A functional CAR-T cell atlas to unravel regulatory mechanisms of CAR-T therapies

Abstract Introduction: Over the past decade, CAR-T therapies have emerged as a groundbreaking treatment for cancer, showing remarkable efficacy in certain hematologic malignancies. However, these therapies still face important limitations, including significant toxicities that compromise long-term efficacy and safety. Despite advances enabled by single-cell RNA sequencing (scRNA-seq), the regulatory mechanisms underlying CAR-T cell function remain poorly understood. Therefore, there is still a lack of integrative studies that leverage the vast amounts of public data to dissect the molecular mechanisms governing the antitumoral response of CAR-T therapies. Methods: In this work, we present a CAR-T cell functional atlas built from large-scale public scRNA-seq datasets to deepen our understanding of transcriptional regulation. We collected transcriptomic information from over 1 million CAR-T cells targeting CD19, BCMA, APRIL, GD2, and HER2, across 13 independent studies. Our atlas includes samples from 102 patients with various hematologic cancers (ALL, NHL, MM) and 14 healthy donors. Importantly, the atlas encompasses both infusion products (IP) and post-infusion CAR-T cells collected at time points ranging from 7 days to 9 years. Importantly, extensive metadata, covering patient demographics, therapy-induced toxicities (CRS, ICANS), and clinical responses, was also compiled. Given the importance of robust dataset integration, we benchmarked several integration strategies (fastMNN, Harmony, RPCA-Seurat, Liger, STACAS, scVI). The scVI embedding was selected for its optimal performance in minimizing batch effects while preserving biological variation. After QC, filtering, and integration, a shared latent space comprising 414,000 high-quality CD3+CAR+ T cells was generated. We annotated 11 distinct T cell phenotypes based on canonical markers, identifying memory, activated, and cytotoxic CD4+ and CD8+subsets among others. Note that all studies contributed to nearly all phenotypes, supporting successful integration. Results: This comprehensive atlas enabled the validation and extension of key CAR-T characteristics. Differential abundance analysis showed that IPs from complete responders were enriched in memory T cells, while regulatory T cells were more prevalent in non-responders. Moreover, we deepen into the impact of sex and age in the IP composition and associated therapeutic response. An age-stratified analysis revealed that, unlike men, women showed an age-related shift from CD8+ cytotoxic to a memory phenotype. Besides, we also examined transcriptional dynamics of CAR-T cells post-infusion, focusing on CD8+ cytotoxic cells, the largest population across all time points. As expected, TCR signaling and cytotoxic pathways were upregulated early after infusion, and then progressively reduced. Intriguingly, selenocysteine synthesis increased post-infusion, consistent with its role in T cell proliferation and cytotoxicity. In a focused analysis comparing the transcriptional landscape between responders and non-responders, we observed that oxidative phosphorylation was significantly downregulated in patients not achieving a complete response, while cholesterol metabolism was strongly repressed in complete responders, suggesting a link between cholesterol accumulation and T cell dysfunction. Finally, our developed atlas allowed the identification of rare cell populations that were not considered in the different studies individually but become relevant when leveraging all datasets. Notably, we identified monocyte-like T cells, also named IACs (ICANS-associated CAR-T cells), in post-infusion samples for the first time (previously described only in the IP). Comparative analysis of pre- and post-infusion states revealed upregulation of inflammatory pathways and IFN-γ-associated genes, offering novel insights into CAR-T cell-associated toxicities. Conclusion: We envision this resource as a foundational tool for understanding CAR-T cells function, deciphering key regulatory mechanisms and guiding the design of the next generation of CAR-T therapies through atlas-based integrative studies. Moreover, our CAR-T cell atlas has been specifically designed to allow straightforward integration of additional data series, thereby increasing the versatility of the tool. To support this effort, we provide open-access computational resources, including a trained CAR-T cell atlas model integrated with the scVI toolkit and an intuitive web application.