Image: generated by METAL LAB
Summary
- Of 1,320 miniprotein binders Claude designed autonomously, 354 bound their targets for a 26.8% hit rate — well above the 10-15% typical for these campaigns.
- A follow-up analysis found that ten structure-prediction models scored only 0.55 to 0.75 in AUC, and many binding failures turned out to actually be expression failures.
1,320 Designs, Tested in the Lab
A new set of results directly tests how often AI-designed proteins actually bind in the lab — and whether it's possible to pick the winners before ever synthesizing them. The findings come from an autonomous protein-binder design campaign Anthropic disclosed, plus a follow-up analysis that dug into the resulting data.
Start with the scale: Claude designed 1,320 miniprotein binders, tested against 15 targets, and 354 of them bound — a 26.8% hit rate.
Anthropic says typical protein-design campaigns today land in the 10-15% range, so this result comes in at roughly double.
Binders are small proteins engineered to attach to a target protein. Because they're much smaller than antibodies, they're called miniprotein binders.
The team originally selected 16 targets, but one had to be dropped because the target protein itself aggregated, making evaluation impossible. Of the remaining 15, binders emerged for 14.
Best Results Came From 24 Hours Per Target
The campaign ran in two modes. In a multi-target mode where several targets were assigned at once with a 48-hour window, the Mythos preview model scored 26.7% and Opus 4.8 scored 22.6%. In a single-target mode where each target got its own 24-hour window, the Mythos preview model rose to 35.1%.
Anthropic also disclosed the compute involved: the multi-target mode used up to 12,500 H100-hours over a 48-hour wall-clock window, while the single-target mode used up to 2,500 H100-hours per target over 24 hours. Within those windows, there was no cap on token usage or the number of sub-agents.
Human involvement was minimal. After the initial instructions, Anthropic says, people only approved access permissions and monitored the infrastructure. Claude had access to the internet, a document collection, Google Drive and Slack, Gmail, a bioRxiv connector, and GPUs, and it chained together structure-design, sequence-design, and co-folding models on its own.
Some Targets Beat Human Competition Records
Results vary widely by target — and that variance is the real story of this campaign.
For RBX1, the hit rate reached 40%. In a public competition on the same target, human participants managed only 3.7%.
The top design bound more tightly than the competition's winning entry. The binding-strength metric KD dropped from 25.7 nanomolar to 3.9 nanomolar — a lower KD means tighter binding.
For 15-PGDH, the previous best of 1.7 micromolar fell to 33.4 nanomolar — more than a fiftyfold improvement. For TREM2, the hit rate climbed to 80%, versus 38.3% in an earlier competition.
There's a flip side, too. For MBP, 90 designs were made and not a single one bound. For TNFα, Opus 4.8 succeeded while the Mythos preview failed, and Anthropic simply noted it doesn't know why.
Overall, at least six targets produced high-affinity binders with KD under 10 nanomolar, and at least four targets matched or exceeded the best previously reported affinity.
| Target | Claude Hit Rate | Benchmark | Best Binding Strength (KD) |
|---|---|---|---|
| TREM2 | 80% | Prior competition: 38.3% | |
| RBX1 | 40% | Human competition: 3.7% | 25.7nM → 3.9nM |
| 15-PGDH | 1.7µM → 33.4nM | ||
| MBP | 0% (90 designs) | No binders | |
| Overall | 26.8% (354 of 1,320) | Typical campaigns: 10-15% | 14 of 15 targets bound |
Targets not shown in the table make up the rest of that 26.8% overall figure. The gap between the successes and the failures is the biggest signal in this campaign.
95% of Designs Actually Expressed in the Lab
Validation was handled independently by two companies, Adaptyv Bio and Twist Bioscience. A case study Adaptyv Bio published lays out the experimental methods in detail.
The designed amino-acid sequences were converted into DNA and synthesized automatically, then the proteins were produced using cell-free protein synthesis and robotic pipetting. Binding was measured by surface plasmon resonance, with duplicate measurements across five target concentrations.
One notable number here is the expression rate: 95% of the designs were actually produced as proteins, matching the top performances seen in competitions. That means the sequences Claude generated weren't difficult to work with in the lab.
But Picking Winners Before Synthesis Is a Different Problem
A separate follow-up analysis of this dataset asks a different question: can computational scores tell binders from non-binders before anything is ever synthesized?
Running ten different structure-prediction models against the same data produced AUC scores scattered between 0.55 and 0.75. AUC measures the probability that a model scores a randomly chosen binder higher than a randomly chosen non-binder — 0.5 is a coin flip, 1.0 is perfect. To trust a single filter as a hard cutoff, you'd want something around 0.9, and these scores fall well short.
Combining multiple predictors didn't change the picture much. A rank-averaged consensus score performed best at 0.75, versus 0.73 for the single best predictor — ensembling only nudges things up slightly.
| Selection Method | AUC |
|---|---|
| Rank-average consensus | 0.75 |
| Single best predictor | 0.73 |
| Median consensus | 0.73 |
| Unanimous consensus | 0.72 |
| Expression level alone | 0.60-0.65 |
| Reference: needed for a stand-alone filter | 0.9 |
The value shows up elsewhere. If your budget only covers synthesizing 50 designs, picking by the consensus score raises the hit rate roughly 1.5x over random selection. It's not a scalpel that can cleanly cut winners from losers, but it's a useful ruler for ranking them.
A significant share of what got recorded as binding failures turned out to actually be expression failures. Median expression titer was about 2.0 mg/mL for binders versus about 0.5 mg/mL for non-binders. Guessing binding status from expression level alone already gets an AUC of 0.60 to 0.65. When a protein simply doesn't express properly, the binding assay never really gets a fair test — but the result sheet just marks it a failure regardless.
The two labs' calls also frequently disagreed. Cohen's kappa, a measure of agreement, fell between 0.40 and 0.50, and KD values differed between the two labs by a median of about 2.2x. When the ground truth itself is noisy, no predictor can outperform that noise.
Editor's Take
The number to pay the most attention to in this campaign isn't 26.8% — it's 3.7%. That's more than a tenfold gap. Protein design has long been a field that demanded a lot of skilled people, and that premise just took a hit.
At the same time, the MBP story shouldn't get left out. Ninety designs, zero binders. And for TNFα, one model succeeded while another failed, and Anthropic wrote plainly that it doesn't know why.
It's rare to see a disclosure that doesn't scrub the failures off the scorecard. Those two lines are exactly what makes the rest of the numbers more credible.
Yesterday's story here was about speed — designs made and bound within 24 hours. Now that the full campaign numbers are out, the picture has shifted. What matters more than speed is that several targets beat human competition records.
For practical purposes, the follow-up analysis is more useful than the headline number. Filter by expression first, don't trust any single predictor but combine rankings, and break results down by target. All three are steps teams can apply today without buying a new model.
The finding that expression failures can masquerade as binding failures is especially worth checking against any biotech team's own pipelines. A good share of AI protein-design scorecards compiled so far may not have made that distinction, which means some failures blamed on the model may actually have been production-stage problems.
Anthropic itself drew clear limits: confirming hit rates and affinities will require broader characterization, miniprotein binders aren't a standard therapeutic modality, and high-affinity binders are only the first step in drug development.
Still, this campaign clearly moves the line. AI protein design has moved past the question of whether it can be done at all. What comes next is deciding what to make — and these numbers together show that choosing is still, for now, work left to human hands.
Comments