从生成式 AI 众包功能性细胞治疗结合域：Bits to Binders 竞赛的评测与启示

Over thirty years ago, structural biologists established the Critical Assessment of Structure Prediction (CASP) to rigorously assess methods for predicting protein structures from amino acid sequences. Participants submitted predictions before the corresponding experimental structures became publicly available (Moult et al, 1995). This framework established a common standard against which methodological advances could be measured. Its value became particularly evident at CASP14 in 2020, when AlphaFold2 produced predictions approaching experimental accuracy for many targets (Jumper et al, 2021; Pereira et al, 2021). CASP’s independent, prospective assessment enabled the community to recognize the scale of this advance by providing rapid, independent validation. The next generation of CASP-like open competitions is similarly poised to help discern the impact of emerging artificial intelligence (AI)-based methods across several domains of biology, from genomics to protein design to therapeutic discovery. The implementation and results of one such effort, the Bits to Binders competition, are described in this issue of Molecular Systems Biology, where Kosonocky, Abel and colleagues evaluated 12,000 designed proteins as recognition domains for chimeric antigen receptor (CAR) T cells (Kosonocky et al, 2026).

Introduction

While antibodies have historically been the primary format of therapeutic protein binders, emerging AI methods predominantly produce highly compacted proteins, commonly termed minibinders, that possess many desirable qualities, including well-folded structures, small size (often 65–80 amino acids), and programmable binding interfaces (Koh et al, 2025). Minibinders comprise alpha helices and beta sheets arranged in structures distinct from the antibody-derived single-chain variable fragments (scFvs) conventionally used for antigen recognition in CAR T cells (Fig. 1A). Though the ultimate therapeutic efficacy of minibinders has yet to be determined, their heterogeneous uses have been demonstrated in the academic literature, from biosensors (Quijano-Rubio et al, 2021) to viral neutralization (Cao et al, 2020) to recognition domains in cell therapies (Xia et al, 2024).

Early candidates from de novo protein design have largely been assessed on biophysical parameters, including predicted stability, solubility, and affinity for a chosen target (Chu et al, 2024). Community competitions, including challenges supported by Adaptyv Bio, have begun to compare design workflows through standardized measurements of protein expression and binding affinity. As the field moves toward repurposing binders as therapeutic biologics, these criteria become necessary but not sufficient, since each therapeutic modality carries additional constraints. Extending the framework to therapeutic applications requires benchmarks that establish whether molecular recognition translates into the intended biological response. Bits to Binders establishes an open community effort to assess the functional potential of AI-designed minibinders in CAR T cells. CARs repurpose antigen-binding domains to focus T cells toward specific target antigens expressed on tumor cell surfaces, including CD20, a surface protein expressed by normal B cells and many B-cell malignancies (Fig. 1A).

Twenty-eight teams submitted designs generated using diverse computational workflows. Each design was incorporated into a common CAR scaffold for evaluation in primary human T cells. Productive CAR function requires receptor expression, accessible antigen engagement and appropriate signaling. Recent characterization of de novo CARs identified signaling without a target, inaccessible binding epitopes and off-target activation as distinct constraints for efficacy beyond binding (Chow et al, 2026). These findings underscore the importance of evaluating designed binders in their intended cellular context. Here, the authors used each binder-encoding sequence to track its corresponding CAR T cells in a pooled proliferation screen (Fig. 1B). After growth, 6811 designs met the sequencing recovery threshold; 707 showed significant enrichment in cultures with CD20-positive lymphoma cells relative to cultures without them. 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 a need for further optimization. Only three showed validated binding in biochemical assays whose conditions differed from the cellular context. Nevertheless, retrospective analyses of the sequences that failed at each step nominate criteria for prospective testing, including screening DNA for high GC content and repetitive sequences, and limiting excessive enrichment of lysine and glutamate in helical structures.

While the overall workflow to aggregate and systematically assess candidate proteins reflects a critical advance, a few caveats signal opportunities to strengthen future competitions. Namely, the pooled screen assessed CAR behavior with or without lymphoma cells, but additional controls, such as isogenic knockouts, are needed to specifically tie activity to CD20. Testing a single receptor scaffold and a single T-cell donor also limits the generalization of sequence features in the binders. Moreover, differences in designed sequences reflect distinct workflows, including sequence selection and refinement, which limit attributing performance to any individual algorithm. Future competitions that expand experimental validation and invite submissions with annotations of designs testing specific pipeline components may better isolate principles for efficacious therapeutic design.

Looking forward, sustaining open competitions will require durable support for experimental testing, independent assessment, and data resources. Long supported by public funding, CASP faced lapses in public funding but secured a one-time private gift in 2025 to sustain the competition. As emerging competitions with more diverse endpoints, including Bits to Binders, expand the scope and promise of AI-enabled community benchmarking, open competitions will similarly require sustained investment in the infrastructure necessary to realize biological breakthroughs.

References

• Cao L, Goreshnik I, Coventry B, Case JB, Miller L, Kozodoy L, Chen RE, Carter L, Walls AC, Park Y-J et al (2020) De novo design of picomolar SARS-CoV-2 miniprotein inhibitors. Science 370:426–431

  Article  CAS  PubMed  PubMed Central  Google Scholar 

• Chow A, Chu H, Li R, Nalbant BN, Dozic AV, Kida LC, Tang Z, Palmeri JR, Lareau CA (2026) Sequence and structural determinants of efficacious de novo chimaeric antigen receptors. Nat Biomed Eng 1–16

• Chu AE, Lu T, Huang P-S (2024) Sparks of function by de novo protein design. Nat Biotechnol 42:203–215

  Article  CAS  PubMed  PubMed Central  Google Scholar 

• Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Žídek A, Potapenko A et al (2021) Highly accurate protein structure prediction with AlphaFold. Nature 596:583–589

  Article  CAS  PubMed  PubMed Central  Google Scholar 

• Koh HY, Zheng Y, Yang M, Arora R, Webb GI, Pan S, Li L, Church GM (2025) AI-driven protein design. Nat Rev Bioeng 3:1034–1056

  Article  CAS  Google Scholar 

• Kosonocky CW, Abel AM, Feller AL, Cifuentes Rieffer AE, Woolley PR, Lála J, Barth DR, Gardner T, Bits to Binders Competitors, Ekker SC et al (2026) Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competitions. Mol Syst Biol. https://doi.org/10.1038/s44320-026-00246-1

• Moult J, Pedersen JT, Judson R, Fidelis K (1995) A large-scale experiment to assess protein structure prediction methods. Proteins 23:ii–v

  Article  CAS  PubMed  Google Scholar 

• Pereira J, Simpkin AJ, Hartmann MD, Rigden DJ, Keegan RM, Lupas AN (2021) High-accuracy protein structure prediction in CASP14. Proteins 89:1687–1699

  Article  CAS  PubMed  Google Scholar 

• Quijano-Rubio A, Yeh H-W, Park J, Lee H, Langan RA, Boyken SE, Lajoie MJ, Cao L, Chow CM, Miranda MC et al (2021) De novo design of modular and tunable protein biosensors. Nature 591:482–487

  Article  CAS  PubMed  PubMed Central  Google Scholar 

• Xia Z, Jin Q, Long Z, He Y, Liu F, Sun C, Liao J, Wang C, Wang C, Zheng J et al (2024) Targeting overexpressed antigens in glioblastoma via CAR T cells with computationally designed high-affinity protein binders. Nat Biomed Eng 8:1634–1650

  Article  CAS  PubMed  Google Scholar 

Download references

Author information

Authors and Affiliations

1. Computational and Systems Biology Program, Memorial Sloan Kettering Cancer Center, New York, NY, 10065, USA

   Abdul Vehab Dozic & Caleb A Lareau

Authors

1. Abdul Vehab Dozic
2. Caleb A Lareau

Corresponding author

Correspondence to Caleb A Lareau.

Ethics declarations

CAL is a consultant for Cartography Biosciences and Neosplice Therapeutics. The remaining authors declare no competing interests.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. Creative Commons Public Domain Dedication waiver http://creativecommons.org/publicdomain/zero/1.0/ applies to the data associated with this article, unless otherwise stated in a credit line to the data, but does not extend to the graphical or creative elements of illustrations, charts, or figures. This waiver removes legal barriers to the re-use and mining of research data. According to standard scholarly practice, it is recommended to provide appropriate citation and attribution whenever technically possible.

Copyright: The Author(s)

Reprints and permissions

About this article

Cite this article

Dozic, A.V., Lareau, C.A. Crowdsourcing functional cell therapy binders from generative artificial intelligence. Mol Syst Biol (2026). https://doi.org/10.1038/s44320-026-00249-y

Download citation

• Received: 15 September 2026

• Accepted: 17 September 2026

• Published: 08 October 2026

• Version of record: 08 October 2026

• DOI: https://doi.org/10.1038/s44320-026-00249-y