Abstract
Chimeric antigen receptor (CAR) T cell therapy holds transformative potential for hematologic malignancies, yet predicting patient-specific treatment efficacy and neurotoxicity remains a major clinical challenge due to the complex and heterogeneous nature of the infused CAR-T cell populations. Here, we introduce CART-GPT, a transformer-based model fine-tuned on a curated atlas of 1.12 million CAR-T single-cell RNA-seq profiles annotated with clinical outcomes. CART-GPT is the first AI model developed for CAR-T therapy that predicts both treatment response and the risk of immune effector cell-associated neurotoxicity syndrome (ICANS), achieving state-of-the-art performance (AUC ~0.8) and marking a significant advance in the field. The model provides interpretable insights, revealing that neither therapeutic efficacy nor neurotoxicity is driven by individual cell types alone, but by the combined influence of discrete, distinct subsets across diverse T cell states and transcriptional programs. A novel cell aggregation strategy links single-cell predictions to patient-level metrics, enhancing both accuracy and biological relevance. As a contribution to this ever-evolving field, we also release a comprehensive, annotated single-cell CAR-T atlas as a community resource to facilitate future research in immunotherapy. These advances demonstrate the potential of foundation models in single-cell biology to inform precision CAR-T treatment planning and facilitate the rational design of next-generation cell therapies.