Fabian Theis describes a practical next step for virtual cells: use a model to propose an intervention, run the experiment, and learn from the result. On Latent Space, the Helmholtz Munich researcher explains why that loop needs better biological questions as much as larger models.

A cell map answers only the first question

Single-cell transcriptomics measures RNA messages, producing a profile of which genes a cell is expressing. Researchers can organize those profiles into maps of cell types and states. Theis describes each cell as roughly 20,000 expression measurements, usually compressed into a smaller learned representation. 08:25–09:47

Prediction adds an input. What happens when a drug, gene disruption, hormone, or neighboring cell changes the system? A virtual-cell model needs both the starting state and the intervention to predict a response. 10:49–12:34; 24:11–25:39

Causal control asks a further question: which intervention will move the system toward a chosen state? That proposal has to survive an experiment. A detailed map or accurate classification score cannot answer it alone.

Bigger datasets need broader experiments

Theis argues that biology has collected many cells while sampling relatively few things done to them. More cells under the same conditions provide a different kind of information from more interventions across tissues and environments. 06:47–08:01

Organoids, small tissue-like systems grown in a dish, expand what researchers can test. They capture interactions missing from single-cell-type cultures. They also have limitations, including absent immune components and incomplete support systems. Representing their neighborhoods remains an open modeling problem. 12:58–16:00

NicheFormer illustrates a harder question: can one cell’s expression profile predict the state of its neighbors? Communication between cells may leave measurable traces. That task demands different information from assigning a familiar cell-type label. 21:41–22:15

Make the baseline earn its place

Theis accepts that simple models remain formidable. Logistic regression can classify many cell types well. In noisy drug-response data, even predicting the average can be hard to beat. 26:39–27:47; 29:39–30:12

But a benchmark score depends on what the test rewards. The discussion of the Arc Virtual Cell Challenge covers exploitable metrics and the risk of confusing a scoring shortcut with biological insight. Theis says more complex questions can justify nonlinear models. He also wants evidence that performance improves as training data grows. 22:58–23:48; 28:35–29:30

His architecture separates learning a useful cell representation from predicting changes under an action. CellFlow then uses flow matching, a generative method for transforming distributions, to model intervention effects. 24:19–25:39

The feedback loop is a collaboration problem

Near the end, Theis describes Lab Compass, an inverse generative model that proposes actions for reaching a desired biological state. He reports three or four iterative experimental cycles with a Cambridge collaborator studying blood-cell development. 48:13–48:53

The difficulty extends beyond selecting the next experiment. Someone has to spend laboratory time on the model’s suggestion. Long assay cycles make that commitment expensive. Theis describes the current approach as active learning, choosing informative experiments. Reinforcement learning across the whole loop remains an ambition. 49:01–49:47

His training model addresses the same coordination problem. Students work with a methods supervisor and a domain supervisor, attending both groups’ meetings. They learn to translate an unresolved biological problem into a question a model can test. 41:41–45:52

Distinctive data gives the model a business

Theis sees biological AI platforms repeatedly pivot toward drug development because investors understand the resulting asset. He offers another route: combine modeling with measurements customers want. 38:18–39:47

The Cancer Organoid Perturbation Atlas brings pharmaceutical partners into a consortium with early access to experimental data. Its value comes from the combination of distinctive data and modeling. That arrangement makes the customer clearer than a general promise to decode biology. 40:57–41:38

The research direction is concrete: connect representations, action-conditioned predictions, experimental choices, and new measurements. Progress depends on whether each proposed intervention teaches the next model something worth acting on.

Caveat: The interview describes research models and limited laboratory feedback cycles. It does not establish clinical deployment or reliable control of human disease.

Source: Latent Space, “Controlling Biology,” November 6, 2026