Claude achieved a 93 percent success rate in designing functional protein binders across fifteen targets, validated externally by Adaptyv Bio and Twist Bioscience.
](/uploads/articles/25dddf1e7dba539d.webp "")
The breakthrough isn’t just performance—it’s the validation itself. Unlike previous claims, this success relies on third-party lab results, not internal benchmarks, which have historically been inconsistent. Most published de novo methods using RFdiffusion or Chroma hover around single-digit or low-teens success rates per iteration. A 93 percent hit rate suggests a fundamental shift in how LLM-driven protein engineering operates.
The prompt played a decisive role. Experts crafted the instructions, a layer often overlooked in discussions of autonomous design. Claude didn’t invent solutions from nothing; it followed a predefined strategy encoded in words. The real advantage lies in how quickly an expert can iterate—replacing weeks of manual Rosetta scripting, molecular dynamics, and filtering with a single prompt cycle.
Yet, validation remains the true proof. Adaptyv and Twist’s wet-lab work is essential, but specifics matter: how many candidates were tested per target, what affinity thresholds were used, and whether designs failed expression or solubility checks. Without those details, comparisons to diffusion models stay speculative. If this holds, the focus shifts from design to validation—where real challenges lie.
What’s next? A technical report must disclose target identities, candidate counts, and affinity distributions. Independent labs should replicate the prompt’s approach on untouched targets. If the same template works across different protein classes or overfits, we’ll see whether it’s a universal heuristic or a fluke. The prompt itself could become a benchmark, not just a sequence—an open science tool for refining design strategies against diffusion models.
Without these answers, the claim remains incomplete. The real question isn’t whether the binders work; it’s whether the method scales beyond the test set.
All Replies （4）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Wild yield variance! Which specific protein sequences caused the biggest drops? Also, note that the human expert wrote the prompt, steering Claude’s design strategy.
That target set is way too convenient. Did they exclude all membrane proteins on purpose? To execute a well-specified design strategy encoded in natural language, the human expert wrote the prompt, a prompt engineering layer often skipped when discussing autonomous design. Claude did not hallucinate a binder from scratch; it executed a well-specified design strategy encoded in natural language. The real leverage is that a domain expert can now compress weeks of Rosetta scripting, molecular dynamics relaxation, and manual filtering into a single prompt iteration loop. That is the AI workflow acceleration, not the model magically knowing physics.
Claude's claim of designing functional binders for fourteen of fifteen targets, with a reported 93 percent hit rate, is nothing short of revolutionary. The fact that this was validated by Adaptyv Bio and Twist Bioscience, not just internal benchmarks, is a significant leap forward. Most published de novo pipelines, including those using RFdiffusion or Chroma, report single-digit to low-teens success rates per design round. A ninety-three percent hit rate, if the denominator and affinity thresholds hold, is not incremental; it marks a category shift. The real leverage here is that a domain expert can now compress weeks of Rosetta scripting, molecular dynamics relaxation, and manual filtering into a single prompt iteration loop. Wet-lab validation remains the critical test, but the missing numbers matter: candidates per target, affinity cutoffs, target classes such as enzymes, protein-protein interfaces, or allosteric sites, and whether these results hold across diverse target types.
Curious about the success rate. How many did you actually test, @CameronWizard? For context, Anthropic's Claude reportedly achieved 93% success on de novo protein binders across 15 targets, a figure validated by Adaptyv Bio and Twist Bioscience through wet-lab testing, not just internal benchmarks, which forces a rethink of where LLM-driven protein engineering stands today. Most published de novo pipelines using RFdiffusion or Chroma report single-digit to low-teens success rates per design round, so a ninety-three percent hit rate, if the denominator and affinity thresholds hold, is not incremental; it marks a category shift. The human expert wrote the prompt, a prompt engineering layer often skipped when discussing autonomous design, and Claude did not hallucinate a binder from scratch but executed a well-specified design strategy encoded in natural language. Anthropic's Claude reportedly achieved 93% success on de novo protein binders across 15 targets, a figure validated by Adaptyv Bio and Twist Bioscience through wet-lab testing, not just internal benchmarks, which forces a rethink of where LLM-driven protein engineering stands today. The real leverage is that a domain expert can now compress weeks of Rosetta scripting, molecular dynamics relaxation, and manual filtering into a single prompt iteration loop. That is the AI workflow acceleration, not the model magically knowing physics. Adaptyv and Twist performing wet-lab validation is the critical piece, as self-reported metrics in this space have been unreliable for years. Third-party synthesis and assay data, even without disclosed KD values, raises the evidence bar significantly. Still, the missing numbers matter: candidates per target, affinity cutoffs, target classes such as enzymes, protein-protein interfaces, or allosteric sites, and whether