Anthropic Claude Scorpion Toxin Binder Design Research

Overview of the Scorpion Toxin Binder Design Campaign

Anthropic utilized an internal research model, operating within Claude Science, to design de novo protein minibinders targeting dangerous scorpion neurotoxins. In a 64-hour autonomous campaign, Claude generated 318,484 protein designs, from which 211 were selected for experimental validation by partners at Adaptyv Bio and Twist Bioscience. The study resulted in 17 confirmed binders against four distinct scorpion neurotoxins. This research aims to explore how AI can accelerate drug discovery for neglected tropical diseases, such as scorpion envenomation, which currently lacks affordable and widely accessible treatment options.

Technical Methodology and Autonomous Design

Claude operated with significant autonomy, directed by an initial prompt and four follow-up instructions. The model was responsible for:

  • Target Selection: Searching the Protein Data Bank (PDB) and AlphaFold Database (AFDB) to identify five neurotoxin targets across four species (Cn2, Ts1, AaH I, AaH II, and LqhIII).
  • Design Strategy: Defining amino acid 'hotspots' for binding, setting quality thresholds, and managing the integration of 12 different protein design and folding tools.
  • Filtering and Screening: Co-folding and screening 553,052 design-toxin pairs to predict structural complexes, ultimately narrowing the field to 180 designs, later supplemented by 31 additional designs.

The campaign consumed approximately 10,000 GPU hours. The resulting designs were tested using surface plasmon resonance (SPR) to measure the dissociation constant (KD), with a strict requirement that hits show binding at both partner labs with a KD of 1 µM or tighter, while exhibiting no off-target binding to the control protein CLEC12A.

Experimental Results and Binding Affinity

Of the 211 designs tested, 17 met the criteria for a binder against a single toxin. Notably, none of the designs intended to target multiple toxins simultaneously met the criteria for more than one target. The strongest candidates, Cn2-B1 and Cn2-B2, demonstrated binding affinities in the single-digit nanomolar range against the Cn2 neurotoxin. Competition experiments suggest these binders likely target the same epitope as the known neutralizing antibody 'LR', indicating potential for neutralizing activity, though this has not yet been verified in cellular or animal models.

Implications for Drug Discovery

While the study successfully produced binders, the research team emphasizes that these molecules are only the first step in developing a therapeutic antivenom. Future work must address:

  • Functional Validation: Testing for neutralization in cellular environments via patch-clamp recordings to confirm the binders prevent toxins from affecting sodium channels.
  • In Vivo Testing: Assessing efficacy and safety in animal models.
  • Developmental Hurdles: Ensuring the molecules are stable, affordable, and accessible for the regions most affected by scorpion stings.

Anthropic has released all 211 designs and their associated binding data to the research community to support further efforts in AI-driven protein design.

Sources