Chimeric antigen receptor (CAR) T cell therapy is a promising cancer immunotherapy; however, several challenges hamper its clinical efficacy. Although new computational models are attempting to explore the vast combinatorial design space of CAR components and suggest novel designs, challenges remain in capturing cellular heterogeneity as well as generalizing to unseen CAR variants. Here, we introduce an optimal transport (OT)-based framework designed to predict responses to CAR expression at the single-cell level, including variants that have not been experimentally tested. Our model accurately captures gene expression changes across diverse CAR variants, outperforming the baseline for in-distribution CARs while reflecting biological characteristics. By embedding CARs using protein language models, we extend our framework to a conditional OT-based model that successfully generalizes our predictions to out-of-distribution CAR designs. Our findings highlight the utility of OT-based modeling in elucidating CAR design-function relationships, enabling the rational design of novel CARs with therapeutic potential. A record of this paper's transparent peer review process is included in the supplemental information.
Driessen et al. (2026) studied this question.