AI antibody design passes its toughest blind test yet—but only for one task

The first large, blinded prospective benchmark of AI antibody design delivered a mixed verdict: artificial intelligence can generate experimentally validated, highly potent antibody candidates, but no single method performed consistently across every design task. Aureka Biotechnologies supplied the clearest success story, with a model-designed SARS-CoV-2 receptor-binding-domain antibody measuring 94.7 pM—roughly a 2,000-fold improvement over the parental antibody and statistically comparable to the best clone produced through conventional maturation. 16

What the AIntibody Challenge tested

The CASP-style AIntibody Challenge was designed to separate model predictions from experimental performance. Twenty-nine organizations submitted 511 antibody sequences for independent, standardized laboratory testing rather than evaluating their own designs. 12

The benchmark focused on antibodies against the SARS-CoV-2 spike receptor-binding domain and covered three discovery scenarios:

1. In-silico affinity maturation: improving a supplied parental antibody.
2. Affinity ranking: identifying the strongest binders within three constrained HCDR3 sequence clusters.
3. Out-of-library design: creating or optimizing CDR sequences outside the supplied experimental libraries. 1

The organizers measured binding with surface plasmon resonance and orthogonal KinExA testing. They also applied a five-assay developability panel, including tests intended to identify properties that can undermine an otherwise potent antibody. That meant entries were judged on laboratory behavior—not just on the scores reported by their models. 1

Aureka’s 94.7 pM result

Aureka’s strongest result came in the affinity-maturation task. Its winning antibody, listed in the paper as AuraBind and associated with Aureka’s AuraIDE platform, reached a measured affinity of 94.7 pM. The result was approximately 2,000 times better than the parental antibody. 146

The comparison with conventional experimentation is important. The best experimentally matured clone, produced after about three months of phage maturation, measured 113 pM. Aureka’s number was lower—and therefore numerically tighter—but the difference between the two top antibodies was not statistically decisive. The fairest interpretation is that the AI design matched the best experimental result in this test, rather than conclusively defeating it. 16

Aureka’s performance was broader than one headline sequence. The company produced six antibodies with affinities below 10 nM that also passed the challenge’s developability criteria, and its entries occupied first, second and fifth place in the affinity-maturation results. 1

Why the winning sequence was notable

The winning antibody was not simply a copy of an existing experimentally selected clone. According to the study summary, it recombined substitutions and mutational patterns inferred from the supplied selection data. A structure-informed affinity model selected the design from 10,000 generated candidates. 1

The platform also combined structural modeling with a preference-trained fitness predictor. Rather than trying to predict an absolute binding constant directly, the predictor learned relative experimental enrichment—an approach more closely aligned with ranking candidates for follow-up testing. 1

That distinction matters in antibody discovery. A useful design system does not necessarily need to know the exact final affinity of every sequence. It needs to improve the odds that a manageable number of candidates will contain strong, usable binders.

Aureka won one task, not the whole benchmark

Aureka was the clear winner of Challenge 1, but its result did not generalize into dominance across all three challenges. In the cluster-ranking task, the leading teams varied by cluster: UCSD/Scripps led for 27F, WashU/Xencor for 28F, and WashU for 47F. Xencor won the out-of-library design task. 1

The paper’s cross-challenge analysis likewise found that every disclosed group had at least one challenge in which its best result was far from the leading submissions. 1

This is the benchmark’s central message. AI antibody design is not one uniform capability. A system that is effective at improving a known antibody with rich target-specific data may be much less reliable when asked to rank clustered sequences or design outside an existing library.

What the result means for laboratory workflows

The strongest practical implication is for lead optimization, not for eliminating the laboratory. In a favorable setting, a capable model could reduce the number of cycles involving antibody-library construction, display selection, clone expression and screening. In Challenge 1, about 13% of submissions achieved at least a 20-fold improvement over the parental antibody, although performance varied substantially between participants. 1

The benchmark therefore supports a lab-in-the-loop workflow:

• use experimental selection and sequencing data to inform candidate generation;
• use models to prioritize a smaller and potentially more diverse set of sequences;
• express and test those sequences under blinded, standardized conditions;
• assess both affinity and developability before advancing candidates.

The final step remains essential. Some highly affine designs failed one or more developability tests, particularly polyreactivity-related checks. The 94.7 pM measurement is impressive, but binding strength alone does not make an antibody suitable for a drug program. 1

The biggest limitation: generalization

The study was a demanding prospective test, but it was also a relatively favorable environment for AI. It used one well-characterized antigen, extensive deep-sequencing and affinity data, and—especially for the second and third tasks—information that may not be available at the same stage of a real discovery campaign. 1

The results reflect that limitation. For the clustered HCDR3 datasets, all but one model showed poor or worse-than-random performance when predicting high-affinity clones. Out-of-library design results were also highly variable. 12

In other words, the benchmark does not establish that AI will produce equally strong results for unfamiliar antigens, sparse datasets, different antibody formats or the broader collection of properties required in therapeutic development. It shows that some systems can work extremely well in defined, biologically grounded regimes.

What the result means for Aureka

For Aureka, the benchmark provides unusually valuable validation because its platform was connected to an independently tested experimental outcome rather than only an internal model score. The company subsequently closed a $100 million Series B in August 2026, bringing reported cumulative funding to nearly $200 million. Granite Asia funded the first tranche, while a strategic investor led a later tranche. 1821

Aureka has also described business-development and NewCo-style collaborations with multiple multinational pharmaceutical companies. 4 Those partnerships and the new capital give the company resources and potential routes to translate model-generated antibody designs into drug programs. 418

But the benchmark is not clinical validation. It does not prove that Aureka’s platform will produce successful medicines, nor that it is broadly superior to experimental antibody-discovery approaches. The evidence supports a narrower—and more useful—claim: Aureka’s AI-assisted workflow can produce highly potent, developable antibody candidates in at least one well-defined optimization setting, while the field still needs broader blind tests across targets and discovery conditions.