Artificial intelligence to forecast efficacy of anti-CD19 directed CAR-T cell therapy

CD19-directed chimeric antigen receptor (CAR) T cell therapy has shown high efficacy and tisagenlecleucel/tisa-cel and axicabtagene ciloleucel/axi-cel rapidly evolved to standard of care for relapsed/refractory large B cell lymphomas (R/R LBCL). However, only 40% of patients experience durable remissions and the determinants of treatment failures are still inconsistently defined. Biological mechanisms explaining lack of response are emerging but they are largely unsuccessful in predicting disease response at the single patient level. Additionally, to maximize the cost-effectiveness of CAR T cell therapy, biomarkers able to predict response and survival prior to CAR T manufacturing would be highly desirable. Many of the parameters used to date to predict CAR T success, are in fact generated at the time of or after CAR T infusion (such as for instance CAR T phenotypes in infusion products, CAR T in-vivo expansion, CAR T phenotypes in patient blood, disease response at 30 days, day 90 PET scan, etc) and do not sufficiently address this issue. It is also unlikely that a single biomarker could robustly profile prognosis and prediction alone. In this study, we present a completely novel approach as we aim at identifying risk categories before CAR T manufacturing combining all clinical and biological features assessable at the time of leukapheresis. To this aim, we will apply artificial intelligence (AI) tools using a training cohort encompassing all R/R LBCL patients treated with CAR

T at our Institution and a validation cohort comprising patients of the Italian CAR T SIE program I am coordinating. We will use comprehensive biological and clinical data from the training cohort to extract and select features and

then exploit the validation cohort to verify the learning results of the AI. These multiscale data include patient-, cancer- and treatment-related information, genomics and radiomics data, immunophenotypic features of the

systemic and tumor microenvironment. Our approach has the following potential implications: 1. The identification of patients not likely to respond before CAR T manufacturing could prompt a more "tailored" use of CAR T on specific subgroups thus significantly impacting the rising problem of CAR T sustainability for health care systems;

2. High-risk patients could be potentially shifted to alternative treatment protocols (bispecific antibodies, etc);

3. These analyses could suggest novel mechanisms for CAR T resistance and indicate strategies to enhance CAR T-cell efficacy