Anthropic Claude protein binder design: 2026 wet-lab results explained
Anthropic says Claude ran binder campaigns with no human design calls. Two CROs wet-tested the output: binders on 14 targets, hit rates about 22–35%. Not a drug. Dual-use tools stay off Fable 5’s general access.
On 18 August 2026 Anthropic posted a life-science experiment. The news was not a new leaderboard. It was Claude finishing protein-binder campaigns, then two contract labs measuring the proteins in wet assays.
The models were Mythos Preview and Claude Opus 4.8. The post is titled How Claude is accelerating protein design and analytical chemistry. A technical report landed the same day. The figures are Anthropic’s. The wet tests were run by Adaptyv Bio and Twist Bioscience.
Which numbers the post prints
| Setup | Wall clock | Reported hit rate |
|---|---|---|
| Multi-target Mythos | 48 h | 26.7% |
| Multi-target Opus 4.8 | 48 h | 22.6% |
| Single-target Mythos | 24 h | 35.1% |
| Typical campaign (Anthropic’s comparison) | — | about 10–15% |
- Targets and n
- Sixteen targets were chosen. Mature GDF-8 aggregated, so both CROs lacked a readable assay. Analyses use 15 targets and 1,320 designs.
- Overall hits
- The report abstract: 354 of 1,320 bound, about 27%. Among designs ranked first in each campaign, 49% bound.
- RBX1
- An open contest had 9 binders in 245 designs. Claude had 28 in 90. Tightest KD 3.9 nM versus 45 nM for the contest winner re-made on the same plate.
- Misses
- None of 90 MBP designs was confirmed to bind. BBF-14 produced only a few modest-affinity binders.
How the campaign was run
- 1
One protocol
A protocol prompt named no epitope, scaffold, or sequence. Humans picked targets, funded GPUs, and placed synthesis orders.
- 2
The model installed tools
Claude built open-source design and structure models from public repos, chose sites, optimized, and ranked 30 designs per target.
- 3
Two labs, as delivered
Adaptyv Bio and Twist Bioscience synthesized and assayed the sequences as delivered. No later design edits.
- 4
A separate chemistry test
Generally available Opus 5 got raw files and two short prompts. NMR took 23 minutes; LC-MS took 19.
Second test: reading spectra
NMR
It built a calibrated spectrum and a table of 18 peaks with hydrogen counts. It flagged broad peaks likely on N or O and proposed a heavy-water check. The lab had run that check three days later.
LC-MS
The vendor format is undocumented. Claude first reproduced the instrument’s totals for all 2,664 scans, then a purity table. Purity 96.4% versus the lab’s 96.33%.
Time
The post says the lab’s finished report arrived about four days after the first spectrum. Claude processed both files in parallel inside about 25 minutes and wrote a report.
What is still missing
No independent journal peer review is cited. No date is given for opening protein design on Fable 5’s general access. Prompts, computational models, and binding data sit on Hugging Face as Anthropic/claude-protein-binder-design.
# Sources
# Anthropic, 2026-08-18: How Claude is accelerating protein design and analytical chemistry
# Shanehsazzadeh, 2026-08-18: Autonomous de novo protein binder design with Claude
# Hugging Face: Anthropic/claude-protein-binder-designIs this a drug?
No. The post says minibinders are not a standard therapeutic modality. A tight binder is an early step. The wet tests confirm binding, not a medicine.
Who measured the numbers?
Claude produced the designs. Adaptyv Bio and Twist Bioscience measured binding. Hit rates and affinities are Anthropic’s compilation. The company says more characterisation will follow.
Can a regular chat user repeat this?
Protein design stays off general access in Fable 5. The chemistry files used generally available Opus 5. The scientist access program has no public date.