A Virtual Cell That Learns Itself

We previously traced the bottom-up approach of explicitly constructing a virtual cell, and the wall it hit. While this approach succeeded for an exceptionally small parasitic organism (M. genitalium), human cells hold far more mechanism than anyone can write down by hand. This post is about the strategy that subsequently replaced the bottom-up paradigm, and about how much of the virtual cell dream this newer strategy has actually delivered.

Investor? Learn about our pre-seed round

If we cannot hand-write the rules of a cell, we can try to learn them instead. Do not tell the model how a cell behaves; show a neural network how cells behave, millions of times over, and let it find the regularities on its own after it sees enough examples to approximate what “normal” looks like. This is the strategy that replaced the bottom-up approach to virtual cells. Interestingly, it borrows directly from the large language models that have reshaped AI and computing in the last few years, so that is where we’ll begin.

A large language model (LLM) like ChatGPT is never taught grammar. It is shown enormous amounts of text and trained on a simple task: predicting the next word (or “token,” roughly equivalent to a syllable). In getting good at that task, LLMs build an internal sense of how language works. Nobody writes the rules of English down for them; the models infer these rules because they are already present in the training data. This same general recipe for training LLMs works on any domain with enough structure, and a cell has a great deal of structure.

For this reason, much of the current virtual cell effort attempts to transfer the same approach that made LLMs successful to the biology world. Take a single cell’s molecular profile, the list of which genes it has switched on and how strongly (its transcriptome), and treat it like a sentence, with genes playing the role of words. Show a neural network tens of millions of these profiles and train it, in the same self-supervised way we did ChatGPT, to learn how to fill in the parts it cannot see from the data’s own structure and without human labels. To do that well, the model has to learn which genes appear together, which programs switch on under which conditions, and what separates one kind of cell from another. It has to build, in effect, an internal model of cellular state by pattern matching at massive scale. A model trained this broadly, on this much data, and then adapted to many narrower tasks, is what the field calls a foundation model.

This only became possible because of the measurement revolution of Series 2. The same single-cell technologies that made cancer legible also produced something new: atlases of tens of millions of individual cells, spanning tissues, organs, donors, and species. Reference maps of healthy human cells were assembled and made public, databases grew into the tens of millions of profiles, and projects were launched with the explicit goal of measuring a billion cells to feed models like these. The fuel for creating single cell foundation models arrived at industrial scale.

As expected, models naturally followed. Some, like Geneformer and scGPT, learned from tens of millions of resting cells. Others were built to place cells from different species into one shared space. More recently, some models were trained not only on cells at rest but on cells that had been deliberately perturbed, with genes switched off or drugs applied, in an attempt to learn how they respond. Each is an effort to distill the accumulated output of the measurement revolution into a single object that knows something general about cells; a model that understands biology.

How do they understand biology? What does that even mean?

The first thing a model like this produces is a representation, or embedding: a way of placing every cell as a point in an abstract space with many dimensions, arranged so that cells in similar states sit close together and the directions in that space line up with real biology. The founding roadmap for the field, published in Cell in 2024, called this a universal representation: a learned coordinate system for cellular state, built from pattern-matching correlations, that everything else can be built on.

That coordinate system has already become useful. Hand it a newly measured cell and it can identify the cell’s type from where the cell lands. It can stitch together datasets from different labs and instruments that would otherwise be impossible to compare. It can surface candidate gene programs and regulatory relationships from the shape of the abstract embedding space itself. These are real capabilities, used in real laboratories today, and they are the kind of thing biomedicine has wanted for a long time. But they are still limited.

The deeper ambition goes beyond snapshot descriptions of cellular state. It is not enough to simply pattern-match at scale; we want to understand cause and effect. The true value of a virtual cell is simulating what will actually happen in real biology if we take some sort of action. If a model has genuinely captured the structure of cellular state, and has seen enough examples of cells being perturbed, then perhaps it can predict the effect of a perturbation it was never shown. Switch off this gene, and the model returns the profile the cell would settle into. Apply this drug, or this pair of drugs, and it predicts the response. That capability is what would turn a static map of cells into something you can run: a way to try an experiment in software (in silico) before committing it to the bench.

This is the frontier the most ambitious groups are pushing on now, and the excitement is easy to understand. A model that reliably predicts how cells respond across the genome would let researchers and clinicians screen millions of possible experiments and intervetnions in silico, then take only the most promising options into the lab or the clinic. The data exists, the models keep improving, and serious institutions have organized themselves around closing the gap. This is the most credible attempt anyone has made to build a virtual cell at the scale of real biology, and it has already produced tools that working scientists rely on.

Step back, though, and look at the data underneath almost all of this. It is overwhelmingly observational: collections of cells as they were found, resting or developing or diseased, but for the most part not deliberately acted upon. A model trained on that data learns, with real accuracy, what goes with what. Which genes appear together, which states sit beside which, what is correlated with what across millions of cells. In essence, it becomes a correlational pattern-matching engine.

But recall what a virtual cell is for. The questions that matter are not questions about mere correlation. They are questions about causation: what happens when you intervene to switch off this gene, give this drug, combine these two. Seeing and doing are not the same thing, and the distance between them is crucial.

It reminds us of a saying that is painfully familiar, yet has somehow has been overlooked by the field: “correlation does not imply causation.”

Can a model that has only ever seen snapshots of biology tell you what happens when you act on it? Can it ever generalize outside the distribution of its training data? That question has a precise answer, and it is where we turn next.

Next in the series: Description Is Not Prediction

The Virtual Cell series is published by Galen, where we are building the computational layer for understanding and programming living systems.