A team of five financial analysts with little to no scientific background has successfully utilized artificial intelligence to design experimental protein binders targeting scorpion venom. Operating on a $20,000 budget for the venture, the group from research firm Citrini Research partnered with Anthropic to tackle a medical issue that disproportionately affects underserved regions with limited healthcare access.
Traditional antivenom production is costly and relies on extracting venom from scorpions, immunizing animals such as horses, and purifying antibodies from their plasma. Seeking an alternative, the team tasked Anthropic’s internal research model, designated as Claude Science, with designing small engineered proteins that could act as molecular clamps to latch onto scorpion neurotoxins and potentially neutralize them.
The project focused on five distinct scorpion venom targets originating from Mexico, Brazil, North Africa, and the Middle East. Globally, more than 1.2 million people are stung by scorpions each year, resulting in over 3,250 deaths, with children facing the highest risk of severe complications. Despite the significant human toll, high development costs and limited commercial returns mean such neglected tropical issues frequently lack sufficient research investment.
Before launching the automated computations, the team worked alongside human researchers at Anthropic to draft an extensive, 35-page prompt detailing research parameters, constraints, and evaluation criteria. Once initialized, the system deployed over 300 concurrent GPUs and more than 300 sub-agents in a Claude Science environment equipped with 12 specialized design tools. Over a continuous 64-hour span, the AI literature searches, tool operations, and predictions generated more than 318,000 preliminary designs.
During the computational phase, the AI interacted with human supervisors to handle practical hurdles, including drafting and sending inquiries to suppliers like Twist Bioscience regarding the manufacturing feasibility of complex proteins. After factoring in supply constraints and filtering the output through 367 total exchanges and four human follow-up prompts costing approximately $49,997 in compute resources across 10,000 GPU hours, the system produced 180 candidate sequences. Researchers manually added 31 candidates to establish a final batch of 211 designs for physical synthesis.
Subsequent laboratory synthesis and testing handled by Twist Bioscience and Adaptyv Bio cost roughly $71,740. Of the 211 tested candidates, 17 successfully met binding standards across four of the five targeted venom regions. Seven candidates bound successfully to the Mexican venom component Cn2, four each to North African toxins AaH II and AaH I, and two to Middle Eastern toxin LqhIII. None of the designs targeting Brazilian venom met the final threshold.
An ambitious attempt to engineer a single protein capable of binding simultaneously to multiple venom types failed entirely, confirming earlier warnings from the AI model that a universal multi-toxin binder was statistically improbable. However, six of those multi-target designs successfully bound to a single venom type and entered the final group of 17.
Among the successful outcomes, a candidate designated as Cn2-B1 demonstrated particularly strong and slow dissociation from its target during observation windows. Further laboratory comparisons using an antibody fragment known as LR, which has previously shown neutralizing efficacy in mouse studies, revealed that pre-occupying the toxin with LR inhibited the AI-designed protein from binding. This competitive overlap suggests the artificial protein targets a functional site capable of disrupting the venom’s interference with neural sodium channels.
Researchers emphasized that while the binding assays demonstrate strong molecular interaction with laboratory-prepared recombinant toxins, the candidates have not yet been tested for cellular or animal protection. All 211 tested designs, computational predictions, and experimental datasets have been released publicly on Hugging Face under a CC BY 4.0 license to allow independent laboratories to verify results and advance the research.
