Artificial intelligence for identifying tumor-reactive CD8⁺ T cells: Biological principles, computational advances, and future directions
image: Integrative translational pipeline for AI-driven tumor-reactive TCR discovery, spanning clinical sample acquisition, multi-omic profiling, structural modeling, functional screening, and therapeutic application.
Credit: Chao Yang1, Yingying Sun2, Yan Zhang1*, Yifen Shi1* and Zhiwen Shi1*
The therapeutic deployment of tumor-reactive (TR) CD8⁺ tumor-infiltrating lymphocytes (TILs) represents the cornerstone of personalized adoptive cell therapies. However, the immense combinatorial diversity of the T cell receptor (TCR) repertoire and the highly personalized nature of the tumor immunopeptidome present formidable biological bottlenecks. Conventional identification strategies—which rely heavily on empirical in vitro co-cultures and known epitope predictions—are severely hindered by restricted antigen coverage, phenotypic distortion during ex vivo expansion, and prohibitively prolonged manufacturing timelines. The integration of artificial intelligence (AI) is currently precipitating a paradigm shift, transitioning TR T cell discovery from a heuristic screening process to a predictive, high-dimensional science. By decoding single-cell transcriptomic states, tracking dynamic clonal architectures, and modeling the structural biophysics of TCR-peptide-major histocompatibility complex (pMHC) interfaces, AI frameworks offer unprecedented sensitivity and scalability. Nevertheless, as computational oncology matures, profound "scientific blind spots" remain. Current algorithms frequently conflate terminal T cell exhaustion with specific tumor reactivity, suffer from severe "shortcut learning" driven by dataset survivorship bias, and optimize for static structural affinity while critically neglecting the mechanotransduction essential for genuine cytotoxic activation. This review critically interrogates the biological rationale and algorithmic evolution of AI-driven TR T cell identification. We dissect the mechanistic limitations of contemporary machine learning strategies and delineate a visionary roadmap—emphasizing spatial context, biophysical dynamics, and multimodal generative AI—to overcome translational hurdles and realize the ultimate goal of programmable, patient-specific immunotherapies.