Arc Institute wants to build a virtual cell with a billion experiments

Four kilos and six hundred grams

On the bench sits a gray box twenty-five and a half centimeters wide. It stands nineteen centimeters tall and fits comfortably in a gym bag. On the scale it reads four kilos and six hundred grams: less than a bowling ball, yet enough to make itself felt in the wrist of whoever lifts it. An object you pick up with one hand and move without calling anyone.

Inside the box runs a calibrated electrical pulse. It is electroporation: the membrane eases open, and in that instant the cargo enters. The molecular scissors that switch off a gene. A flash, an opening, a gene gone silent. The gesture is all here, and there is no escaping it.

The plan is to repeat it more than a billion times.

Arc Institute has put it in writing together with its funder: the Technology Centers will conduct "more than a billion cellular experiments" to establish cause-and-effect relationships. A billion shocks. A billion genes switched off one at a time. A billion cells asked what happens when a piece of them is taken away. And at the end of that chain stands a goal with a name that sounds like it came out of a novel: the first high-utility virtual cell.

Here lies the disproportion that holds the whole thing up. On one side, a number that refuses to be imagined, all of biology reduced to a loop that turns on itself and learns. On the other, an object that fits in a gym bag, with a button, a tray, one shock at a time.

Because the scale, however large it is written into the contracts, must pass materially through that box. Between the promise and the bench there is no shortcut: every single event of that billion is a real cell, in a real well, crossed by a real current for a fraction of a second. The figure is abstract. The gesture is not. And the gesture has a weight: four kilos and six hundred grams.

From Three to Three Hundred

In the summer of 2022 the first lab spaces opened. A year after launch, the institute told of growing from a handful of people to more than a hundred, scientists and operations staff together. Today Arc declares more than three hundred. They divide into two cross-cutting initiatives, nine core labs and five technology centers. The internal vocabulary has already annexed them: in there they are called Arconauts.

The technology centers are more than a line on an org chart. In the classic academic model every lab builds its own instruments, its own protocols, its own daily grind: everything in-house, and every house redoing from scratch the work of the house next door. Here, none of that. Here data production is a centralized, industrial service, and the labs use it the way they use electricity. It is the arrangement of an organization that expects to have to generate numbers in quantities no single group could sustain.

In 2025 the leap in scale found its money. Arc joined the annual cohort of The Audacious Project, the philanthropic initiative connected to TED that raises funds around long-range projects. That is where the billion experiments stopped being a hallway phrase and became a commitment with a date attached: by 2029, Arc's model will have to be accurate enough.

From three founders to three hundred people: that is a pace. And a pace, when that billion sits at the end of the chain, leads straight to a single question. Where the count stands.

Quanto è lontano un miliardo

The count around Arc runs on a stopwatch: a billion experiments, and the race already under way. But before the finish line, it helps to look at the numbers along the track. Three numbers alone, and three different worlds.

The first comes from academic labs. The largest Perturb-seq ever published covers about two and a half million cells, across more than two thousand genetic perturbations. All in one go.

The second is the figure Arc sets beside its own model: over a hundred million perturbed cells. Only most of those perturbations are not CRISPR cuts. They come from Tahoe-100M, an archive of cells exposed to drugs. Switching off a gene means tearing a page from the book and watching what changes in the story. Bathing a cell in a molecule means pouring something over it and watching how the ink reacts. Two different questions, put to the same cell.

The third is the finish line: the billion. Between the starting point of two and a half million and that figure lie four hundred times as much, and then some more. This is no growth curve — it is a change of category, like moving from a small province to the population of half a continent.

Then there is the word holding everything up. In this field an experiment can mean a sequenced cell, a perturbation condition tested, or a single well on a plate: the same activity, counted the three ways, gives numbers far apart from one another. Change the unit of measure, and the billion drifts several lengths closer or farther away.

Nianzhen Li, who runs the experimental platform, points meanwhile to a nearer milestone: results from more than a hundred million single cells by year's end. Cells sequenced, one by one. Yet another unit, another boundary on the same map.

And this is where the count stops being bookkeeping. Between two and a half million and a billion, the distance cannot be read by lining up the digits — the zeros all look alike, and the eye slides past them. It can only be read once it is clear that every leap is a multiplication, never an addition, and that none of the three numbers is counting the same thing as the others.

Quando la somma batte la macchina

On June 23, 2025, Arc put its first virtual cell model on the table. It is called STATE, and the manuscript appeared on bioRxiv as a preprint: born that way, and still that way today, without the stamp of a journal. Its data diet is disclosed, and worth looking at squarely. On one side, 167 million observational cells: cells watched while they go about their business, untouched by anyone. On the other, more than a hundred million perturbed cells, spread across seventy different cellular contexts.

The ratio between the two piles says something precise. This virtual cell has watched far more than it has touched. The figures Arc claims are two, and both are comparisons with earlier models: a 50 percent improvement in telling the effects of one perturbation from another, and double the accuracy in spotting the genes that, after an intervention, turn out more switched on or more switched off than before. Arc calls it the first model to consistently beat simple linear baselines.

On June 18, 2026, VCBench arrived, a shared test bed for single-cell models. On the subset of 71 perturbations common to the various models, the additive baseline scores 0.890. It is an elementary rule: to predict the effect of two interventions together, add the two individual effects, with no artificial intelligence involved. And it beats every foundation model in the test, STATE included. Geneformer stops at 0.627, scGPT at 0.545. It is not an isolated stumble: already in 2025 a study in Nature Methods had shown five foundation models unable to beat simple linear baselines in predicting perturbation effects. The sum beats the machine.

And here Arc made its most surprising move: it opened up the contest instead of defending the score. The Virtual Cell Challenge asks entrants to predict what happens to an H1 stem cell when a gene inside it is switched off. The first edition closed in December 2025: more than five thousand entrants from 114 countries, more than 1,200 teams submitting results, over three hundred reaching the final submission. The hundred-thousand-dollar prize went to the BioMap team; the generalist category was won by Altos Labs, and the awards were presented at NeurIPS, the big machine-learning conference. The contest metrics remain the technical ones: genes more switched on or off after the intervention, the ability to tell one perturbation from another, error on expression counts. None of them measures a therapeutic outcome.

Four or five years, she says

Silvana Konermann sets herself a short horizon. Four years, maybe five: that much time, in her view, before virtual cell models become accurate enough to genuinely serve anyone hunting therapies against complex diseases. This is no promise to hand down to whoever comes after. It is a thing she wants to see herself, with her own eyes, while she is still at the bench.

On safety the tone does not change; it stays dry. A tool designed for human cells, she says, would be hard to bend to any dangerous use. Building the same tool for viruses, though, is an idea to avoid. She says it just that way, without beating around the bush.

It is a line the institute has drawn on its own, unprompted. No regulator has asked for it yet.

Opinion reported as such, not verified by the editorial team.

The well and the whole body

Meanwhile the flow of data keeps widening. The institute's cell atlas has passed 600 million cells, spanning observations and perturbations — an archive anyone can open and query.

On January 12, 2026, Tahoe Therapeutics, Arc and the Biohub announce they are working together. The partnership will yield more than 120 million single-cell data points across 225,000 drug-patient interactions, released openly for anyone who wants them. Opening the archives today carries as much weight as opening the model.

Yet a gap remains that none of those figures close. Every measure in play says one thing only: how well a cell's reaction can be predicted in culture. And a cell in culture is a plastic well, a warm liquid, a population behaving the only way it knows how inside it.

Between that number and a therapy that works in a person lie all the steps medicine has always known. The whole body. Time. The liver breaking molecules down one after another. And people, each responding in their own way.

Due geni che nessuno vedeva

Lo scetticismo più tagliente sulle classifiche dei modelli arriva da dentro casa. Hani Goodarzi lavora ad Arc, e su Science ha messo giù la frase che nessuno voleva scrivere: ogni volta che compare un nuovo modello, è sempre il migliore. Ci sta dentro tutti i giorni, in quel settore. Lo dice uno che ha il banco in quella stanza.

Poi c'è la storia dei due geni dell'emoglobina. Setacciando le interazioni fra geni, i ricercatori hanno visto la stessa coppia riaffiorare da previsioni di modelli diversi: HBG2 e HBZ, sempre loro, come un nome che ricorre in due lettere scritte da mani lontane. Sono andati a guardare da vicino.

Quello che hanno trovato è un buco. Quasi tutti i modelli si aspettavano un effetto quasi nullo quando i due venivano spenti insieme. Eppure ciascuno dei due, spento da solo, colpiva forte. La sinergia fra quella coppia era il punto cieco: ciò che quasi nessuno dei modelli stava vedendo.

Alla prossima Virtual Cell Challenge la classifica dirà se la cellula virtuale sta imparando davvero. Poi arriverà il 2029, la data che l'istituto ha messo per iscritto. E si saprà se quella scatola grigia da quattro chili e sei etti stava costruendo una previsione o una promessa.