Background Adoptive cell therapy (ACT) has revolutionized cancer treatment. However, optimizing its design for consistent long-term benefits remains challenging. Mechanisms determining ACT efficacy are not fully understood, making experimentally explaining immune-tumor interactions difficult. Multiple aspects contribute to pharmacological variances, such as cellular heterogeneity and dynamic tumor microenvironment (TME). While widely used to investigate these perturbations, cell-line and animal models are resource-heavy, cross-sectional, and provide limited data granularity, hindering the translation to clinical practices. In silico modeling can flank in vivo/vitro experiments by providing additional insights and cost-effective hypothesis testing, though accuracy, complexity, and interpretability vary. To systematically evaluate ACT variables and bridge biological mechanisms with experimental observations, we hypothesized that a data-informed, rule-based microscale model can simulate cell population dynamics in ACT and thereby predict differential tumor control effects. Given the scarcity of computational models on natural killer cells, in this project, we developed ABMACT to understand, simulate, and predict TME dynamics in Chimeric Antigen Receptor Natural Killer (CAR-NK) ACT for B cell lymphoma.

Methods Agent-based models (ABM) are well-suited to emulate spatial interactions at the single-cell level and retain stochasticity, flexibility, and interpretability. ABMs encode the biological rules of individual cells and simulate the collective behaviors of cell populations over time. ABMACT integrated data from in vitro/vivo experiments and scRNA-seq to capture variable cytotoxicity in NK cells, delineate NK cell fate transitions, and predict tumor control. We validated our models on a Raji-NK cell coculturing mouse model experiments¹ and benchmarked it against other modeling methods such as ordinary differential equations (ODE).

Results By cross-evaluating multiple NK cell characteristics and dosages, we found that NK cell proliferation rate most significantly influenced time to tumor clearance, followed by effector-to-target ratio, and serial killing capacity. Higher baseline cytotoxicity, proliferation rate, serial killing capacity, and lower death rate in the CD19IL15 CAR-NK group accelerated tumor reduction. ABMACT substantially outperforms ODE models in the accuracy and interpretability of NK cell dynamics. In addition, ABMACT allows investigations of multiple dosing, which can help optimize ACT treatment schedules and dosages.

Conclusions By developing ABMACT, we provide a computational framework for testing the impact of CAR-NK ACT design and conditions on long-term tumor control and patient responses. Integrating biological knowledge with multi-modal data, ABMACT offers mechanistic explanations of immune-tumor interactions. In silico models reduce the need for extensive laboratory work and can adapt to other cell types and ACTs, aiding hypothesis generation and testing for ACT design and personalized treatment planning.

Reference

1. Li L, et al. Loss of metabolic fitness drives tumor resistance after CAR-NK cell therapy and can be overcome by cytokine engineering. Sci Adv 2023;9:eadd6997.