Profile · AI Drug Discovery
Numerion Labs
The company that taught a neural network to design medicine - now hunting drugs across a chemical universe most software never sees.
The platform, rendered. No lab coat required.
The company that taught a neural network to design medicine - now hunting drugs across a chemical universe most software never sees.
On a single graphics card, Numerion Labs sorts through ten billion possible molecules and ranks the best million in less time than it takes to read this sentence. That is not a metaphor. It is a benchmark, published in October 2025 alongside NVIDIA. The company calls the protocol APEX, which is a tidy acronym for a stubborn idea: search chemistry exhaustively, not selectively.
The name is new. The pedigree is not. Numerion Labs is the company formerly known as Atomwise, the outfit that, back in 2012, first pointed deep convolutional neural networks at the problem of finding drug-like molecules. The rebrand arrived in 2025 with a new CEO, a new platform, and the same audacious premise it started with - that most of medicine's best molecules are hiding in plain sight, and nobody has had the tools to look properly.
Traditional virtual screening evaluates less than 0.1% of available compounds. Numerion built its whole company around the other 99.9%.
- The bottleneck, stated plainlyThe number of possible drug-like molecules is often estimated at more than 10 to the 60th power - a figure so large it stops meaning anything. For decades, drug discovery handled this the way you handle a library you'll never finish: you read the spines you already recognize. Chemists screened the compounds they had on a shelf, or the ones that looked familiar, and called it a search.
The trouble with reading only the spines you recognize, of course, is that the best book is usually the one you've never heard of. The molecules that become first-in-class medicines tend to be the novel ones - the structures no one has made yet. Screening the familiar 0.1% is efficient, comfortable, and quietly self-defeating.
Everything we wished we could do as chemists 30 years ago, we can now actually do.
- From the Numerion platform manifestoAbraham Heifets and Izhar Wallach met as computational-biology PhD students at the University of Toronto. Their wager - made years before "AI" became a pitch-deck reflex - was that deep learning could predict how a small molecule binds to a protein, and that this prediction could replace a great deal of slow, expensive bench chemistry. They co-founded Atomwise, went through Y Combinator, and spent the next decade proving the idea against real targets.
In February 2025, the company brought in Steve Worland as CEO. He is not a machine-learning native; he is a drug developer with three decades and several approved medicines behind him, including scientific contributions to Paxlovid and Inlyta. The pairing is the bet, restated: marry the people who know how to ship a drug with the people who built the search engine for one.
Heifets and Wallach pioneer deep convolutional neural networks for structure-based drug design.
Roughly $45M led by B Capital, with DCVC, Baidu Ventures, Tencent and Y Combinator joining.
Co-led by B Capital Group and Sanabil, lifting total funding to nearly $175M.
A veteran drug developer takes the helm to push discovery toward the clinic.
Peer-reviewed research: 10 billion compounds screened in under 30 seconds, code open-sourced.
New brand, new platform - COSMOS, APEX and EXPO - same founding obsession.
Numerion's pitch is that drug hunting is a single workflow, not a pile of point tools. Its AI chemistry superplatform breaks into three named pieces, each handling a different stage of the hunt - predict, enumerate, optimize.
A universal chemistry foundation model that predicts drug function directly from chemical structure - the company's claim to finding diverse, genuinely novel compounds rather than familiar-looking ones.
Approximate-but-Exhaustive Search. Pairs deep-learning surrogates with GPU-accelerated enumeration to evaluate billions of starting points in seconds, so promising matter isn't skipped.
Expert optimization algorithms that deploy project-specific know-how without huge datasets or retraining, improving the odds of turning a novel binder into a clinical candidate.
Beyond software, Numerion runs its own proprietary pipeline of small-molecule candidates, with a stated focus on immune and inflammatory disease.
A chemistry domain 10,000x larger than rival tools, with screening efficiency the company measures in billions-fold gains.
- The platform's own scorecard, handle with curiosityAmbition is cheap; Numerion offers a few things you can check. The platform has delivered hit-identification success across more than 230 academic and collaboration projects, spanning a wide range of protein classes. The APEX work was co-authored with NVIDIA, peer-reviewed, posted to arXiv, and shipped with open-source code - which is a more falsifiable kind of claim than a press release. In its Atomwise era, the company ran drug-discovery collaborations with pharmaceutical names including Eli Lilly and Bayer.
Figures are company-reported and directional - the point is the order of magnitude, not the decimal places.
"Drug discovery has always faced a fundamental challenge of searching chemical space at vast scale and speed. With APEX, we've demonstrated that it is now possible to virtually evaluate billions of molecules in seconds."
- Steve Worland, CEO, Numerion LabsStrip away the acronyms and the mission is short: accelerate the discovery of life-saving medicines by searching chemistry well enough that the best molecules stop slipping through. Numerion frames its own work as "the art of drug hunting, powered by ML" - a phrase that gives away the company's self-image. The machine does the searching at impossible scale; the scientists decide what's worth chasing.
It is a useful tension to keep. A model that scores ten billion molecules is only valuable if a chemist can tell which of the top million is worth making in a flask. Numerion's whole structure - a foundation model bolted to a real drug pipeline, a software platform run by people who have shipped approved drugs - is an argument that neither half works alone.
Return to where we started: one graphics card, ten billion molecules, thirty seconds. A decade ago that ranking would have taken months of compute, or simply wouldn't have been attempted - chemists would have trimmed the question down to something a shelf of compounds could answer. Numerion's bet is that shrinking the question was always the mistake.
If the approach holds, the interesting consequence isn't speed for its own sake. It's that the molecules worth testing change - away from the familiar, toward the novel structures that exhaustive search surfaces and selective search never could. Whether that turns into approved medicines is the question every AI-drug company is still being graded on, Numerion included. But the company has done the unglamorous thing of publishing a benchmark anyone can check. In a field thick with promises, a reproducible number is its own kind of statement.