A Yale professor and a former CERN engineer are teaching language models to read cells. Their bet: simulate human biology well enough, and the industry can fail its worst drugs in software - not in animals or in patients.
Every drug that fails does so twice. First on paper, where a hypothesis looks good enough to fund. Then in the real world - in a mouse, in a dish, or worst of all, in a person - where it turns out human biology never agreed. The second failure is the expensive one, and it happens far more often than the industry likes to admit. CellType, a two-person company out of Y Combinator's Winter 2026 batch, has a blunt proposal: move the second failure back into software, where it costs almost nothing.
The company calls itself "the agentic drug company," and its one-line description is unusually literal: we simulate human biology. Behind that line sits a foundation model that reads cells the way a language model reads text, and a set of AI agents built to run the grind of early drug discovery - hypothesis, screening, prioritization - on their own, overnight, while the humans sleep.
The technical trick at the center of CellType is called Cell2Sentence, and once you hear it, it is hard to un-hear. A single cell can be described by which of its genes are switched on and how strongly. Rank those genes from most expressed to least, line them up, and you have something that looks a lot like a sentence - a sequence where order carries meaning. Feed millions of those "cell sentences" to a large language model, and the model starts to learn the grammar of biology: which patterns go together, which shift when a cell gets sick, which change when a drug is introduced.
The method was developed in David van Dijk's lab at Yale and presented at ICML 2024, in collaboration with researchers at Google DeepMind. That academic pedigree is not a footnote. It is the reason a two-person startup gets its calls returned by large pharmaceutical companies.
The model that came out of this, C2S-Scale, is a 27-billion-parameter system built on Google's Gemma architecture. It was trained on billions of cell sentences spanning tissues, disease states, and drug perturbations. The scale matters because the questions matter: not just "what is this cell," but "how will this specific cell respond to this specific drug at this specific dose."
Plenty of companies promise a "virtual human." CellType did something narrower and more convincing: it made a prediction in software and then watched it hold up at the lab bench. In a computational screen of more than 4,000 compounds, the model flagged Silmitasertib - the cancer drug CX-4945 - as a candidate to make tumors more visible to the immune system. The specific prediction was roughly a 50% increase in antigen presentation when paired with a low dose of interferon. When the experiment was run in living cells, the effect showed up.
The ordering is the whole point. A model that explains results after the fact is a curiosity. A model that calls the shot before the experiment is run is a tool you can plan around.
"It's not too often that a tenured professor steps away to build a company around their own work."Dave Messina, Pioneer Fund
The economics of drug development are famously grim. A single approved medicine can carry a bill north of a billion dollars, much of it spent on candidates that never make it - a long tail of compounds that looked promising in animal models but stumbled once they met human biology. Mice are not small humans. Cell lines drift. The gap between preclinical promise and clinical reality is where fortunes and years disappear.
CellType's pitch is to shrink that gap by simulating the human context up front - de-risking targets, predicting human-specific toxicity, stratifying virtual patients before a trial, and surfacing biomarkers. Get it right, and a discovery cycle that used to take years compresses toward weeks.
David van Dijk, the CEO, is a Yale professor with more than 11,000 citations and work in Cell, Nature, NeurIPS, and ICML. He reportedly turned down an offer from Google to start CellType around his own research - the kind of move that says he thinks the commercial version can go further than the paper. His co-founder, Ivan Vrkic, co-developed Cell2Sentence at Yale and, in an earlier life, built control software for CERN's Large Hadron Collider. From steering particle beams to steering cells is a strange career arc, but both jobs share a temperament: enormous data, unforgiving physics, no room for hand-waving.
The agentic drug company. We simulate human biology.CellType
Today the model is service-shaped: pharmaceutical companies pay for discovery runs, hypothesis validation, and access to the platform. The company says it is fielding inbound interest from Top 10 pharma - the sort of attention a two-person startup rarely commands, and a direct dividend of van Dijk's academic standing. In March 2026 it signed an MOU with Senhwa Biosciences to fold its platform into the clinical development of CX-4945, the very asset its model had flagged.
CellType sits in a crowded and fast-moving neighborhood - AI-for-biology and "virtual cell" efforts from the likes of Recursion, Insilico Medicine, Isomorphic Labs, and a wave of single-cell foundation model teams. Its defensibility argument is not the algorithm, which is published, but the harder-to-copy stack around it: trained model weights that cost millions of dollars in compute, curated proprietary data, experimental validation in living cells, and pharma relationships built over an academic career. The method is open. The trust is not.
For a drug hunter, the promise is practical: point the agents at a disease, let them chain hypothesis to compound screen to lead prioritization to mechanism and off-target checks, and wake up to a ranked, reasoned shortlist instead of a blank whiteboard. For a clinical team, it is a way to ask "will this hurt humans in a way the mouse won't show" before the money is committed. And for the broader industry, it is a wager that the cheapest place to be wrong about a drug is inside a simulation.
Whether CellType becomes the "virtual human" that pharma consults before every animal study, or one useful tool among many, is unsettled. What it has already done is unusual for a company this young: made a specific, checkable prediction, and been right when the cells were counted.