Claude designs proteins from scratch and beats human benchmarks

Designing a brand-new protein that can lock tightly onto a biological target is usually a grind. For computational biologists, working through a single target can take months of scripting, tweaking, and screening.

Anthropic just demonstrated that general-purpose AI can run that entire early-stage playbook on its own. Using Claude Mythos Preview and Opus 4.8, the models designed minibinders against 15 biological targets, succeeding on 14 of them.

To keep things honest, Anthropic had Adaptyv Bio and Twist Bioscience physically synthesize and test the molecules in real wet labs. The results held up.

Crushing the industry baseline

In standard de novo protein design, getting 10% to 15% of your computationally designed molecules to actually bind in a lab dish is considered standard. Claude beat that by a wide margin.

When tasked with designing against all targets at once over a 48-hour run, Claude hit success rates between 22.6% and 26.7%. When focused on one target at a time across 24-hour sessions, Mythos Preview reached a 35.1% hit rate.

Across the campaign, Claude produced 354 confirmed binders from 1,320 designs. For at least six targets, it found high-affinity binders. On four targets, it matched or outperformed the best previously published affinities.

Anthropic also threw in two novel competition targets, 15-PGDH and GDF-8, to make sure the model wasn't simply regurgitating past research from its pretraining data. Claude handled both.

In a separate chemistry test, Claude Opus 5 was handed raw NMR and LC-MS instrument files with a two-sentence prompt. In under 25 minutes, it delivered a purity reading of 96.4%, virtually identical to the analytical lab's own 96.33% report that typically takes days to turn around.

The model as lab coordinator

What makes this experiment different from earlier AI biology breakthroughs is the workflow.

Claude wasn't just predicting a shape. It was acting as the orchestrator. Given an initial prompt, web access, and connections to tools like BioRxiv and cloud storage, the model worked autonomously.

It picked binding sites, called existing open-source folding and sequence models, iterated through optimization cycles, and filtered candidates for solubility and novelty. Human intervention was limited to monitoring servers and approving network requests.

It did take serious horsepower, though. The multi-target run consumed up to 12,500 Nvidia H100 hours, while single-target runs used up to 2,500 H100 hours each.

Currently, Anthropic still keeps advanced life-science features blocked inside its flagship model for safety reasons, but it plans to roll out a dedicated access program for vetted scientists soon.

Why it matters

For biotechs and academic labs, this shifts the bottleneck from computational bandwidth to wet-lab throughput.

Early-stage drug discovery has traditionally required specialized computational teams writing custom pipelines. If a frontier model can coordinate open-source tools with a 35% hit rate off a text prompt, smaller biopharma teams can screen targets faster and cheaper without maintaining massive in-house software teams.

For the AI landscape, it shows the power of using frontier reasoning models as executive coordinators rather than building narrow scientific tools from scratch. Claude didn't replace specialized folding networks; it ran them better than an overworked human researcher might.

We think this agentic approach is the right direction, but the resource barrier is the obvious catch. Burning thousands of H100 hours per campaign puts this workflow out of reach for casual users, and balancing automated bio-design capabilities with biosecurity safeguards will remain a tricky line for Anthropic to walk.