Claude Designed Working Protein Binders on 14 of 15 Lab Targets

Anthropic reported that Claude models designed working protein binders against 14 of 15 targets and processed contract lab analytical files in minutes, results the company frames as early evidence that language model agents can shoulder real drug discovery labor.

The August research post describes two experiments run inside Claude Science. In the first, Claude Opus 4.8 and a Mythos preview executed autonomous binder design campaigns lasting 24 to 48 hours per target. Contract research organizations Adaptyv Bio and Twist Bioscience synthesized the sequences and measured binding without knowing which designs came from AI.

Hit rates above typical human campaigns

Anthropic said 354 of 1,320 submitted designs bound successfully, a 27 percent hit rate overall. When Mythos tackled targets one at a time, the rate rose to 35.1 percent. The company compares that to the 10 to 15 percent hit rates common in human led protein design programs.

Standout cases include RBX1, where Mythos reached a 40 percent success rate versus 3.7 percent for human entrants in an Adaptyv competition, and TNFα, a target relevant to anti inflammatory drugs such as Humira. Anthropic cautions that binders are not finished medicines. Affinity is only the first gate before developability, toxicity, manufacturing, and trials.

The agents used open source structure prediction and design tools, which Anthropic notes makes similar campaigns reachable for academic labs with enough compute.

Chemistry analysis without vendor software

In a second test, Claude Opus 5 received raw nuclear magnetic resonance and liquid chromatography mass spectrometry files plus a two sentence prompt. Working without proprietary vendor GUIs, it returned purity and hydrogen count results in 19 and 23 minutes that matched the contract lab's numbers, including 96.4 percent purity versus 96.33 percent reported manually.

That workflow targets a mundane bottleneck. Medicinal chemists spend hours interpreting spectra after each synthesis step. Automating the readout does not replace physical experiments, but it can compress iteration loops if reliability holds across compound classes.

Verification limits

These are company run studies, not yet independent peer review. Anthropic acknowledges life sciences verification is expensive, which slows AI progress relative to math benchmarks where answers check quickly. External labs tested binding, yet the overall campaign design and success metrics still come from Anthropic's write up.

Regulatory and IP questions remain separate. Other August coverage highlighted how AI proposed drugs can still list only human inventors on patents under U.S. rules requiring natural persons.

For the industry, the practical takeaway is narrower: Claude demonstrated end to end binder campaigns with measurable lab confirmation, not slide deck biology.

Sources

Anthropic research blog post on Claude protein design and chemistry analysis (August 2026)

Pharmaphorum summary of Adaptyv Bio and Twist Bioscience validation (August 2026)

The Next Web reporting on hit rates and competition comparisons (August 2026)