Scaling CAR-T Targeting of HLA-presented Intracellular Antigens with AI-Driven Experimentation

The majority of cancer-driving proteins are intracellular, and so can only be recognized by immunotherapies through short peptide fragments displayed on the human leukocyte antigen (HLA). We developed a lab-in-the-loop system to learn the rules of scFv-pHLA protein-protein interactions on human cells. We use generative models of proteins and of screens to design, synthesize, and test interactions between tens of millions of scFvs and 100 pHLAs in a single multiplexed experiment, producing large scale training datasets. Transformers trained on the data predict unseen interactions and exhibit reliable scaling laws, with steady model improvements against seen and unseen pHLAs as experiments continue. Overall, AI-driven experimentation enables models to systematically learn to design TCR mimicking antibodies.