Engineering de novo binder CAR-T cell therapies with generative AI

Abstract

Chimeric antigen receptor T cell (CAR-T) therapies have revolutionized cancer treatment, with six CAR-T products currently in clinical use1-4. Despite their success, high resistance rates due to antigen escape remain a major challenge5,6. In silico design of de novo binders (DNBs) has the potential to accelerate the development of new binding domains for CAR-T, possibly enabling personalized therapies for cancer resistance7,8. Here, we show that DNBs can be used for CAR-T, targeting clinically relevant cancer antigens. Using a DNB against the epidermal growth factor receptor (EGFR), we demonstrate comparable cytotoxicity, cytokine secretion, long-term proliferation, and lysis of primary patient-derived cancer organoids with single-chain variable fragment (scFv)-based and DNB-based CAR-T cells. Moreover, we use generative artificial intelligence (AI) guided binder design with RFdiffusion9 to target the B cell maturation antigen (BCMA), a key antigen in multiple myeloma treatment10-17. We confirmed the activity of our AI-designed BCMA CAR-T in short- and long-term effector readouts, including a xenograft mouse model of multiple myeloma. Notably, our AI-guided CAR-T approach also successfully targets a mutated BCMA protein variant resistant to the clinically used bispecific antibody teclistamab. In sum, we demonstrate a proof-of-concept for engineering new, bespoke cellular immunotherapies targeting cancer resistance with the help of generative AI. This approach may further accelerate the development of new CAR-T therapies addressing cancer resistance.