Claude designed working protein binders against 14 of 15 targets and completed a contract lab's analytical chemistry workflow in under 25 minutes, according to new results from Anthropic. The protein design campaign, run with minimal human input, achieved hit rates between 22.6% and 35.1% depending on the setup - well above the 10-15% typical in the field. Some designs bound their targets several times more tightly than the best previously published results.
The results point to a shift in how AI can support experimental science. Protein design campaigns that historically took specialists weeks or months per target now ran end-to-end in 24-48 hours. The chemistry work - interpreting raw NMR and LC-MS instrument files - took 23 and 19 minutes respectively, matching what a contract lab produced over several days.
Claude ran the protein design campaigns itself
Anthropic gave Claude a prompt, access to GPUs, and a list of 15 protein targets drawn from standard benchmarks plus two novel ones. The model then handled everything: choosing where to design against each target, generating candidate structures, running multiple rounds of computational optimization, and screening for designs that would express and bind.
External evaluators Adaptyv Bio and Twist Bioscience produced and tested Claude's designs in the lab. Claude succeeded against 14 of 15 targets, producing 354 confirmed binders from 1,320 total designs. That volume alone is notable - the two largest public collections of de novo protein designs contain roughly 770 binders against 40 targets.
The model used publicly available specialist protein design and folding models, orchestrating them across multiple design cycles. "Claude conducted all of the work that goes into designing a binder, which can take a human operator weeks," the company said. "It chose where on each protein target to design against; generated candidate structures and sequences by orchestrating several structure design, sequence design, and co-folding models; ran the designs through multiple cycles of in silico optimization; and computationally screened for novel, diverse candidates."
Some targets proved harder than others. Claude produced only modest binders against BBF-14, a de novo designed protein used as a benchmark precisely because it is novel, and failed entirely against maltose-binding protein, a large flexible bacterial protein with a smooth surface that gives a binder little to grip. In one case, Opus 4.8 succeeded on a target where the newer Mythos Preview failed: TNFα, a therapeutically important target that requires binding a groove formed by two protein copies. Opus 4.8 designed cross-reactive binders that worked across human, monkey, and mouse versions of the protein.
Anthropic noted that agentic biological discovery is dual-use and that protein design capabilities remain unavailable for general access in its most capable model while it builds trusted access programs for scientists. The company said it plans to release the full prompts and experimental data. For researchers exploring how these tools apply to their own work, AI for Science & Research covers the latest developments in this area.
Claude processed raw instrument files without vendor software
The second experiment tested a different skill: interpreting data from instruments chemists use daily. When a chemist synthesizes a compound, they must confirm its identity and purity using NMR spectroscopy and liquid chromatography-mass spectrometry. The instrument runs take minutes; analyzing the output takes far longer.
Anthropic gave Claude Opus 5 raw files from a contract lab - the proprietary binary formats that normally require the instrument manufacturer's software - plus a two-sentence prompt. Claude converted the NMR data into a calibrated spectrum, identified 18 peaks, and counted hydrogens for each. It flagged four broad peaks as likely nitrogen- or oxygen-bound hydrogens and proposed the standard follow-up: add heavy water to confirm. The contract lab had independently run that exact test three days later.
For the LC-MS data, Claude reverse-engineered the undocumented vendor format, then verified it had read the file correctly by reproducing the instrument's own recorded totals across all 2,664 scans before analyzing anything. It returned the separation trace, mass and UV spectra, a purity table, and the compound's molecular mass - plus reusable code for reading such files and caveats about the limits of the instrument's precision.
The results matched the lab's own analysis. Hydrogen counts were within 0.08 of the lab's figures, and Claude measured purity at 96.4% versus the lab's 96.33%. The lab's finished report arrived four days after the first spectrum was acquired; Claude produced its results within 25 minutes.
For scientists who want to build these workflows themselves, the AI Learning Path for Research Scientists provides structured training on applying AI to research tasks.
Why this matters for research scientists
The two experiments address different bottlenecks. Protein design has been gated by expertise: running specialist models well enough to produce binders has required computational protein engineers. Claude's campaign suggests the orchestration layer can be automated, letting scientists specify a target and evaluate what comes back.
Chemistry analysis is gated by time. Every compound a chemist makes requires the same tedious interpretation of spectra. Claude's ability to process raw instrument files and propose the same follow-up experiments a human would run suggests this work can be compressed from hours or days to minutes - with the caveat that scientists should verify results and understand the limits of what the model reports.
Neither result means the wet lab is obsolete. Every protein design still had to be physically tested, and the chemistry analysis was performed on a routine quality-control sample, not a novel compound. But the bottleneck in both cases shifted from the human expert to the instrument. That is where AI is changing the pace of research.
Your membership also unlocks:
