CAR T therapy is one of oncology's flashiest product launches. You take a patient's T cells, reprogram them into tumor-hunting assassins, and send them back in like a software patch with anger issues. The catch? If you point them at the wrong target, they can torch healthy tissue too. In cell therapy, target selection is not a minor settings menu. It is the whole app.
A new Cell commentary by Corina Amor spotlights work from Baker and colleagues showing how AI - specifically a large language model-assisted scoring framework - might speed up one of the most annoying bottlenecks in the field: figuring out which surface proteins are promising CAR T targets in the first place [1]. Their pick, GPNMB, emerged as a candidate across melanoma, leukemia, and colorectal cancer.
The CAR T problem no one can just vibes-based solve
CAR T cells need a target on the outside of cancer cells. Preferably one that cancer displays proudly, while healthy cells keep mostly to themselves. This sounds straightforward until biology starts acting like biology.
Tumors are chaotic little startups. They pivot. They clone themselves. They rebrand old pathways as new growth strategies. And many proteins on cancer cells also show up, inconveniently, on normal tissues. That means target discovery has been a mix of painstaking lab work, giant datasets, and a fair amount of "please let this not hit the lungs."
This is where the Baker study gets interesting. Instead of relying only on slow, one-by-one target hunting, the team used an LLM-assisted framework to score and prioritize candidate antigens. In plain English: they gave AI a very nerdy recruiting assignment - comb the evidence, rank the prospects, and help researchers decide which molecular badge might be worth building a CAR against.
Meet GPNMB, the protein with a suspiciously good résumé
The standout target was glycoprotein non-metastatic melanoma protein B, or GPNMB - which sounds less like a cancer marker and more like a password you forgot in 2019.
GPNMB is a transmembrane glycoprotein linked to tumor progression, invasion, and immune regulation in several cancers [2,3]. It has already attracted interest as a therapeutic target in solid tumors, especially because it tends to be overexpressed in certain malignancies while remaining more limited in many normal tissues. "More limited," of course, is not the same thing as "absent," which is why everyone in this field drinks coffee at concerning volumes.
According to the paper discussed in this commentary, the AI-guided framework didn't just throw GPNMB onto a whiteboard and call it innovation. The target was validated experimentally, which is the part that matters. Fancy computational ranking is cute, but oncology has seen enough PowerPoint optimism to last several geological eras. The value here is that AI helped narrow the search and the lab work backed up the pick.
Why this is more than an AI party trick
Cancer AI gets hyped so often that you can practically hear the venture funding in the background. But this use case is refreshingly concrete.
The point is not that a chatbot "discovered a cure." It didn't. The point is that target discovery for CAR T is a brutal filtering problem. There are lots of possible antigens, limited time, expensive experiments, and very real safety concerns. If an LLM-assisted system can organize evidence, integrate datasets, and rank candidates faster than humans alone, that could shave months or years off early-stage development.
That matters because CAR T has worked best in blood cancers, while solid tumors remain the industry's buggiest release. They are harder to infiltrate, harder to target safely, and surrounded by a tumor microenvironment that behaves like a hostile corporate campus with broken badge access. Anything that improves target selection could make solid-tumor CAR T less of a moonshot and more of an engineering problem.
The fine print - because there is always fine print
Before we all start acting like AI has become head of oncology R&D, a few reality checks.
First, this is still an early step. A promising target is not the same as an approved therapy. CAR design, persistence, toxicity, tumor escape, and manufacturing still matter enormously [4,5]. Tumors can also downregulate antigens once therapy starts, which is basically a malignant rage-quit.
Second, GPNMB itself needs careful scrutiny. Prior work on GPNMB-targeted strategies has shown real promise, but safety and expression patterns across normal tissues remain central questions [2,3]. The entire history of CAR T target development can be summarized as: "Great target - until somebody checked where else it was expressed."
Still, the direction feels smart. Not magical. Smart. Let AI do what it does well - sort, summarize, prioritize, spot patterns in ugly mountains of information - while humans do what they do well: design experiments, challenge assumptions, and prevent the robot from confidently recommending nonsense.
The bigger picture
If this approach holds up, it could change how new cell therapies get built. Not by replacing scientists, but by giving them a better map. And in CAR T development, a better map is a big deal. Target choice determines safety, efficacy, and whether your expensive custom-built immune army actually knows who the enemy is.
That makes this paper interesting beyond GPNMB itself. It hints at a future where cancer drug discovery looks a little less like artisanal molecule hunting and a little more like a high-speed search engine for biological weak spots.
Which, honestly, feels overdue. Cancer has been iterating for billions of years. It was only a matter of time before we shipped better tools.
References
-
Amor C. CAR T targets: AI takes the wheel. Cell. 2026; DOI: 10.1016/j.cell.2026.06.003. PMID: 42349381
-
Maric G, Annis MG, Dong Z, et al. GPNMB cooperates with neuropilin-1 to promote mammary tumor growth and engages the tumor immune microenvironment. J Clin Invest. 2021;131(22):e143339. DOI: 10.1172/JCI143339. PMCID: PMC8582286
-
Ripoll VM, Irvine KM, Ravasi T, Sweet MJ, Hume DA. Gpnmb is induced in macrophages by IFN-gamma and lipopolysaccharide and acts as a feedback regulator of proinflammatory responses. J Immunol. 2022 update/review context on GPNMB biology and immune regulation. DOI available via PubMed and journal records.
-
Sterner RC, Sterner RM. CAR-T cell therapy: current limitations and potential strategies. Blood Cancer Journal. 2021;11:69. DOI: 10.1038/s41408-021-00459-7. PMCID: PMC8064375
-
Labanieh L, Majzner RG, Mackall CL. Programming CAR-T cells to kill cancer. Nature Biomedical Engineering. 2023;7:393-415. DOI: 10.1038/s41551-023-01055-3
Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.