COMPUTATIONAL METHODS FOR IDENTIFYING SINGLE AND COMBINATORIAL CAR T-CELL TARGETS FROM SINGLE-CELL TRANSCRIPTOMICS DATA

Abstract

Chimeric antigen receptor (CAR) T cell therapy has demonstrated remarkable success in treating hematological malignancies, but faces significant challenges in solid tumors due to antigen heterogeneity and on-target, off-tumor toxicities. This dissertation presents computational approaches leveraging single-cell RNA sequencing data to identify both single and combinatorial CAR-T cell targets with optimal safety and efficacy profiles across multiple solid tumor types.

This dissertation presents three contributions. First, we established a comprehensive framework for evaluating CAR target safety and selectivity using patient tumor single-cell transcriptomics data. Through a pan-cancer analysis, we demonstrated the near-optimality of existing CAR targets in most cancers while identifying novel promising targets for head and neck squamous cell carcinoma. These findings validated current clinical approaches and highlighted the challenge of finding superior single antigens in most cancer types.

Building on this foundation, we developed LogiCAR designer, a novel genetic algorithm that identifies tumor-specific logical combinations of surface antigens using AND, OR, and NOT operators. When applied to breast cancer cohorts comprising nearly 2 million cells from 342 patients, LogiCAR designer successfully identified triplet antigen combinations that discriminate between tumor and normal cells more effectively than clinically approved single-antigen targets. We then applied this framework to a new paradigm: individualized circuit design. We identified circuits tailored to the unique tumors of patients and showed the striking potential to achieve >99% tumor-targeting efficacy for most patients, when applied to a novel multi-ethnic cohort of 82 patients. Importantly, LogiCAR designer demonstrated robust convergence properties and computational efficiency across gene combination sizes of up to 5 genes, making it practical for potential clinical deployment of next-generation cell therapies.

Finally, I apply our computational framework to lung cancer in non-smokers (LCINS), a critical unmet clinical need with distinct biology from smoking-associated lung cancer. While smoking rates have declined, LCINS incidence has been rising. Patients have limited treatment options: tumors inevitably evolve resistance to molecularly targeted therapies, and they tend to respond poorly to immune checkpoint inhibitors due to low mutational burden of LCINS. This creates an urgent need for alternative treatment strategies. Using single-cell transcriptomics data from LCINS patients, we systematically evaluated potential CAR targets and their combinations, and identified promising candidates with superior efficacy and safety profiles specific to this patient population.

This work advances the field of CAR-T cell therapy by providing computational tools for rational target selection for cell therapy design. The methods presented enable systematic identification of safer and more effective CAR targets through data-driven analysis of single-cell transcriptomics data, addressing critical challenges in extending CAR therapy to diverse malignancies. We demonstrate applications in head and neck cancer, breast cancer, and lung cancer in non-smokers, establishing a generalizable framework for CAR target discovery that can be adapted to additional cancer types, patient populations, and even other pathologies.