Precision CAR-T Cell Therapy Towards Computationally Guided Curative Immunotherapy: A Multi-Layered Adaptive Intelligence Framework

Abstract

Background: Chimeric antigen receptor (CAR) T cell therapy has revolutionized the treatment of relapsed and refractory haematological malignancies, with complete response rates ranging from 62% in B cell lymphoma to 83% in relapsed multiple myeloma. However, long-term remission is still not achieved in a significant proportion of patients: in adult B cell acute lymphoblastic leukaemia patients, the median event-free survival is typically around 13.3 months, with an initial response rate of over 80%, and progression-free survival curves still fall with prolonged follow-up in relapsed/refractory multiple myeloma. These data indicate a basic shortcoming: the lack of an intelligent predictive layer that can predict relapse, personalise the infusion product, and modify the therapeutic plan in near-real time.

Framework: Here, we introduce AI-CART (Adaptive Intelligence to CAR-T Therapy), a new three-level computational architecture that combines multi-modal patient data via a transformer-based multi-omic intelligence engine that implements DeepSurv (deep Survival - a deep learning–based survival prediction model), GNN (graph neural networks) to estimate antigen escape probability, SHAP-driven (SHapley Additive exPlanations) explainability, and a patient-specific digital twin simulated by ODE (Ordinary Differential Equation–parameterized) immune-tumour dynamics and scenario simulation by reinforcement learning.

Performance: Evaluated on a conceptually harmonized dataset from six landmark clinical trials [ZUMA-1 (A clinical trial of axicabtagene ciloleucel (CAR-T therapy) in lymphoma), ZUMA-3 (A clinical trial of brexucabtagene autoleucel in acute lymphoblastic leukemia), ELIANA (A clinical trial of tisagenlecleucel in pediatric/young adult B-cell ALL), JULIET (A clinical trial of tisagenlecleucel in diffuse large B-cell lymphoma), TRANSCEND (A clinical trial of lisocabtagene maraleucel), CARTITUDE-1 (A clinical trial of ciltacabtagene autoleucel in multiple myeloma), LEGEND-2 (A clinical trial of LCAR-B38M CAR-T therapy in multiple myeloma); n=1,847 patients], the AI-CART (Artificial Intelligence–enabled Chimeric Antigen Receptor T-cell) framework achieves 2-year PFS (Progression-Free Survival) prediction C-statistics (Concordance Statistic (also called C-index; measures predictive discrimination) of 0.78 to 0.91 across malignancy subtypes, exceeding standard clinical models by 17 to 28 percentage points. The sensitivity of the Toxicity surveillance modules ranges from 0.74 to 0.88 and specificity from 0.77 to 0.84 against grade 3+ CRS (Cytokine Release Syndrome), ICANS (Immune Effector Cell–Associated Neurotoxicity Syndrome) and chronic cytopenias.

Implications: AI-CART offers a conceptual architectural roadmap to the transition of CAR-T therapy to computationally-driven personalised immunotherapy. An equity-based federated learning system generalizes the system to multi-continental deployment without sharing patient-level data.