On September 9, 2026, *Nature Biomedical Engineering* online published work from the lab of Caleb Lareau at Memorial Sloan Kettering Cancer Center (MSK). Using a generative protein design pipeline, the team created 1,758 entirely new protein binding domains targeting three CAR-T-relevant antigens — BCMA, CD19, and CD22 — and carried them all the way through protein binding assays, CAR activation tests, primary T cells, single-cell sequencing, and mouse xenograft models. The most useful part of the results is not that "AI designed a better CAR," but that they systematically answered a different question: why some AI-designed binding domains fail once built into a CAR.
🎯 The starting point: CAR-T covers too few patients
CAR-T shows strong efficacy in blood cancers, but its coverage is narrow. Background data from MSK: only about 4% of patients with advanced or metastatic cancer in the US are eligible for CAR-T therapy, and a recent analysis suggests the fraction who actually benefit is close to 3%.
The binding domains in approved therapies are single-chain variable fragments (scFvs) — essentially repurposed pieces of existing antibodies. AI-designed binding proteins take a different path: built from scratch rather than derived from existing antibodies, they can be designed to bind specific regions of a target protein and are smaller and more compact than scFvs.
Lareau's analogy: it is like generating a million answers in ChatGPT and having a second AI model evaluate them. The team first used two generative tools to produce roughly 1 million candidate binders, then used a second model to rank them against various criteria, finally selecting 100–200 candidates worth testing in the real world. DNA sequences encoding these candidates were sent to an external lab for synthesis, returning in about a week.
🧬 Round one: binding is only the beginning
The team first targeted BCMA with both RFdiffusion and BindCraft pipelines, then used yeast surface display (YSD) to screen for proteins that actually bind.
| Pipeline | Binding rate at 100 nM BCMA | |---|---| | BindCraft | ~10.7% of induced yeast bound | | RFdiffusion (five rounds) | up to ~2.4% |
Bio-layer interferometry (BLI) confirmed that some of the new binding domains reached affinities of roughly 2–200 nM. This step showed that zero-shot de novo design can directly deliver clinical-grade affinities.
📉 Built into CARs, 17 binding domains split into three classes
After placing the binders into a second-generation CAR architecture, the team measured two things: whether they activate in the presence of antigen, and whether they activate without antigen. The 17 BCMA binders fell into three classes:
| Class | Count | Behavior | |---|---|---| | Weak activation | 9 | Insufficient CAR activation | | Clear tonic signalling | 3 | Self-activation without antigen | | Low background, strong BCMA-dependent activation | 5 | Ideal |
B5 performed best, inducing cytokine and effector molecule levels in primary CAR-T cells that matched or exceeded the clinical BCMA scFv C11D5.3.
🤖 CARPNN: letting a model learn the rules from experimental data
The next tool, CARPNN (CAR Protein Neural Network), uses ProteinMPNN-style methods to redesign sequences at both the binding interface and non-interface regions while preserving structure, then filters candidates with Boltz-2.
Starting from B5, the team generated 4,000 sequences and selected 21 for CAR experiments. A clear pattern emerged: the more positively charged the binder's net charge, the stronger the tonic signalling — correlation rho = 0.65. Metrics like hydrophobic residue count or hydrogen bond count showed no comparable correlation.
This places a direct constraint on AI protein design: you cannot optimize only "how to bind the tumor antigen" — you must also manage the binder's own surface charge. Otherwise CAR-T cells may activate persistently without target antigen and become exhausted before ever meeting the tumor.
The optimized variant, B5.I0, showed higher expression of effector genes IL2 and IFNG than the parent in single-cell RNA sequencing.
🐭 A controlled test under high tumor burden
The mouse xenograft model used RPMI-8226 cells. Under standard conditions, both B5.I0 and a clinical BCMA CAR drove clear tumor regression.
The team then raised the difficulty: 1 CAR-T cell facing 10 tumor cells. In this high-tumor-burden model, B5.I0 achieved near-complete tumor control; neither the clinical-grade binder C11D5.3 nor the dual-epitope VHH1-VHH2 binder used in CARVYKTI (cilta-cel) achieved comparable control under the same conditions.
Lareau's own words: the AI-designed binding domain essentially achieved complete control of the cancer, while existing clinical products failed to suppress tumor growth.
An important boundary: this is still a preclinical mouse xenograft model and cannot be equated directly with clinical efficacy.
🕳️ Pitfall one: too much positive charge makes CARs activate without antigen
This is tonic signalling. When a binder's net charge skews positive, the CAR activates without antigen, exhausting T cells before they encounter the tumor and reducing efficacy. CARPNN's sequence diversification quantified this rule at rho = 0.65.
🚧 Pitfall two: AI picked the right target (CD19) but the wrong spot
Applying the same pipeline to CD19 produced a fascinating failure: the AI did design binders that bound recombinant CD19 protein, but once built into a CAR they barely activated.
The problem was not "can it bind" but where it bound. Nearly all the AI designs favored a region of CD19 that, in the real membrane environment, is also the binding site of CD81 — which blocks that epitope.
In vitro, with an isolated recombinant CD19 protein, the binders bound fine; in real cells, CD81 was in the way and the CAR could not reach. When researchers knocked down CD81, the previously inactive C1 and C2 CARs regained CD69 activation and killing capacity.
The lesson for AI protein design: an epitope that is computationally bindable is not necessarily accessible in a physiological environment.
⚠️ Pitfall three: targeting CD22 but attacking CD22-negative cells
The third target was CD22. Binders such as D1 and D2 activated normally on CD22-positive cells — but then a danger signal appeared: in unambiguously CD22-negative RPMI-8226 cells, D1 and D2 still produced about 42.9% and 59.6% CD69 activation. This was off-target activation.
The fix was somewhat counterintuitive. The team again used CARPNN, generating 10,000 sequence variants based on D1 — but this time focusing mutations on non-binding-interface residues. The resulting D1.N0 showed:
| Metric | D1 | D1.N0 | |---|---|---| | CD22-dependent activation | Strong | Retained 65–74% CD69⁺ | | RPMI-8226 (CD22-negative) off-target | ~42% | Reduced to 19% | | HeLa (CD22-negative) off-target | 48.8% | Reduced to 8.2% |
Binder optimization usually fixates on the binding interface. This paper shows the non-interface surface can equally determine whether a protein ultimately works — and that such edits can reduce off-target risk while preserving on-target activity.
📐 Three failure modes and one workflow
The paper groups the main obstacles to using de novo binding domains as CARs into three categories: tonic signalling, occluded epitope engagement, and off-target activity.
On computational filtering, the team compared ipAE, ipTM, and ipSAE and found ipSAE ≥ 0.85 to be one of the relatively stable screening thresholds across multiple design tasks, useful for prioritizing candidates more likely to yield functional binders. They also noted that the AI models tend to favor alanine, and high alanine content causes binding problems.
What the authors recommend is a closed loop — not "pick the binder with the best computational score and stop":
1. AI generates binders 2. Validate binding with recombinant protein / yeast display 3. Build into CARs 4. Check three things simultaneously: on-target activity, tonic signalling, off-target activity 5. Optimize sequences with ProteinMPNN/CARPNN 6. Validate in primary T cells and animal models
🔁 Significance: the evaluation standard expands from "how strongly it binds" to four questions
What this work really changes is how the question is asked. Binder evaluation used to ask how strongly a design binds its target. Now it must ask four:
1. Can it reach the target on a real cell membrane? (CD19 case) 2. Does it self-activate without target? (tonic signalling, rho = 0.65) 3. Does it mistakenly recognize other cells? (CD22 case) 4. Can computational metrics predict this in advance? (ipSAE ≥ 0.85)
Lareau says they are turning designs traditionally driven by intuition into more scientific, evidence-based designs. The approach is not limited to CAR-T and can be applied to designing binding domains for other therapeutic uses, including precision drug conjugates and other cancer therapies. His long-term vision: a patient with a rare blood cancer or sarcoma for whom existing therapies have failed could have a therapy designed with AI within weeks.
The lab is currently in the third year of this ten-year plan. MSK has filed a provisional patent on this work, with Lareau and four other authors listed as inventors; Lareau also consults for Cartography Biosciences.
🧭 What to watch next
- Whether the BCMA results can reach the clinic. B5.I0's high-tumor-burden data come from mouse xenografts — a long road remains from animal models to humans;
- Generality of the three failure modes. These rules were derived from just three targets (BCMA, CD19, CD22); whether they hold for other targets remains to be validated;
- Method spillover from CARPNN. Can the idea of fixing off-target activity via non-interface residues 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 designed chimaeric antigen receptors, 2026-09-10