A global AI contest designed 12,000 proteins against lymphoma: only 4 kill tumor cells in real tests

By Bioengineer.org · October 8, 2026. Twenty-eight teams from 42 countries went head to head in Bits to Binders, an open AI protein design contest that pushed the field to its toughest test yet: it was not enough for a protein to bind its target well in a simulation, it had to work inside a living human T cell. The result, published in Molecular Systems Biology, is a reality check for a discipline used to triumphant headlines: of 12,000 computer-designed proteins, only four managed to destroy tumour cells specifically in the lab.

The goal was to design the recognition domain of a CAR receptor (chimeric antigen receptor), the part an engineered T lymphocyte uses to locate and destroy cancer cells. In the CAR-T therapies already approved, that domain is an antibody fragment —the scFv, between 230 and 260 amino acids— taken from living organisms. The contest proposed replacing it with a "minibinder" generated entirely by AI, just 80 amino acids long, capable of recognising CD20, a surface protein characteristic of lymphomas.

The scale of the experiment is what sets it apart from earlier validations, almost always limited to a handful of candidates. The organisers cloned the 12,000 sequences into a library of CAR receptors, using each sequence as its own genetic barcode, integrated them into the genome of primary T lymphocytes and subjected the resulting cells to a two-week assault with CD20-bearing lymphoma cells (the Raji line). According to the study's own data, 98% of the designs passed DNA synthesis quality control, but only 56.8% were recovered with sufficient sequencing depth after culture, a sign that a huge fraction never even got expressed or folded inside the cell. From there emerged 707 designs (5.9%) with specific enrichment against CD20, with hit rates ranging from 0.6% to 38.4% depending on the team, a gap that by itself measures the distance between the best and worst current design practices.

Only ten of those 707 moved on to full individual validation, and there the filter became even harsher. Seven of the ten outperformed unmodified T cells in proliferation and in cytokine production (the chemical signals indicating that a T lymphocyte has been activated), and three of them —identified as 1506, 5300 and 5981— matched or even beat the reference antibody on that metric. But of the ten as a whole, only four also managed to kill tumour cells specifically above background noise, a function that did not always coincide with the best cytokine producers. Measured by surface plasmon resonance on the isolated binders, barely three of the ten showed appreciable binding to their target, and the best did so with an affinity hundreds of times weaker than the natural antibody; several designs that did work inside the cell showed no measurable binding on their own, suggesting they operate through avidity mechanisms that only emerge within the complete receptor.

The study's most uncomfortable finding, however, is not the low success rate —to be expected in such a young technology— but the inability of the AI's own confidence metrics to anticipate it. Scores such as ipTM or pLDDT (the indicators these models use to estimate the reliability of their own structural prediction) barely correlated with real functional success, and for several teams the correlation was outright negative. What did predict failure with remarkable accuracy —0.91 out of 1 in the authors' statistical analysis— was a far simpler sequence statistic: an excess of alpha helices loaded with lysine and glutamate, a pattern that, according to the team, interferes with the cell's translation machinery so that the protein never gets built. Put another way: the computer knew how to design well on paper, but not whether the cell would be able to build it.

As industry context, the spread of tools used by the 28 teams captures precisely the current state of AI protein design: generative diffusion models such as RFdiffusion, which creates new protein shapes out of random noise (used by 18 teams), programs such as ProteinMPNN to decide which specific amino acids hold that shape together, and AlphaFold2 to check that the final design folds as expected. One telling detail: the only team that also verified how its design folded inside the complete CAR receptor —and not just as an isolated molecule, a computationally trivial step— was the one that produced the best-performing binder in the entire contest. The good news is that the whole process, with the more than 400 sequence parameters analysed, has been published openly, on Zenodo and GitHub, a level of transparency about one's own failures that is unusual in biotech.

Our reading is that this study is worth more for what it dismantles than for what it achieves. At Zendoric we have long argued that you have to measure AI's real capability rather than trust its narrative, and here the same principle appears applied to biology: a flashy confidence score on a screen is not the same as a protein that works inside a living cell, just as a saturated benchmark is not the same as a genuinely safe language model. The difference is that here getting the metric wrong does not cost a failed funding round, but years of work on a drug candidate that was never going to work.

In the short term the message is one of caution: no design from this contest is anywhere near a clinical trial, the real hit rate is around 0.03% of everything tested, and validation is needed in other donors, other targets and other receptor scaffolds before an AI-generated binder reaches a patient. But in the long run, the experiment points in exactly the direction that strikes us as inevitable: a rational design of cancer therapies that today takes years and costs fortunes, and that with datasets like this one —large, open and honest about their own failures— can start to be compressed into months. The promise of an AI capable of helping eradicate disease will not be fulfilled by a single model that solves everything at once, but by this kind of methodical work: thousands of attempts, a handful of hits, and a reusable lesson about what to avoid next time.