Anthropic's Claude hits 26.8% on de novo protein design, beating the 10-15% industry norm

Anthropic just ran its Claude models through the full protein design stack and came back with numbers that could make biotech VCs sit up. In two experiments, Claude agents designed minibinders—small proteins that lock onto disease targets—and 354 of 1,320 lab-tested candidates actually bound. That’s a 26.8 percent hit rate, roughly double the 10 to 15 percent the company cites as typical for publicly documented campaigns. When Claude ranked its own designs first, the success rate jumped to 49 percent.

Here’s the twist that separates this from a typical AI-for-science demo: Anthropic didn’t build a protein model. Claude installed and orchestrated open-source tools the field already uses—RFdiffusion, SolubleMPNN, ESMFold2, and others—running them through a 16,000-word protocol prompt that covered everything from scientific guidance to budget discipline. Human involvement was minimal after targets were selected. The compute bill came to $50,000 per multi-target campaign and $10,000 per single target on Modal. Independent validation was handled by paid contract labs Adaptyv Bio and Twist Bioscience.

The most striking result came from RBX1, a target tied to protein degradation. An open design contest had produced just 9 binders from 245 candidates. Claude managed 28 of 90. The contest’s winning design bound at 45 nM; Claude’s best hit 3.9 nM, roughly ten times tighter. On TNFα—a notoriously difficult immune target behind five approved drugs, including Humira—Claude produced 12 binders from 150 designs, though all traced back to just four scaffolds, meaning they’re less independent than the raw count suggests.

The caveat is obvious: an independent review is still pending, and these are early-stage results, not a drug. But the direction matters. If a general-purpose language model can run the entire design stack with this level of autonomy, the bottleneck shifts from software availability to lab capacity and target selection. That’s a fundamentally different economic picture for early-stage discovery than the one biotech has operated under for decades.