Claude Designs Proteins: Anthropic's Wet-Lab Validation Lands at 35% Hit Rate

Yesterday, Anthropic published something rare in the AI-biology crossover: a paper with wet-lab results. Not simulations. Not in-silico benchmarks. Actual proteins synthesized, expressed, and tested in a lab by Adaptyv Bio and Twist Bioscience.

The headline: Claude Mythos Preview designed protein binders against 15 targets and succeeded against 14 of them. Between 22% and 35% of its designs bound successfully—roughly double the typical 10–15% hit rate in protein design campaigns today. Some designs bound several times more tightly than the best previously published results.

This isn't a paper about protein design algorithms. It's a paper about whether an LLM can orchestrate the protein design workflow end-to-end, with minimal human input. That distinction matters.

The Setup: 15 Targets, One Prompt, Zero Human Tuning

Anthropic gave Claude one prompt (~30,000 tokens describing protein design methodology), access to open-source protein design and folding models (RFdiffusion, ProteinMPNN, AlphaFold2, ESMFold), internet search, GPUs, and connectors to Google Drive and BioRxiv. Then they walked away.

Claude operated autonomously in Claude Science sessions: 48 hours in multi-target mode (all 15 targets at once) using both Opus 4.8 and Mythos Preview, plus additional 24-hour single-target runs for Mythos Preview (one target per session, all running in parallel).

For each target, Claude chose where on the protein to design against, generated candidate structures by orchestrating multiple specialist models, ran cycles of in-silico optimization, and computationally screened for novel, diverse candidates that would express, stay soluble, and bind. It produced 30 designs per target—1,320 designs total.

graph TD
    A[One Prompt - 30K tokens] --> B[Claude Science Session]
    B --> C[RFdiffusion - Structure Design]
    B --> D[ProteinMPNN - Sequence Design]
    B --> E[AlphaFold2 - Co-folding]
    B --> F[ESMFold - Screening]
    C --> G[In-silico optimization cycles]
    D --> G
    E --> G
    F --> G
    G --> H[30 designs per target]
    H --> I[Adaptyv Bio / Twist Bioscience wet lab]
    I --> J{14/15 targets - binders confirmed}
    style A fill:#27272a,stroke:#a78bfa
    style J fill:#1e3a5f,stroke:#60a5fa

The Numbers That Matter

The single-target mode outperformed multi-target by nearly 10 percentage points—expected, since focused sessions let Claude dedicate full attention to one binding problem at a time. But even the multi-target results are remarkable: a model that's never been trained on protein structure prediction is orchestrating specialist models better than most human practitioners.

The resulting 354 confirmed binders (from 1,320 designs) is a significant addition to the public corpus. For comparison, the two largest public collections—proteinbase.com and the Overath et al. collection—together contain ~770 binders from ~5,700 designs across 40 targets.

The Three Results That Stand Out

1. Competitive with expert-led competitions

Against RBX1 (a target from Adaptyv Bio's BenchBB competition), Mythos Preview in single-target mode achieved a 40% hit rate—compared to 3.7% among competition participants. Its top-ranked design was a high-affinity binder that outperformed the winning submission.

2. Species cross-reactive binders for a therapeutically relevant target

Opus 4.8—not Mythos Preview—succeeded against TNFα, a multimeric target that multiple expert groups have struggled with. It produced binders that cross-react across human, cynomolgus monkey, and mouse TNFα, which is critical for animal studies. The fact that a less capable model outperformed a more capable one on this specific target is a finding the field should note: capability is not uniform, and the most capable model isn't always the best tool for every task.

3. β-sheet binders

Most computationally designed binders are α-helix bundles—the easy mode. Claude designed 15 confirmed binders across six targets containing at least 20% β-strand, which is harder to design and more prone to misfolding. This demonstrates structural reasoning, not just pattern matching.

Where It Fell Short

Claude failed against maltose-binding protein (MBP), a large, flexible bacterial protein with a smooth, water-loving surface—"a binder very little to grab on to," as Anthropic notes. None of the 90 designs was confirmed to bind, though one showed a weak reproducible signal. Against BBF-14 (a de novo designed β-barrel that doesn't exist in nature), the hit rate was modest: three binders with sub-micromolar to micromolar affinity.

These failures aren't limitations of the protein design models—they're challenges in the design space that expert humans also struggle with. What's notable is that Claude didn't hallucinate success. It generated designs, they were tested, and they failed transparently. That's the process working as designed.

The Analytical Chemistry Result

Alongside the protein work, Anthropic tested Claude Opus 5 on routine analytical chemistry: interpreting raw NMR and LC-MS data from a contract lab. Given only the raw instrument files and a two-sentence prompt, Claude returned finished results in 23 and 19 minutes respectively—matching the lab's own analysis on hydrogen counts (within 0.08 ¹H) and purity (96.4% vs 96.33%).

The NMR result is particularly impressive because Claude recovered the raw signal from the instrument's proprietary binary format—an undocumented vendor format—then applied phase correction, baseline correction, peak picking, and integration. It even proposed the same follow-up experiment (heavy water exchange) that the lab had independently run.

Why This Matters for People Building Things

Two things make this paper different from most AI-biology results:

1. Wet-lab validation is the gold standard. Most AI protein design papers report in-silico metrics (RMSD, pLDDT, binding energy predictions). Those are useful proxies. They are not binding (literally). Anthropic paid for independent wet-lab testing at scale—1,320 designs synthesized and tested. That's expensive and slow, and that's precisely why most labs don't do it. The fact that Claude's designs held up under real experimental conditions raises the credibility bar for everyone in this space.

2. The orchestration model works. Claude didn't design proteins from first principles. It coordinated existing specialist models (RFdiffusion, ProteinMPNN, AlphaFold2) in a workflow that a human protein engineer would execute over weeks. The LLM isn't replacing the specialist models—it's replacing the human operator sitting in front of them. This is the agent thesis applied to biology: the bottleneck isn't the algorithms; it's the orchestration. And Claude can now orchestrate as well as a mid-level protein engineer, with zero training in structural biology.

Anthropic is careful to note that these are minibinders, not therapeutic antibodies or small molecules—designing a binder is "just the first step" toward a drug. But they're explicitly working to extend this to the full end-to-end pipeline across all drug modalities. That timeline is accelerating.

The dual-use implications are real and Anthropic acknowledges them: these capabilities remain blocked in Claude Fable 5 for general access. Protein design capability is dual-use research. The question isn't whether it exists—it's who controls access.