AI-designed antibody outperforms lab results in first blinded global benchmark

An AI-designed antibody has outperformed the best result from months of traditional laboratory experiments, according to the results of the first international blinded benchmark for AI antibody design, published in Nature Biotechnology on August 25, 2026.

Aureka Biotechnologies, a California- and Shanghai-based AI drug discovery company, won the top spot in the AIntibody Challenge with an antibody designed by its foundation model AuraIDE that achieved a binding affinity of 94.7 picomolar — an approximately 2,000-fold improvement over the parental antibody and tighter than the 113 picomolar achieved by the best experimentally derived clone from three months of phage maturation.prnewswire+1

A CASP-Style Test for Antibody AI

The AIntibody Challenge, organized by Andrew Bradbury and M. Frank Erasmus, attracted 29 organizations that submitted 511 AI-designed or AI-predicted antibodies across three tasks. Modeled after the CASP competitions that validated AlphaFold's protein structure predictions, the benchmark aimed to provide the first objective, prospective evaluation of AI in antibody discovery.aintibody+2

In Challenge 1 — deemed the most practically relevant task — teams received only first-round NGS sequencing data from a parental antibody's affinity maturation against the SARS-CoV-2 spike protein's receptor-binding domain. They then had 14 days to design up to 10 new antibody sequences, which organizers independently expressed and tested under uniform blinded conditions.prnewswire

"The AIntibody Challenge was established to provide the first rigorous prospective evaluation of the ability of artificial intelligence to improve antibody discovery using independent third-party validation," Bradbury said.phys

Beyond Recombination

AuraIDE's winning sequence was not a simple recombination of mutations already present in the experimental data. It differed from the parental antibody at 19 amino acid positions and from the nearest sequence in the dataset by at least 12 amino acids. The model placed first, second, and fifth on the Challenge 1 leaderboard and produced six antibodies below 10 nanomolar affinity that also passed all five developability assessments.prnewswire

Aureka also placed third in Challenge 2 (ranking existing candidates by affinity) and ninth in Challenge 3 (designing novel CDR combinations), making it the only team to finish in the top 10 across all three tasks.prnewswire

Implications for Drug Discovery

The results suggest that AI can now replace at least one round of combinatorial library construction and wet-lab screening in antibody optimization — a process that typically requires months. Still, the AIntibody organizers noted that success did not consistently transfer across tasks and that "broad, reliable generalization remains an important challenge for the field".aintibody

Aureka, which closed a $100 million Series B round earlier in August, has raised nearly $200 million to date and maintains collaborations with several global pharmaceutical companies.biospace+1