ANALYSIS

Claude Designed Working Proteins for 14 of 15 Targets

Two iridescent three-dimensional protein structures on a black background, one dissolving into digital pixels, representing an AI-designed protein
Claude's designs were synthesized and tested at the bench, not scored in simulation. Source: MIT Technology Review
TLDR

Claude ran the whole design pipeline as an agent, then labs confirmed the results

Anthropic reported that Claude did not just suggest protein sequences. It ran the campaign. Given a long prompt, internet access, tool connectors, and a large pool of GPUs inside Anthropic's research environment, Claude Science, the model chose where on each target protein to bind, orchestrated existing open-source design and folding models, screened the candidates, and selected which to send for synthesis. It behaved as an autonomous agent operating the pipeline, not the autocomplete inside it.

The compute behind that autonomy was real: up to 12,500 Nvidia H100 hours across a single 48-hour session for one approach, and up to 2,500 H100 hours per target for another. The output then left the computer and went to the bench. Two independent laboratories, Adaptyv Bio and Twist Bioscience, synthesized 1,320 designs and tested them by surface plasmon resonance across five target concentrations. 95 percent of the designs expressed, and 354 came back as confirmed binders across 14 of the 15 evaluated targets.

Source: @adaptyvbio, the lab that ran the physical validation of Claude's designs.

Why beating a 10 percent hit rate rewrites drug-discovery economics

Designing a protein that binds a chosen target is the first step in building antibodies, diagnostics, and many therapeutics. The reason it is slow and costly is the failure rate. At a 10 percent hit rate, a lab synthesizes and tests roughly ten candidates to find one that works, and each test burns reagents, instrument time, and weeks. Claude's headline configuration reached 35.1 percent on single targets and 26.7 percent across a harder multi-target mode, while Opus 4.8 landed at 22.6 percent. Moving the hit rate from one in ten to one in three does not make the work three times faster in a straight line. It changes which projects are affordable to attempt at all, because the ratio of wet-lab spend to usable output shifts by a factor that compounds across every target a company chases.

Bar chart comparing wet-lab confirmed protein binder hit rates: field standard 10 to 15 percent, Claude Opus 4.8 at 22.6 percent, Claude multi-target at 26.7 percent, Claude single-target at 35.1 percent, and Claude on RBX1 at 40 percent versus 3.7 percent for human entrants
Wet-lab confirmed hit rates from Anthropic's protein design campaign, validated by Adaptyv Bio and Twist Bioscience. Source: Anthropic research, August 2026.

Where Claude won, and where it failed

The result is not uniform, and the range is the most honest part of it. On RBX1, a component of the cell's protein-disposal machinery and a prized drug target, Claude confirmed binders 40 percent of the time and produced a binder roughly ten times tighter than the previous best. On TREM2, an immune receptor tied to Alzheimer's research, it hit 80 percent. On the enzyme 15-PGDH it beat the prior competition winner's binding strength by more than fiftyfold. It also failed outright on maltose-binding protein, could not surpass the best existing binder for the Nipah virus target, and one target was dropped for measurement-quality reasons.

TargetRelevanceClaude resultPrior benchmark
RBX1Cancer, protein degradation40% hit, 3.9 nM affinity3.7% human hit, 25.7 nM
TREM2Alzheimer's immunology80% hit rate10 to 15% field norm
15-PGDHFibrosis, aging33.4 nM affinity1.7 uM competition winner
MBPBenchmark control0% (no binders)field norm
Nipah virusAntiviralDid not beat best bindercompetition best held
Selected targets from 15 evaluated. Source: Adaptyv Bio wet-lab benchmarking, August 2026.

The chemistry result matters as much as the biology

The less-noticed half of the disclosure is analytical, not biological. Claude Opus 5 was handed raw spectroscopy files and asked to do the interpretation a chemist does by hand.

Claude, working within Claude Science, returned processed NMR and LC-MS results in 23 and 19 minutes, respectively, working in parallel.
Anthropic research, August 18, 2026

Its purity reading came out at 96.4 percent against the lab's 96.33 percent, and during the run it caught and corrected one of its own errors. A human chemist typically needs 30 minutes to an hour per sample. A system that can propose a molecule and also read the instrument output that verifies the molecule removes the human bottleneck from both ends of the routine design-test cycle.

A model that can both propose a molecule and interpret the instrument that checks it is not acting as a lab tool. It is acting as a lab member.

What a general model in the wet lab means for biotech

For most of the AI-for-biology era, the sharpest results came from specialized systems built for one job: the AlphaFold lineage for structure, newer models like Evo and Chai for design. Anthropic's campaign is a general reasoning model walking into that arena and posting competitive numbers by orchestrating those same specialized tools rather than replacing them. That reframes the build-versus-buy question for every biotech deciding whether to train a bespoke model or point a frontier agent at the problem, and it hints at why Anthropic is investing in a life-sciences workbench at all.

By the Numbers
Targets with confirmed binders14 of 15 evaluated
Designs synthesized and confirmed354 of 1,320, 95% expressed
Peak hit rate35.1% vs a 10 to 15% field norm
RBX1 hit rate40% vs 3.7% for human entrants
Compute per sessionup to 12,500 Nvidia H100 hours
Spectra analysisunder 25 minutes at 96.4% measured purity
Source: Anthropic research and Adaptyv Bio wet-lab benchmarking, August 2026.

The caution is worth stating plainly. Fifteen targets is a small set, a binder is not a finished drug, and two of the hardest targets defeated the model. But for two years the promise of AI in biology was prediction, telling you what a protein already is. Anthropic's campaign points somewhere else: the model designed proteins that did not exist, and independent labs confirmed the designs work, which turns a frontier model from a description of biology into a way to manufacture it.

Santage is committed to independent, transparent journalism. This article is produced in accordance with Santage's Editorial Standards and aims to provide accurate and timely information. Readers are encouraged to verify information independently.