On October 20, 2026, Anthropic announced the results of the largest AI protein design experiment to date: within 64 hours, Claude generated 318,484 protein binder designs targeting scorpion venom neurotoxins; 211 were selected and sent to independent laboratories for testing, and 17 ultimately passed binding tests. The strongest candidates, Cn2-B1 and Cn2-B2, showed single-digit nanomolar affinity for the target toxin. Cross-checking by two independent contract research organizations, review by two external experts, and the simultaneous release of all 211 sequences and a 69-page technical report together form the most complete reproducible closed loop in AI protein design to date.
The choice of scorpion venom as a target was not random. According to Anthropic's official research report, more than 1.2 million people worldwide suffer scorpion stings each year, causing more than 3,250 deaths, mostly among children in areas with limited medical resources. Existing scorpion antivenom is extremely costly to produce, and it is very difficult to recover R&D investment; such diseases have been classified as “neglected tropical diseases.” It is precisely this structural market failure that makes scorpion venom an ideal setting for testing AI drug discovery capabilities. The question for this experiment was jointly proposed by Anthropic and Citrini Research, a financial investment research institution.
How 310,000 Designs Were Completed in 64 Hours
According to Anthropic's official report, the entire design workflow was run by an internal research model in the Claude Science environment. It began with a 35-page prompt drafted jointly with Citrini Research and four subsequent supplementary instructions—the rest of the work was completed entirely autonomously by Claude. During the process, Claude retrieved target toxin structures from the Protein Data Bank and the AlphaFold database, identified five key amino acid hotspot positions for each of five neurotoxins, and then called 12 specialized protein design tools to generate candidate sequences, filtering out all designs similar to known proteins or toxins. The system ultimately completed co-folding calculations for 553,052 design-toxin pairs, coordinating more than 300 parallel computational agents; the entire process consumed about 10,000 hours of H100 GPU compute.
The key here is not whether Claude “understands” protein chemistry, but that it acts as an orchestration layer capable of linking heterogeneous specialized toolchains. Standalone structure-generation models and folding-prediction models have long existed; the real difficulty has always been: which site is worth trying, which method suits the current target, and how intermediate results should guide the next round of screening. In the past, computational protein design experts filled this role; now Claude fills it—and far faster than humans.
What the 17 Hits Really Mean: Breaking Down the Hit Rate
The overall hit rate of 17/211 is about 8%. This figure needs to be broken down. In the design group focused on a single toxin (60 candidates), the hit rate rose to 18% (11/60); among candidates designed specifically to target multiple toxins simultaneously, none passed the binding test. This stratified result clearly points to the current method's boundary in task complexity.
Measured against industry context, according to public data cited in Anthropic's research report, the typical industry hit rate for de novo protein binder design is about 10% to 15%. The 18% for single-target design this time is slightly above that benchmark, but not by a dramatic margin. Two things truly stand out: first, the absolute design throughput—310,000 designs completed within 64 hours, which no human team could match; second, the single-digit nanomolar affinity achieved by the two binders Cn2-B1 and Cn2-B2, and according to laboratory data, their binding sites overlap heavily with those of a known antibody capable of neutralizing Cn2 in mice, a structural clue of significant value for downstream development.
In August this year, Anthropic published a preliminary study: Claude designed binders against 15 protein targets, and experimentally supported binders were obtained for 14 targets. Mythos Preview and Opus 4.8 achieved hit rates of 26.7% and 22.6%, respectively, and up to 35.1% in single-target focused mode—all higher than the overall figure in the scorpion venom experiment. The difference lies in the nature of the targets: the August experiment dealt with commonly used laboratory proteins, whereas this one involved a neurotoxin with real medical need. The difficulty levels differ, and comparability is limited. What matters more this time is the upgraded testing approach: cross-checking by two independent CROs, rather than a single internal result.
Independent Testing: What Makes This Experiment Truly Hard to Ignore
The most common criticism of AI life-science research is that models perform brilliantly computationally but falter once they enter the wet lab. This experiment deliberately built a three-layer structure to guard against that criticism: Adaptyv Bio and Twist Bioscience, two independent contract research organizations, tested the 211 candidates; the results were sent for review to two external experts, Susana Vázquez Torres of the Spanish National Cancer Research Center and Timothy P. Jenkins of the Technical University of Denmark; and all sequences and experimental data were released simultaneously, allowing any team to reproduce them independently.
This combination of public sequences, independent testing, and external expert review is quite rare in AI protein design research. It provides a data benchmark that can be cited repeatedly in future debates.
Implications for Different Parties: Gains and Losses Vary by Stakeholder
For early-stage R&D departments at pharmaceutical companies, this experiment suggests a substantial possibility of process compression: from target identification to candidate screening, AI can complete in days high-throughput computational design that previously took months. Previously, speed was mainly limited by manpower and the size of specialized teams; the cost of 10,000 hours of H100 GPU compute is no longer an industry barrier, and speed itself is becoming a competitive advantage.
For contract research organizations (CROs), AI generating massive numbers of candidates means demand for wet-lab testing will expand in tandem rather than being replaced. The deep involvement of two CROs in this experiment illustrates this complementary relationship: AI compresses the computational design cycle, while laboratory demand actually rises because of the surge in candidate volume.
The implications for non-expert teams are more disruptive. Citrini Research, which participated in defining the question, is a financial investment research institution with no direct background in biopharmaceutical R&D. This shows that, when the AI toolchain is mature enough, launching a well-structured protein design project no longer has to be led by a team with a full computational protein background—provided they can design the right testing workflow and partner with organizations capable of carrying out wet-lab experiments.
Known Boundaries: What This Experiment Proved and Did Not Prove
It is necessary to distinguish clearly between facts and inferences. What has been proven: AI can generate protein binder candidates at high throughput within a reasonable compute budget; some binders have measurable nanomolar affinity; computational predictions match wet-lab results; and effective binders were obtained for four of the five neurotoxins. What has not been proven: whether these 17 binders can neutralize scorpion venom in a cellular environment, whether they are effective in live-animal experiments, and their clinical feasibility. Anthropic explicitly states in the report that it has not established that these binders can neutralize scorpion venom in a cellular environment and does not claim to have produced an antivenom. From binding to antivenom, multiple hurdles remain, including cell-based neutralization testing, toxicological safety assessment, animal models, production scale-up, and clinical trials.
The complete failure of the multi-toxin design direction also deserves separate mention. Among candidates designed specifically to bind multiple toxins simultaneously across the five toxins, none passed testing. The difficulty of developing broad-spectrum antivenom—one product covering multiple toxins—persists even in the AI-led design stage and has not been automatically solved.
Regulatory Gaps and the Key Signals Ahead
The speed of AI-driven candidate generation is approaching a level that human regulatory capacity struggles to keep pace with. An AI proposed 318,484 design candidates within 64 hours; at the design stage this is legal and requires no approval. But once it enters clinical translation, neither the FDA nor the EMA currently has a complete approval pathway specifically for AI-designed molecules. This is not unique to this experiment, but the scale of this experiment makes the problem more concrete.
The nearest observable signal is where these 17 binders go next: data from cell-based neutralization experiments may appear in the coming months, at which point the gap between “binding” and “efficacy” will be quantified. If Cn2-B1 and Cn2-B2 also pass cell-level tests, this would become a key trigger for investment in animal experiments and would upgrade AI drug discovery from a proof of concept to the starting point of an actual development pipeline. The broader signal is that this experiment demonstrates that high-quality protein design projects can be initiated by non-traditional actors. In the future, who can integrate AI candidate generation, wet-lab testing, and preclinical assessment into a complete pipeline will shape the competitive landscape sooner than the question of who has a larger laboratory—and the current depth of integration between AI infrastructure and biological knowledge will be the main variable in the speed of this shift.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接