A lab that argues with biology's blind spot
In an office in South San Francisco, a machine is doing something stubbornly difficult: predicting how a strand of RNA folds in on itself. Not protein - RNA. The molecule most of drug discovery has politely ignored for decades. Atomic AI, about thirty people deep, runs its own wet-lab assays, feeds the data to deep learning models, and asks a question the pharmaceutical industry mostly stopped asking: what if the floppy, shape-shifting molecule everyone wrote off is actually the next great drug target?
The company is not selling certainty. It is selling structure - literally. Its bet is that if you can see the shape of RNA accurately enough, you can find the pockets where a small molecule fits, and you can drug things that have been labeled "undruggable" for a generation.
Most of the genome is RNA. Almost none of the drugs are.
Here is the inconvenient arithmetic. The overwhelming majority of the human genome is transcribed into RNA. Only a sliver becomes protein. And yet the overwhelming majority of approved medicines target - you guessed it - proteins. The industry built its tools around the smaller, tidier problem and called the rest undruggable.
RNA is harder for an honest reason. Proteins tend to hold a shape; RNA is restless, folding and unfolding, refusing to sit still for its portrait. Without a reliable picture of that 3D structure, you cannot find the grooves where a drug molecule would bind. The map was missing, so the territory stayed unexplored.
From a Science cover to a company
Raphael Townshend did not stumble into RNA. He did his PhD at Stanford on the geometric deep learning of biomolecular structure - work that landed on the cover of Science in 2021 and earned him a spot on the Forbes 30 Under 30 list a year later. The thesis, in plain terms: machines can learn the rules of molecular shape directly from geometry, without being told the rules first.
The bet he made was that this was not a paper. It was a company. He founded Atomic AI in 2021 and assembled a team that refuses to pick a side in the usual machine-learning-versus-wet-lab argument - it does both. Co-founder Brent Townshend leads the data generation that feeds the models, keeping the chemistry, conveniently, in the family.
Raphael Townshend
Founder & CEO. Stanford CS PhD, UC Berkeley EECS. Forbes 30 Under 30. Research featured on the cover of Science.
The team
~30 people spanning machine learning, chemistry, biology and engineering - plus a CSO recruited from Bristol Myers Squibb and Ribometrix.
Two engines, one obsession with shape
Atomic AI's platform has a name with a backronym only a structural biologist could love: PARSE - Platform for AI-driven RNA Structure Exploration. The idea is a closed loop. Custom wet-lab assays generate enormous amounts of chemical-mapping data; the models learn from it; the predictions point the experiments at better questions; repeat.
ATOM-1 Foundation Model
A foundation model for RNA structure and function, trained on in-house chemical mapping data across millions of RNA sequences and over a billion nucleotide-level measurements. The company says it predicts RNA secondary and tertiary structure more accurately than previously published methods.
PARSE Discovery Platform
The integrated R&D engine that pairs deep learning with custom wet-lab biology to find structured, ligandable RNA motifs - the rare pockets where a small molecule can actually grab on.
The pipeline Therapeutics
Internally, Atomic AI develops selective RNA-targeted small molecules. Through partnerships, PARSE is aimed at new RNA-based medicines and tools - a two-track business model around a single capability.
The Short, Eventful Life of Atomic AI
Five years from a thesis to a watchlist
The receipts, such as they are
A pre-clinical biotech does not get to wave a blockbuster drug around. What Atomic AI can point to instead: real capital from investors who do this for a living, a foundation model with a published preprint, and the scale of data underneath it.
Funding, from seed to Series A
USD raised by round
Bars scaled to total raised. The Series A did most of the talking; the seed just opened the door.
The backers are a who's-who of deep tech and biotech: Playground Global, 8VC, Greylock, Factory HQ, Not Boring Capital, AME Cloud Ventures, and angels including Nat Friedman and Patrick Hsu. The leadership bench has BMS pedigree. None of it guarantees a drug. All of it buys time to find one.
Make the undruggable, druggable
Strip away the platform names and the funding rounds, and the mission is almost old-fashioned: treat diseases that current medicine cannot reach. Atomic AI's wager is that the limiting factor was never biology's stubbornness but our inability to see RNA clearly. Solve the seeing problem, and the drugging problem becomes tractable.
It is a patient bet. Foundation models do not become approved therapies overnight, and RNA has humbled smarter approaches before. The company's stated values - bold innovation, collaboration, care, and a growth mindset - read like a reminder to itself that this is a long game played by a small team.
The map is the product
If Atomic AI is right, the value is not any single molecule. It is the map itself - a reliable, machine-built atlas of RNA structure that makes a previously dark region of biology navigable. Map a continent and you do not just get one road; you get every road anyone builds afterward.
That is the quiet ambition under the press releases. Not one drug, but the infrastructure that makes a class of drugs possible. Back in that South San Francisco office, the machine is still folding RNA, still arguing with biology's blind spot. The difference now is that the argument has investors, a foundation model, and a billion measurements behind it. The molecule everyone ignored finally has someone reading its shape - carefully, and out loud.