Researchers at the University of Pennsylvania and Abramson Cancer Center reported an AI-enabled, human-in-the-loop framework for nominating CAR T targets, then validated the top pick in preclinical models. The Cell study used large language models alongside single-cell RNA-seq datasets to filter thousands of potential antigens for tumor composition, tissue specificity, and clinical feasibility. In repeated simulation runs, the framework converged on GPNMB as its leading candidate. The team engineered a GPNMB-directed CAR T product and showed tumor elimination across multiple preclinical settings, including melanoma, monoblastic leukemia, and colorectal adenocarcinoma. The authors described the approach as complementary to expert review: LLMs perform broad scanning, while investigators “go deep” through experimental evaluation. The work aims to address a key bottleneck in expanding CAR T beyond blood cancers, where identifying safe, selective solid-tumor targets has been slow and labor-intensive. Why it matters: the platform could accelerate solid-tumor CAR target discovery pipelines and reduce search costs by tightening the link between computational prioritization and translational feasibility.