Claude Designed 17 Scorpion Toxin Binders in 64 Hours — Oct 23, 2026

0:00Claude designed 300,000 proteins in 64 hours. 17 worked. Welcome to MedAI Times podcast, your daily update on medical AI. Don't forget to like and subscribe. Claude designed scorpion toxin binders.

0:14OpenAI sends nearly 2,000 agents through a mouse brain atlas, a face photo that predicts cancer survival, two FDA clearances, one in breast cancer and one in cardiology, and Anthropx draft IPO numbers leak.

0:28On Wednesday, we covered OpenAI buying Orchestra Bio. This week, both OpenAI and Anthropx showed what their biology work actually produces. Anthropx says Claude designed protein binders against five scorpion toxins in 64 hours.

0:42Scorpion stings kill thousands of people a year, and antivenom is costly and often in short supply, so a cheap design route would matter. Claude generated 318,000 designs and picked 180 for testing.

0:56The human team added 31 more. Two outside labs, Adaptive Bio and Twist Bioscience, tested 211 designs. 17 bound their target, about 8%. The two best against a Mexican scorpion toxin called CN2 bound in the single-digit nanomolar range near the site where a known neutralizing antibody sits.

1:17Now the limits. Humans wrote the first prompt and four follow-ups, and Claude ran on its own for the first 24 hours. Every test used lab-made toxin, not venom. Nobody checked neutralization in cells or animals, and not one multi-toxin design worked.

1:33Anthropx itself writes that it does not claim to have produced an antivenom. Binding is the first gate, not the last. This is a signal for protein design and noise for antivenom because nothing here has been shown to stop a sting.

1:48The other lab claim came from OpenAI with PerturbAi and Scripps. Nearly 2,000 agents each took one gene disruption from a mouse brain CRISPR atlas covering about 7.7 million cells and finished the analysis in 32 minutes.

2:03Blinded expert reviewers preferred the SWARMS reports to single-agent reports about three times in four. The headline lead links a gene called DCTN, one to cholesterol production, which may matter in ALS.

2:16The design idea is the useful part. One gene per agent keeps each context window clean, so literature search gets room. At the largest workload, a single agent pulled about 11 references, and the SWARM pulled more than 700.

2:33Remember, it is a preprint. Eight of 14 authors work for the companies. The comparison is against single agents rather than human analysts, and no lead has been confirmed at the bench. In the Lancet Digital Health, Mass General Brigham published a foundation model trained on more than 40 million face photographs.

2:51One version estimates biological age. The other predicts cancer survival. The two were only weakly correlated, and combining them improved survival modeling in an external cancer cohort. Coverage cites a hazard ratio around three for the highest-risk quartile, but I could not check that against the paper.

3:10The authors say it supports treatment decisions and needs prospective validation. A selfie as a prognostic test is a long way from a clinic workflow. The FDA cleared Symbiosis' tumor site risk on October 15th.

3:24It estimates five- and 10-year recurrence risk from breast MRI plus clinical data for hormone receptor-positive HER2 negative stage 1 to 3A disease. The company's validation is a precision study on 120 images and a usability study with 30 staff.

3:42That shows the software runs consistently. It does not show the score changes outcomes. The world's first claim is the company's own. In cardiology, Elucid's FFR-IQ got its clearance on September 29th, announced this week.

3:56It estimates fractional flow reserve from coronary CT scans using plaque as well as anatomy. In a trial of 347 patients, the company reports 92% sensitivity and 93% specificity against invasive wire measurements.

4:12It is the third vendor offering this in the United States. Company figures, so wait for independent data. In pharma, AstraZeneca signed an AI target discovery deal with In-sitro for liver and heart disease.

4:25The headline reads $1.5 billion. The cash that is certain is 40 million up front. The rest is milestones and no candidate has been named. In-sitro builds disease models from lab-grown cells and machine learning to find targets.

4:39Then AstraZeneca takes over and funds any trials. Target selection is where most programs fail. So this is a bet on the earliest and least certain step. One regulatory note. The FDA's comment window on generative AI devices closed October 19th.

4:56The agency's list of AI-enabled devices reached 1,614 records through June, with 181 added in the first half of this year alone. On CLAWD, Anthropic opened CLAWD Duet to a wait list of US Pro and Max users.

5:11It is a proactive agent on Opus 5.5 that checks your apps every few hours, sends a morning brief, and runs in a cloud machine. It asks approval before consequential actions. Anthropic admits it may misjudge what is urgent.

5:25Team and enterprise plans are excluded, which keeps it out of most hospitals for now. Also on Anthropic, draft IPO papers reported by the Wall Street Journal and Reuters show $16.3 billion of first-half revenue and an $80.7 billion net loss, mostly accounting charges tied to convertible notes.

5:46About a quarter of revenue came from two customers. That is leaked draft reporting, not a public filing. Spotlight. When a lab says an AI designed a binder, ask what the test measured. Surface plasma on resonance measures whether a protein sticks to a target and how tightly.

6:03It cannot tell you whether sticking blocks the toxin's harm. That needs neutralization assays in cells, then animals. Binding is necessary, not sufficient. Links and sources are in the description.

6:16Would you want animal neutralization data before calling an AI designed antivenom candidate promising? Or is test tube binding enough to start? Thanks for listening. Find us on YouTube and your favorite podcast app.

6:28See you tomorrow.