive clones for next-generation solid tumor immunotherapy.