Summary
- Claude was given only information about a target protein and, with no human involvement, designed binding proteins over a 24-hour period
- Two labs, Adaptyv Bio and GenScript, independently synthesized and measured the results, and nearly half of the 12 targets showed actual binding
- Compared with an August experiment that used human prompting (14 designs across 15 targets, 22-35% success rate), this run matched similar results without any human input
Anthropic's AI model Claude designed new binding proteins within 24 hours, working entirely on its own after being given nothing but a target protein, and two independent labs confirmed that nearly half of them actually attached to their targets. The key point here is that a human specified only the goal, and Claude handled the entire rest of the design process by itself.
To put this in context, Anthropic reported back in August that Claude, working from prompts written by human experts, designed binding molecules for 14 out of 15 target proteins, and that Adaptyv Bio and Twist Bioscience verified a 22-35% actual binding rate. This time is different: humans supplied only the target, and Claude carried out the entire design process alone over 24 hours. The verification lineup also changed, with GenScript joining as a new testing partner.
What was done, and how
Once given a single target protein, Claude repeatedly generated candidate proteins using open-source biomolecular models, predicted their folded structures, and scored them—all on its own. It then used bioinformatics tools to check the diversity and novelty of the candidate pool before selecting the designs most likely to bind the target. The only human input was naming the target; Claude handled the rest of the design process alone over the full 24 hours.
Two labs split the verification work
The designs Claude selected were actually synthesized and measured in two separate labs, Adaptyv Bio and GenScript. Among the 12 disclosed targets, nearly half of Claude's designed proteins showed actual binding activity. This verification step matters because a design that looks convincing inside a computer simulation and one that actually works on the lab bench are two very different things.
How it compares with last month's results
The August experiment covered in Claude Beats Industry Average in Binding-Protein Design for Drug Discovery relied on human experts steering the process through prompts. Placing the two experiments side by side makes clear what changed.
| Item | August experiment (expert-prompted) | This experiment (fully autonomous) |
|---|---|---|
| Number of targets | 15 | 12 |
| Verification labs | Adaptyv Bio, Twist Bioscience | Adaptyv Bio, GenScript |
| Binding success rate | 22-35% | Nearly half |
| Human involvement | Provided expert prompts | Specified target only |
Why this result matters
One of the toughest steps in finding a drug candidate is identifying a binding protein that actually attaches to its target. No matter how precisely computer simulations predict folding structures and binding strength, the rate at which those predictions hold up in real experiments tends to be low—industry norms have generally put success around 10-15%. Last month's experiment, guided by human experts, reached a 22-35% success rate. This time, without any human guidance, that figure climbed to nearly half. That shift is what makes this result notable.
Editor's take
The short version is that AI pulled off a near-production-grade result in the early design stage of drug discovery without human guidance. The August experiment succeeded because human experts steered it through prompts; this time, Claude was handed a single target and left alone for 24 hours, and it still produced a comparable binding rate. The fact that this gap has narrowed lines up with Anthropic's apparent push to automate the entire drug discovery pipeline, extending eventually to antibodies and small-molecule compounds.
Companies that have tried putting open-source protein language models into real practice keep running into the same wall: simulation scores look promising, but the results often don't hold up in actual binding experiments. That's exactly why the verification process here—two separate labs, Adaptyv Bio and GenScript, checking the results independently—deserves more attention than the headline number itself. An in silico result that can't be reproduced ends up as a single paper and nothing more, but a result confirmed independently in two places gives you actual grounds to move to the next stage.
For biotech and pharma teams in Korea, the practical move right now probably isn't adopting a fully autonomous pipeline like this outright, but rather building out collaboration between open-source biomolecular models and in-house verification labs first. The faster AI accelerates design output, the more wet-lab verification capacity becomes the bottleneck. Over the coming weeks, it seems likely that Anthropic will extend this autonomous design approach to other molecule types, including antibodies and small-molecule compounds.
Comments