FULL STORY

Claude's Autonomous Protein Binder Claim Draws Replication Debate

Anthropic claimed Claude nearly autonomously designed protein binders hitting 14/15 targets. A computational biologist soon noted the agent independently replicated his team's Protein Hunter protocol, adding a critical follow-up.

2026-08-19 ~ 2026-08-20 · 2 episodes · 48 posts

Episode 1 · Anthropic Says Claude Designed Protein Binders for 14 of 15 Targets (2026-08-19, 45 posts)

On August 19, Anthropic reported that Claude, given only a 30k-word background protocol written by human experts, near-autonomously designed de novo protein binders for specific disease targets—a key step in drug development. In wet-lab validation run with Adaptyv Bio and Twist Bioscience, Claude produced effective binders for 14 of 15 targets, with a molecular-level success rate of 22%–35%, well above the typical 10%–15% for human experts; some designs matched or exceeded best-known binding affinities. Many commenters framed it as AI moving from assisted analysis to autonomous research, though a credit-attribution controversy remains.

Confirmed

  • The task was de novo binder design: 14 of 15 targets hit (@AnthropicAI, @ncfrey); molecular-level success rate 22%–35% (@AnthropicAI; @DrSingularity cites 23%–35%; @Hesamation rounds to 35%). The “93% success” relayed by @ycombinator is the 14/15 per-target hit rate, not the molecular-level figure; model names such as “Claude 3.5 Sonnet” or “Claude 3.7” in some posts are inaccurate—the models involved were Mythos Preview and Opus 4.8 (@ncfrey)
  • Claude orchestrated multiple open-source protein design and co-folding models, iterating like a human expert; community breakdowns list RFdiffusion, ESMFold2, PXDesign, Genie, BoltzGen (@karstenkreis, @chaitjo, @teortaxesTex)
  • With no design or prediction tools pre-installed, Claude handled the full pipeline—target research, epitope selection, installing/running open-source tools, candidate generation, filtering; 1,320 measurable designs were tested by Adaptyv Bio and Twist Bioscience (@ncfrey, @rohanpaulai)
  • Some designs matched or beat best-known binding affinities (@DrSingularity); challenges remain, e.g. no successful binder among 90 designs for maltose-binding protein (MBP) (@AllThingsApx)
  • Anthropic published the prompts used, such as “no need to rest” and “do not wait for approval” (@Hesamation)
  • Per @anshulkundaje, Anthropic responded to criticism that binder design is not drug discovery and to calls for head-to-head comparison with phage libraries; per @Polymarket, Anthropic also disclosed Claude's automated chemistry research capabilities

Unconfirmed

  • Expert Andrew Iaginine (via @teortaxesTex) and Ruxandra Teslo (via @AllThingsApx) argue the binders were actually produced by specialized open-source models like RFdiffusion, with Claude only orchestrating; Anthropic has not directly addressed this in detail

Why it matters

  • If the results hold, LLMs in science extend from analysis into creative R&D, potentially compressing early-stage drug design timelines
  • The core dispute—whether orchestrating specialized models counts as “autonomous design”—will shape how AI scientific capability is evaluated going forward

25 more related posts →

Episode 2 · Claude Protein Design Agent Accused of Replicating Protein Hunter Protocol (2026-08-20, 3 posts)

Computational biologist Sergey Ovchinnikov found that a Claude-based protein binder design agent replicated his team's earlier Protein Hunter protocol without attribution, with the protocol even referenced in its prompt.