Anthropic Claims Claude Agents Beat Industry Hit Rates in Autonomous Protein Design Trials

Anthropic says its Claude models can now run the entire early-stage protein design pipeline without human intervention in design decisions, according to two technical reports published by the company. The claims center on de novo protein binder design and lab data analysis, with no independent peer review yet available.

The core claim

In the first experiment, Claude models Mythos Preview and Opus 4.8 designed "minibinders" — small proteins engineered to lock onto a target protein — against 16 target proteins. Fifteen produced usable measurements, and Claude succeeded on 14 of those 15, according to Anthropic.

Of 1,320 designs synthesized and tested by contract labs Adaptyv Bio and Twist Bioscience, 354 bound to their targets — a 26.8 percent hit rate. Among designs Claude itself ranked highest, the rate rose to 49 percent. Anthropic cites a typical industry range of 10 to 15 percent, drawn from the proteinbase.com database of documented campaigns, as its comparison baseline.

Running single targets individually (rather than all 16 in parallel within a 48-hour window) pushed Mythos Preview's hit rate to 35.1 percent, but required 2.8 times the compute budget per target — a tradeoff the authors acknowledge confounds the comparison.

No new model, new orchestration

Anthropic did not build a specialized biology model. Instead, Claude installed and operated existing open-source tools: backbone generators including PXDesign (358 designs), RFdiffusion3 (267), Genie 3 (185), FreeBindCraft (135), BoltzGen (134), RFdiffusion (118), and Proteina-Complexa (100); sequence design via SolubleMPNN; and structure prediction/ranking through ESMFold2, ESMFold2-Fast, and Protenix v2. AlphaFold-3, Rosetta, and ESM3 were excluded for licensing reasons.

A roughly 16,000-word system prompt governed each agent run, with about two-thirds devoted to scheduling, sub-agent delegation, verification, and budget management rather than science. Claude chose which surface epitope to target on each protein — a decision not specified in the prompt. Compute budgets ran $50,000 per multi-target campaign and $10,000 per single target on Modal's cloud infrastructure.

Humans selected targets, wrote the prompt, ordered synthesis, and interpreted final measurement data — but did not touch individual design decisions.

Standout and failed cases

Results varied sharply by target. Against RBX1, part of a cellular protein-degradation complex, Claude produced 28 binders out of 90 designs, versus 9 of 245 in a public design contest for the same target. Anthropic had the contest's winning design re-tested on the same assay plate: it bound at 45 nM versus Claude's best design at 3.9 nM — roughly ten times tighter.

Against TNFα, a difficult inflammation target with five approved drugs including Humira, Opus 4.8 produced 12 binders from 150 designs (Mythos Preview produced none), though all 12 trace to just four scaffolds.

Two targets failed outright: BBF-14, a synthetic protein with no natural evolutionary precedent, yielded only three weak binders; MBP, a smooth bacterial protein, yielded zero of 90. Confidence scores from the folding models did not flag either failure in advance.

In a second, separate experiment, Claude model Opus 5 interpreted raw NMR and LC-MS chemistry files — including decoding a proprietary format with no available reader — producing analysis in 19 and 23 minutes, reproducing instrument-stored values across 2,664 measurement points.

Caveats Anthropic acknowledges

The reports explicitly note there was no parallel human-expert control campaign, so Anthropic does not claim Claude's designs beat what specialists using the same tools and budget would produce. For four of six contest-benchmark targets, prior contest results were included in Claude's reading material. Only binding was measured — not actual 3D structure confirmation or biological effect in living systems.

What this means

The individual components here — AlphaFold-derived structure prediction, RFdiffusion, ProteinMPNN — are established, multi-year-old tools. What Anthropic is actually claiming is novel is the orchestration layer: a general-purpose language model independently installing, sequencing, and cross-validating a dozen specialized scientific programs across 24 workflows without a human choosing intermediate steps. Because every tool used is open-source, Anthropic argues equivalent campaigns are reproducible by any lab with cloud compute access — the barrier is agent orchestration, not proprietary infrastructure. The claimed hit-rate improvement over industry baselines is significant if it holds up, but it comes from Anthropic's own report, without a controlled human-expert comparison, and with partial overlap between Claude's reading list and benchmark contest data. Independent replication will determine whether this represents a genuine capability shift in computational biology or a favorable framing of results from a company with an interest in showcasing agentic Claude deployments.