Fabian Theis on Why Virtual Cells Mean Perturbation Prediction

Key Takeaways

• A virtual cell functions as an input-output predictor rather than an observational atlas.
• High-throughput screens are expanding from passive single-cell profiling to full CRISPR knockouts across 100 cell lines and batteries of 1,000 drugs.
• Pharmaceutical companies cannot currently screen drugs against specific cell-type contexts, which drives demand for predictive biological world models.
• Generative tools like CellFlow and foundation models like scConcept and Nicheformer bridge the gap between static transcriptomics and combinatorial perturbation response.

Cell Atlases Are Not Enough

Fabian Theis wants researchers to stop treating virtual cells like digital anatomical diagrams. For years, computational biology poured resources into mapping cell types: sequencing RNA from tissue samples to build massive catalogs of what cells look like in steady or diseased states. Those observational catalogs gave researchers millions of cells, but they failed to answer the practical question every drug developer actually cares about: what happens when you strike the cell with a chemical or genetic shock?

“Whenever people talk about virtual cells, maybe also your question, I think what you actually ask is how does Can you predict how a cell would behave under all kind of perturbations?”▶ 4:28 (play from 4:28) Theis explained. In classical experiments, the number of sampled perturbations was tiny compared to the sheer volume of observed cells. You might have had thousands of cells profiled across only two conditions: a healthy control and a sick donor. That imbalance left computational models blind to cause and effect. A virtual cell cannot simply describe biology. It must predict biological responses to shocks it has never seen in the wet lab.

The Combinatorial Explosion of Perturb-Seq

The bottleneck shifted because experimental scale caught up with machine learning architectures. Instead of testing one chemical compound at a time on an immortalized cancer line, high-throughput Perturb-seq and CRISPR libraries now produce dense intervention data across diverse genetic backgrounds.

“Now you sort of have, I don't know, 100 different cell lines where you do sort of full perturbsic like full CRISPR, or you have like batteries of 1,000 drugs or so added, and suddenly this space gets really rich, right?”▶ 7:27 (play from 7:27) said Theis. That richness allows models like scConcept, Nicheformer, and generative frameworks like CellFlow to learn the underlying mechanics of cellular state transitions.

The commercial pressure here comes from drug discovery pipelines. Today, pharmaceutical screening usually happens in generic in vitro assays that flatten biological complexity. A drug that works in a standard flat dish often fails in human tissue because it acts differently inside a microglia than it does inside a cardiomyocyte. “Pharma is really desperately uh needing to be cell type context-specific in their screens, which at the moment they can't do yet,”▶ 5:42 (play from 5:42) Theis said. As Theis noted, “like a virtual cell model should take all of it, right? and also combinations thereof, of course.”▶ 12:30 (play from 12:30) When you test combinations of three or four molecules across dozens of cell types, physical testing becomes physically impossible. Only predictive models can search that space.

What to Do With This

Audit your biological ML roadmap this week. If your models predict static properties like binding affinity without conditioning on cell type and genetic background, you are building for 2018 pharma. Pick one target pipeline, pull an open-source Perturb-seq dataset covering multiple cell lines, and test whether your existing architecture can predict the transcriptional readout of an unseen gene knockout.