Claude just designed something that worked in a real laboratory. Not code. Not an essay. Proteins.

Tech Times covered Anthropic's autonomous protein design campaign, and the numbers hold up to scrutiny.

Protein binders are small engineered proteins designed to attach tightly to a specific target protein -- and that attachment is the basic mechanism behind a huge share of modern medicine. A binder can inhibit a target, activate it, or deliver something to it, which is the working principle behind drug classes like monoclonal antibodies. Historically, designing even one working binder for one target has meant months of computation, optimization, and lab screening by trained protein engineers, per target, with no guarantee of success.

Across 15 clinically significant protein targets, Claude's autonomous design campaigns produced 354 confirmed binders from 1,320 total designs -- an overall hit rate of 22.6% to 35.1%, against an industry baseline of 10% to 15%. Fourteen of 15 targets yielded at least one successful design. The targets included well-known drug-development names like PD-L1, TNFα, EGFR, VEGF-A, and Cas9, alongside disease-specific targets like TREM2 (Alzheimer's) and Nipah virus Glycoprotein G. Several targets were pulled directly from recent open-design competitions specifically so Claude couldn't have simply memorized a published answer.

This wasn't Claude suggesting ideas for a human to execute -- it ran the workflow itself, largely without a human in the loop. According to Anthropic, Claude worked through four stages autonomously: selecting and analyzing each target using internet access and scientific literature, designing candidate structures and sequences by orchestrating several specialized open-source biology models, running iterative rounds of computational optimization, and then screening its own candidates for novelty, diversity, and likely solubility before producing a final shortlist of about 30 candidates per target. Anthropic gave Claude substantial infrastructure to do this -- up to 12,500 NVIDIA H100 GPU-hours for the largest runs, connectors into tools like Google Drive, Slack, Gmail, and BioRxiv, and a roughly 30,000-token instruction set describing the protein design process. The human role was limited to approving access requests and monitoring the infrastructure, not directing the science.

On RBX1, a target with an actual open human design competition behind it, the contrast is stark: that competition drew 245 human entrants and produced a 3.7% hit rate, with the winning entry binding at approximately 45 nM. Claude's top design for the same target bound at approximately 3.9 nM -- roughly ten times more tightly than the best human-designed submission, and Claude's own single-target run against RBX1 achieved a 40% hit rate.

Against TREM2, a target directly relevant to Alzheimer's disease, 72 of 90 Claude-designed proteins bound -- an 80% hit rate. That's a sharp jump from the 38.3% hit rate achieved in a prior Adaptyv Bio competition against the same target, giving a rare apples-to-apples comparison on identical scientific ground.

Anthropic itself is explicit about the limits here: "Protein minibinders, even high-affinity ones, are not by themselves drugs, they are one early step in a multi-stage development process that also includes affinity maturation, pharmacokinetics testing, formulation, and multi-phase clinical trials." No AI-discovered drug has received full FDA approval yet. What this result actually demonstrates is where the bottleneck in AI-assisted drug discovery is shifting -- from the computational design phase, which Claude just compressed from months to days, toward the wet-lab and clinical testing stages downstream, which still move at the same pace they always have.

This wasn't a computational claim taken on faith. Two independent, specialized contract labs -- Adaptyv Bio and Twist Bioscience -- physically synthesized and tested every one of Claude's designs, with no computational shortcuts and no modifications to what Claude produced. That's the same kind of wet-lab validation any human-designed candidate would go through.

Anthropic has been upfront that this capability cuts both ways: the same autonomous biological design ability that can speed up real medicine "could help a bad actor build a bioweapon." Because of that, Anthropic has deliberately kept protein design capability off its most capable publicly available model, running this work on earlier models (Opus 4.8 and a Mythos preview) instead, and is building a vetted access program for scientists rather than opening the capability broadly. That's a meaningful signal about how seriously the dual-use risk is being taken here -- the same underlying capability that produced 354 working proteins is also one Anthropic has chosen not to hand out freely.

Source: https://www.techtimes.com/articles/325081/20260820/claude-runs-autonomous-protein-design-campaign-wet-lab-confirms-twice-industry-hit-rate.htm

Additional background: https://www.anthropic.com/research/Claude-accelerates-protein-design