Abstract
Introduction: Chimeric antigen receptor T-cells (CAR-T) are increasingly being utilized for the treatment of lymphomas.
Cytokine secretion contributes not only to CAR-T therapeutic efficacy but also to toxicities, including cytokine release syndrome (CRS), immune effector cell–associated neurotoxicity syndrome (ICANS), and cytopenia known as immune effector cell–associated hematotoxicity (ICAHT), all of which are driven by increased cytokine secretion.
Thus, we performed comprehensive analyses on post infusion cytokine profiles in lymphoma patients treated with CAR-T cell therapy, to identify specific cytokine profiles associated with key clinical outcomes, including CRS, ICANS, cytopenia of any type and overall response.
Methods: Data were collected from the electronic medical records of B-cell lymphoma patients treated with CAR-T therapy at two tertiary medical centers. Plasma from peripheral blood samples was isolated for cytokine analysis on Day 7 following CAR-T cell infusion. Thirty cytokines, chemokines, and growth factors levels were measured using a multiplex bead-based immunoassay MILLIPLEX® Panel A (48-plex) (Millipore). “Cytopenia” was defined as reduction in any blood count (absolute neutrophil count<500/ µL, platelets< 20,000/ µL, hemoglobin< 8 gr%) Occurring until 15 days post CAR-T cells infusion.
Descriptive statistics were used to characterize the cohort. Pearson correlation analysis was performed to assess relationships between cytokine levels and clinical outcomes (Overall survival, ICANS, CRS and cytopenia subtypes).
Statistical modeling using Logistic Regression (LR) was conducted on all possible cytokine groups of size 1 to 6 (overall 768,181 groups). Groups were analyzed for statistical inference over the entire dataset, and for predictive power of Cytopenia.
Based on statistical significance, 97 cytokine groups were used for analysis using LR, AdaBoost and Random Forests (RF) to predict outcomes and toxicities.
Model performance was evaluated using accuracy, area under the curve (AUC), and 5-fold cross-validation cross-model analysis identified the most reliable predictive cytokines across all toxicity types.
Results: A total of 54 patients were included in the clinical analysis [29 males, 25 females, median age 63 (range 20-81) years], with complete cytokine data available for 38 patients (70.4%). Most patients (57.4%) had de novo diffuse large B-cell lymphoma, while 42.6% had transformed lymphoma subtypes. Twenty-six, 23 and 5 patients were treated with tisa-cel, axi-cel and brexu-cel, respectively.
The overall incidence of CRS was 89%, with 13% grade 3–4. ICANS occurred in 26% of patients with 7% grade 3-4. Cytopenia was observed in 36 patients, with the most frequent type being pancytopenia (18 patients). Thirty-two patients (59%) patients achieved complete response 1 month post infusion. The onset of ICANS highly correlates with IL-6 (0.82 correlation coefficient) and MIP-1a (0.81), TNFa (0.73) values, CRS (levels 3-4) highly correlate with IL-8 (0.68), G-CSF (0.65) and IL-6 (0.61). Cytopenia does not show a strong correlation to any cytokine.
Statistical inference was conducted using an LR model fitted over the entire dataset using every possible group of size 1-6. Out of which, 97 groups were selected for further analysis, based on significant p-values (<0.05) for all model coefficients.17 of these groups fitted to an accuracy larger than 0.92, AUC for these models range between 0.82-0.98.
From predictive capabilities analysis on the same 97 groups, we listed the top scoring groups for each model type (LR, RF, AdaBoost) and gathered the 7 groups which appeared in 2/3 top scoring lists. The models' accuracy ranges from 0.814 to 0.818, with AUC 0.83-0.92, indicating predictive capability of cytopenia based on those groups of cytokines.
Example of such cytokine groups (from the top 3 results) are:
[EGF, IL-10, IL-15, VEGF-A] – AUC - 0.9233, [IL-5, IL-17A, MCP-1] – AUC - 0.9167, [IL-4, IL-5, IL-17A, IL-18, IL-22] – AUC - 0.8933
No cytokine group was found to predict CAR-T treatment response or differentiate between CRS and infection.
Conclusions: Machine learning approaches can be employed to evaluate the predictive potential of day-7 CAR-T cytokine signatures. We were able to demonstrate a specific cytokine signature that may help to predict cytopenia's.
Larger cohorts and other time point cytokine profiles are essential to enhance the predictive potential of these models.
