Researchers developed and experimentally validated a human-in-the-loop AI-guided framework to identify candidate chimeric antigen receptor (CAR) T-cell targets, according to a study published in Cell.

By integrating single-cell RNA sequencing data sets from human skin cancer and healthy tissue, the approach identified GPNMB as the top candidate for CAR T-cell targeting. As proof of concept, the investigators engineered a GPNMB-directed CAR T-cell therapy that demonstrated antitumor activity in mouse models of multiple tumor types.

“Discovering a good CAR target is like trying to find a needle in a haystack, except the haystack keeps growing as more sequencing data become available," explained lead study author Daniel J. Baker, PhD, of Perelman School of Medicine at the University of Pennsylvania, Philadelphia. "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.”

“Our goal was to show how LLMs could be used in scientific discovery to efficiently find new targets and build new therapies," added senior author Carl H. June, MD, Director of the Center for Cellular Immunotherapies and of the Parker Institute for Cancer Immunotherapy at the University of Pennsylvania Perelman School of Medicine, and the Richard W. Vague Professor in Immunotherapy in the Pathology & Laboratory Medicine Department. 

Building the Framework

The investigators integrated four publicly available single-cell RNA sequencing skin cancer data sets with a healthy human skin data set to construct a malignant-plus-healthy atlas for identifying genes compositionally enriched in the malignant samples.

Following data set integration, the investigators extracted additional features considered critical for CAR T-cell development for each potential target from public resources, including UniProt, the Human Protein Atlas, the Genotype-Tissue Expression Project, Open Targets, ClinicalTrials.gov, and ProteomicsDB. Four categories of features were considered: subcellular localization, tumor specificity, cancer composition, and clinical translatability. They then used a human-in-the-loop approach combining expert-defined criteria with three independent LLMs (ChatGPT-4o, Claude 3.7, and Gemini 2.5 Pro) to determine feature weights.

The investigators applied this weighting schema to prioritize the top 100 candidates for further evaluation. Each LLM then nominated 10 genes from the shortlist and provided explicit reasoning for each selection based on predefined criteria. Consensus across the models informed the top hits for CAR T-cell development.

“By 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,” Dr. Baker said.

Target Validation and Preclinical Testing

The most frequently nominated target was GPNMB. Orthogonal validation using bulk transcriptomic data, immunofluorescence microscopy, and surface-level protein confirmation via flow cytometry supported that GPNMB were transcriptionally upregulated in melanoma, expressed at the protein level, and present on the cell surface. Other highly ranked candidates included ERBB3 and IGF1R. The investigators subsequently selected GPNMB for therapeutic development due in part to its surface expression across hematologic and solid tumors and low or absent expression in healthy tissues.

To build on the AI-nominated target, the investigators engineered a second-generation CAR T-cell construct using the antigen-binding domains of the antibody-drug conjugate glembatumumab vedotin, which had demonstrated in studies the feasibility of targeting surface GPNMB. Primary human T cells transduced with a lentiviral vector successfully expressed the GPNMB-directed CAR on their surface and demonstrated robust activation and expansion across multiple donors, the investigators wrote.

The GPNMB-directed CAR T-cell therapy demonstrated in vitro activity in hematologic malignancy (monoblastic leukemia and acute myeloid leukemia) and solid tumor (melanoma, colorectal cancer, pancreatic cancer, and renal cancer) models, as well as in xenograft models of monoblastic leukemia, melanoma, and colorectal adenocarcinoma.

Insights and Opportunities

“These findings establish a scalable pipeline for CAR T-cell target discovery and support the translation of GPNMB-directed CAR T cells as a multicancer therapeutic,” the investigators concluded.

Supporting that conclusion, they designed the framework to be modular and disease-agnostic, enabling its adaptation to other cancer types and diseases, diverse data sets, and future LLMs. The framework is included in the study’s methods to facilitate its use by other researchers, and the team at the University of Pennsylvania plans to expand its application while continuing to refine the GPNMB-directed CAR T-cell therapy for potential clinical trials.

Senior author Zoltan Arany, MD, PhD, the Samuel Bellet Professor of Cardiology at Perelman School of Medicine and Chair of the Physiology Department, speaking from an AI standpoint, said, “This work highlights how AI can unlock the vast and growing wealth of bioinformatic data in a systematic and data-driven way. This is only the tip of the iceberg, as agentic AI is on the rise.”

DISCLOSURES: The study was supported by the National Institutes of Health, the Centurion Foundation Innovation Fund, the Parker Institute for Cancer Immunotherapy, and the Norman and Selma Kron Endowed Fellowship. For full disclosures of the study authors and data and code availability, visit cell.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.