---
title: An AI-Designed Antibody Matched Months of Lab Work in the First Blinded Test of Its Kind
description: In the first blinded international contest, an AI-designed antibody matched three months of lab work. Field-wide, AI has not yet beaten the bench.
author: Darie Nani (Editor-in-Chief)
date: 2026-08-27T10:30:00.000Z
updated: 2026-08-27T11:22:01.848Z
canonical: https://www.sovereignmagazine.com/article/aintibody-ai-antibody-design-nature-biotech-212438
image: https://cdn.nanimediahouse.com/aintibody-ai-antibody-design-212438.webp
categories: Artificial Intelligence
content_type: Analysis
region: Global
publication: Sovereign Magazine
schema_type: Article
---

In the first international competition to test AI antibody design against laboratory methods under identical, blinded wet-lab conditions, a computer-designed antibody matched and slightly beat the best result that about three months of bench work produced. The full findings were [published on Aug. 26 in Nature Biotechnology](https://www.nature.com/articles/s41587-026-03238-6), and they mark the clearest head-to-head evidence yet on where AI-designed antibody discovery actually stands.

The contest, called AIntibody, tested 511 antibody designs from 29 organizations. Each team had 14 days per task and could submit up to 10 sequences, all of them then made and measured under the same conditions so no group could grade its own homework. Andrew Bradbury and Frank Erasmus at Specifica, an IQVIA business, designed the competition and laid out its structure in a 2024 Nature Biotechnology paper before the results were run. It covered three tasks: improving an existing antibody's binding strength, ranking candidates by affinity, and designing the binding loops from scratch.

## Aureka's antibody narrowly beat three months of phage-display work

The standout number came from the affinity maturation task, where the goal was to take an existing antibody against the SARS-CoV-2 receptor-binding domain and make it bind more tightly. Aureka Biotechnologies took first, second and fifth place. Its best antibody measured 94.7 picomolar affinity, roughly a 2,000-fold improvement over the antibody it started from.

The strongest antibody produced by conventional laboratory work in that task measured 113 picomolar and had taken about three months of phage-display maturation to reach. The two results are close enough to count as a statistical tie, with the AI design numerically slightly ahead. Aureka also reported six developable antibodies binding below 10 nanomolar that passed the contest's developability screen across five measures, including thermal stability and aggregation. The company built the designs with an in-house foundation model, AuraIDE, trained on protein co-evolution data, and maintains an open-source model called OpenDDE.

Aureka says it has raised close to $200 million to date, including a $100 million Series B in August 2026, on the strength of pharmaceutical partnerships and revenue from live drug-discovery programs.

## Across the whole contest, AI did not beat the bench

The broader picture is more sober. Independent analysis of the full results found that, taken across all three tasks, computational design did not outperform experimental methods on a field-wide basis. Performance varied widely from team to team and task to task, and no single approach dominated. A group that excelled at one task tended not to carry that strength to the others, so success did not transfer.

## AI-designed antibodies are already being tested in people

AI can already design a working antibody, and those molecules have moved past benchmarks into patients. Generate Biomedicines' GB-0895, aimed at severe asthma, is in Phase 3 with two global trials enrolling roughly 1,600 patients. Absci has two AI-designed antibodies, ABS-101 and ABS-201, in Phase 1/2a with early safety data it describes as positive.

No AI-designed antibody has yet been cleared by the U.S. Food and Drug Administration, and every candidate remains in trials.

## The clinical trials still take the same years and money

Most of the cost and time in drug development comes after the design stage, in the years of trials that turn a binder into an approved medicine. Independent benchmarking using the FLAb2 antibody dataset found that AI models fail to correlate with about 80 percent of the developability properties that determine whether an antibody can become a medicine, and that the models lose roughly 40 percent of their apparent predictive power once you account for how much they simply echo antibodies already in their training data. Immunogenicity and aggregation remain persistent failure points.

The clinical numbers point the same way. For AI-discovered drugs generally, Phase 1 success runs high, around 80 to 90 percent. Phase 2 success drops to about 40 percent, the same rate the industry has posted for years. The pattern suggests AI is compressing the design stage, the fast part, while the slow, expensive work of proving a drug safe and effective in people stays exactly as slow as it always was.

## FAQ

**Q: What is the difference between maturing an antibody and designing one from scratch?**
AIntibody ran both as separate tasks. Affinity maturation starts from an antibody that already binds and improves its grip, which is what Aureka won. De novo design builds the binding loops from nothing, a harder problem where results across teams were weaker and less consistent.

**Q: What does developability mean for an antibody?**
It covers the practical properties that decide whether a binder can survive being made into a drug: thermal stability, resistance to clumping, low stickiness, and manufacturability. A tight-binding antibody that scores badly on these is unlikely to become a medicine, which is why the contest screened for them.

**Q: Why does it matter that every entry was tested blind under identical conditions?**
Most AI antibody claims come from the company that built the model, testing its own designs in its own way. AIntibody made and measured every entry in the same lab under the same protocol, with no team able to grade its own work, so the numbers can be compared directly. That is what makes it the first clean head-to-head between AI design and the bench.

**Q: Why do so many candidate drugs still fail in Phase 2 even when AI designs them?**
AI speeds up the design stage, where a molecule is invented and optimized. Phase 2 tests whether the drug actually works in patients, a question no design tool answers. That is why AI-discovered drugs pass early safety trials at high rates but still fail Phase 2 at about the historical average.

**Q: What does an antibody's binding affinity in picomolar mean?**
Affinity measures how tightly an antibody grips its target, and a lower number means a tighter grip. Picomolar binding is very strong. A 94.7 picomolar antibody binds more tightly than a 113 picomolar one, which is why the contest's top result counted as edging ahead of the laboratory-made version.