Claude designs protein binders against scorpion neurotoxins

Overview

Updated 1 hour ago

Anthropic's Claude model designed small proteins that bind to scorpion neurotoxins, the molecules behind thousands of deaths each year from Mexico to North Africa. Two contract labs tested 211 designs; 17 bound to four of five target toxins, including two with single-digit nanomolar affinity.

The new part is the autonomy. Claude picked targets from protein databases, generated 318,484 designs, filtered them, co-folded them against their targets, and drafted the research report with only four follow-up prompts over 64 hours. The binders haven't been tested for neutralization, so this is a capability proof, not yet an antivenom.

Why it matters

If Claude's binders survive neutralization tests, antivenom discovery for neglected tropical diseases could move from years to weeks.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

Validated toxin binders

Designs meeting binding criteria at both labs with KD of 1 µM or tighter against a target toxin.

318,484

Protein designs generated by Claude

Claude generated this many designs and selected 180 for testing; 31 more were added later.

Overall hit rate

17 of 211 tested designs met binding criteria; single-toxin designs hit at 18%.

64 hours

Autonomous campaign duration

Claude ran the design campaign mostly on its own, with four human follow-up prompts, using about 10,000 GPU hours.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Susana Vázquez Torres

Independent expert who reviewed Claude's scorpion binder designs

Called the work a promising first step

Timothy P. Jenkins

Independent expert who reviewed Claude's scorpion binder designs

Called the work a promising first step

Organizations Involved

Anthropic

AI research laboratory

Ran the autonomous protein design campaign with Claude

The AI lab that built Claude and ran the scorpion toxin binder campaign.

Citrini Research

Financial investment research firm

Co-developed the design prompt; backed life sciences stocks after results

The financial research firm that collaborated with Anthropic on the design prompt.

Adaptyv Bio

Contract research organization

Tested designs by cell-free expression and surface plasmon resonance

One of two labs that experimentally validated Claude's designs.

Twist Bioscience

Contract research organization (DNA synthesis and biopharma)

Synthesized designs and tested binding in mammalian cells

Produced Claude's designs and tested their binding in mammalian cells.

Timeline

August 2026 October 2026

3 events Latest: Today

October 2026
Anthropic publishes scorpion binder results
Today Publication

Anthropic reports 17 of 211 tested designs bound four of five scorpion toxins and releases all designs and data.
Today
August 2026
Citrini gives Claude the scorpion toxin design prompt
Campaign launch

Citrini Research hands Claude a roughly 35-page prompt; the model begins a 64-hour autonomous design campaign.
August 2026
Claude completes its design campaign
Capability demonstration

Claude generates 318,484 designs and selects 180 for testing; four follow-up prompts expand the pool to 211.
August 2026

Scenarios

Claude's Cn2-B1 binder neutralizes scorpion venom in lab and animal tests

Possible Resolves by Oct 20, 2027

Discussed by: Anthropic (outlined the testing path); outside experts Vázquez Torres and Jenkins

Anthropic would test Cn2-B1 against venom-purified toxin, use patch-clamp recordings to see whether it blocks sodium-channel activity, then run mouse protection studies. Competition experiments suggest the binder attaches near the site of a known neutralizing antibody, which supports the possibility it works. Success would make it a candidate for antivenom development.

A single binder validated against multiple scorpion toxins

Unlikely Resolves by Oct 20, 2028

Discussed by: Anthropic (attempted this in the campaign); Citrini Research

A binder that latches onto two or more toxins would cut the number of molecules needed for a universal antivenom cocktail. All 30 multi-toxin designs in this campaign failed to meet criteria. Anthropic notes the failure could stem from prompt design, tool limits, or the biological feasibility of multi-toxin binding. Success would require new design approaches.

Autonomous AI design campaigns expand to other disease targets

Possible Resolves by Oct 20, 2027

Discussed by: Anthropic (says it could speed discovery for envenomation and neglected diseases)

If the method generalizes, other labs adopt agentic AI to design binders against other hard-to-target proteins. Anthropic's earlier campaign already produced binders against 14 of 16 targets, and this campaign shows the workflow. Watch for new research publications on different therapeutic targets.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

December 2020

AlphaFold (2020)

DeepMind's AlphaFold predicted protein structures to near-experimental accuracy in the CASP14 competition, solving a problem biologists had worked on for 50 years.

Then

Structural biology labs gained instant access to high-quality predictions for proteins they hadn't crystallized.

Now

The breakthrough shifted the field's bottleneck from prediction to experimental validation — the same bottleneck this scorpion campaign faces.

Why this matters now

AlphaFold proved AI could crack hard biology problems. Claude's campaign extends that from prediction to design, automating the whole workflow.

2023

RFdiffusion (2023)

David Baker's lab at the University of Washington published RFdiffusion, a generative model that designs de novo protein binders from scratch. Hit rates were often in low single digits per target.

Then

Researchers could design binders against targets without existing antibodies, but most designs failed in the lab.

Now

De novo binder design became routine, with hit rates anywhere from single digits to around 10%.

Why this matters now

Claude's 18% hit rate on single-toxin designs compares favorably to this baseline, though the scorpion binders still need functional validation.

2022

Insilico Medicine's INS018_055 (2022)

Insilico Medicine's AI-designed drug for a fibrotic lung disease, INS018_055, entered Phase I clinical trials — the first AI-designed drug candidate to reach that stage.

Then

The trial showed AI-designed molecules could pass early safety and tolerability checks in humans.

Now

It set a precedent that AI-designed drugs are testable candidates, not just predictions.

Why this matters now

The scorpion binders are steps behind this — they're validated for binding but not for function. The path from design to therapy remains long.