What's this about?
Many drugs work by binding to a specific target in the body and blocking or changing what it does, so an early step in drug development is designing a molecule that binds tightly to that target. Traditionally this means weeks or months of expert work per target, sifting through many candidates to find the few that work. Anthropic tested whether Claude could do this from scratch. Given a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 of 15 targets. The designs were then built and tested in the lab by Adaptyv Bio and Twist Bioscience, independently of Anthropic.
What exactly is "de novo protein design"?
Proteins are chains of amino acids that fold into three-dimensional shapes, and the shape determines what they stick to. De novo design means inventing a new protein — a sequence nature never produced — whose folded shape will grip a chosen target tightly, like designing a key for a lock you can see but have never had a key for. The traditional approach is largely search: generate enormous numbers of candidates, filter computationally, then test survivors in a lab. What makes 14 out of 15 notable is the hit rate. Wet-lab validation typically kills most designs, so a high fraction binding on the first pass is the meaningful number.
Why is this worth noticing?
- Independent wet-lab validation is what separates this from a benchmark. Two outside companies built and tested the molecules physically. This isn't a simulation score or a held-out dataset — the proteins either bound or they didn't.
- The human expert wrote the prompt, and that matters both ways. It means this isn't autonomous science; it also means the expert's contribution was a prompt rather than months of candidate screening, which is the actual claim about time saved.
- It lands the same week as the "AI will cure diseases" argument. Several people, including a clinician, spent this week pointing out the gap between designing a candidate molecule and an approved drug — phase 3 trials alone take about five years. This result is real and sits squarely on the early side of that gap.