CAR T therapy is already a tiny science-fiction heist.

Doctors remove your T cells. Rewire them with a synthetic receptor. Send them back in with a wanted poster.

It works best when the villain wears an obvious badge, like CD19 on many B-cell cancers. Solid tumors are messier. More fake mustaches. More innocent neighbors.

The new Cell paper from Baker et al. asks a cleaner question: can AI help find a better badge?

Its top answer: GPNMB.

Not "AI cured cancer." Please unclench.

Something narrower. Better. A human-guided AI pipeline nominated GPNMB. Then lab work showed GPNMB-directed CAR T cells could attack several cancer models in mice [1].

The Target Is The Trap

A CAR T cell needs a target on the outside of a cancer cell.

Too rare, and the therapy misses.

Too common in healthy tissue, and the therapy may hit the wrong address.

Too patchy inside the tumor, and cancer cells slip away wearing sweatpants and a fake ID.

That is why target discovery hurts.

The team started with single-cell RNA sequencing from human skin cancer and healthy tissue. Single-cell RNA-seq is cellular roll call. Instead of blending tissue into molecular soup, it asks each cell what genes it is using.

Then came filters.

Is the gene found in tumors? Is it on the cell surface? Is it low in vital healthy tissue? Could a CAR realistically bind it?

If you mapped this pipeline in mapb2.io, it would look less like magic and more like a very picky funnel. Good. Science needs more picky funnels and fewer crystal balls.

The LLM Walks Into The Lab

The researchers screened more than 10,000 possible targets. Then they asked frontier large language models to rank candidates.

Not once.

Repeatedly.

Penn Medicine says they ran the nomination process 1,000 times to reduce hallucinations, which is wise because LLMs can sound very confident while handing you a banana labeled "tumor antigen" [2].

GPNMB kept showing up.

That matters because GPNMB is not random alphabet soup, though it does look like a Wi-Fi password. It is a transmembrane glycoprotein already linked to several cancers.

The team validated GPNMB expression across blood and solid tumors. Then they built a human GPNMB-directed CAR T cell.

In mouse models of monoblastic leukemia, melanoma, and colorectal adenocarcinoma, those cells showed potent anti-tumor activity [1].

Mouse data is not a patient.

It is a starting gun.

Why This One Lands

The clever part is the pipeline.

Other recent work points the same way. SCAN-ACT used single-cell and multi-omics data to nominate CAR and TCR targets in sarcoma and glioblastoma [3]. Reviews keep repeating the same uncomfortable truth: CAR T needs safer targets, better tumor entry, and better persistence in hostile tumor neighborhoods [4,5]. AI is also moving into CAR design, manufacturing, and clinical prediction [6].

Baker et al. stitched those threads together.

Public datasets. Single-cell resolution. Expert rules. LLM ranking. Wet-lab validation.

The LLM did not get a lab coat and a parking space. Humans still chose constraints and checked the biology.

Good.

The model is the fast librarian. The scientists are the adults confirming the book exists.

The Catch, Because Biology Always Has One

GPNMB may be promising.

It may also be complicated.

A useful CAR target must clear a nasty safety bar. If normal tissues need GPNMB, the therapy could cause collateral damage. If tumors lose GPNMB under pressure, escape happens. If the tumor microenvironment suppresses T cells, the cells may arrive heroic and leave exhausted, like everyone after assembling flat-pack furniture.

There are early human efforts around GPNMB-targeting CAR T therapy, including GCAR1 trials for selected GPNMB-expressing solid tumors [7].

Still, this Cell paper reports preclinical mouse work.

Dose, toxicity, persistence, antigen escape, and patient selection remain open files.

The Bigger Bet

If other labs reproduce and expand this, the impact is not just one new CAR T.

It is faster target hunting.

That could matter for rare tumors, where samples are scarce and time is cruel. It could help teams reuse public cell atlases instead of starting every search from zero.

No, AI did not replace immunologists.

It gave them a metal detector.

Now the field has to prove the beep is treasure, not a soda can.

References

1. Baker DJ, Frommer LM, Uslu U, et al. "AI-driven discovery of GPNMB CAR T cells as a multi-cancer therapy." Cell. 2026;189(13):3871-3882.e12. DOI: 10.1016/j.cell.2026.06.002. PMID: 42349383.

2. Penn Medicine. "AI framework aids target discovery for CAR T cell therapy." 2026. Link.

3. Testa S, Pal A, Subramanian A, et al. "SCAN-ACT: adoptive T cell therapy target discovery through single-cell transcriptomics." Genome Medicine. 2025;17:89. DOI: 10.1186/s13073-025-01514-9.

4. Labanieh L, Mackall CL. "CAR immune cells: design principles, resistance and the next generation." Nature. 2023;614:635-648. DOI: 10.1038/s41586-023-05707-3.

5. Zugasti I, Espinosa-Aroca L, Fidyt K, et al. "CAR-T cell therapy for cancer: current challenges and future directions." Signal Transduction and Targeted Therapy. 2025;10:210. DOI: 10.1038/s41392-025-02269-w.

6. Luciani F, Safavi A, Guruprasad P, Chen L, Ruella M. "Advancing CAR T-cell Therapies with Artificial Intelligence: Opportunities and Challenges." Blood Cancer Discovery. 2025;6(3):159-162. DOI: 10.1158/2643-3230.BCD-23-0240.

7. ClinicalTrials.gov. "GCAR1, a Chimeric Antigen Receptor (CAR) T-CELL Therapy for Relapsed/Refractory GPNMB-Expressing Solid Tumours." NCT07297667. Link.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.