Daniel J. Baker had a problem that vexed CAR T cell researchers for years: among the thousands of proteins on the surface of cancer cells, which one could serve as a universal target across the many forms the disease takes?
The answer, published June 25 in Cell, came not from a decade of painstaking laboratory work but from an AI pipeline that completed its search in weeks. The system, built at the University of Pennsylvania’s Center for Cellular Immunotherapies, screened more than 10,000 potential targets and nominated one, glycoprotein non-metastatic melanoma protein B (GPNMB), as the most promising candidate for a new class of CAR T cells.
When the team engineered CAR T cells targeting GPNMB and tested them in mouse models, the cells showed potent anti-tumor activity against three distinct cancer types: monoblastic leukemia, melanoma, and colorectal adenocarcinoma.
“A modular, disease-agnostic AI pipeline for target discovery in cell and gene therapy,” the authors write, describing an approach that could change how CAR T researchers decide which antigen to chase.
CAR T cell therapy works by engineering a patient’s own T cells to recognize a specific protein on cancer cells. The approach has been transformative for blood cancers, several CAR T therapies are now FDA-approved for leukemias, lymphomas, and multiple myeloma. But finding the right target remains the hardest part of the process.
An ideal target must be present on the surface of cancer cells, absent or minimal on healthy tissue, and expressed broadly enough to cover multiple cancer types. Discovering such a target manually, through painstaking proteomic screens and validation, can take years.
Baker and his colleagues took a different approach. They fed four public single-cell RNA-seq datasets from human skin cancer and healthy tissue into an AI system built around multiple frontier large language models. The system scored each of more than 10,000 potential targets on features critical for CAR T therapy: surface expression, tumor specificity, tissue distribution, and clinical feasibility.
To reduce the risk of AI hallucinations, a well-known failure mode of LLMs that could produce plausible but false candidate targets, the team ran the simulation 1,000 independent times, keeping only the targets that consistently appeared at the top.
GPNMB emerged as the consensus choice.
Why GPNMB?
GPNMB is not a completely unknown protein. Prior clinical safety data existed from glembatumumab vedotin, an antibody-drug conjugate that targets GPNMB and had been tested in breast cancer, melanoma, and osteosarcoma trials. This prior safety record made it an especially attractive candidate for CAR T engineering.
The Penn team designed human GPNMB-directed CAR T cells and tested them in mouse xenograft models. The results showed robust tumor cell killing across hematologic and solid tumor types, a combination that CAR T cells rarely achieve, as most effective CAR T therapies to date have been limited to blood cancers.
“It is the first-of-its-kind use of large language models for target discovery in cell and gene therapy,” the researchers note.
The paper’s 33 authors span multiple institutions including Penn’s Center for Cellular Immunotherapies, the Icahn School of Medicine at Mount Sinai, RWTH Aachen University, and the University of Michigan. Carl H. June, a pioneer of CAR T cell therapy who led the first successful human CAR T trial at Penn in 2010, is the senior author. Baker, who earned his PhD from Penn in December 2025, is the lead author.
AI Democratization
The approach is designed to be reproducible by any research group. The pipeline uses only publicly available datasets, GTEx, TCGA, Tabula Sapiens, ProteomicsDB, Open Targets, and is not tied to any specific LLM. The team included the full methods for the AI workflow in the paper, making it adaptable to future models and other disease targets.
The same issue of Cell includes a companion commentary titled “CAR T targets: AI takes the wheel” (DOI: 10.1016/j.cell.2026.06.003), contextualizing the approach as a shift in how CAR T researchers will identify targets going forward.
The Penn team now plans to refine GPNMB-directed CAR T cells for eventual clinical trials, and to apply the framework to other cancers and potentially non-cancer diseases.
Source: Baker DJ, Frommer LM, Uslu U, Patel KK, et al. “AI-driven discovery of GPNMB CAR T cells as a multi-cancer therapy.” Cell 2026; 189(13):3871-3882.e12. DOI: 10.1016/j.cell.2026.06.002