Scorpion toxin binders: 17 of Claude’s 211 designs worked

An internal research model, run inside Claude Science, generated 318,484 protein designs against five scorpion neurotoxins over 64 hours and four follow-up prompts; of the 211 designs that Adaptyv Bio and Twist Bioscience then tested in the lab, 17 met the criteria for a binder, a protein that latches tightly onto a chosen target, Anthropic said on October 20. What the labs measured is attachment to the toxin; Anthropic says it has not confirmed whether any of the 17 neutralizes venom in a cellular environment, and makes no antivenom claim.
Claude’s final list ran to 180 designs, and Anthropic’s team added 31 more, for 211: 60 against a single toxin, 12 per toxin, 30 against the two β-family toxins Ts1 and Cn2, 30 against the three α-family toxins AaH I, AaH II and LqhIII, and 91 against all five at once. It used 12 design tools, co-folded and screened 553,052 design-toxin pairs, and the campaign consumed on the order of 10,000 GPU hours. The 17 binders come out as seven against Cn2, four against AaH II, four against AaH I, two against LqhIII and none against Ts1. At Anthropic’s request, Claude emailed the sequences to Twist Bioscience from one of the company’s accounts, in emails marked as written by Claude.
One toxin at a time
The multitoxin designs were the cheap path: a single molecule would cut the number needed for a universal antivenom cocktail, which Anthropic describes as the most cost-effective and regionally neutral treatment. The five targets were chosen as a hard test of that idea. They share one fold, the overall shape of the protein, yet any two are only 26% to 62% identical in sequence. The share of designs labeled binders fell with every extra target: 11 of 60 single-toxin designs (18%), 5 of the 60 double-β and triple-α designs (8%), and 1 of the 91 all-five designs, the fusion AaHI-B3. Across all 151 designs aimed at more than one toxin, six were binders, 3.97%, and none of them bound more than one target.
Within the single-toxin sets the spread is wide: Cn2 yielded five binders out of 12 designs (41.7%), AaH I, AaH II and LqhIII two each, and Ts1 none. Anthropic lists its prompting, the limits of the available design tools and the biological feasibility of multitoxin design among possible explanations, alongside the model’s current capabilities. Its follow-ups were aimed mainly at getting a multitoxin binder, and Claude selected most of the 180 designs, including the strongest binder, Cn2-B1, only after the third of them.
Scanning all 211 rows of the released table, we found four designs that were hits against two or more toxins at both labs. Three are fusions from Set 9 that also hit CLEC12A, an unrelated human protein used as a stickiness control, in both runs. The fourth, AaHII-B4, was designed against AaH II alone, has a clean CLEC12A record, and was a hit against Ts1 at both labs, with a KD of 9.14 nM at Adaptyv Bio and 28.1 nM at Twist Bioscience. It is labeled a binder of AaH II only because its fitted maximum response at Twist came out 4.2 times the theoretical maximum. The report says it may bind Ts1, which is documented rather than measured.
What counts as a binder
Anthropic’s bar had four parts. A design had to be a hit at both labs in every replicate. Every KD, the dissociation constant that puts a number on binding strength, had to be 1 µM or tighter, lower being tighter. The fitted maximum response at Twist Bioscience had to stay within the theoretical maximum that one toxin molecule per binding site can produce. And the design had to show no above-background signal against CLEC12A. LqhIII is the exception to the first part: it was assessed only by Adaptyv Bio, so the two LqhIII binders rest on one lab’s data.
Both labs read binding by surface plasmon resonance, with toxin flowing over the design while the instrument tracks how fast it attaches and comes off. Adaptyv Bio made the proteins in a cell-free system, using protein-making machinery extracted from bacteria; Twist Bioscience made them in human kidney cells, fused to the stem of an antibody. Scorpions sting 1.2 million people or more a year, with more than 3,250 deaths, mostly among children in places with limited access to medical care, and conventional antivenom is made by milking venom, immunizing horses over months and purifying their antibodies, which usually covers only a small number of neurotoxins.
Anthropic sent the work to two outside researchers, Susana Vázquez Torres of the Spanish National Cancer Research Centre and Timothy P. Jenkins of the Technical University of Denmark, both of whom called it a promising first step. “We can now obtain very tight binders remarkably quickly, and binding itself is no longer the main bottleneck,” Vázquez Torres said. All 211 designs and the lab data are published on Hugging Face.
Our analysis
What we checked
On October 21 we loaded the released design table, 211 rows, re-derived the binder set from the raw per-lab result, KD, Rmax and CLEC12A columns, and joined it to the origin table to see which tool each design came from.
The headline figures reproduce exactly: the 60/30/30/91 split, the 17 binders, the 7/4/4/2 spread and the zero for Ts1. What the table adds is how thin the line around 17 is. Design cf8652bc clears the both-labs, 1 µM and Rmax criteria against AaH I, with a KD of 260 nM at Adaptyv Bio and 134 nM at Twist Bioscience, and is excluded only by a near-background CLEC12A signal in one Twist run, which the fourth criterion treats as off-target binding. Read that signal as noise and the set is 18.
Across the 211 tested designs, the binders came from five named tools plus one two-domain fusion, and two of the 12 tools Claude listed as used produced nothing that reached a lab.
| Tool | Tested designs | Binders |
|---|---|---|
| RFdiffusion3 | 17 | 5 |
| Genie 3 | 27 | 5 |
| Proteina-Complexa | 12 | 2 |
| PXDesign | 48 | 2 |
| Mosaic | 55 | 2 |
| FreeBindCraft | 12 | 0 |
| Fusions | 32 | 1 |
These are selected designs, not a randomised comparison, and the counts are single digits, so this is the composition of the tested set and not a ranking of the tools. Three of the 17 binders sit on a backbone first built against a different toxin: Cn2-B5 on a Ts1 backbone, AaHII-B2 and AaHII-B3 on AaH I backbones.
Up next
Jev answers in 195 ms, OpenAI Decisions in 208 ms
A refund-checking bot gets Jev’s yes/no in a typical 195 ms, OpenAI Decisions’ in 208, but OpenAI’s first call after a lull took nearly two seconds in Mumbai. Would that two-second first call push you off OpenAI Decisions, or is 208 ms enough?
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

