Fabian Theis told Latent Space that biology has moved from mapping cells to predicting them and now steering the next experiment.
The obvious read is renewed pharma interest in virtual cells, with the hosts citing Biohub's $1.8 billion data push with DOE, NIH, Meta and Google as the trigger. I am skeptical because Theis himself kept flagging the gap between metaphor and measurement. He leads AI and computational biology at Helmholtz Munich, and his lab built Scanpy, AnnData and SCverse that working biologists actually use, so his caution carries weight, and he said perturbation prediction started with a 2019 model called SEGAN that still looks competitive because the questions asked were too simple.
Theis stated the claim fairly before testing it: a virtual cell should predict output behavior given any input, from CRISPR knockouts and human genetics to media cocktails, signaling gradients and time, including combinations. That would give pharma the cell type context specific screens he said it desperately needs and cannot yet run at scale. What would have to be true is a perturbation space as rich as the cell space, plus a model that holds across RNA and other molecular layers.
That space is not there yet. Theis said cell counts exploded to 10 million transcriptomes in a single PacBio run he called a giga lab, while perturbation sampling stayed scarce at control versus disease. He said only now is sampling growing to around 100 cell lines with full CRISPR or batteries of 1,000 drugs. Each cell is roughly a 20,000 dimensional transcript vector that clusters into types or trajectories, Theis said, but he admitted there is no agreed representation for spatially connected organoids so the field falls back to distributions of dissociated cells.
Mini brains that compute and beating heart clumps in a dish impressed him, Theis said, yet those organoids still miss immune context and feeding and other physiology. The modeling debate is the tell. Theis said strong prior structure can cost predictive power, a lesson he said large scale machine learning has shown over decades, and he said his field is not yet in the data rich regime where pure scale wins so hybrids may still rule. That fits his outside critique that current perturbation benchmarks are not hard enough to justify complex models, per Latent Space companion notes.
His fix is Lab Compass, an inverse generative model that takes a forward world model and asks what intervention reaches a target state, with full reinforcement learning in the loop as the long term vision, though he called that vision a bit out there. If you are building in this space, sell one prediction that survives a harder benchmark and a real organoid, not a virtual cell.
