Skip to main content
In August 2026 Anthropic had 1,440 Claude-designed miniproteins synthesized and tested independently by two labs, Adaptyv Bio and Twist Bioscience, neither able to see the other’s data. The two labs ran genuinely different formats: cell-free monovalent designs on single-cycle SPR here, HEK293-expressed Fc fusions on a high-throughput SPR array there. On the 1,235 designs measured at both labs, our binder and non-binder calls agreed with Twist’s 89% of the time, with a Cohen’s κ of 0.71. The raw sensorgrams, fits and per-design labels from both labs are public under CC BY 4.0, so the comparison can be redone rather than taken on trust.

Autonomous de novo protein binder design with Claude

The full report, 18 August 2026, including the assay conditions at both labs and the adjudication rule used where they disagreed.

Anthropic/claude-protein-binder-design

Raw traces, fits and per-design labels from both labs, plus the computational models for all 1,440 designs.
Agreement here is on binder versus non-binder, not on KD. The two formats differ in valency, so the affinity values are not directly comparable, and the report does not claim otherwise.