Anthropic has reported two experiments that place its Claude models inside parts of the early drug discovery workflow. The larger claim is not that Claude invented a new protein model, but that a general language model agent can coordinate existing specialist tools, make design choices, and return lab-testable candidates with limited human intervention.
The results are striking, but they also come with clear boundaries. Anthropic says the protein design hit rates were above the typical range it cites from publicly documented campaigns, while an independent review of the results is still pending.
What Claude was asked to design
The first experiment focused on minibinders, small proteins designed to attach tightly to a target protein. In drug discovery, that kind of binding can block or alter a protein's function, which is why the task matters.
The work used de novo design, meaning the binders were generated computationally rather than found in nature. According to the technical report described in the source, this kind of project normally involves expert judgment, specialized software, compute orchestration, and several days of work.
Anthropic used Mythos Preview and Opus 4.8 against 16 target proteins. Usable measurements came back for 15 targets, and Claude succeeded on 14 of those 15. In the lab, 354 of 1,320 designs bound to their targets, giving a hit rate of 26.8 percent.
The ranking result was also notable. Among the designs Claude placed first on its own lists, 49 percent bound to the target. Anthropic compares the overall result with a typical range of 10 to 15 percent from publicly documented campaigns in the proteinbase.com database.
The agent layer is the main point
Anthropic did not present this as a new foundational protein-design system. Claude used open-source tools already known in the field. The protein backbones came from tools including PXDesign (358 designs), RFdiffusion3 (267), Genie 3 (185), FreeBindCraft (135), BoltzGen (134), RFdiffusion (118), and Proteina-Complexa (100), among others.
The amino acid sequence work was mostly handled by SolubleMPNN, a variant of ProteinMPNN. For filtering and ranking, Claude used ESMFold2, ESMFold2-Fast, and Protenix v2 to predict how each binder and target might fold together and to score confidence in the binding result.
Some tools were deliberately excluded. AlphaFold-3 weights, Rosetta, and ESM3 were not used because of licensing reasons.
The system prompt was about 16,000 words. Only about a third covered scientific guidance and a reading list. The rest addressed scheduling, delegation to sub-agents, verification, and budget discipline. The prompt also did not tell the model which epitope to target on any protein surface.
That makes the experiment less about replacing protein-design software and more about automating the control layer above it. Claude researched each target, selected a docking site, installed public-code tools, combined them across 24 different workflows, and produced a final ranking without humans making individual design decisions.
Where the results were strongest
Anthropic ran multi-target campaigns with a $50,000 compute budget and single-target work with a $10,000 budget, using Modal as the cloud provider. Humans selected the targets, wrote the prompt, ordered synthesis, and interpreted measurement data. Between those steps, the report says only short, non-technical instructions were needed after infrastructure outages.
Validation was carried out by paid contract labs Adaptyv Bio and Twist Bioscience. They produced the designs biologically without modification and measured whether each one bound to its target and how tightly. Binding strength was reported as KD in nanomolar (nM), where a smaller value means tighter binding.
Performance varied sharply by target. For TREM2, 72 of 90 designs bound. For VEGF-A, 54 of 90 bound. On RBX1, an earlier open design contest had produced 9 binders from 245 newly designed candidates. Claude produced 28 binders from 90.
Anthropic also rebuilt the winning design from that contest and tested it on the same assay plate. The contest winner bound at 45 nM, while Claude's best design bound at 3.9 nM, roughly ten times tighter.
TNFα showed a more mixed but still important result. The source describes it as especially hard, noting that several earlier design approaches reported zero hits. Claude produced 12 binders from 150 designs, all from Opus 4.8 and none from Mythos Preview. However, those 12 binders came from just four different scaffolds, so they were not fully independent examples.
Claude also produced cross-species binding in some cases. Although binding to animal versions of targets was only a secondary prompt goal, 130 of 233 tested binders also bound the mouse counterpart of their target.
The failures matter too
Two targets exposed important limits. BBF-14 is described as a computer-invented, barrel-shaped protein that does not occur in nature and has no evolutionary history for a design method to use. Against BBF-14, Claude produced only three weakly binding designs.
The bacterial maltose-binding protein MBP was another failure case. Its smooth, water-loving surface gave binders little to attach to, and none of 90 designs worked.
Anthropic says the folding-prediction confidence scores did not warn about those failures. Designs for MBP and BBF-14 received scores similar to designs for targets where the campaigns worked, showing that the ranking layer could miss hard target-specific problems.
Chemistry analysis was faster, but narrower
The second experiment moved from protein design to chemistry data interpretation. Opus 5 analyzed raw outputs from two standard measurements provided by a contract lab: nuclear magnetic resonance spectroscopy (NMR) and liquid chromatography with mass spectrometry (LC-MS).
NMR helps chemists check whether they made the intended substance. LC-MS is used to assess sample purity. Both instruments produce proprietary file formats that are often reviewed manually in manufacturer software.
Using raw files and prompts of one to three sentences, Claude returned results in 23 and 19 minutes. For the LC-MS file, it could not find a suitable reader program and decoded the file format itself. As a check, it exactly reproduced the stored summary values for all 2,664 measurement points.
In the NMR task, Claude proposed the same follow-up experiment that the lab had independently run three days after the first measurement. It also corrected itself: an initial report of four missing signals became two after an internal check. The report says purity values of 96.4 versus 96.33 percent used different baselines, and that using the lab's method would put Claude at 98.8 percent.
The broader message is specific rather than sweeping. Claude did not replace the lab, the instruments, or the specialist scientific tools. It acted as an agent that could connect them, run workflows, make selections, and interpret data. If the findings hold up under independent review, the experiment points to a future where language model agents make early discovery work more automated while still depending on human framing, lab validation, and careful review.