AuraBind

Affinity maturation normally takes two rounds of wet-lab work. The first screens separate libraries that vary one or two CDRs at a time, and sequencing reveals which substitutions were enriched in each. The second recombines those substitutions into a combinatorial library and screens again, because the useful combinations are not simply the most frequent residue at each position — enrichment in a single-CDR library says little about how that residue behaves alongside changes elsewhere in the paratope. AuraBind answers the second round computationally: from the first round's sequencing alone, it scores which cross-CDR combinations will bind more tightly while remaining developable.

AuraBind is the antibody design model of Aureka Biotechnologies, an AI-native TechBio company based in Laguna Hills, California and Shanghai. It is the system the company entered in AIntibody, the first prospective, blinded, wet-lab-validated benchmark of computational antibody discovery, published in Nature Biotechnology in August 2026. The paper's Methods name it AuraBind; Aureka calls the same system AuraIDE in its own announcements, presenting it as the proprietary counterpart to OpenDDE, the company's openly released co-folding model.

Architecturally, AuraBind extends Protenix, ByteDance's open reimplementation of the AlphaFold 3 design, from structure prediction into sequence design by grafting a fitness prediction adapter onto the folding trunk: one forward pass yields both a predicted antibody–antigen complex and a score for how tightly that pairing should bind.

#Key Features

• Structure-informed fitness scoring: Antibody and antigen sequences are encoded into shared pairwise representations capturing interchain geometry and epitope–paratope relationships, which the fitness head reads instead of sequence alone.
• Preference-optimized ranking: The adapter is trained with a direct preference optimization objective to reproduce the relative ordering of variants seen in high-throughput enrichment assays, rather than regressing absolute affinity values.
• Constrained generative sampling: For each design task the model samples 10,000 variants conditioned on the target antigen, with the parental framework and non-variable regions fixed and only the designated CDRs masked.
• Joint structure and confidence output: A diffusion module produces 3D coordinates alongside predicted local distance difference test and predicted alignment error metrics, used to filter training complexes by structural confidence.

#Technical Details

Training data combined the AIntibody sequencing datasets with Aureka's internal synthetic co-evolution datasets, which use different antibody frameworks and unrelated targets. Reads were collapsed to unique sequences per library and sort bin, aligned to parental frameworks under IMGT numbering, and discarded where framework residues deviated. The supervised label was log10 of read redundancy, standardized within library and round, with phase-aware masking and library-level train/validation/test separation.

In AIntibody Challenge 1 — in silico affinity maturation against the SARS-CoV-2 receptor-binding domain — 25 organizations submitted 165 designs, all expressed as full-length IgG and measured under uniform conditions by SPR and KinExA. AuraBind's best design measured 94.7 pM by KinExA, roughly a 2,000-fold improvement over the parental antibody and the only submission in the challenge to reach sub-100 pM. The best antibody from the withheld experimental combinatorial round measured 113 pM; the two are statistically indistinguishable, with overlapping 95% confidence intervals. Six of Aureka's designs were developable below 10 nM. The winning sequence sat 19 substitutions from the parental antibody across the concatenated CDRs and at least 12 from any experimental sequence — a wider departure than the other top entries, which stayed within 2–15 substitutions.

#Applications

The target use is antibody lead optimization: replacing the combinatorial library round of an affinity maturation campaign with a single in silico step, which the benchmark authors estimate would save two to three weeks. Aureka pairs the model with a single-cell functional screening platform and program-specific post-training, applying it to antibody classes that resist conventional discovery such as GPCR targets and dual-target bispecifics. The model is proprietary and not offered as a product; what is public is the challenge submission code under an MIT license and a single task-specific checkpoint archived on Zenodo under CC BY 4.0.

#Impact

AuraBind is one of the few antibody design systems validated prospectively and blind by third-party wet labs rather than retrospectively, and its result shows that a structure-aware model trained on enrichment rankings can match a full round of combinatorial screening on a well-characterized target. The benchmark's own conclusions bound that claim tightly. Success did not transfer across tasks: by the company's account Aureka placed third in the affinity-ranking challenge and ninth in out-of-library CDR design, and the paper reports that AI ranking strategies generally underperformed random clone picking. A consensus sequence derived by simple statistics, with no machine learning at all, ranked third by KinExA in the challenge AuraBind won. The RBD is among the most extensively characterized proteins in existence and participants received unusually deep sequencing data, so the authors read these results as an upper bound on current capability for this target class rather than evidence of general-purpose antibody design.