SCIENCE & TECH
For decades, designing a molecule that can grab onto a specific protein in the human body — step one of making almost any modern drug — has been slow, expensive, expert-driven work. Teams of PhD chemists spend weeks or months hand-crafting candidate molecules, testing them, tweaking them, and testing again, hoping to hit a success rate in the low teens.
This month, Anthropic says its Claude AI models did that job themselves — with no human scientist steering the process in real time — and did it better than the industry average. According to results the company published this week, Claude autonomously designed novel protein binders that succeeded against 14 of the 15 biological targets it was given, posting hit rates roughly double what human-led campaigns typically achieve. Independent labs built and tested the molecules in the physical world to check the AI's homework. Some of them passed with flying colors. And now, drug-discovery scientists, AI skeptics, and at least one infamous former pharmaceutical executive are arguing about what it all actually means.
WHAT CLAUDE ACTUALLY DID
To understand why this is a big deal, it helps to understand what a "protein binder" is and why chemists care so much about them. Many drugs — from cancer treatments to autoimmune therapies — work by binding tightly to a specific protein in the body and either blocking it or triggering a desired biological response. Finding a molecule, often itself a small protein, that binds precisely and strongly to a chosen target is one of the very first and most labor-intensive steps in developing a new medicine. Get the binder wrong, and everything built on top of it — years of further drug development, clinical trials, manufacturing — is wasted effort.
Traditionally, this "de novo" binder design work has required expert computational biologists running specialized software, iterating through rounds of design and lab testing, and drawing on years of hard-won intuition about protein folding and molecular interactions. Anthropic's new experiment set out to see whether Claude could do that work largely on its own.
According to details Anthropic published in its research writeup, the company set two of its models — the publicly available Claude Opus 4.8 and an experimental system called Mythos Preview — loose on a batch of 15 different target proteins. Working from a protocol designed by human experts but without direct human guidance during execution, the AI agent handled the entire pipeline itself: researching each target, selecting which publicly available computational design tools to use, generating candidate protein sequences, narrowing thousands of options down to a shortlist, and ranking which sequences were most likely to actually work — before handing its final picks off to a real laboratory for synthesis.
The numbers, once the physical testing was done, were striking. Across a 48-hour session where the models tackled all 15 targets simultaneously, Mythos Preview and Opus 4.8 posted overall hit rates of 26.7% and 22.6%, respectively — meaning roughly a quarter or more of the AI's proposed molecules actually bound to their targets when built and tested in real cells. When Mythos Preview was allowed to focus on individual targets one at a time over 24-hour sessions, that hit rate climbed to 35.1%. For comparison, typical hit rates in human-led de novo protein design campaigns run in the 10% to 15% range, according to figures cited in Anthropic's own report and echoed across independent coverage of the results.
In total, the campaign produced 354 confirmed working binders out of 1,320 designs submitted across all 15 targets — and Claude succeeded against 14 of them. Some of its best-performing designs, Anthropic said, bound to their targets several times more tightly than the best previously published de novo binder for that class of molecule.
WHO CHECKED THE HOMEWORK
Anthropic didn't just take its own AI's word for it. Two outside biotechnology companies, Adaptyv Bio and Twist Bioscience, independently synthesized the protein sequences Claude generated and ran the physical lab tests to see whether they actually bound as predicted. That independent verification step matters enormously in a field that has, in recent years, seen plenty of bold AI claims that didn't hold up once outside labs tried to reproduce them. Having two separate synthetic biology companies confirm the binding results gives the numbers considerably more credibility than a purely in-house claim would carry.
Anthropic has also framed the experiment as evidence of a broader capability shift — not just that Claude can design proteins, but that it can run substantial parts of a real scientific research pipeline with minimal hand-holding, making its own decisions about tools, strategy, and prioritization along the way. In a related piece of the same research push, the company said its Opus 5 model had also processed raw NMR and mass spectrometry chemistry data with purity readings within 0.1% of what human lab scientists measured themselves — another sign, Anthropic argues, that its models are creeping into the actual mechanics of laboratory science, not just answering questions about it.
NOT EVERYONE IS IMPRESSED
Not everyone is buying the hype. Martin Shkreli — the former pharmaceutical executive notorious for hiking the price of a life-saving drug by more than 5,000% and later convicted of securities fraud, who has reinvented himself in recent years as a vocal biotech commentator — publicly tore into Anthropic's results, calling the work "not impressive." Shkreli argued that the binding affinities Claude achieved for peptide-based targets were actually fairly weak by field standards, and pointedly noted that none of Claude's successful binders targeted proteins found inside cells — arguably a harder and more valuable class of target than the extracellular proteins Claude focused on. He also argued that for many of the extracellular targets Claude tackled, scientists could achieve similar or better results simply by taking the working "business end" of an existing monoclonal antibody and adapting it — a far more established, and arguably simpler, technique than training an AI to design something from scratch.
It's a fair scientific critique, and Anthropic itself has been notably careful not to overclaim. In its own writeup, the company explicitly cautioned that "protein binders are not drugs," and that finding a molecule that binds tightly to a target is merely "the first step in the process of developing a drug-like molecule" — a process that still requires extensive additional work on safety, manufacturability, delivery into the body, and clinical testing before anything resembling an actual medicine emerges. In other words: Claude built a promising set of molecular Lego bricks. It didn't build a drug, and it's not close to one yet.
WHY WALL STREET AND SILICON VALLEY ARE PAYING ATTENTION ANYWAY
Caveats aside, the results are landing at a moment when Anthropic is aggressively positioning itself as more than just a chatbot company — it's trying to establish Claude as a genuine tool for scientific discovery, in direct competition with specialized biotech AI players and in a market where investors are hungry for evidence that large language models can do more than write essays and summarize emails. The company is reportedly gearing up for a potentially historic IPO, and demonstrations like this protein-design campaign serve a dual purpose: they're genuine scientific results, but they're also a marketing showcase for potential investors and pharmaceutical partners.
That commercial dimension hasn't gone unnoticed. Shares and sentiment around companies in the synthetic biology supply chain — firms like Twist Bioscience that manufacture the DNA and proteins AI systems like Claude design — got a visible bump in attention following the announcement, as analysts speculated that wider adoption of AI-driven protein design could mean steadily rising demand for the physical lab infrastructure needed to actually build and test what the AI dreams up. If pharmaceutical companies increasingly trust AI systems to generate first-draft drug candidates, the thinking goes, the bottleneck shifts from computational design to physical synthesis and testing capacity — a boon for the companies that supply it.
THE BIGGER PICTURE: AI CREEPING INTO THE LAB
This protein-design campaign doesn't exist in isolation. It lands amid a broader wave of AI companies racing to demonstrate real scientific and research capabilities, not just conversational skill. A separate benchmark published this month, called Reconstruction, found that even frontier language models could only reconstruct the core idea of a scientific paper from its bibliography alone 3% to 15% of the time — a humbling reminder that AI's scientific reasoning still has clear limits, even as headline results like Claude's protein work grab attention. Taken together, the picture that emerges is one of an AI industry making real, measurable progress on narrow, well-defined scientific tasks — like designing a binder for a known target — while still falling well short of the kind of open-ended scientific insight and creativity that human researchers bring to a field.
WHAT THIS MEANS FOR YOU
If you or someone you love has ever waited years for a promising new drug to clear clinical trials, this is the part of the story that matters most: even modest, verifiable improvements in the earliest stages of drug design can ripple downstream into faster, cheaper development pipelines. A protein-design step that used to take a specialized team weeks to complete, at a roughly 10-15% success rate, being done by an AI agent in 24 to 48 hours at double that hit rate — if it holds up at scale, across many more targets and many more labs — could meaningfully compress how long it takes a promising molecule to go from computer screen to laboratory bench to, eventually, a patient's bloodstream.
That said, patients shouldn't expect AI-designed miracle cures anytime soon. As both Anthropic and its critics agree, a working protein binder is a long way from an approved medicine. The years of safety testing, toxicology studies, manufacturing scale-up, and human clinical trials that stand between "this molecule binds in a dish" and "this drug is approved and on pharmacy shelves" haven't gotten any shorter — only the very first step in that long chain has, potentially, gotten faster.
WHAT COMES NEXT
Anthropic says it plans to continue expanding this line of research, testing its models against a wider and more challenging range of biological targets, including the harder intracellular targets that critics like Shkreli have pointed out were notably absent from this round of results. Independent researchers and pharmaceutical companies are likely to attempt their own replications and extensions of the approach in the coming months, which should help settle exactly how much of Anthropic's reported success is a genuine leap in AI-driven science, and how much reflects a carefully chosen, favorable set of test conditions.
For now, the episode stands as one of the clearest public demonstrations yet of an AI system operating with real autonomy inside an actual scientific workflow — deciding what tools to use, what molecules to try, and what to hand off for physical testing, largely without a human scientist looking over its shoulder in real time. Whether that proves to be a genuine turning point for drug discovery or simply an impressive but narrow party trick may take years, and a great many more independently verified experiments, to fully answer.
