A Hundred Million Cells, Four People, Three Days: Biology's ImageNet Moment

Sarah Guo opens this episode of *No Priors* on a release day. In the room are Vevo Therapeutics co-founders Nima Alidoust and Johnny Yu, and, from the Arc Institute, co-founder Patrick Hsu, CTO Dave Burke, and core investigator Hani Goodarzi — who is also a Vevo co-founder. The object on the table is Tahoe-100, which they call the largest single-cell drug-perturbation dataset ever created. The other object is why biology still needs a virtual cell, not just another protein-structure model, and when anyone should expect a treatment from the machine-learning boom.

Johnny works on single-cell sequencing at Vevo. Nima is a quantum chemist who converted into a computational chemist and likes playing with biological data; Vevo's job, as he puts it, is to predict how chemicals interact with cells in different biological contexts. Some people call that the virtual cell. Hsu describes Arc as working at the interface of biology and machine learning to understand, and one day treat, complex human diseases — most of the major killers. Burke builds novel AI models for biology. Goodarzi works with both of them on Arc's virtual-cell initiative.

The comparison that makes the week land is numerical. Nima later puts publicly available perturbational single-cell data, if you are generous, at one to two million points. Tahoe is a hundred million. In the same week Arc is putting out a Virtual Cell Atlas whose observational half, assembled by crawling public repositories, is about 230 million cells. Add them and you have 330 million. Protein structures, functions, and drug-protein interactions have been stacking up for twenty years. What the field has not had, Nima says, is how different cells behave in different contexts, and how genes function in the presence of the other genes. That, they believe, is the era that just opened.

Burke reaches for the canonical scene in AI history. Progress is punctuated by datasets. Fei-Fei Li's ImageNet in 2009 produced a nonlinear jump in machine vision. The hope here is that perturbational sets — data that elicit a cellular response — can do the same thing one level up from proteins. Hsu notes that AlphaFold-style structure models stood on CASP and the PDB. The open question is how you do that for cells and cellular dynamics, which is what actually tells you about health and disease. The bet is to study higher abstractions: not just the molecular machines, but how they operate in an entire cell.

Johnny translates the spreadsheet into English. A decade of little academic and industry sets is a mess. Batch effects are the first problem: even one person running the same cells on two different days produces data that do not look the same. If you want an "internet of biology" at ChatGPT scale, a poorly labeled, super-batchy corpus is only moderately useful. Tahoe, he says, roughly doubles all the data accumulated over the past decade. It covers 50 cancer models from 50 different patients and 1,200 drug treatments. It is deep, rich, and "effectively has no batch effects." He does not call it just another dataset. He calls it the first one that will actually enable machine learning in the space.

Nima later names the production numbers that make the cleanliness possible: a hundred million single-cell points, but really 60,000 drug–patient or drug–cell-line interactions, run by four people from Vevo over about three days. Biologists have a phrase, "in our hands," for results that only reproduce on one bench. Computer scientists have the embarrassing cousin, "on my machine." Few hands, consistent work, almost no automation yet — that is how you avoid what Goodarzi calls infecting a corpus. Tools, software versions, and genome builds have changed over a decade. Collating already-processed public sets contaminates the training data with analytical batch effects. Four people and three days is the opposite of that infection.

Sarah asks Arc, which works on both virtual cells and protein language models, to say why the field needs both and where each actually is. Burke, who has an engineering-computer background, reaches for hardware. DNA lives in ROM: read-only memory, coding for the cell. RNA lives in RAM: working memory. Expression levels constantly change, "almost like one of those 1980s graphic equalizers where you got like 20,000 bars for each gene," adjusting to environment, stress, aging, disease, or health. A virtual-cell model is an attempt to infer a notional CPU: given an edited gene or a drug, how does the transcriptomic profile move. Once you have that model, you can pose the inverse question. Given a disease-state profile, what perturbation — edit or drug — pushes the cell back toward healthy.

Goodarzi splits the map into data-poor and compute-poor regions. DNA language models are not very data-limited: decades of genome sequencing took care of that. The bottleneck is compute, especially context length, how much DNA you can consume in one input. Cell-state models are the other extreme. Profiling cells at single-cell resolution is a new technology, really exploding only in the last five or six years. Before Arc's SC Basecamp — the public crawl released with Tahoe on the Virtual Cell Atlas — curated human cells sat in the order of 45 to 60 million. Scale, he says, is one thing. Information content is another.

They already trained early "single-cell foundation models" on those tens of millions. Down-sample by 99 percent, train on 1 percent, and performance does not fall that much. The training data's information content is not amazing. What the model needs is data from very different biological contexts. That is why Burke and Hsu keep returning to perturbations: they create new contexts and new cell states the model can actually learn from. Observational public data, Nima adds, is almost entirely healthy tissue, almost entirely a snapshot of a liver sample. It does not give you causality, and it does not tell you what a new genetic or chemical perturbation will do. Tahoe is built for that second job.

Hsu's short version of why perturbational data matters: most biological search is descriptive. You stare, you notice what else changes when you poke, you collect associations. Genetic or chemical perturbations give you a clean before and after, a set of causal changes that can drive a cell state. Do it across many cell types and tissues, or the model never learns a general sense of what cell states are possible. Burke puts it in topology. The model is trying to lay down a manifold in a high-dimensional latent space. To explore that manifold it has to see lots of perturbations and responses; then it can predict for points it has not sampled that still sit on the manifold. Healthy cells alone will not show you how the manifold warps in disease. Many cancer tissue types are one way to get at the states drug discovery actually cares about.

Sarah asks how to think about 100 million or 230 million points against scaling laws. Goodarzi's short answer is that they will not know until they get there. Inspiration comes from language models and DNA models, where a comfortable mark is around a trillion training tokens — GPT-3 at roughly half a trillion, ESM-3 at 700 billion. Cells are not sentences, but if you count genes and their expression as tokens, this collection gets them close enough to start asking. A cell in these sets is 2,000 to 5,000 genes; a hundred million single-cell points is akin to 200 to 300 billion tokens. How many of those tokens are actually informative is a finer point he flags and does not pretend to close.

Choosing the arrows is its own design. For cancer, Johnny says, you match the toolkit to the biology: cancer-relevant genes, growth, DNA regulation, drugs against key pan-cancer pathways. Those pathways are conserved enough that the same data can still feed neuroscience or immune-cell development. The factory is Vevo's Mosaic platform. Cells from many patients — lung, pancreatic, and so on, each with their own genetics — are pooled into a single mosaic tumor, then screened reproducibly against hundreds or thousands of drugs. Instead of one cancer model at a time, tens or hundreds. Tahoe-100 was generated that way. They covered the biggest cancer types by annual incidence and want rare disease and chemical space next, with the models themselves pointing at the gaps.

Nima's punchline is about not having to choose. Generate 50 times more perturbational data than is publicly available, in five weeks, against a public corpus that took ten years, and you do not have to prioritize as much. Go large on chemical space, go large on patient samples, and you do not have to come up with an a-priori hypothesis about what to feed the models. "Let the data surprise you." Sarah walks the Mosaic loop back in lay language — a super-tumor of patient variation, each mouse treated, the extracted signal the drug-by-patient interaction — and asks whether anyone else thinks this is crazy. Not crazy, they tell her: it happens every day in their lives. It is also, honestly, science fiction.

Nima thinks hypothesis-driven culture has been slowing biology, for a boring reason: experiments were expensive. Sequencing cost is down, single-cell sample-prep cost is down, compute cost is down. It is time, he says, to change the mentality, be a little more courageous, a little more freewheeling about what samples you put together. Burke maps tokens per experiment over fifty years of biomedical research and gets the hockey stick every founder loves. The rolling fight of mechanism versus large-scale profiling, hypothesis-driven versus hypothesis-free, will wash out with scale. You do not have to choose. Most mechanistic data today was made to ask very specific, well-scoped questions. More tokens per experiment is just going to be the way.

Goodarzi's version: biology treated humans as the foundation models that ingest information and come up with hypotheses, biases included. At UCSF they sometimes joke that medical chemists like Kevan Shokat are the last layer of a neural network. An AI-generated molecule lands on the desk; the chemist knows whether it looks real and cannot quite verbalize why. People criticize models for hallucinating. Scientific research, Nima answers, involves hallucination. That is what creativity is. Sarah asks if they are Sutton bitter-lesson people — intuition baked into the system is the wrong long-term bet, just scale the data. They hope they do not have to make that choice. They are already seeing scaling-law hints in proteins, in Arc's Evo DNA models, and in inference-time scaling in a recent study, with the usual caveat that they will need benchmarks.

Nima has watched this movie. Every time you take a success from field A into field B, people — including inside your own organizations — produce a list of a hundred reasons it will not transfer. Then you get surprised, and the same list appears from B to C. The models that learned human language learned structural biology; with Evo they are learning DNA. That is not trivial, and plenty of people said protein language models would never work, that you needed domain-specific architectures. The ethos they want for systems biology is to try what worked elsewhere first. Maybe it works, maybe it does not. If you do not try, you never know. Sarah, whose fund's rare hard conviction is that these techniques scale in domains experts are still unsure about, calls that music to her ears.

Vevo generated Tahoe as a private, venture-backed startup. When Johnny told Nima they could generate a hundred million single-cell points, Nima asked if they could, then said let's go. Within hours — Johnny, Hani, and Nima are co-founders — they also said let's open-source it. Two reasons. First, put a new stick in the ground: show the community the game has changed, and that people should move on from a million observational points, or a hundred thousand. Second, Vevo's DNA is a small team of superstars rather than a hundred hires. Paradoxically, open source lets them stay that small. They mentioned Tahoe to Burke sometime between Christmas and New Year; the Arc team got excited. Without the open-source piece it would not have been as exciting. The community plays with the data, tells them what is good and what is not, and a three- or four-person in-house team gets the rest of the people who want virtual cells on the quest. Data was the bottleneck. They are trying to remove it.

The serendipity on Arc's side was a mission-driven atlas they were already conceiving. Burke describes the Arc Virtual Cell Atlas launching the same week. Tahoe is the star of the perturbational half. The observational half is SC Basecamp, which he analogizes to a Google crawler and index: an agent that mines public single-cell RNA data, curates it uniformly, and yields about 230 million cells. Pre-train on the observational set, then bring in Tahoe's dynamics. Capabilities of AI agents for science are still early, but this is a useful example. All dry-lab workflows, he thinks, will get automated with agents or copilots. The Sequence Read Archive is the largest repository of next-generation sequencing; NIH grants and journal deposits dump data there with no uniformity requirement, poorly annotated, sprawling. A team of computational biologists would have slaved over that. The agent crawled, organized, processed, and stripped a lot of prior methods' batch effects and biases.

Goodarzi emphasizes the analytical infection. Datasets span a decade. Tools changed, versions changed, genome builds changed. Taking processed sets and collating them contaminates training. They cannot erase every experimental chemistry change, but removing the analytical component was, to their surprise, very observable and quite helpful. Foundational sets for a whole field need that hygiene if people are going to write papers on top. Sarah's question after the production story is the obvious next one: if scientists and entrepreneurs can now go build virtual cells on this, how do you tell the quality of a model.

Burke's quality metric is predictive ability. Perturb a cell — suppress or upregulate genes, or apply a drug — and see whether the model forecasts the differentially expressed genes, the DEGs. Today's best models are very poor at this, on the order of 10 percent. There is no accepted benchmark; the industry would benefit from one. One conjecture in the room is that the problem is not model structure. Machine learning has rich architectures. The issue is data quality. The hope for the Arc atlas plus Tahoe is a starting point where models can actually get high predictive value. Hsu's prior on why you would bother, when you have real cells, is that biology is slow. In the last year of his PhD his adviser tried to talk him onto an aging project: age animals for two years, that is one experimental round. He declined. "May I please, sir, graduate." Coming from engineering, being constrained by biological time still strikes him as completely crazy. Massively parallel in-silico simulation sounds great. At 10 percent accuracy you are simulating noise.

If the models work, Nima wants to predict how a new chemical structure interacts with cells from different patient models — whether it takes a cancer cell from a diseased state to a healthy one, or, for cancer, kills it without killing healthy cells. Johnny's future-vision line is a moment when a virtual cell spits out a drug that turns a diseased cell healthy again. Generalizability has two axes: cell and chemical. Every disease is unique; you cannot Mosaic every tumor that arises, so the model has to transfer. In-silico libraries hold tens of millions of compounds and, if you put your mind to it, infinite biologics, most of which will never exist. A model that can traverse that space and tell you what to synthesize leapfrogs the slow screen of a couple of hundred thousand well-behaved fragments. Ninety percent of drugs fail in clinical trials, Hsu notes. Either the matter is wrong — potency, binding, toxicity, pharmacokinetics — or you drugged the wrong target. Virtual cells shrink the target search space so you can point the cannon and then spend time on the right chemical.

That is the gist of why virtual cells sit next to protein language models instead of replacing them. Protein models speak structural biology: how a protein folds, how a ligand binds, how an antibody docks. Nima bet his PhD on physics-based quantum-mechanical simulations of those bindings and found biology more complex. The target lives in a cell, the cell in a tumor, the tumor in a broader system. Virtual cells are how you venture into systems biology. Sarah asks about multi-cell, aggregates, organoids. Burke's first question is the right abstraction: they think transcriptomic, because complex gene pathways show up there when a cell reacts. You could make an arbitrarily complex model of the machinery; they think the genetic level is right. Organoids of 20,000 cells can take the same perturbations and start capturing intercellular dynamics. Nima adds that Tahoe includes in-vivo models and humanized mice with some of the mouse immune system. Call it single-cell, but the point of so many points is context: this immune environment, this tumor, this mutation. The environment is filtered through the cell. Observe the cell at enough resolution and you can even add spatial data.

Sarah's first hot take is why Vevo insisted on being a platform company rather than a single-hypothesis shop, like 99.9 percent of biotechs. Nima's difference is the team you build and the incentives. A single-hypothesis company is the human-as-foundation-model again: come up with a hypothesis, test the hell out of it, and become heavily incentivized to make that hypothesis work. You see drugs go to clinic after three patient samples. A platform tries to have enough hypotheses, generated in a hypothesis-free way, that you are not wedded to one — which, he argues, is actually more scientific about what you take to the clinic. They had plenty of early hypotheses that could have been a one- or two-asset company. They chose the platform so they could be more rigorous about what ships.

Then the water-cooler question: the rise of Chinese biotechs, threat or not. Johnny does not soften it. Their cost basis is more competitive. Packages look good — safety, tox, IND-enabling studies. People were surprised at pipelining efficiency and the ability to manufacture antibodies. He thinks that is great for the industry: patients, investors, and companies all want a lower cost basis. The United States has a wide array of CROs you can chain into a "virtual biotech" that looks good on paper and is incredibly slow, or you fully vertically integrate and it is incredibly expensive. The Goldilocks middle is competent vendors plus lean companies. Nima is blunter: it hit us in the face, and Johnny's phrase, which Nima repeats, is that it slapped us. Three years ago, when they were founding Vevo, Johnny was already talking about China and the commoditization of things like molecular design. Two responses exist. Regulatory capture, lobby limits. Or make it part of the ecosystem and change business models and team construction. He prefers the second. His Reaganesque closer: it is morning in bio. Play a different game.

The old-school way, he says, is a lot of planning. If he had a penny for every massive organization that announces something extraordinary and promises it in three to five years, he would be rich. That is the ethos in bio, and he is done with it. The tools are here; it is time to build now. Evo 2 went from the first Evo paper to the next generation in a matter of months. Tahoe got created the same way. Second piece: small, focused teams of superstars. Vertically integrated giants are not just capital-intensive; they are slow and bureaucratic. Third: naysaying. In everything you want to do in bio, strong biologists will tell you why it will not work. That culture has to change. Sarah reports the pharma-CEO version of the same freeze: most of the top-line budget is clinical development, so the AI they can feel today is summarizing trial documents and cohorting patients. Target discovery's ledger lights up only when the models get better. Point the cannon, measure twice, cut once, and the cost basis should fall.

She has been pitched AI-for-biotech for a decade, and treatments still take on the order of eleven years. Why now. Burke goes back to neural nets bouncing around until compute, data, and models hit a nonlinear point — ImageNet, then AlexNet, after a long period when people thought only humans could recognize images. Single-cell sequencing, if you are not a biologist, is already mind-blowing; add data generation and more sophisticated training and things start happening. Evo 2 was trained on 9.3 trillion nucleotides with no biology lectures. It learned where ribosome-binding sites are, codon degeneracy, and it can predict BRCA1-variant pathogenicity zero-shot at an AUROC of about 0.94. Nobody taught it. Hsu, Nima, and the others locate the field somewhere between GPT-1 and GPT-4 overall. Protein models, Nima says, are past GPT-3. Virtual-cell models are GPT-1 to GPT-2, closer to GPT-1. The catch, he adds, is that GPT-4 announced itself the day it shipped. A GPT-4 of cell-state models still has to prove itself through drug development: a platform that moves success from 10 percent to 30 percent is amazing and still 30 percent, hanging off a cycle on the order of ten years. Sarah's closer is to treat the bottleneck as a system. If it was terribly debilitating at the beginning, this week is supposed to be the breakthrough.