> On September 9, 2026, *Nature Biomedical Engineering* published work from Caleb Lareau's team at Memorial Sloan Kettering Cancer Center (MSK). Using a generative protein design pipeline, they created 1,758 novel protein binders targeting three CAR-T antigens—BCMA, CD19, and CD22—then carried them through protein binding, CAR activation, primary T cells, single-cell sequencing, and mouse xenografts. The most useful result is not that "AI designed a better CAR," but that the study systematically answers a different question: why some AI-designed binders fail once installed in a CAR.
The starting point: CAR-T covers too few patients
CAR-T works well in blood cancers but reaches a narrow population. MSK's background figures: only about 4% of US patients with advanced or metastatic cancer are eligible for CAR-T, and a recent analysis suggests the fraction who actually benefit is close to 3%.
Approved therapies use single-chain variable fragments (scFvs), essentially repurposed parts of existing antibodies. AI-designed binders take another route: built from scratch rather than adapted, they can be designed to target specific regions of a protein and are smaller and more compact than scFvs.
Lareau's analogy: it is like generating a million answers in ChatGPT and having another AI model evaluate them. The team used two generative tools to produce about 1 million candidate binders, ranked them with a second model, and selected 100–200 candidates worth testing experimentally. DNA encoding the candidates was sent to an external lab for synthesis and returned in about a week.
Round one: binding is only the beginning
The team first used RFdiffusion and BindCraft pipelines against BCMA, then screened for genuine binders with yeast surface display (YSD).
| Pipeline | Binding at 100 nM BCMA | |---|---| | BindCraft | ~10.7% of induced yeast bound | | RFdiffusion (five rounds) | up to ~2.4% |
Biolayer interferometry (BLI) confirmed affinities of roughly 2–200 nM for some designs. This step showed zero-shot de novo design can directly deliver clinical-grade affinity.
17 binders in CARs split into three classes
After installing binders into a second-generation CAR, the team measured two things: whether the CAR activated with antigen, and whether it self-activated without antigen. The 17 BCMA binders fell into three classes:
| Class | Count | Behavior | |---|---|---| | Weak activation | 9 | Insufficient CAR activation | | Strong tonic signalling | 3 | Self-activation without antigen | | Low tonic, strong BCMA-dependent activation | 5 | Ideal profile |
B5 performed best, inducing cytokine and effector molecules in primary CAR-T cells at levels matching or exceeding the clinical BCMA scFv C11D5.3.
CARPNN: letting models learn rules from experimental data
The next tool, CARPNN (CAR Protein Neural Network), uses ProteinMPNN-style sequence redesign of both binding and non-binding surfaces while preserving structure, with Boltz-2 filtering.
Starting from B5, the team generated 4,000 sequences and took 21 into CAR assays. A clear rule emerged: the more positively charged the binder's net charge, the stronger the tonic signalling, with a correlation of rho = 0.65. Hydrophobic residue count and hydrogen bond count showed no comparable correlation.
This directly constrains AI protein design: you cannot only optimize binding to the tumor antigen; the binder's surface charge matters too. Otherwise the CAR-T may stay activated without target antigen and become exhausted before meeting the tumor.
The optimized B5.I0 showed higher effector gene expression (IL2, IFNG) than the parent in single-cell RNA sequencing.
A controlled test under high tumor burden
In the RPMI-8226 xenograft model under standard conditions, B5.I0 and the clinical BCMA CAR both caused clear tumor regression.
The team then raised the difficulty: 1 CAR-T cell per 10 tumor cells. In this high-burden setting, B5.I0 achieved near-complete tumor control; the clinical binder C11D5.3 and the dual-epitope VHH1-VHH2 binder used in CARVYKTI (cilta-cel) did not achieve comparable control.
Lareau's quote: with the AI-designed binder they essentially achieved complete control of the cancer, while existing clinical products failed to suppress tumor growth.
A necessary caveat: this is still a preclinical mouse xenograft model and cannot be equated directly with clinical efficacy.
Pitfall 1: too much positive charge makes CARs self-activate
This is tonic signalling. When a binder's net charge skews positive, the CAR activates without antigen, T cells become exhausted before encountering the tumor, and efficacy drops. CARPNN's sequence diversification quantified the rule at rho = 0.65.
Pitfall 2: right target (CD19), wrong epitope
Applying the same workflow to CD19 produced an instructive failure: the AI did design binders that bound recombinant CD19 protein, but CARs built from them barely activated.
The problem was not binding ability but epitope choice. Nearly all AI designs converged on a region of CD19 that, in the real membrane environment, is also the binding site of CD81—which blocks access. With isolated recombinant CD19 in vitro, the binders bound; on real cells, CD81 was in the way. When the researchers knocked down CD81, the previously inert C1 and C2 CARs regained CD69 activation and killing.
Lesson: an epitope that is computationally bindable is not necessarily accessible in a physiological context.
Pitfall 3: targeting CD22 but attacking CD22-negative cells
The third target was CD22. Binders D1 and D2 activated normally on CD22-positive cells, but a danger sign appeared: on CD22-negative RPMI-8226 cells they still produced ~42.9% and ~59.6% CD69 activation—off-target activation.
The fix was counterintuitive. Using CARPNN again, starting from D1 and generating 10,000 variants, the team mutated primarily non-binding-surface residues. The resulting D1.N0:
| Metric | D1 | D1.N0 | |---|---|---| | CD22-dependent activation | Strong | Retained 65–74% CD69+ | | RPMI-8226 (CD22-negative) off-target | ~42% | Down to 19% | | HeLa (CD22-negative) off-target | 48.8% | Down to 8.2% |
Binder optimization usually focuses on the binding interface. This paper shows non-interface residues can determine whether a protein works, and that such edits can reduce off-target risk while preserving target engagement.
Three failure modes and one workflow
The paper groups the main obstacles to using de novo binders in CARs into three classes: tonic signalling, occluded epitope engagement, and off-target activity.
On computational screening, the team compared ipAE, ipTM, and ipSAE and found ipSAE >= 0.85 to be a relatively stable threshold across design tasks for prioritizing candidates likely to yield functional binders. They also noted AI models tend to favor alanine, and high alanine content causes binding problems.
The recommended approach is a closed loop, not "pick the highest-scoring binder and stop": generate binders with AI → validate binding on recombinant protein/yeast → install in CAR → check on-target activation, tonic signalling, and off-target activity → optimize sequences with ProteinMPNN/CARPNN → validate in primary T cells and animals.
Significance: the evaluation standard expands from one question to four
The real change is in the questions asked. Previously, an AI-designed binder was judged by how strongly it bound its target. Now there are four:
1. Can it reach the target on a real cell membrane? (CD19 case) 2. Does it self-activate without antigen? (tonic signalling, rho = 0.65) 3. Does it misrecognize other cells? (CD22 case) 4. Can computational metrics predict success in advance? (ipSAE >= 0.85)
Lareau describes the goal as turning traditionally intuition-driven design into more scientific, evidence-based design. The method is not limited to CAR-T and could be used to design binders for other therapeutic purposes, including precision drug conjugates and other cancer therapies. His long-term vision: a patient with a rare blood cancer or sarcoma, out of standard options, could get an AI-designed therapy within weeks.
The lab is in year three of this ten-year plan. MSK has filed a provisional patent on the work, with Lareau and four other authors listed as inventors; Lareau also consults for Cartography Biosciences.
What to watch next
- Can the BCMA results reach the clinic? B5.I0's high-burden data come from mouse xenografts; there is a long road from animal models to humans.
- Generalizability of the three failure modes. The rules were derived on BCMA, CD19, and CD22; whether they hold for other targets needs testing.
- CARPNN's spillover. Whether repairing off-target activity via non-interface mutations can generalize to other generative protein design tasks.
References
1. Chow A., Chu H., Li R., Nalbant B. N., Dozic A. V., Kida L. C., Tang Z., Palmeri J. R., Lareau C. A. Sequence and structural determinants of efficacious de novo chimaeric antigen receptors. Nature Biomedical Engineering, 2026-09-09, DOI 10.1038/s41551-026-01790-9 2. Memorial Sloan Kettering Cancer Center: Using AI to design better cancer therapies, 2026-09-10 3. Medical Xpress: AI-designed proteins outperform existing CAR T designs against BCMA tumors in mice, 2026-09-10 4. Nature Biomedical Engineering Research Briefing: Defining attributes of effective binders for AI-assisted CAR design 5. Bioon: Sequence and structural determinants of de novo chimeric antigen receptors, 2026-09-10