Live·Open questions in longevity research
All news
AI in medicineScientific ComputingScience Research

Retesting confirmed binding for 29 of 107 AI-designed nanobodies reported as binders in published studies

19 November 2026· 261119005

Retesting confirmed binding for 29 of 107 AI-designed nanobodies reported as binders in published studies

A nanobody is a miniature antibody that binds to a target protein. On November 17, the Shanghai company Spiritein published a report describing how it retested nanobodies reported as binders in published studies using a single standardized procedure. The company then designed its own nanobodies against two target proteins: nine of 326 candidates bound to their target.

Spiritein sponsors the “Yangtze Cup”, an AI protein design competition in Shanghai where 7000 AI-generated sequences will undergo laboratory testing free of charge. It also runs AbGround, a platform for comparing AI models, where it has pledged to release the code and raw data.

Binding was detected for 29 of 107 nanobodies (27,1%). The remaining 78 showed no binding, although some had been described as strong binders in the papers: 77 purified samples produced no signal even at 2 μM, and one could not be purified. Confirmation rates varied by AI model: binding was confirmed for 1 of 12 from Germinal, 15 of 76 from PPIFlow, 5 of 8 from BoltzGen, 5 of 6 from Latent-X2, 1 of 3 from Latent-Y, and 2 of 2 from IgGM. Six of the 29 nanobodies form large aggregates: four from PPIFlow and both from IgGM. In a blog post, Stanford graduate student Fang Wu recommends separate tests to establish whether these samples actually bind to the target or simply aggregate.

Germinal is a model from Arc Institute whose authors obtained binders against all four targets. “Germinal's result is particularly poor: retesting detected binding for only one of 12 published binders. This result needs an explanation,” Fang Wu wrote on X. He also notes that another experiment could yield a different result because of differences in the protein construct and how it is produced.

To compare the models on equal terms, the company selected a single target, IL-7Rα, a receptor on immune cells. Design proceeded in two steps: one of four open models generated the nanobody structure, and one of three methods selected an amino acid sequence to fit it. BoltzGen, one of the four models, can also perform sequence design itself. This produced 13 combinations and about 960 thousand nanobody designs. A shared computational filter selected up to 20 candidates from each combination with the highest confidence in their predicted fit to the target, and 221 nanobodies were sent to an external laboratory. Nine bound to the target (4,1%): eight from BoltzGen, with 4 of 20 for each of two sequence design methods, and one from PPIFlow. BoltzGen yielded 0 of 20 with each of the remaining options, the third sequence design method and its built-in method.

For the second target, the company chose the well-studied PD-L1, a protein that helps tumors suppress immune T-cells. According to the report, more than 20 approved drugs target this protein or its partner, PD-1. Four models produced 105 candidates, of which 100 could be purified, but none bound to the target. BoltzGen, which had produced eight of the nine binders against IL-7Rα, yielded 0 of 40. Germinal yielded 0 of 20, whereas 7 of 101 variants bound to PD-L1 in the Nature Biotechnology paper, according to the report's summary.

Across four completed open competitions run by Adaptyv Bio, where one laboratory tests proteins from different designers using the same protocols, binder rates ranged from 2,5% to 14%. The 4,1% rate for IL-7Rα falls within that range.

Spiritein interprets the zero-binder results as evidence of a broader shortcoming in the models: they still struggle to reliably generate nanobodies that bind and to identify them before synthesis. Fang Wu recommends that papers report the number of candidates at every selection stage. In his words, “the label ‘experimentally validated’ needs to be accompanied by the experimental conditions and results for each candidate.”

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#ai-protein-design#nanobodies#binder-reproducibility#boltzgen#il-7rα#pd-l1