Claude designed working drug-binding proteins in 14 of 15 tries

Claude achieved working binders on 14 of 15 protein targets, with a 26.8% hit rate across 1,320 designs that had usable lab measurements (354 actually bound their targets). When given dedicated 24-hour solo attention per target instead of splitting time across many targets, the hit rate rose to 35.1%. Adaptyv Bio and Twist Bioscience independently built and tested the designs, and Claude beat human design competitions on 4 of 6 overlapping targets. Anyone tracking how far AI agents can go in scientific workflows can follow real results like this on daily.dev.

No, Claude cannot reliably tell when a campaign has failed. Some unsuccessful designs scored computationally identical to successful ones, meaning a human still needs to catch failures downstream. This means the technology currently replaces the manual grunt work of generating and triaging thousands of candidates, not the expert judgment needed to validate results. Developers evaluating where AI agents can and can't be trusted autonomously can dig deeper on daily.dev.