The central challenge of CAR T-cell therapy has never been engineering the immune weapon itself — it has been knowing precisely what to aim at. Solid-tumor targets that are abundant on cancer cells but absent from healthy tissue are vanishingly rare, and the hunt for them has largely relied on slow, intuition-driven lab work. A new computational framework published in Cell may meaningfully accelerate that search.
Researchers built an AI pipeline that integrated single-cell RNA sequencing data from human skin cancer and matched healthy tissue, then cross-referenced public multi-cancer datasets to filter candidates by tumor enrichment, tissue specificity, and clinical tractability. Large language models were layered on top to rank and nominate the most therapeutically promising targets. The top-ranked output was glycoprotein non-metastatic melanoma protein B (GPNMB) — a cell-surface protein previously noted in isolated cancer studies but never systematically elevated to a pan-cancer CAR T target. Expression validation confirmed GPNMB is broadly present across both hematologic and solid tumor types. A fully human GPNMB-directed CAR T construct was then engineered and tested in mouse models spanning monoblastic leukemia, melanoma, and colorectal adenocarcinoma, demonstrating potent anti-tumor activity across all three.
The implications extend beyond GPNMB itself. GPNMB is not an unknown molecule — it has been explored as an antibody-drug conjugate target (glembatumumab vedotin), lending some biological credibility to the AI's selection. However, the mouse efficacy data, while encouraging across three distinct cancer models, cannot yet predict human tolerability or on-target/off-tumor toxicity in normal tissues that express GPNMB at low levels. The study's real contribution may be methodological: demonstrating that AI-integrated single-cell transcriptomics can compress what typically takes years of target identification into a reproducible, scalable pipeline. If replicated and refined with human clinical data, this approach could substantially broaden the addressable landscape of CAR T-cell oncology.