Global AI protein design contest puts 12,000 computer-made cancer binders to the test in living T cells

Artificial intelligence has become remarkably good at dreaming up proteins that stick to chosen targets, but a striking question has lingered behind the field’s headline successes: does a protein that binds on a computer actually work inside a living cell? A massive open competition called Bits to Binders has now delivered one of the most direct answers yet. Twenty-eight teams from 42 countries submitted 12,000 artificially designed minibinders, each exactly 80 amino acids long, intended to recognize the lymphoma surface antigen CD20 when deployed as the recognition domain of a chimeric antigen receptor, or CAR, in human T cells. Rather than scoring the designs by binding affinity alone, the organizers pushed every single one through a functional pipeline in primary human T cells, measuring whether the engineered cells proliferated, expanded, released cytokines, and killed tumor cells. The results, published in Molecular Systems Biology, reveal both the promise and the hidden failure modes of current AI protein design.

The competition’s premise deliberately raised the bar beyond standard binder benchmarks. A CAR binding domain must do far more than dock onto its antigen: it must be synthesized as DNA, expressed on the T cell surface within a fixed receptor scaffold, and trigger signaling that is strong enough to activate the cell but not so strong that the T cell becomes exhausted. Teams were told their 80-amino-acid binders would replace the single-chain variable fragment, or scFv, of a second-generation CAR28z construct, a domain that in natural antibodies typically runs 230 to 260 amino acids. Success ultimately depended on CAR-T cell function, not just molecular stickiness. Organizers provided primers on CAR-T biology, the exact scaffold, and assay details, then let competitors loose with whatever AI-driven workflows they preferred, creating an unusually diverse and honest snapshot of how the field actually designs proteins today.

The methodological landscape that emerged was dominated by generative diffusion models. Eighteen teams used RFdiffusion, two used sequence-structure co-design tools such as Chroma, and others turned to hallucination-based approaches like BindCraft, EvoBind, and ColabDesign, or protein language models including PepMLM, ESM-2, and RayGun. Nearly twenty teams employed ProteinMPNN or its soluble variant to design sequences onto fixed backbones, and seventeen used AlphaFold2 to verify that their creations were predicted to fold as intended. Twelve teams essentially followed the default RFdiffusion-ProteinMPNN-AlphaFold2 pipeline, while nine practiced iterative diffusion, trimming low-scoring segments and regenerating until scores improved. The result was extraordinary sequence diversity: more than 83 percent of submissions showed no significant alignment to any sequence from another team, although within-team redundancy was high, with a fifth of sequences nearly identical to a teammate’s design.

Testing 12,000 designs individually would have been impractical, so the organizers built a pooled screening system with elegant internal bookkeeping. Each codon-optimized binder sequence was cloned into the CAR plasmid library, where the binder itself served as a unique DNA barcode. Using the GeneWeld targeted integration method, the library was knocked into the T cell receptor alpha chain locus of primary human T cells, generating a pool of TRAC-negative, CAR-positive cells. After methotrexate selection using a co-expressed DHFR variant, the cells were split: one arm cultured alone, the other challenged repeatedly with mitotically inactivated CD20-positive Raji tumor cells. Over two weeks, binders that enabled productive antigen recognition drove proliferation of their host cells, and next-generation sequencing of the barcode region revealed which designs expanded. Comparing tumor-challenged cultures against the no-target controls allowed antigen-dependent growth to be quantified robustly for every design simultaneously.

The screen produced a sobering attrition curve. Nearly 98 percent of designs passed DNA synthesis quality control, but only 56.8 percent were recovered at sufficient sequencing depth after growth, implying that many binders failed to yield a viable CAR, likely because of poor expression, folding, or membrane localization. Ultimately, 707 designs, just under 6 percent, showed significant CD20-specific enrichment of at least twofold. Hit rates varied dramatically between teams, from 0.6 percent to 38.4 percent, a spread that itself encodes lessons about which strategies work. The ten best non-redundant designs advanced to individual validation, where each was rebuilt as a standalone CAR construct and subjected to a full battery of functional assays against CD20-positive Raji cells and CD20-negative K562 controls.

Seven of the ten leading designs outperformed untransfected control T cells in proliferation, expansion, and cytokine production, with flow cytometry confirming CAR expression in most of them. Designs 1506, 5300, and 5981 were the strongest cytokine producers, generating as much interleukin-2 and interferon-gamma as the scFv positive control, and design 5300 exceeded it. When it came to the most demanding readout, specific killing of CD20-positive targets, only four designs, 1506, 2383, 3494, and 3718, achieved significant target-specific lysis above background, and two of those matched the scFv control in specificity. Notably, several broadly functional designs showed no measurable binding in isolation, suggesting they may have relied on avidity or multivalent mechanisms within the receptor context. The pooled hit rate of 707 thus appears to be an upper bound on true functional success.

Surface plasmon resonance against detergent-solubilized CD20 added a crucial biophysical dimension. The scFv control bound with a dissociation constant of 1.75 nanomolar, but only three of the top ten designs showed appreciable binding: design 2383 at 643 nanomolar, and designs 1506 and 5981 with weak but detectable signals between roughly 200 and 600 nanomolar. The authors caution that this apparent preference for moderate or weak affinity is consistent with CAR biology, where excessively tight binding can drive trogocytosis and T cell exhaustion, and lower-affinity receptors have sometimes eliminated tumors better. However, the binding assay tested isolated 80-mers produced by bacterial cell-free synthesis rather than full receptors in mammalian cells, so some failures may reflect structural context rather than design quality. The team recommends that future efforts fold candidate binders within the complete CAR sequence, a computationally trivial step that only one team actually performed, and that team produced the top-performing design.

The most consequential findings came from mining the entire dataset for predictors of success and failure. The organizers computed more than 400 features per sequence, from GC content and Shannon entropy to predicted secondary structure and interface energies. Strikingly, 98.9 percent of designs that vanished from the proliferation screen had been created with ProteinMPNN or SolubleMPNN, and the failed designs were almost pure alpha helices, averaging 83 percent helical content, densely packed with lysine and glutamate. The ratio of residues in lysine-plus-glutamate helices was by far the strongest predictor of recovery failure, reaching 0.91 ROC-AUC in a cross-validated logistic model, averaging 35 percent of residues in failed designs versus 15 percent in recovered ones. The likely mechanism is translational interference: glutamate repeats are known to stall ribosomes, and lysine and glutamate codons produce adenosine-rich DNA, with poly-A tracks also disrupting translation. Failed sequences carried nearly four glutamate-glutamate repeats on average versus 1.6 in recovered designs, and adenosine dinucleotide repeats predicted recovery with 0.88 ROC-AUC. An independent phage display study has since confirmed the lysine-glutamate association, suggesting the signal generalizes well beyond CAR-T cells.

Other failure modes emerged with equal clarity. Designs that caused CD20-specific depletion of T cells, possibly reflecting exhaustion, were enriched in cysteines predicted to form intramolecular disulfide bonds, which may have exposed hinge sulfhydryls and amplified tonic signaling. DNA synthesis failures clustered among sequences with high GC content and low entropy. Perhaps most humbling for the structure-prediction community, the standard confidence metrics, including ipTM, pLDDT, ipSAE, LIS, SAP, PDockQ, and language-model likelihoods, showed little predictive power for functional outcomes, and most teams’ own confidence rankings correlated poorly, or even negatively, with experimental success. Simple sequence statistics beat sophisticated structure-based scores. Applying retrospective filters for GC content, entropy, adenosine repeats, glutamate repeats, and lysine-glutamate helix content removed 4,600 sequences and nearly doubled the proportion of designs showing CD20-specific proliferation, from 5.9 to 7.6 percent, rising to 10.6 percent with a cysteine filter that would also have eliminated two of the non-functional top-ten designs.

The authors are careful about the study’s limits: a single CAR scaffold, a single donor isolate, a comparison of Raji-challenged cells against no-target rather than CD20-negative controls, and a design space constrained to 80 amino acids all shape the conclusions. No design is remotely clinic-ready, and all underperformed the scFv control in most readouts. Yet the significance of the work lies less in any single binder than in the demonstration that generative protein design can now be evaluated end-to-end in a therapeutic context, at scale, in the open. The dataset has been released on Zenodo with code on GitHub, and the practical recommendations, avoid lysine-glutamate helix overload, watch for cysteines, fold within the full receptor, and treat moderate affinity as a feature rather than a bug, offer the field its first empirically grounded playbook for designing proteins that do not merely bind, but function inside living cells.

Subject of Research: Experimental validation of AI-designed de novo protein binders as CAR-T cell recognition domains targeting CD20

Article Title: Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition

Article References: Kosonocky, C. W., Abel, A. M., Feller, A. L., Cifuentes Rieffer, A. E., Woolley, P. R., Lála, J., Barth, D. R., Gardner, T., Bits to Binders Competitors, Acosta, D., Angioletti-Uberti, S., Annapure, V., Anthony, N., Barghout, R. A., Beining, M., Bibekar, P., Biton, D., Bryant, P., Bushuiev, A., … Marcotte, E. M. (2026). Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition. Molecular Systems Biology. https://doi.org/10.1038/s44320-026-00246-1

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00246-1

Keywords: CAR-T cells, protein design, artificial intelligence, CD20, ProteinMPNN, RFdiffusion, de novo binders, pooled screening, protein engineering, immunotherapy, translational stalling, benchmarking