Controlling Biology - Fabian Theis, Helmholtz Munich

From cell atlases to steering biology's next experiment

Single-cell biology is moving from mapping cell states to steering them. Theis's lab built human cell atlases, then prediction models, and now Lab Compass, an inverse model that proposes the perturbation to reach a wanted state. The field is split over whether strong priors or more data get a virtual cell there first.

Takeaways

• From mapping cells to steering the next experiment. Theis's lab built human cell atlases, moved to prediction, and now runs Lab Compass, an inverse generative model that proposes what to do to reach a wanted cell state. The stated goal is reinforcement learning with the lab in the loop.
• A virtual cell is more than perturbation prediction. Theis calls the virtual cell the holy grail, the prediction of how a cell behaves under any perturbation, and says reducing it to perturbation prediction is too simplistic. He argues the route needs world models of state and action, using flow matching on latent spaces.
• Data and models scaled by orders of magnitude. Parse Bio measured 10m transcriptomes in one experiment. Perturbation space filled to about 100 cell lines under full CRISPR and batteries of roughly 1,000 drugs. A 350-million-cell, 100-million-parameter model covering 16 species is on Hugging Face.
• Strong priors can cost predictive power. Theis splits the field between people who build priors in for explainability and robustness and people who trust bigger data for expressiveness. AlphaFold had rotational symmetry built in early and learned it from data later. He expects flexible, data-driven models with interpretation added afterwards.
• Baselines and metrics still decide results. Logistic regression remains competitive for cell type prediction, CellTypist works well, and scTab showed nonlinearity adds. A flow-matching approach won the ARC Virtual Cell Challenge, while linear models hold up for simple knockouts. A Nature paper argues the metric was the problem.
• Experimentalists' trust gates closed-loop biology. Lab Compass has run three or four lab-in-the-loop iterations with a University of Cambridge collaborator. Theis says the cultural difficulty is getting experimentalists to trust model predictions, and that robotics is premature because assay cycle times remain long.

Inside Helmholtz Munich

• Theis built the computational health center at Helmholtz Munich, Europe's largest research organisation by funding. It has about 45 group leaders and professors and 500 to 600 people, the largest academic health AI centre in Europe.
• Helmholtz works like a distributed NIH, spread across Germany. Staff hold affiliations at the Technical University of Munich and the University of Munich, with no undergraduate teaching, only PhD students and postdocs.