Researchers at the University of Pennsylvania and Abramson Cancer Center unveiled an AI-driven, human-in-the-loop framework to nominate CAR T cell antigens across solid-tumor contexts. In a study published in Cell, the team combined large language models with single-cell RNA sequencing datasets to filter and rank antigens designed for CAR T safety and feasibility. Rather than relying on AI alone, the approach ran 1,000 independent LLM simulations to reduce noise and mitigate hallucination risk, then used expert review to select top candidates. The researchers highlighted Glycoprotein non-metastatic melanoma protein B (GPNMB) as the lead target. Preclinical testing showed engineered GPNMB-directed CAR T cells eliminated tumors not only in melanoma models but also in monoblastic leukemia and colorectal adenocarcinoma settings. The work targets a known bottleneck in solid-tumor CAR T development: identifying selective targets that preserve therapeutic windows.