Virtual Biology

Biological research normally takes place in the lab. You grow cells, add a compound, and observe over days what happens. Virtual Biology follows a different path: the behavior of cells, protein molecules, or entire tissues is recreated on a computer. Such a model calculates how a biological system is likely to react before anyone sets up an experiment. Experts also speak of research “in silico” — that is, in the silicon of computer chips, as opposed to the glass dish in the lab. The goal is not to abolish the lab, but to use it more purposefully.

Why labs put models first

Biological experiments are slow and expensive. A single cell experiment can take weeks and cost thousands of euros. It often takes ten to fifteen years before a new drug reaches the market. The vast majority of candidates fail along the way, many of them only after years of work.

A computer model can test through millions of variants in just a few days. It doesn’t provide certain answers, but it does deliver a ranking. Researchers then test only the most promising hundred candidates in the lab instead of searching blindly. This pre-selection is the actual gain.

There is also an ethical dimension. Many experiments today are conducted on animals because human cells in the lab behave differently than they do in the body. The better the models become, the more animal testing can be replaced. The US Food and Drug Administration (FDA) relaxed the requirement in 2023 that every drug must undergo mandatory animal testing. Computer models are explicitly named there as an alternative.

From measurement data to computational cell models

The foundation consists of huge measurement datasets. With modern methods, it is possible to determine which genes are currently active in individual cells. Such datasets now comprise dozens of millions of individual cells from humans, mice, and plants. They are the raw material from which the models learn.

Training works similarly to how language programs are trained. A language model learns from texts which word follows which. A cell model learns from measurement data which gene activity is linked to which other activity. You can think of this as the grammar of the cell: certain patterns occur together, others exclude each other. The model recognizes these regularities without anyone explaining them to it.

After that, you can ask the model questions. What happens if you switch off a particular gene? How does a cancer cell react to this compound? The model predicts a state it has never actually seen. It is important to understand the difference from classical simulation: there, the biological rules are programmed in by hand. In Virtual Biology, the system derives the rules itself from data. A common misconception is that such predictions are proof. They are hypotheses and must be confirmed in the lab.

Pharma companies, start-ups, and Nobel Prizes

The best-known example is AlphaFold from Google DeepMind. The program predicts how a protein chain folds in space — a question that research had worked on for fifty years. In 2024, it earned the Nobel Prize in Chemistry. The predicted structures of over 200 million proteins are freely available online.

In the field of virtual cells, things are younger. The Arc Institute, a research institute backed by Nvidia and several universities, is working on a model of the human cell. Large pharmaceutical companies such as Roche and Novartis are building their own departments for this. In business news, the term usually appears when a start-up raises capital or announces a partnership with a pharmaceutical company.

As a patient, you notice little of this for now. No drug comes solely from the computer; clinical trials in humans remain mandatory. A realistic outcome is a gradual acceleration of the early research phases. Anyone reading headlines should therefore look closely: a model prediction is not proof of efficacy.