Nature Biotechnology
The results of AIntibody, the first international AI antibody design competition, have been published in Nature Biotechnology, with Aureka Biotechnologies claiming first place in Challenge 1 through its proprietary model AuraIDE, cited in the publication as AuraBind.
The winning antibody achieved an affinity of 94.7 picomolar (pM) as measured by KinExA, representing an approximately 2,000-fold improvement over the parental antibody. This result surpassed the best experimental antibody obtained through traditional laboratory methods, which had an affinity of 113 pM. AuraIDE placed three entries among the top five positions in Challenge 1, with antibodies ranking first, second, and fifth.
A Rigorous Benchmark for AI Antibody Design
The AIntibody competition established a blinded benchmark similar to the CASP protein structure prediction challenge that validated AlphaFold2. All 511 submitted antibodies from 29 participating organizations were tested under identical conditions through independent wet-lab validation, eliminating variations in experimental conditions and selective reporting.
Challenge 1 focused on in-silico affinity maturation, requiring participants to design new antibodies based solely on next-generation sequencing (NGS) data from the first stage of affinity maturation of a parental antibody targeting the SARS-CoV-2 Spike protein receptor-binding domain. Teams had 14 days to submit up to 10 sequences, which were then expressed as full-length IgG antibodies and tested for affinity using surface plasmon resonance and KinExA.
Beyond affinity, each antibody had to pass five developability assessments covering hydrophobicity, polyreactivity, self-interaction, thermal stability, and aggregation propensity. The entire process remained blinded until the competition concluded, ensuring objective evaluation.
AI Design Replaces Experimental Cycles
The competition's structure simulated a critical step in antibody development. Conventional workflows would require researchers to combine the best mutations from separate CDR libraries into a combinatorial library and conduct additional wet-lab screening rounds. However, competing models saw only first-round NGS data and had to predict which mutation combinations would yield high-affinity, developable antibodies without access to combinatorial screening results.
Twenty-five organizations submitted 165 antibody designs to Challenge 1. AuraIDE's winning sequence differed from the parental antibody at 19 amino acid positions and varied by at least 12 positions from the most similar sequence in the experimental dataset, indicating the model explored sequence space beyond existing experimental data rather than simply recombining frequent mutations.
Aureka submitted six antibodies with affinities below 10 nanomolar that met developability criteria. The company also placed third in Challenge 2, which focused on ranking existing candidates by affinity, and ninth in Challenge 3, which required designing novel CDR combinations, making it the only team to finish in the top 10 across all three challenges.
Implications for Drug Discovery
The results demonstrate that AI models can now perform optimization work that previously required repeated rounds of library construction, screening, and experimentation. The winning antibody was designed in under a week, while the best experimental result took three months of phage maturation experiments to achieve.
Aureka Biotechnologies is an AI-native biotechnology company that has raised nearly $200 million. The company combines its AuraIDE foundation model with a proprietary single-cell functional screening platform and has established collaborations with several global pharmaceutical companies to develop antibody therapeutics for difficult targets including GPCRs and dual-target monoclonal antibodies.
