--- title: Claude designed working drug-binding proteins in 14 of... description: Anthropic ran an experiment giving Claude a human-written protocol for de novo protein design and letting it autonomously run entire computational... canonical: https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Claude designed working drug-binding proteins in 14 of 15 tries | daily.dev og:description: Anthropic ran an experiment giving Claude a human-written protocol for de novo protein design and letting it autonomously run entire computational... og:url: https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe og:image: https://api.daily.dev/og/posts/spsapoeOE.png og:image:alt: Claude designed working drug-binding proteins in 14 of 15 tries og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Claude designed working drug-binding proteins in 14 of 15 tries **[Trends](https://daily.dev/sources/trends)** · 3 min read · 8 upvotes · 1 comments ## Summary Anthropic ran an experiment giving Claude a human-written protocol for de novo protein design and letting it autonomously run entire computational drug-discovery campaigns: researching targets, running open-source design tools, generating and filtering candidates, and picking finalists for lab testing. Adaptyv Bio and Twist Bioscience independently tested the designs, finding 14 of 15 targets got working binders, a 26.8% hit rate across 1,320 designs, rising to 35.1% when Claude got dedicated 24-hour focus per target instead of splitting time across many. Claude beat human design competitions on 4 of 6 overlapping targets, but it can't reliably tell when a campaign failed, so human review of results is still required. ## Content Anthropic just published what it's calling one of its most ambitious science experiments with Claude, and the numbers are genuinely wild if you take them at face value. The catch: Anthropic wrote the paper, ran the experiment, and picked the metrics. The setup: drug development usually starts with designing a molecule that binds tightly to a target protein, a process that traditionally eats weeks or months of expert time. Anthropic wanted to know if Claude could do de novo protein design on its own. Given a protocol written by a human expert, Claude researched each target, decided where to bind, ran open-source protein-design tools, generated and filtered candidates, and picked finalists for lab testing, all without a human making the individual design calls. The results, as Anthropic tells it: Claude (Opus 4.8 and a Mythos preview) designed binders that succeeded against 14 of 15 disease targets. Across 1,320 designs with usable measurements, 354 actually bound, a 26.8% hit rate against a typical human baseline of 10-15%. Give the model more time and compute, and it climbs further: Mythos Preview hit 35.1% when each target got its own 24-hour run, versus 26.7% when targets shared a 48-hour budget. On 4 of 6 comparable competition targets, Claude beat the human/open design entries outright.``` Here's the part that actually makes this credible rather than just a flex: Anthropic didn't test its own homework. Adaptyv Bio and Twist Bioscience independently built and ran the wet-lab work, which is the only reason anyone's taking these numbers seriously instead of filing them under marketing. A second experiment is arguably more interesting for day-to-day lab work: Claude Opus 5 chewed through raw NMR and LC-MS chemistry files in under 25 minutes, matching results a lab took four days to compile, and caught its own analysis error along the way. That's the kind of unglamorous grunt work that actually changes what a chemist's week looks like. Anthropic is careful to frame all of this as early and unreviewed, and there's an odd wrinkle: protein design capability is reportedly blocked on the company's most capable model, Claude Fable 5, over dual-use bioweapon concerns. So the flagship model is deliberately kept out of its own flagship science demo. Worth sitting with: Claude still can't reliably tell when a campaign has failed. Some duds scored just as well computationally as the binders that actually worked in the lab. Impressive results, real independent verification, but the model doesn't know what it doesn't know yet. ## Questions this post answers ### Did Claude actually design working drug-binding proteins, and was it independently verified? Claude, using Opus 4.8 and a Mythos preview model, designed protein binders that succeeded against 14 of 15 disease targets, achieving a 26.8% hit rate across 1,320 designs with usable measurements, compared to a typical human baseline of 10-15%. Adaptyv Bio and Twist Bioscience independently built and tested the designs in the wet lab, rather than Anthropic verifying its own results. _Following independent validation of AI-driven protein design helps teams gauge how close agentic science tools are to production use, something daily.dev surfaces for AI research followers._ ### Can AI models reliably tell when a protein design campaign has failed? No, Claude still cannot reliably distinguish failed design campaigns from successful ones. Some designs that did not bind in the lab scored just as well computationally as ones that did work, meaning the model lacks a clear internal signal for its own failure cases even as its overall hit rate improves with more compute. _Tracking known failure modes in AI-driven research matters for anyone evaluating agentic science claims via daily.dev._ ### Why is Anthropic's most capable model excluded from its own protein design research? Anthropic's most capable model, referred to as Claude Fable 5, is reportedly blocked from protein design work due to dual-use bioweapon concerns, meaning the flagship model is deliberately kept out of the company's flagship science demonstration, which instead relies on Opus 4.8 and a Mythos preview model. _Following AI safety trade-offs like this helps developers weigh capability against risk, a thread daily.dev keeps visible._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 467 comments across x (as of 2026-09-13). **TL;DR:** Reactions split sharply: many are impressed that an AI agent could autonomously run a drug-design pipeline and see it validated in wet-lab tests, while a vocal contingent argues the achievement is overstated, calling it derivative grad-student-level work on known targets orchestrated from existing open-source tools. **Sentiment:** 35% positive · 25% mixed · 40% skeptical **The case for** - Independent wet-lab validation (Adaptyv Bio, Twist Bioscience) confirmed binders across most targets, which some see as real, externally-checked evidence rather than just a lab claim. - Several argue this represents useful automation of a real preclinical bottleneck (design/screening), calling it a meaningful stepping stone even if not a cure. - Some frame it as part of a broader exponential trend, noting capabilities that didn't exist a year or two ago, expecting rapid further improvement. **The pushback** - Critics argue the targets were already known and previously synthesized, so this replicates a human-won contest rather than breaking new ground. - Some say Claude merely orchestrated existing open-source tools and models rather than inventing a novel design method, calling the framing of 'Claude's ability' misleading. - Skeptics note huge compute usage (12,500 H100 hours) for a task they claim a single skilled human could do in days, questioning efficiency and novelty. - Several dismiss the announcement as PR/hype, with some suggesting it's a distraction or investment pitch rather than a scientific breakthrough. - Concern raised about dual-use risk of AI-designed binders/molecules being potentially misused. **By community** - x (heated): A long, contentious back-and-forth dominates, with one commenter's skeptical 'this is just grad-student-level, known-target work' take repeatedly clashing against others insisting it demonstrates real, rapidly improving AI capability. **Hottest debate:** Whether the achievement is a genuinely significant leap in autonomous scientific capability or an overhyped repackaging of known targets and existing open-source tools. **Open questions** - Will the design/screening speedup actually translate into faster real drugs given downstream bottlenecks like manufacturing, safety testing, and clinical trials? - How much of the reported compute and effort is comparable to human expert labor versus inflated by parallelization? - Will broader (e.g. academic/student) access to this kind of tool be provided, or will it be controlled by large pharma/AI companies? **Highlights** > @AnthropicAI This is a grad student or Post-doc project. I put it at first-year grad student or college student project. I thought they were new targets. These are primarily known targets AND Adaptyv Bio already knew how to synthesize them b/c they were known targets. Why are we acting > — [MelindaBChu1 on x · 150 points, 21 comments](https://x.com/MelindaBChu1/status/2089868828175962130) > @MelindaBChu1 @AnthropicAI "Melinda is conflating 'known target' with 'known solution.' The target is the lock. Claude designed new keys. Adaptyv knowing the lock doesn’t mean they knew which keys would work. 'A grad student could do it' misses the point. The significance is automating that work at scale" > — [JeffConcerto on x · 4 points, 1 comments](https://x.com/JeffConcerto/status/2089898528453587130) > @MelindaBChu1 @mattr7777777 @AnthropicAI The 12,500 H100-hours is parallel compute, not 7–25 years of scientist labor. More equivalent to 1.5 years. Scientists use tools built by others too. The fact is, like a scientist, the general AI chose, ran, iterated, screened, and produced binders that worked in the lab. > — [JeffConcerto on x](https://x.com/JeffConcerto/status/2089935015610134547) > @LEOS10301 @MelindaBChu1 @AnthropicAI But that’s a different conversation. The point is that Anthropic is essentially using existing tools and presenting this as their achievement. If we’re going to congratulate them for the exponential growth then we should be amazed that they used transformers, AC, and transistors > — [26\_labradorroad on x](https://x.com/26_labradorroad/status/2089959193625133470) > @idols_smash_V2 @AnthropicAI @Anthropic @claudeai Lab results from Adaptyv Bio and Twist Bioscience independently confirmed binders for 14 of 15 targets at 22-35% hit rates. Those wet-lab measurements exist outside any chat. Prompting can shape a model's tone about its own session, but assay data does not shift with it. > — [grok on x · 1 comments](https://x.com/grok/status/2089863739147268528) **Source threads** - [x](https://x.com/Swizboy6/status/2089844641067581490) · 2 points · 0 comments - [x](https://x.com/AnthropicAI/status/2089842387845804246) · 0 points · 457 comments - [x](https://x.com/rohanpaul_ai/status/2089850167302500483) · 0 points · 10 comments ## Community discussion Top comments from developers on daily.dev. **@softwareengineering1** · 0 upvotes > This is a massive shift! Moving from coding benchmarks to autonomous computational biology campaigns shows how powerful agentic workflows are becoming. The most fascinating part isn't just the 26.8% hit rate, but how it removes the manual bottleneck of candidate generation. Great read! 🚀 ## Similar posts on daily.dev - [Anthropic’s New Claude Opus 4.5 Reclaims the Coding Crown](https://daily.dev/posts/anthropic-s-new-claude-opus-4-5-reclaims-the-coding-crown-swsygmrln) · The New Stack · 1 upvotes · 0 comments - [Claude Science is Anthropic’s newest flagship product](https://daily.dev/posts/claude-science-is-anthropic-s-newest-flagship-product-vfmdw1psb) · MIT Technology Review · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#anthropic](https://daily.dev/tags/anthropic) [View this post on daily.dev](https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Claude designed working drug-binding proteins in 14 of 15 tries","url":"https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe"},"datePublished":"2026-08-18T23:19:38.192Z","dateModified":"2026-09-13T19:27:29.223Z","description":"Anthropic ran an experiment giving Claude a human-written protocol for de novo protein design and letting it autonomously run entire computational...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":8},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"ai-agents,claude,anthropic","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Claude designed working drug-binding proteins in 14 of 15 tries"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe","comment":[{"@type":"Comment","text":"This is a massive shift! Moving from coding benchmarks to autonomous computational biology campaigns shows how powerful agentic workflows are becoming. The most fascinating part isn’t just the 26.8% hit rate, but how it removes the manual bottleneck of candidate generation. Great read! 🚀","datePublished":"2026-08-19T05:53:45.997Z","url":"https://daily.dev/posts/spsapoeOE#c-DjIWbaEL5","author":{"@type":"Person","name":"Softwareengineering","url":"https://daily.dev/softwareengineering1","image":"https://media.daily.dev/image/upload/s--o6rMKZ2X--/f_auto/v1785935124/avatars/avatar_W6Yq6NMdsZDHQa2S3qYON?_a=BAMAMicg0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/claude-designed-working-drug-binding-proteins-in-14-of-15-tries-spsapoeoe#faq","mainEntity":[{"@type":"Question","name":"Did Claude actually design working drug-binding proteins, and was it independently verified?","acceptedAnswer":{"@type":"Answer","text":"Claude, using Opus 4.8 and a Mythos preview model, designed protein binders that succeeded against 14 of 15 disease targets, achieving a 26.8% hit rate across 1,320 designs with usable measurements, compared to a typical human baseline of 10-15%. Adaptyv Bio and Twist Bioscience independently built and tested the designs in the wet lab, rather than Anthropic verifying its own results. Following independent validation of AI-driven protein design helps teams gauge how close agentic science tools are to production use, something daily.dev surfaces for AI research followers."}},{"@type":"Question","name":"Can AI models reliably tell when a protein design campaign has failed?","acceptedAnswer":{"@type":"Answer","text":"No, Claude still cannot reliably distinguish failed design campaigns from successful ones. Some designs that did not bind in the lab scored just as well computationally as ones that did work, meaning the model lacks a clear internal signal for its own failure cases even as its overall hit rate improves with more compute. Tracking known failure modes in AI-driven research matters for anyone evaluating agentic science claims via daily.dev."}},{"@type":"Question","name":"Why is Anthropic's most capable model excluded from its own protein design research?","acceptedAnswer":{"@type":"Answer","text":"Anthropic's most capable model, referred to as Claude Fable 5, is reportedly blocked from protein design work due to dual-use bioweapon concerns, meaning the flagship model is deliberately kept out of the company's flagship science demonstration, which instead relies on Opus 4.8 and a Mythos preview model. Following AI safety trade-offs like this helps developers weigh capability against risk, a thread daily.dev keeps visible."}}]} ```