CAR-Tography: Longitudinal Mapping of CAR-T Toxicities across B-Cell Malignancies Reveals Biologically Distinct and Prognostically Relevant Toxicity Phenotypes

Background: CAR T-cell therapy (CAR-T) results in a unique spectrum of toxicities, affecting morbidity, mortality, and quality of life. Traditionally, immunotoxicities like cytokine release syndrome (CRS) and immune effector cell-associated neurotoxicity syndrome (ICANS) are studied independently. However, the overall toxicity landscape of CAR-T is complex and interdependent, with patterns that remain incompletely understood. To address this, we mapped early CAR-T toxicities day-by-day across multiple dimensions to understand their biological and clinical implications. Study Design: This international multicenter observational study included 558 adult patients treated with standard-of-care CD19 or BCMA CAR-T therapies for relapsed/refractory B-cell malignancies (77% B-NHL, 7% B-ALL, 16% multiple myeloma). The product breakdown was: axi-cel (39%), tisa-cel (22%), liso-cel (17%), cilta-cel (9%), ide-cel (7%), brexu-cel (6%). A team of trained investigators evaluated four major toxicity domains daily, from lymphodepletion until 30 days post-infusion. CRS (1) and ICANS (2) were graded according to ASTCT criteria. Infections (3) were clinically or microbiologically defined and graded per Hill et al., Blood 2018. Hematotoxicity grades (4) were derived computationally using serial neutrophil counts (EHA/EBMT ICAHT grading). Results: We comprehensively reviewed over 20,000 patient days, resulting in 80,352 unique toxicity datapoints. To quantify the overall burden of toxicity at a patient-level, we first devised a cumulative toxicity index (CTI) reflecting the aggregated severity and duration of the four major toxicity domains. Across all patients, the median CTI was 23 (IQR 10-51). We identified product-specific variations of CTI. Among CD19 CAR-T products, brexu-cel had the highest CTI, followed by axi-cel, tisa-cel and liso-cel (45 vs. 27 vs. 24 vs. 18, p=0.003). For BCMA CAR-T products, cilta-cel and ide-cel had a comparable CTI (17 vs. 14, p=0.6). Next, we employed a topographical imaging approach to visualize patient-level toxicity profiles across four dimensions of toxicity and time. Visual inspection revealed complex interactions between these five elements, which are challenging to interpret directly. To elucidate these complexities, we applied an unsupervised learning technique-latent trajectory class analysis (Hart & Fei et al., Biometrics 2020)-to the longitudinal toxicity data, focusing on the large B-cell lymphoma cohort (n=384). We identified three distinct toxicity phenotypes: low-tox (LT), mid-tox (MT), and high-tox (HT). The LT phenotype (n=175) was characterized by mild-to-moderate CRS, short duration of cytopenias, rare and mild ICANS, and notably, the absence of infectious complications. The MT phenotype (n=134) exhibited similar rates of CRS and ICANS, but showed longer and deeper cytopenias, and an increased infection signal throughout the first 30 days. The hallmark of the HT phenotype (n=75) was protracted and severe toxicity across all four domains. In terms of CTI, we noted a gradual increase with each phenotype (L/M/HT: 10 vs. 37 vs. 85, p<0.001). Cytokine profiling revealed divergent inflammatory signatures among the toxicity phenotypes, with the HT phenotype showing upregulation of IL-6, IL-10, and TNF-α. The newly defined phenotypes were strongly linked to morbidity and mortality. HT patients were hospitalized longer (L/M/HT: 12 vs. 15 vs. 22 days, p<0.001), required more ICU admissions (4% vs. 5% vs. 47%, p<0.001), and had increased non-relapse mortality (1-year NRM: 3.7% vs. 4.4% vs. 12.6%, p=0.003). Notably, HT patients showed inferior PFS (1-year PFS 53% vs. 47% vs. 30%, p<0.001) and OS (1-year OS, 78% vs. 70% vs. 52%, p<0.001). In a multivariable Cox regression model accounting for age, pre-lymphodepletion LDH levels and product, the toxicity phenotype was independently associated with both PFS (HT vs. LT: HR 2.0, p=0.003) and OS (HT vs. LT: HR 1.7, p=0.007). Conclusions: Leveraging over 80,000 longitudinal data points from a large, deeply annotated international cohort, we developed a new metric, the Cumulative Toxicity Index (CTI), to quantify the main toxicity burden in CAR-T patients. Additionally, employing an unsupervised approach, we defined biologically distinct and prognostically relevant toxicity phenotypes, underlining that CAR-T toxicities should be understood as more than the sum of their parts.