You're right to be suspicious. "AI discovers new cancer therapy" has the same energy as "this one weird trick changes everything" - which usually means someone wants your attention, your hope, or your venture capital. Fair. But this Cell paper actually does something more useful and less magical: it uses AI to help solve one of the most annoying problems in CAR T therapy - figuring out what, exactly, these engineered immune cells should attack.
And yes, that problem matters. A lot.
CAR T: amazing idea, inconvenient reality
CAR T cells are your own T cells, retooled in the lab to recognize a chosen target on cancer cells. In blood cancers, that trick has produced some genuinely dramatic results. In solid tumors, though, the story gets messier fast. The main issue is target selection. Pick the wrong protein and your CAR T cells either miss the tumor, hit healthy tissue, or wander around looking busy while accomplishing very little - basically the corporate middle management of immunotherapy.
That is the setup for this paper by Baker and colleagues. They built an AI-driven pipeline to hunt for better CAR T targets by combining single-cell RNA sequencing data from tumors and healthy tissues, then filtering candidates for tumor enrichment, tissue specificity, and practical drug-development considerations. They also used large language models to help prioritize the shortlist. Which sounds a little like asking a very well-read intern to sort the pile, but in this case the workflow seems structured rather than gimmicky.
The winner was GPNMB - glycoprotein non-metastatic melanoma protein B, a name that only a committee could love.
Why GPNMB caught their eye
GPNMB is not brand new to cancer biology. Researchers have been circling it for years because it shows up in several cancers and often tracks with aggressive disease biology, immune suppression, or both.12 What makes it interesting here is the breadth. The authors found evidence for GPNMB expression across multiple solid and hematologic tumors, which raises the tantalizing possibility of one CAR T strategy being useful in more than one disease.
That "multi-cancer" angle is the shiny object here, and for once it may not be pure glitter. The team validated GPNMB expression and then engineered a human GPNMB-directed CAR T cell. In mouse models of monoblastic leukemia, melanoma, and colorectal adenocarcinoma, those CAR T cells showed anti-tumor activity.
So the paper is doing two things at once:
- showing off a scalable way to discover CAR T targets, and
- making the case that GPNMB itself is worth pursuing.
That's smart. Also efficient. Clinic has made me appreciate efficient.
The part where AI is useful instead of obnoxious
The most interesting feature of this study is not "AI beat the scientists." It didn't. The useful part is that AI helped sift a gigantic haystack of expression data to find needles that human teams could then judge, rank, and test.
That is a much more believable role for AI in oncology. Not oracle. Not robot oncologist in a Patagonia vest. More like an extremely fast pattern-finder that never needs coffee and does not complain when you hand it 14 awful datasets at 6:30 p.m.
Single-cell transcriptomics is especially suited to this kind of triage. It lets researchers ask: which cells in a tumor express a target, how much, and how often compared with normal tissues? For CAR T, that matters because the dream target is on cancer cells and absent from organs you would prefer to keep. An incredibly high bar, unfortunately set by biology, which remains a deeply unreasonable collaborator.
Why this could matter in the real world
If this approach holds up, it could speed up one of the slowest parts of cell-therapy development: target discovery. That matters because CAR T has outgrown the old blood-cancer playbook. People want workable targets for solid tumors, myeloid malignancies, maybe even non-cancer diseases. The bottleneck is not imagination. It's finding antigens that are common enough, specific enough, and safe enough.
A target like GPNMB could be appealing because it may cover multiple cancers with one platform. That could make development more efficient and expand who gets considered for cell therapy. It also fits with broader efforts to use computational methods and atlas-scale datasets to nominate therapeutic targets more rationally rather than just poking around and hoping for destiny.34
There's also a broader lesson here: AI may be most helpful not when it tries to replace experiments, but when it helps choose which experiments are worth doing. Less "the machine has spoken." More "the machine narrowed the list from 4,000 to 4 so the humans can stop screaming."
Before we all start ordering victory banners
A few caution flags. First, this is still preclinical. Mouse activity is encouraging, but mice have spent decades curing cancer with an enthusiasm not always shared by human biology.
Second, target expression breadth is both the appeal and the risk. If GPNMB is present in normal tissues at meaningful levels, on-target off-tumor toxicity becomes the nightmare scenario. CAR T does not do subtle. Once activated, it tends to kick down the door first and ask questions never.
Third, tumor heterogeneity remains a pain. Even a promising antigen can be patchy within tumors or change under treatment pressure. Cancer cells are slippery little opportunists. Give them one route blocked and they start looking for a side alley.
So yes, exciting. But the adult-in-the-room version of exciting.
Bottom line
This paper offers a credible template for finding better CAR T targets and puts GPNMB on the shortlist of antigens worth serious clinical attention. The headline is not that AI invented a cure. The headline is that AI may help us get less lost while looking for one.
And in oncology, where progress often arrives in sturdy shoes rather than a cape, that counts.
References
-
Baker DJ, Frommer LM, Uslu U, et al. AI-driven discovery of GPNMB CAR T cells as a multi-cancer therapy. Cell. 2026;S0092-8674(26)00602-X. doi:10.1016/j.cell.2026.06.002
-
Rose AAN, Siegel PM. Emerging therapeutic targets in breast cancer bone metastasis. Future Oncol. 2010;6(1):55-74. doi:10.2217/fon.09.136
-
Jogalekar MP, Serrano LMS, Vallabhajosyula P, Gangadaran P. CAR T-cell therapy targeting solid tumors: current challenges and emerging approaches. J Exp Clin Cancer Res. 2024;43:54. doi:10.1186/s13046-024-02908-4 PMCID:PMC10867193
-
Labanieh L, Majzner RG, Mackall CL. Programming CAR-T cells to kill cancer. Nat Biomed Eng. 2023;7:393-415. doi:10.1038/s41551-023-01045-7
-
Pan J, Tan Y, Wang G, et al. Advances in identifying ideal targets for CAR T-cell therapy. Signal Transduct Target Ther. 2024;9:44. doi:10.1038/s41392-024-01732-5 PMCID:PMC10824187
-
Zheng Y, Chen Y, Li X, et al. Single-cell transcriptomics in cancer target discovery and immunotherapy development. Cancer Lett. 2023;566:216244. doi:10.1016/j.canlet.2023.216244
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.