Anthropic's Claude Ran 64 Hours to Design 17 Scorpion Antivenom Proteins

Anthropic has published a case study in which an internal research version of Claude ran most of a protein-design campaign against five scorpion neurotoxins. Starting from one prompt and four follow-ups, the model generated and screened hundreds of thousands of candidates, selected 180 proteins for synthesis, and drafted the report. Human researchers added 31 candidates, and laboratory tests confirmed 17 binders across four toxins.

The project paired Anthropic with Citrini, an investment research firm, to test whether non-specialists using Claude could pursue a credible biomedical target. Scorpions sting at least 1.2 million people each year and cause more than 3,250 deaths, primarily among children in regions with limited medical care. Conventional antivenom production requires milking thousands of scorpions, injecting venom into horses, and purifying the resulting antibodies. The process takes months and leaves many affected regions with inadequate supplies.

Claude’s 64-hour design run

The five targets came from four regions: Mexico’s Cn2, Brazil’s Ts1, North Africa’s AaH I and AaH II, and the Middle East’s LqhIII. Each toxin binds to sodium channels in nerve cells. The campaign sought minibinders, compact proteins designed from new amino-acid sequences to occupy the toxin surface that normally contacts the channel.

Claude served as the planner and tool operator, while specialized databases and protein models performed structure retrieval, sequence design, folding, and scoring. Its recorded tasks included:

• Choosing structures: Claude searched the Protein Data Bank and AlphaFold database, then selected five amino-acid hotspots on each toxin.
• Generating candidates: Twelve design tools produced 318,484 proteins. Filters removed candidates resembling toxins or known proteins.
• Predicting complexes: Claude co-folded and scored 553,052 candidate-toxin pairs, predicting the three-dimensional structure of each complex.
• Ranking results: The model set quality thresholds, assigned weights to the design tools, selected 180 candidates, and drafted a report with figures.

The computational run finished 64 hours after the first prompt and consumed roughly 10,000 GPU hours. Researchers then added 31 selections of their own, bringing the wet-lab set to 211 proteins.

Seventeen binders clear the bar

Adaptyv Bio and Twist Bioscience produced the proteins using two complementary systems, one cell-free and one based on human kidney cells. The labs measured binding with surface plasmon resonance, or SPR, which tracks how quickly molecules associate and separate in real time.

A design qualified as a binder only if both labs detected it in every replicate, its dissociation constant was below 1 micromolar, and it produced no signal against an unrelated human protein used to detect nonspecific stickiness. Lower dissociation constants indicate tighter binding.

The 17 binders represented 8.1% of the 211 tested designs. Among the 60 proteins designed for exactly one toxin, 11 passed the criteria, an 18% rate that the report describes as consistent with conventional protein-design campaigns. The other six successful proteins had been designed for multiple toxins but validated against only one. No candidate achieved the intended multi-toxin binding profile.

Cn2-B1 and Cn2-B2 produced the strongest results, binding the Mexican Cn2 toxin with single-digit nanomolar affinities. Competition experiments indicate that they attach on or near the site recognized by LR, an antibody known to neutralize Cn2 in mice. That overlap supports a plausible blocking mechanism, although the new proteins have not undergone neutralization tests.

Binding is several steps from treatment

The experiments used recombinant toxins produced from gene sequences. The study included no purified scorpion venom, sodium-channel function tests, cultured-cell neutralization, animal protection studies, or human testing. Anthropic therefore describes the proteins as early-stage binders and makes no therapeutic claim.

Advancing a candidate such as Cn2-B1 would require several additional stages:

1. Functional testing: Patch-clamp recordings would measure whether the binder prevents the toxin from disrupting electrical currents through sodium channels.
2. Protection studies: Cell assays and mouse experiments would test activity against purified venom and determine whether the molecule prevents injury or death.
3. Drug development: Researchers would need to establish stability, dosing, immune response, toxicity, manufacturing yield, storage requirements, cost, and regulatory compliance.

Two outside experts consulted by the project described the results as a credible first step. For neglected diseases with weak commercial incentives, a faster route to initial binders could reduce early discovery costs. Affordability and clinical usefulness will depend on the downstream work that this campaign did not address.

The missing comparison

The campaign demonstrates that an LLM agent can coordinate existing protein-design systems and deliver experimentally verified hits on four of five selected targets. Claude chose resources, configured 12 tools, managed a long-running workflow, filtered results, and prepared synthesis recommendations from limited human instructions.

The study included no matched campaign run by expert protein designers or by the same tools without Claude. It therefore cannot determine whether the agent improved hit rate, speed, or cost relative to an experienced team using the identical software stack. The failed Ts1 designs and unsuccessful multi-toxin strategy also limit how broadly the results can be generalized.

Compute sets the practical limit

The full campaign’s roughly 10,000 GPU hours exceed the computing budgets of many academic laboratories. Direct reproduction also depends on an internal Claude research system and access to the same collection of design and folding tools.

Anthropic reports that a simplified single-agent pipeline running on one GPU for 24 hours can achieve comparable computational scores. That comparison covers in silico metrics rather than laboratory binding results. Wet-lab validation will be necessary to determine whether the cheaper pipeline produces a similar hit rate.

Prompts, scores, and structures are public

• All 211 designs, predicted structures, scores, and SPR measurements are available in the campaign dataset.
• The technical report includes the prompts, follow-ups, tool configurations, methods, and results.
• Anthropic’s earlier protein-design work documents a related single-prompt experiment.

The released artifacts give researchers enough detail to inspect Claude’s decisions, reanalyze its rankings, study failed candidates, and compare future agent-driven campaigns against the same experimental record.