Researchers at the University of Pennsylvania unveiled an AI-assisted pipeline to nominate and validate CAR-T targets for solid tumors, spotlighting GPNMB as a multi-cancer candidate. The work, published in Cell, uses a human-in-the-loop framework that pairs large language models with single-cell RNA-seq datasets, then iteratively refines target shortlists for expert testing. The team applied the workflow to skin cancer datasets, filtering more than 10,000 potential antigens by tumor composition, tissue specificity, and CAR-T feasibility constraints. Multiple LLM runs (1,000 independent simulations) were used to reduce noise and hallucinations, resulting in a consensus list that was experimentally evaluated. In preclinical mouse models, GPNMB-directed CAR-T eliminated tumors not only in melanoma but also in models of monoblastic leukemia and colorectal adenocarcinoma, supporting broader therapeutic exploration beyond skin cancer. The approach targets a core bottleneck in solid-tumor CAR-T development: finding antigens that balance selectivity and clinical tractability.