Anthropic's Claude Designs Drug-Binding Proteins at 35% Hit Rate, Crushing Human Experts

Designing a molecule that latches onto a specific protein target is one of the ugliest bottlenecks in drug discovery. It normally takes an expert weeks or months per target, sifting through thousands of candidates. Anthropic just handed that job to Claude, and the results have real teeth.

In a new campaign detailed in Anthropic's research post, Claude was given a protein design prompt and left to run autonomously. Claude (Mythos Preview and Opus 4.8) designed protein binders against 15 targets, and succeeded against 14 of them. The hit rates land well above what the field currently produces.

The numbers that matter

A quick primer: a minibinder is a small protein engineered to grab tightly onto a target protein. Binding is the mechanism behind a huge fraction of modern drugs, which either block, activate, or deliver payloads to their targets. Designing one from scratch is called de novo design, and until recently it was the domain of specialists running week-long computational pipelines.

The headline metrics from the campaign:

• Mythos Preview and Opus 4.8 achieve overall hit rates of 26.7% and 22.6% respectively when designing against all targets simultaneously in a 48-hour session
• Mythos Preview achieves an overall hit rate of 35.1% when designing against each target separately using multiple 24-hour sessions
• Industry baseline sits at 10 to 15% in typical protein design campaigns today
• Includes high-affinity binders (sub-10 nM dissociation constants) against at least six targets, and binders matching or exceeding the best reported affinity against at least four targets

Wet lab validation was outsourced to keep Anthropic honest. External evaluators Adaptyv Bio and Twist Bioscience independently produced and tested Claude's designs in the lab.

How the campaign was actually run

This was not a chatbot spitting out sequences. Claude was operating inside Claude Science as an agent, orchestrating a full pipeline of specialist tools. It chose where on each protein target to design against, generated candidate structures and sequences by orchestrating several structure design, sequence design, and co-folding models, ran the designs through multiple cycles of in silico optimization, and computationally screened for novel, diverse candidates that would express, stay soluble, and bind.

The compute envelope was serious. Opus 4.8 and Mythos Preview ran in multi-target mode with 48 hours of wall time and up to 12,500 NVIDIA H100 hours of compute for running specialized protein design and folding models. Mythos Preview also ran in single-target mode with 24 hours of wall time and up to 2,500 NVIDIA H100 hours per target.

The prompt itself is enormous. According to the open-sourced dataset on Hugging Face, it runs to roughly 30,000 tokens and is bundled with all the in silico and in vitro data. Beyond that prompt, Claude got internet access, a corpus of protein design papers, connectors for Google Drive and Slack, GPU access, and no token budget cap within the allotted time. After giving Claude the prompt, Anthropic left the model to execute autonomously, providing no additional scientific, technical, or operational guidance after initiating the campaigns.

Where Claude embarrassed the humans

One target stands out. RBX1 is a small protein used to drive targeted destruction of regulatory proteins, and Adaptyv Bio ran a public competition against it. Against RBX1, Mythos Preview in single-target mode achieved a 40% hit rate, compared to a 3.7% hit rate among participants. Its top-ranked design was a high-affinity binder that outperformed the winning design, which was among 245 designs entered.

The other headline win was TNFα, the inflammation signaling protein that anchors drugs like Humira. TNFα is a challenging target to design against because of its multimeric structure, which requires targeting a binding site in the groove formed by two proteins. Although Mythos Preview was unsuccessful, Opus 4.8 designed multiple binders, including some that worked across species, binding human, cynomolgus monkey, and mouse TNFα, which is important for conducting animal studies. Cross-species binding matters because it means the same molecule can be used in animal safety studies before human trials.

Claude also managed to produce binders with β-sheets, extended strands of amino acids that must line up side by side. These structures are harder to design and more prone to misfolding and aggregation. Claude designed 15 confirmed binders across six targets that contain at least 20% β-strand. Most computationally designed binders are α-helical bundles, so this is a meaningful expansion of what agentic design can cover.

Where it fell over

Two targets gave Claude trouble. BBF-14 is a β-barrel-shaped protein that does not exist in nature; it was itself de novo designed, and is now used as a benchmark for binder design precisely because of its novelty. Maltose binding protein (MBP) is a large, flexible bacterial protein with a smooth, water-loving surface that makes it a good lab reagent, leaving a binder very little to grab on to. Claude still managed to produce three independent BBF-14 binders with modest affinities, but against MBP none of the 90 designs was confirmed to have bound to the target.

One interesting failure mode: on TNFα, Mythos Preview (the newer model) failed while Opus 4.8 succeeded. Anthropic isn't sure why. Given the inherent complexity of protein design, it's unsurprising that specific areas exist where an overall less capable model could outperform a generally more capable one.

The other experiment: analytical chemistry in 25 minutes

Buried in the same post is a second result. Claude Opus 5 was pointed at raw analytical chemistry files, the kind that normally require vendor-specific software plus a human chemist. Supplied with only a contract lab's raw files for a routine quality-control sample and a short plain-language prompt, Claude returned processed NMR and LC-MS results in 23 and 19 minutes respectively, working in parallel. Its results matched the lab's own processing, with hydrogen counts within 0.08 ¹H and purity measured at 96.4% versus the lab's 96.33%.

The LC-MS piece is the more impressive engineering feat. The instrument files use an undocumented vendor format. Claude worked out how the data was encoded, then confirmed it had read the file correctly by reproducing the instrument's own recorded totals for all 2,664 scans before analyzing anything. Reverse-engineering a proprietary binary format on the fly, then validating the parse, is exactly the kind of tool-use scaffolding people have been trying to build explicitly.

Access, safety, and what it means

Here is the catch on availability. Life science research tasks are currently blocked in Anthropic's most capable model, and one of the company's highest priorities is to launch an access program for scientists. The protein design results were obtained with Mythos Preview and Opus 4.8, neither of which is broadly available. Opus 5 is the most capable generally available model right now, and the chemistry experiment used it directly.

The reason for the gate: agentic biological discovery is dual-use, and without robust safety measures such capabilities could enable bad actors to perform dangerous research, such as the development of bioweapons.

Anthropic has published the full technical report alongside the Hugging Face dataset containing prompts, computational models of the designed complexes, and experimental data. If you want to see what a 30,000-token protein design prompt looks like, or verify the hit rates yourself, everything is there under a CC BY 4.0 license.

The broader assumption that needs updating: de novo binder design has been framed as a specialist workflow requiring weeks of orchestration by computational biologists. That framing may not survive contact with general-purpose reasoning agents that can drive existing open-source protein design models autonomously. Claude did not invent new folding models. It just used the ones the field already has, better and faster than the humans typically do.