Inflammatory Biomarker Clusters Are Predictive of Response and Toxicity in Large B-Cell Lymphoma Treated with CD19 CAR-T Cell Therapy

Disease recurrence and toxicity are common sequelae of CD19-directed chimeric antigen T-cell (CAR-T) cell therapy in large B-cell lymphoma (LBCL). We have developed a machine learning approach that is informative of adverse CAR-T outcomes and can support personalized treatment approaches.

In a single-center cohort of 198 LBCL patients treated with CD19-CAR-T cells (axicabtagene ciloleucel [axi-cel] 50%, tisagenlecleucel [tisa-cel] 32%, lisocabtagene maraleucel [liso-cel] 18%), we applied unsupervised k-means clustering on standard laboratory and cytokine measurements collected within the day prior to CAR-T infusion. Three distinct biomarker clusters emerged (Fig. A) – inflammatory (n=27, 14%), neutral (n=121, 61%), and favorable (n=50, 25%). The inflammatory subgroup was enriched for patients with high levels of inflammatory markers such as IL-6, TNFa, and ferritin, while the favorable cluster tended to have higher complete blood counts and albumin; intermediate values were characteristic of the neutral subgroup.

To determine the clinical implications of these subgroups, we studied their association with day 100 best response to CAR-T, significant toxicity by day 30 (defined as life-threatening complications such as immune toxicities requiring intervention, bloodstream infections, or early death), and overall survival (OS) after multivariate adjustment for CAR-T costimulatory domain (CD28 vs 41BB), patient age, primary refractory disease, and pre-lymphodepletion LDH. Compared to those in the favorable biomarker cluster, patients in the inflammatory cluster (Fig. A) had increased odds of not achieving CR by day 100 (odds ratio [OR] 4.23, 95% confidence interval [CI] 1.37-14.4, p < 0.05) and a high toxicity profile (OR 12.5, 95% CI 3.66-49.5, p < 0.001). OS was also reduced with the inflammatory subgroup in univariable analysis (Fig. B) and a multivariable Cox regression model (hazard ratio [HR] 4.03, 95% CI 1.99-8.16, p < 0.001). Patients in the neutral subgroup, compared to those in the favorable group, had reduced overall survival (HR 1.80, 95% CI 1.00-3.24, p < 0.05) and increased odds of high toxicity (OR 3.40, 95% CI 1.56-7.79, p <0.01).

Serial assessment of the uncommon biomarkers and cytokines in our day 0 panel may not be readily available to many centers. We developed a random forest prediction model for cluster type based on patient, disease, and widely available pre-lymphodepletion laboratory features. Our random forest model outperformed gradient boosting and logistic regression model alternatives and achieved a high discrimination (AUC 0.81) for cluster prediction. Next, we applied the random forest model to pre-lymphodepletion data from an independent LBCL cohort (n=155) from a different center (Fig. B). Most patients were assigned to the neutral subgroup (n=108, 70%) followed by the inflammatory (21%) and favorable (9%) subgroup. Assignment to the inflammatory cluster was strongly associated with inferior OS compared to the favorable subgroup (HR 2.89, 95% CI 1.12-7.48, p < 0.05) after multivariate adjustment for CAR-T costimulatory domain (CD28 vs 41BB), patient age, primary refractory disease, and pre-lymphodepletion LDH.

In conclusion, we identified three pre-infusion (day 0) biomarker clusters, which are tightly associated with CAR-T outcomes. These clusters could guide decision-making regarding hospitalization after CAR-T infusion and to preempt early inflammation and disease relapse in high-risk subgroups. Finally, to improve accessibility to cluster assignment, we developed a prediction model for cluster type based on readily available pre-lymphodepletion data and applied it towards an external center cohort as a proof-of-principle.