Researchers at Penn Medicine have developed an AI-assisted framework for identifying new targets for CAR T-cell therapy, using the approach to nominate and validate GPNMB as a potential multi-cancer target.
The work used a human-in-the-loop strategy that combined public single-cell RNA sequencing datasets, large language models, expert review, and experimental validation. The aim was to address one of the central bottlenecks in expanding CAR T-cell therapy beyond blood cancers: finding surface antigens that are sufficiently expressed on tumors while limiting the risk of damage to healthy tissue.
“Discovering a good CAR target is like trying to find a needle in a haystack, except the haystack keeps growing as more sequencing data becomes available,” said lead author Daniel Baker, who completed the work under the mentorship of Carl June and Zoltan Arany at Penn.
The team focused first on skin cancer, integrating four publicly available single-cell RNA sequencing datasets with public protein and tissue-expression resources. More than 10,000 potential targets were filtered for features relevant to CAR T development, including tumor expression, tissue specificity, and clinical feasibility. Several large language models were then used to prioritize candidates, with repeated simulations designed to reduce the risk of hallucination or unstable model output.
Glycoprotein non-metastatic melanoma protein B, or GPNMB, emerged as the most frequently nominated candidate. The researchers then confirmed its surface expression across several tumor types and engineered a human GPNMB-directed CAR T cell. In mouse models, the CAR T cells showed antitumor activity in melanoma, monoblastic leukemia, and colorectal adenocarcinoma.
Target discovery remains particularly difficult in solid tumors, where candidate antigens are often shared with healthy tissues, tumors are heterogeneous, and the tumor microenvironment can suppress immune-cell activity. The authors argue that their framework could help democratize target discovery by relying on publicly available data, making it useful for groups without access to large clinical sample collections or proprietary sequencing programs.
“To our knowledge, this study represents one of the first uses of large language models in the field of cell and gene therapy, including CAR T cell therapy,” said June. “Our goal was to show how LLMs could be used in scientific discovery to efficiently find new targets and build new therapies.”
The framework was designed to be modular and disease-agnostic, meaning it could be applied to other cancers or non-malignant diseases as datasets and language models improve.
"This paper is a useful example of where AI may actually help: not replacing biology, but organizing vast datasets, nominating candidates, and accelerating the path to experiments," said Bruce Levine, who was not involved in the research, on X.
Although the GPNMB CAR T cells showed activity in laboratory and mouse models, further work will be needed to assess specificity, safety, manufacturing feasibility, and clinical translation.
Newsletters
Receive the latest pharmaceutical news, personalities, education, and career development – weekly to your inbox.
