Computational antibody design has accumulated an enormous body of claims, with conference presentations, preprints, and company white papers describing model after model capable of producing high-affinity binders at speeds no wet-lab pipeline can match. The field has lacked a systematic, prospective test of those claims: one with every organization's predictions measured under identical conditions by an independent laboratory. The AIntibody Design Challenge addressed that gap, and its final results, published in Nature Biotechnology, represent the most rigorous head-to-head evaluation of AI-driven antibody discovery to date.
Zymo Research entered the competition against 28 other organizations, spanning academic laboratories, large pharmaceutical companies, and AI biotechnology startups, and ranked among the top four teams in the affinity maturation task. The company's computational protein engineering platform generated antibody variants with up to 75-fold improvements in binding affinity over their parent molecules and produced candidates with picomolar-level binding affinities, according to company-reported results.
AIntibody challenge design: three tasks, one antigen
The AIntibody challenge structured its evaluation around three distinct computational tasks applied to a single, well-characterized antigen: the SARS-CoV-2 receptor-binding domain (RBD). Inspired by the Critical Assessment of Structure Prediction (CASP) framework, the competition tested 511 artificial intelligence (AI)-designed or predicted antibodies from 29 organizations across in silico affinity maturation from phase 1 sequencing outputs, affinity ranking within heavy-chain complementarity-determining region 3 (HCDR3) clusters, and complementarity-determining region (CDR) design using proteins absent from the selection output.
Each participating team submitted up to 10 antibody sequences per task, and the organizers synthesized those sequences as full-length immunoglobulin G (IgG) molecules measured under standardized conditions in an independent laboratory. SPR and KinExA quantified binding; five developability assays covering hydrophobicity, polyreactivity, self-interaction, thermal stability, and aggregation added a multiparametric filter that retrospective benchmarks, which score models against pre-existing datasets, cannot replicate.
Challenge 1, the affinity maturation task, asked teams to improve RBD binding using next-generation sequencing (NGS) data from only the first stage of a parental antibody's maturation campaign. Across 165 submissions from 25 organizations, only 13 percent of entries achieved a 20-fold or better improvement in binding affinity, and 63 percent produced developable binders. The best AI entry and the best conventional experimental comparator showed statistically indistinguishable results, which the organizers cited as evidence that AI-driven optimization can compete with bench-based methods under the right data conditions.
What did Zymo Research's platform achieve?
Zymo Research's entry produced antibody variants with up to 75-fold binding improvement over the parental molecule; several candidates reached picomolar-range affinities, placing the company among the top four teams in Challenge 1. The company did not disclose sequence or structural details of the submitted antibodies, and the algorithmic basis for its platform remains proprietary.
Key parameters from Zymo Research's Challenge 1 submission, as reported by the company:
- Binding affinity improvement: Up to 75-fold over the parental antibody, with several candidates reaching picomolar range
- Ranking: Top four among 25 organizations in the affinity maturation task
- Target antigen: SARS-CoV-2 RBD, applied uniformly across all Challenge 1 submissions
- Platform: AI-driven computational protein engineering; specific algorithmic details not disclosed
- Prior product lines: DNA and RNA purification reagents, epigenetics tools, and NGS sample preparation; Challenge 1 is the company's first publicly benchmarked protein design result
"Achieving top-tier performance in a blinded, experimentally validated challenge is an important demonstration of what our computational protein engineering platform can deliver," said Larry Jia, Founder and CEO of Zymo Research.
AI affinity maturation is competitive, but CDR ranking and design lag behind
AI-driven affinity maturation matched conventional experimental pipelines in Challenge 1, but the ranking and de novo design tasks exposed clear limits in current computational approaches. Several teams demonstrated that the optimization task can approach bench-based performance, which matters because it is routine, resource-intensive, and a credible target for computational acceleration in drug discovery programs.
Challenge 2 tested HCDR3 affinity ranking across 176 submissions from 26 organizations. Most models performed below the threshold set by randomly picking clones from the clustered dataset. Only one group (WashU) exceeded the 39 percent random-picking baseline, at 50 percent; a purely consensus-based approach that selected the most common amino acid at each CDR position without machine learning placed third by KinExA in Challenge 1, which the paper cited to argue that AI models require benchmarking against strong statistical controls.
The out-of-library CDR design task drew 168 submissions from 23 organizations; the best developable entry reached 8.69 pM, but 30 percent of submissions failed to bind the target at all. Performance did not transfer consistently across organizations or tasks, reinforcing the paper's conclusion that current AI platforms are narrowly optimized rather than generally capable.
Andrew Bradbury, MD, PhD, Chief Scientific Officer at Specifica (an IQVIA business), led the international AIntibody Challenge alongside first author M. Frank Erasmus. "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. Built upon a real-world in vitro antibody discovery pipeline, and by comparing AI predictions and designs under blinded experimental conditions, the challenge provided an objective benchmark to judge the effectiveness of AI in antibody discovery that should accelerate the development of more reliable and effective AI approaches in future therapeutic discovery," Bradbury said.
The following table summarizes all three tasks:
| Task | Submissions | Key result | Notable finding | Zymo Research result |
|---|---|---|---|---|
| Challenge 1: in silico affinity maturation | 165 from 25 organizations | Several developable antibodies at <100 pM; 63% developable binders | Best AI entry statistically indistinguishable from best experimental result | Top-four finish; up to 75-fold affinity improvement; picomolar candidates |
| Challenge 2: HCDR3 affinity ranking | 176 from 26 organizations across three clusters | 9.8-13.8% of AI submissions outperformed the cluster control | Only WashU exceeded the 39% random-picking baseline | Not reported |
| Challenge 3: CDR de novo design | 168 from 23 organizations | Best developable entry: 8.69 pM; 30% nonbinders | Performance varied widely and did not transfer across tasks | Not reported |
The organizers chose the SARS-CoV-2 RBD specifically because abundant structural and sequence data would give AI models the best possible conditions. Future iterations will test whether those results hold against targets where training data are limited.
Unresolved questions heading into the next AIntibody iteration
Cross-task generalization and performance on targets with limited training data represent the two core gaps the inaugural competition identified, and both are primary targets for the planned follow-on challenge. Zymo Research has not announced whether its Challenge 1 capability will become a commercial service or the basis for biopharma partnerships; the company has said it continues to invest in AI-guided protein engineering infrastructure.
The AIntibody organizers plan a follow-on challenge that requires participants to work from sequencing data alone, without affinity labels, and will apply stricter developability criteria that disqualify any candidate failing a single assay. The Nature Biotechnology paper frames the inaugural competition as a first step toward CASP-caliber benchmarking for antibody discovery, with future rounds testing whether AI optimization advantages extend beyond the SARS-CoV-2 RBD to antigens with sparser training data. Whether top-performing platforms from the inaugural round can maintain their rankings under those more restrictive conditions is the defining question the next iteration will answer.
This article is based on a press release issued by Zymo Research and was produced under Drug Discovery News' AI Editorial Guidelines.
