Eighteen. That is how many known RNA structures he had to teach his model. It beat labs swimming in data — and put RNA on the cover of Science. Now he is building the company to drug it.
Walk into the room where biology met machine learning around 2020 and the conversation was all about proteins. AlphaFold had just bent the field. Raphael Townshend was in that room - literally, on DeepMind's AlphaFold team for a spell - and he walked toward the harder thing instead. RNA. The molecule with fewer solved structures, messier folds, and almost no drugs pointed at it.
Today he is founder and CEO of Atomic AI, a South San Francisco company built on a single contrarian conviction: that the 3D shape of RNA can be predicted, and that once you can see the shape, you can design small molecules to grab it. Most of the human genome gets transcribed into RNA. Almost none of it has ever been a drug target. Townshend treats that gap not as a footnote but as the whole opportunity.
"We are creating an entirely new field of drug discovery."— Raphael Townshend, on launching Atomic AI
The flagship is ATOM-1, which Atomic AI describes as a foundation model for RNA structure. It does not work from sequence alone in a vacuum. It ties together machine learning, structural biology, and wet-lab assays so the model and the bench correct each other - predictions go out, experiments come back, the model sharpens. The downstream goal is a pipeline of RNA-targeted small molecules aimed at targets the rest of the industry has written off.
That phrase - "traditionally undruggable" - does a lot of heavy lifting in biotech pitch decks. In Townshend's case the receipts came before the company. His 2021 Science paper showed a model that could pick out accurate RNA structural models without being told in advance what "accurate" looks like, and it did so trained on a dataset most researchers would call hopelessly small.
The trick was learning from scarcity. Where protein structure prediction feasted on hundreds of thousands of solved structures, RNA had a near-empty cupboard. A method that needs less data is not a compromise here - it is the entire point, because the molecules whose structures are hardest to solve experimentally are exactly the ones medicine most wants to reach.
Atomic AI has assembled a roughly 30-person team spanning machine learning, chemistry, biology, and engineering, plus scientific advisors drawn from Stanford, UC San Diego, the University of Michigan, and Bristol-Myers Squibb. The culture line the company puts forward - bold innovation, collaboration, care, growth - reads like every startup's, but the staffing tells the real story: this is a place where a chemist and a deep-learning researcher are expected to argue productively at the same whiteboard.
Protein folding had an embarrassment of data. RNA had almost nothing. Townshend's bet was that the right geometry-aware model could learn more from a handful of structures than a brute-force model could from a flood. The bars below are illustrative of that gap in available training structures - not exact counts.
Illustrative scale · Source: "Geometric Deep Learning of RNA Structure," Science (2021)
"There is a significant need to develop tools that can accurately predict 3D RNA structures."— Raphael Townshend
Lead author on geometric deep learning of RNA structure - a method that found accurate folds without hand-fed assumptions.
ATOM3D: Tasks On Molecules in Three Dimensions - a benchmark suite for learning on 3D molecular data.
Learning from Protein Structure with Geometric Vector Perceptrons - geometry-aware neural nets for biomolecules.
Recognized in the Science category for turning a research insight into a company.
Department of Energy Computational Science Fellowship and an NSF Graduate Research Fellowship.
Founded the inaugural workshop on machine learning for structural biology - convening a discipline that barely had a name.
Strip away the platform names and the funding rounds and the ambition is plain: make RNA a place where medicines get made. Proteins have had decades and an army of crystallographers. RNA has had Townshend and a model that learned to do more with less. If the bet pays off, the list of "undruggable" diseases gets shorter - not because anyone found a clever workaround, but because the field finally learned to see the shapes it was aiming at.
It is the kind of swing that looks obvious only in retrospect. For now it is a 30-person company in South San Francisco, a foundation model, and a founder who keeps pointing at the part of the genome everyone else skipped.