AI-driven approach identifies promising new target for CAR T cell therapy

Leading CAR T cell therapy researchers have developed a human-in-the-loop artificial intelligence (AI) framework that firmly centers scientists' expertise to find viable target antigens for CAR T cell therapy. The work was led by experts from the Perelman School of Medicine at the University of Pennsylvania and Penn's Abramson Cancer Center and published today in Cell.

As proof-of-concept, the team developed a CAR T targeting glycoprotein non-metastatic melanoma protein B (GPNMB), the top candidate nominated by this AI-driven approach, which showed robust tumor-killing activity in mouse models of multiple cancer types.

CAR T cell therapy, a personalized form of immunotherapy that was developed at Penn Medicine and has revolutionized care for several types of blood cancer over the last decade, is increasingly being tested in other solid cancers and even non-cancerous conditions. However, identifying the best antigens for the CARs to target remains a challenge in applying CAR T cell therapy beyond blood cancers. The current FDA-approved CAR T cell therapies target surface antigens that are widely expressed in blood cancers, but not other cancer types. Finding the right targets for new CAR T applications is an incredibly time-consuming and labor-intensive process, compounded by the ever-expanding amount of data.

"Discovering a good CAR target is like trying to find a needle in in a haystack, except the haystack keeps growing as more sequencing data becomes available," explained lead author Daniel Baker, PhD, who earned his doctorate from Penn in December 2025 and completed this work under the mentorship of CAR T cell therapy pioneer Carl June, MD and Zoltan Arany MD, PhD, chair of Physiology at Penn. "We thought this would be a strong use-case for AI because one of the strengths of large language models (LLMs) is the amount of data they can consider. Human experts excel at going deep, while LLMs are good at looking across a broad range of data. So, we created a framework that combines these strengths to build a systematic way to nominate and prioritize potential targets."

Speeding up target discovery in skin cancer

To build and test their AI framework, the research team chose to focus on skin cancer. Unlike other solid tumors, broad immunotherapy strategies, such as immune checkpoint inhibitors and, more recently, tumor-infiltrating lymphocyte (TIL) therapy, have shown efficacy in melanoma, indicating that other immune strategies, like CAR T cell therapy could make a clinical impact, if a good CAR target could be identified.

The researchers integrated four publicly available single-cell RNA sequencing skin cancer datasets along with data from public databases, with specific guidelines to prioritize the 10,000+ potential targets for critical CAR T cell target features. They then used several frontier LLMs to nominate ideal targets from that prioritized list. These simulations were then independently repeated 1,000 times to weed out some of the inherent risks and known issues with AI, such as hallucinations. The results were combined to create a final short list of priority targets for expert review and biological validation.

Daniel Baker, PhD, lead authorBy building this AI framework to work with public data sets, we hope to democratize target discovery so that it's broadly available beyond teams who have access to clinical samples or major institutions that are able to do their own sequencing."

Once the framework was built, the entire process took less than a few weeks, far quicker and less expensive than the current manual methods for target discovery, which can take several months to several years. The research team then validated the targets in laboratory tests to confirm that they were expressed on the surface of cancer cells and built a CAR targeting their lead target, GPNMB. Further preclinical testing of the GPNMB CAR T in laboratory models showed efficacy not only in melanoma, but also in models of leukemia and colorectal cancer.

An AI advance, now available to all

"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, the Richard W. Vague Professor in Immunotherapy at Penn. "Our goal was to show how LLMs could be used in scientific discovery to efficiently find new targets and build new therapies."

Although the team developed the framework using skin cancer data, it was specifically designed to be modular and disease-agnostic, meaning that the same approach could be used for any cancer type or even other diseases. It's also designed to work with different types of datasets. Similarly, the framework was not designed for a specific LLM or model, so it can be applied to future models as LLMs continue to evolve and advance.

"This work highlights how AI can unlock the vast and growing wealth of bioinformatic data in a systematic and data-driven way," said Arany, the Samuel Bellet Professor of Cardiology at Penn. "This is only the tip of the iceberg, as agentic AI is on the rise."

The AI framework is included in the paper's methods section so that other scientists can use and adapt it for their own research. The Penn research team plans to apply the framework to other cancer types and diseases, and to continue refining the GPNMB-targeted CAR T cell therapy for potential future clinical trials.

Sikander Hayat, PhD, of the Icahn School of Medicine at Mount Sinai and RWTH Aachen University, is a co-corresponding author of the manuscript, with Baker, June, and Arany.

The study was supported by the National Institutes of Health (CA248315, 1P01CA214278 and R01CA226983), the Centurion Foundation Innovation Fund, the Parker Institute for Cancer Immunotherapy, and the Norman and Selma Kron Endowed Fellowship.

Source:

University of Pennsylvania School of Medicine

Journal reference: