Did an AI design a better CAR-T component? The lab result came from people, several models and three test rounds
A company preprint reports a signalling domain less than one-third the usual length with stronger laboratory readouts. It was built through scientist-directed screening, not autonomous discovery, and has not been tested in animals or patients.
The short version
- What happened
- Generate Biomedicines screened 339 unique intracellular-domain designs over three rounds; the lead design was 45 amino acids, against 154 in the clinical 4-1BB–CD3ζ reference.
- Who it affects
- For patients and clinicians, the relevant promise is a future CAR-T platform that may carry more functions or use less genetic payload. This experiment remains several evidence stages away from showing a safer, more effective or more accessible treatment.
- What remains unproven
- Unreviewed company preprint
Keep in mind
Unreviewed company preprint
Commercially interested collaborators
Cell assays rather than animal or patient outcomes
Some experiments used few blood-cell donors
No matched expert-only or computational-design baseline
No direct exhaustion, toxicity, manufacturing-cost or clinical-outcome evidence
Multi-source analysis · 2 linked sources. Primary analysed source date: 27 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The compact component worked better in several laboratory tests
Generate Biomedicines says a human-directed system of AI agents helped produce a much shorter signalling component for CAR T cells while improving several laboratory readouts. The lead intracellular domain, krgr–ζ1–40, was 45 amino acids long—less than one-third of the 154-amino-acid 4-1BB–CD3ζ reference used in clinical CAR designs. In primary human T cells it produced more receptor at the cell surface, stronger antigen-driven interferon-gamma release and sustained target-cell killing after temporary mRNA delivery.
That matters because shorter genetic instructions could leave room for additional receptors or functions when a therapy is delivered by mRNA, a lipid nanoparticle or a viral vector. But the measured outcomes were receptor expression, cytokine release and killing of cultured target cells. The study did not test a treatment in an animal, establish safety or benefit in a patient, or show that a smaller signalling domain will make manufacturing cheaper.[1][2]
This was a multi-model, scientist-steered search—not autonomous discovery
The first screen began with 186 designs selected after scientists reviewed proposals from Claude Opus 4.6, Claude Sonnet 4.6, GPT-5.2, Kimi K2.5 and MiniMax M2.5. Thirty-eight matched or exceeded the reference signal in a Jurkat reporter-cell assay; 78 were retested in primary T cells. Later rounds used Claude Code with newer Opus models, earlier measurements and scientist-defined questions. Across all three rounds, the paper records 339 unique designs with activity calls.
That sequence is more informative than the shorthand that 'Claude designed' the component. People set the objectives, recovered and reviewed prompts, chose which candidates the laboratory would build, interpreted plates and decided what to test next. The final lead also built on an earlier candidate proposed by Claude. Experimental feedback did the decisive filtering: the authors say current models did not reliably predict which designs would work.[1][2]
The distinction that matters
What the experiment actually tested
The result came from a repeated human–model–laboratory cycle.
Propose
Scientists directed several AI models
Candidate signalling sequences
Measure
Laboratory screens tested cells
Expression, cytokine and killing readouts
Refine
People and agents used earlier results
A compact lead—not a tested medicine
A useful design result is not yet evidence of a better therapy
The preprint reports stronger expression across several binders and delivery methods, including tests using blood cells from five donors for one lipid-nanoparticle experiment. Yet other experiments used few donors, the advantage varied with the antibody binder, and the mechanism behind higher surface expression remains unresolved. The study did not measure T-cell exhaustion or performance in a living organism. It is company-authored, unreviewed and commercially aligned with Generate and Anthropic.
Most importantly, the team did not compare its workflow with expert-only design or another computational method under the same laboratory budget. So the evidence supports a compact candidate and a traceable lab-in-the-loop process; it does not establish that agents discover better CARs faster than skilled scientists. The assessment would strengthen with independent replication, a prespecified matched baseline, animal efficacy and toxicity studies, and eventually a clinical programme showing that the compact design improves outcomes rather than only cell assays.[1][2]