Claude's Autonomous Protein Design: From Leak to Lab Results

Episode 1 · Claude Autonomously Designs 318k Proteins in 64 Hours; 17 Bind Scorpion Toxins in Wet Lab (2026-10-20, 22 posts)

Anthropic partnered with finance research team Citrini on a protein-design experiment codenamed Citrini Therapeutics: an internal Claude biology research model was tasked with designing binder proteins against five scorpion neurotoxins (Cn2, Ts1, AaHI, AaHII, LqhIII) for an experimental antivenom. Over roughly three months, Claude worked largely autonomously for 64 hours from a single initial prompt (reportedly 35 pages) plus four follow-up questions, generating 318,484 protein designs; 211 were sent for testing and 17 proved effective, covering 14/15 targets, with all data open-sourced. It stands as an end-to-end, verifiable demonstration of AI-driven drug design.

Confirmed

• Starting the morning of August 12, 2025, when a Slack notification signaled the start of the autonomous drug-design run, Citrini worked directly with Anthropic's internal research model for about three months. Per @sholtodouglas, more than 300 sub-agents were spawned to run the design campaign; per Reddit discussion relayed by @1411337, the team was five investors with almost no biology background.
• Per @kimmonismus citing the official blog, the models were unreleased Mythos Preview and Opus 4.8; per @chrisalbon, the 64-hour run used 300+ sub-agents, 300+ GPUs and cost $49,997.
• Wet-lab details vary across accounts: @citrini mentions Twist, while Adaptyv (via @CatAstroPiyush and @AllThingsApx) reported its automated wet lab tested the 211 designs and found 17 binding to toxins (8% hit rate); @alvelda says two labs tested the designs. All data was released; @DrSingularity reports the successful binders targeted four dangerous scorpion toxins.
• Anthropic published the results on its science blog on October 21; per @新智元, the paper's first author is credited as ClaudeScience. Per @新智元, over 1.2 million people are stung by scorpions each year with over 3,250 deaths, making cheap antivenom a concrete unmet need; Anthropic's Nick Frey offered positive commentary per @sholtodouglas.

Unconfirmed

• @apoorvasriniva noted post-announcement questions about who validates the work; third-party replication and full experimental details remain outstanding, and the exact toxin coverage (four vs. five) and division of labor between Twist and Adaptyv differ across accounts.

Why it matters

• The experiment provides a quantified, end-to-end validation of AI-driven drug design: 318,484 designs → 211 tested → 17 effective, 14/15 targets hit, at under $50,000 total cost, with open data for community verification.
• @sholtodouglas emphasizes that non-biologists produced functional novel proteins with Claude's help, offering a reproducible test case for whether AI can genuinely do biological research rather than just hype.

• Citrini Used Anthropic's Internal Model to Design Scorpion Antivenom Proteins Overnight — citrini · 2026-10-20
• Citrini spent 3 months with Anthropic's internal model to design a scorpion antivenom — citrini · 2026-10-20
• Claude runs 64-hour autonomous research: 318k protein designs, 17 binders validated — kimmonismus · 2026-10-20
• Claude autonomously generates 318,000 protein designs in 64 hours, hits 14 of 15 targets — kimmonismus · 2026-10-20
• Citrini Spent 3 Months with Anthropic's Internal Model to Create Scorpion Antivenom — citrini · 2026-10-20
• Claude-designed scorpion antivenom proteins: 211 designs in 64 hours, 17 worked in lab — dr_alphalyrae · 2026-10-21
• Anthropic uses Claude to design scorpion toxin binders: 17 of 211 lab-tested designs work — AnthropicAI · 2026-10-21
• Citrus... Citrini used Claude to design antivenom proteins overnight with 300 sub-agents — _sholtodouglas · 2026-10-21
• Citrini spends 3 months using Claude to design scorpion antivenom, 17 of 211 designs bind — _sholtodouglas · 2026-10-21
• 3 months testing whether Anthropic's internal research model can design scorpion antivenom — apoorvasriniva · 2026-10-21
• Citrini spent 3 months using Anthropic's internal model to design scorpion antivenom, then got grilled on validation — apoorvasriniva · 2026-10-21
• Claude autonomously designs 318,484 proteins in 64 hours; 17 bind scorpion toxins in lab — Dr_Singularity · 2026-10-21
• 5-person team uses Claude and 300+ agents to design 318,484 proteins in 64 hours — 141_1337 · 2026-10-21
• Anthropic's internal Claude bio model spent $49,997 and 64 hours designing scorpion toxin binders — chrisalbon · 2026-10-21
• Claude autonomously designs scorpion-antidote proteins in 64 hours; 17 candidates validated in wet labs — AI寒武纪 · 2026-10-21
• Deep dive: how Claude designed scorpion-toxin-binding proteins in 64 hours — 机器之心 · 2026-10-21
• Claude designs 318,484 proteins in 64 hours; 17 bind scorpion toxin — 新智元 · 2026-10-21
• Claude-designed proteins neutralize scorpion toxins: 17 of 211 designs bind in wet lab — CatAstro_Piyush · 2026-10-21
• Anthropic's unreleased bio model ran 300 sub-agents, picked 211 drug candidates, 17 bound — ZeroStateReflex · 2026-10-21
• Claude autonomously designs protein binders targeting scorpion neurotoxins — VectorWireAI · 2026-10-21

2 more related posts →

Episode 2 · Rumor: Anthropic to announce major biology breakthrough within two weeks (2026-10-21, 2 posts)

Leaker ChrisGPT claims Anthropic will announce a biology breakthrough far bigger than its antivenom research within two weeks, adding that Claude previously produced 17 validated proteins in 64 hours. The claim remains unverified.

• Scoop: Anthropic to announce bigger biology breakthrough after Claude designed 17 validated protein binders — ChrisGPT · 2026-10-21
• Scoop: Anthropic to announce an even bigger biology breakthrough within two weeks — ChrisGPT · 2026-10-21

Episode 3 · Claude Designs Protein Binders, 17 Validated Against Scorpion Toxins (2026-10-31, 2 posts)

Anthropic partnered with Citrini Research to test whether Claude could help non-experts tackle real health problems: from 318,000 candidates, 211 AI-designed proteins were lab-tested and 17 successfully bound scorpion toxins, though results remain early-stage.

• Claude-designed proteins bind four scorpion toxins: 318,484 candidates yield 17 hits — jiqizhixin · 2026-10-31
• Claude designed scorpion toxin binders: 17 of 211 lab-tested designs bound — shoumikgoswami · 2026-11-01