Anthropic published results on 20 August from an autonomous protein design campaign run inside Claude Science. The model was given 15 drug-relevant targets, among them PD-L1, TREM2, TNF-alpha and EGFR, and produced confirmed binders against 14 of them. Adaptyv Bio and Twist Bioscience synthesised and tested the designs in the wet lab without modifying them first.
The hit rates were 26.7% for a preview model and 22.6% for Opus 4.8, against an industry baseline of 10% to 15%. On one target, RBX1, the preview model reached 40%, where human entrants in an Adaptyv Bio competition had managed 3.7%, and its best design outperformed the winning human entry.
The mechanism deserves precision, because the headline invites a wrong reading. The model did not itself predict protein structure. It operated publicly available specialist design and co-folding tools that the field already uses, selecting targets, running the tools, interpreting results and iterating on them. What was automated is the expert judgment wrapped around the software, historically weeks or months of specialist work for each target, rather than the underlying science.
Binding is also the first step and the smallest one. A molecule that sticks to its target is not a drug: it still needs selectivity, stability, a delivery route, manufacturability and years of trials, and no AI-discovered compound holds full FDA approval as of today. What this compresses is the computational front of the pipeline, which moves the constraint downstream onto wet-lab validation rather than removing it. The interesting number in a year will be how many of these 14 survive contact with a cell.
