Biotechnology · global
From Computer Design to the Laboratory: 17 Claude-Designed Proteins Pass Scorpion Toxin Binding Tests
Of 211 designs, 17 met the research team's binding criteria, and the complete data were released at the same time. This provides experimental evidence that can be scrutinized for AI-designed molecules, but whether they can neutralize toxins and protect animals remains unanswered.
Finding molecules that can bind toxins is one starting point for developing new treatments against scorpion venom. Anthropic announced on October 20 that 17 of the 211 proteins designed using Claude and submitted for laboratory testing met the team's toxin-binding criteria. The results take AI design a step beyond computer predictions; whether binding can prevent toxins from harming the body has yet to be demonstrated in this research.
The designs targeted five scorpion neurotoxins: Cn2, Ts1, AaH II, AaH I, and LqhIII. According to Proteinbase's project description, Claude used protein design and evaluation tools to generate more than 310,000 candidate molecules, then selected 211 for experimental testing. The biomedical goal of this approach is to find new proteins for antivenom research, while the experiments tested whether they could actually bind the specified toxins.
Qualification cannot be determined solely by whether an instrument produces a signal. The criteria listed in Anthropic's publicly released materials included consistency across repeated measurements, a dissociation constant no higher than 1 micromolar, and no binding signal above background for the unrelated protein CLEC12A; for most toxins, designs also had to pass testing at both Adaptyv Bio and Twist Bioscience. LqhIII was measured only by Adaptyv Bio, using the corresponding repeat-measurement and affinity criteria, plus CLEC12A screening. Of the final 17, seven targeted Cn2, four each targeted AaH II and AaH I, and two targeted LqhIII. No designs targeting Ts1 met the criteria.
The “success counts” reported on different platforms also need to be understood according to their respective definitions. Proteinbase contains 679 measurements of design–toxin pairings from Adaptyv Bio, with 51 pairings showing positive binding; this is a count of pairings, not 51 proteins that passed the full screening process. That dataset does not include Twist Bioscience's measurements or an additional off-target screen, so it cannot directly replace Anthropic's more stringent list of 17 proteins.
The key pharmacological hurdles still lie ahead. Hunterbrook, which participated in the project, reported that the experiments had not established toxin neutralization, and broader testing also suggested that some molecules might be prone to sticking to other substances. Passing a screen against a single unrelated protein is therefore insufficient to establish comprehensive selectivity, much less infer protection in animals or safety in humans. Hunterbrook also disclosed its own participation in the project and that an affiliated investment firm held long and short positions in AI biotechnology-related securities. Its reporting must be read in light of these interests.
Between design and testing, there are also waits that algorithms cannot eliminate. According to Hunterbrook's account of its participation, obtaining scorpion toxins and other reagents slowed experimental progress. Even if computational tools can rapidly propose candidate molecules, material supplies, protein preparation, and actual measurements still determine when research yields answers.
A more concrete contribution of these results is that the team also released the sequences, predicted structures, binding measurements, and experimental curves for all 211 designs, allowing outside observers to examine both successful results and those that did not meet the criteria. These data provide a traceable case of AI participation in protein design. To advance toward antivenom drugs, the next step remains to demonstrate that the molecules can block toxin activity and show protective effects and acceptable safety in animal studies.