AI-Designed Protein Binders Face Real-World Test in Landmark CAR T Cell Competition

More than three decades ago, structural biologists created the Critical Assessment of Structure Prediction, or CASP, to rigorously evaluate methods for predicting protein structures from amino acid sequences. Participants submitted predictions before the corresponding experimental structures became publicly available, establishing a common standard against which methodological advances could be measured. The value of that framework became unmistakable at CASP14 in 2020, when AlphaFold2 produced predictions approaching experimental accuracy for many targets. Independent, prospective assessment allowed the scientific community to grasp the scale of the breakthrough almost immediately, because the results could not be explained by hindsight or selective reporting. According to a News and Views commentary published in Molecular Systems Biology by Abdul Vehab Dozic and Caleb A Lareau of Memorial Sloan Kettering Cancer Center, the next generation of CASP-like open competitions is now poised to do the same for emerging artificial intelligence methods across biology, from genomics to protein design to therapeutic discovery.

One such effort, described in the same issue of the journal, is the Bits to Binders competition, in which Kosonocky, Abel and colleagues evaluated roughly 12,000 designed proteins as recognition domains for chimeric antigen receptor, or CAR, T cells. The scale of the undertaking is remarkable: twenty-eight teams submitted designs generated using diverse computational workflows, and each design was inserted into a common CAR scaffold for evaluation in primary human T cells. This design choice matters enormously, because it isolates the variable of interest, the binder itself, from the many other components that determine whether a cell therapy works. By holding the scaffold constant, the organizers could ask a deceptively simple question with profound implications: which AI-generated proteins can actually redirect a living immune cell to kill a cancer cell?

The molecular players at the center of the competition are minibinders, highly compact proteins that generative AI methods produce with increasing reliability. Unlike the antibody-derived single-chain variable fragments, or scFvs, conventionally used for antigen recognition in CAR T cells, minibinders typically comprise alpha helices and beta sheets arranged in small structures of roughly 65 to 80 amino acids, with well-folded architectures and programmable binding interfaces. While antibodies have historically been the dominant format for therapeutic protein binders, minibinders have already demonstrated heterogeneous uses in the academic literature, ranging from biosensors to viral neutralization, including picomolar inhibitors of SARS-CoV-2 designed de novo, to recognition domains in cell therapies targeting antigens in glioblastoma. Their ultimate therapeutic efficacy, however, has yet to be determined, which is precisely why a functional benchmark matters.

Earlier candidates from de novo protein design were largely assessed on biophysical parameters such as predicted stability, solubility, and affinity for a chosen target. Community competitions, including challenges supported by the company Adaptyv Bio, have begun to compare design workflows through standardized measurements of protein expression and binding affinity. But as the field moves toward repurposing these binders as therapeutic biologics, such criteria become necessary yet insufficient, because each therapeutic modality carries additional constraints. Extending the benchmarking framework to therapeutic applications requires tests that establish whether molecular recognition actually translates into the intended biological response. Bits to Binders was built to answer exactly that question, establishing an open community effort to assess the functional potential of AI-designed minibinders inside CAR T cells, the context in which they would ultimately need to perform.

The biological target chosen for the competition was CD20, a surface protein expressed by normal B cells and many B-cell malignancies. CARs work by repurposing antigen-binding domains to focus T cells toward specific target antigens on tumor cell surfaces, and the competition used the well-characterized antibody rituximab as a benchmark scFv alongside the de novo designs. Productive CAR function, however, demands far more than binding. A receptor must be expressed on the T cell surface, its binding epitope must be accessible on the target cell, and it must trigger appropriate signaling. Recent characterization of de novo CARs identified signaling without a target, inaccessible binding epitopes, and off-target activation as distinct constraints on efficacy that go well beyond simple binding, underscoring the importance of evaluating designed binders in their intended cellular context rather than in a test tube.

To capture these complexities at scale, the organizers used each binder-encoding sequence as a molecular barcode to track its corresponding CAR T cells in a pooled proliferation screen. After growth, 6,811 designs met the sequencing recovery threshold, and 707 of those showed significant enrichment in cultures containing CD20-positive lymphoma cells relative to cultures without them. The funnel then narrowed sharply. Individual testing of ten leading designs confirmed increased proliferation, expansion, and cytokine production for seven compared with unmodified T cells. Four showed significant selective killing in comparisons involving CD20-positive and CD20-negative cell lines. The designs generally underperformed the antibody-derived CAR control, indicating that further optimization remains necessary before AI-generated minibinders can match the refined binding domains that decades of antibody engineering have produced.

Perhaps the most sobering finding came from the biochemical side of the validation. Only three of the designs showed validated binding in biochemical assays, whose conditions differed from the cellular context of the proliferation screen. That disconnect between cell-based activity and cell-free binding measurements is itself an important result, because it suggests that computational predictions of affinity do not reliably predict function inside a living cell. Nevertheless, the retrospective analyses of sequences that failed at each step of the pipeline nominated concrete criteria for prospective testing, including screening DNA sequences for high GC content and repetitive elements, and limiting excessive enrichment of lysine and glutamate residues in helical structures. These are exactly the kind of actionable design rules that open competitions are uniquely positioned to generate, since no single laboratory could test such a diverse portfolio of approaches on its own.

The commentary authors are careful to note several caveats that signal opportunities to strengthen future competitions. The pooled screen assessed CAR behavior with or without lymphoma cells, but additional controls, such as isogenic knockout cell lines, would be needed to specifically tie activity to CD20 rather than to some other feature of the tumor cells. Testing a single receptor scaffold and a single T-cell donor also limits how far the sequence features identified in the binders can be generalized. Moreover, differences in the designed sequences reflect distinct workflows, including sequence selection and refinement steps, which makes it difficult to attribute performance to any individual algorithm. Future competitions that expand experimental validation and invite submissions annotated with the specific pipeline components being tested may better isolate the principles that separate efficacious therapeutic designs from expensive failures.

Looking forward, sustaining open competitions of this kind will require durable support for experimental testing, independent assessment, and shared data resources. The history of CASP offers a cautionary tale: despite decades of public funding, the competition faced lapses in support and secured a one-time private gift only in 2025 to keep running. As emerging competitions with more diverse endpoints, including Bits to Binders, expand the scope and promise of AI-enabled community benchmarking, they will similarly require sustained investment in the infrastructure necessary to realize biological breakthroughs. The stakes are considerable. If generative AI can reliably produce functional therapeutic binders, the design-build-test cycle for cell therapies could compress from years to months. Bits to Binders has now shown, with unusual transparency, both how far the technology has come and how much distance remains between a computationally elegant protein and a medicine that works in a patient.

Subject of Research: Benchmarking AI-designed minibinders as functional recognition domains in CAR T cell therapy through an open competition

Article Title: Crowdsourcing functional cell therapy binders from generative artificial intelligence

Article References: Dozic, A. V., & Lareau, C. A. (2026). Crowdsourcing functional cell therapy binders from generative artificial intelligence. Molecular Systems Biology. https://doi.org/10.1038/s44320-026-00249-y

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00249-y

Keywords: generative AI, protein design, minibinders, CAR T cells, Bits to Binders, CASP, CD20, immunotherapy, protein engineering, benchmarking, Molecular Systems Biology, computational biology