Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich
Fabian Theis argues virtual cells are best seen as perturbation-prediction world models, that data scale and honest baselines now matter most, and that the next step is lab-in-the-loop steering of experiments.
Watch on YouTube ↗Charlie's summary
- ▶ 4:31
A virtual cell is a world model: f(x, a) predicts the next cell state from the state plus an action
Fabian thinks you understand a system when you can predict its output given different inputs, which is why virtual cell now often means perturbation prediction. He calls that a bit too simplistic. A perturbation model is a world model over the transcriptome: state x plus action a gives next state x'. His lab learns a representation first, then adds a generative model such as flow matching to push the state forward along the action.
- ▶ 6:02
Perturbation prediction is taking off because perturbation space is finally getting rich
Fabian says cells were always measured in large numbers, but perturbations were scarce, often just control versus disease. Now there are experiments with 10 million transcriptomes, full CRISPR screens across around 100 cell lines, and batteries of 1,000 drugs. That density is what makes fancy prediction models possible. Pharma, he says, urgently needs cell-type-context-specific screens, which it can't yet do.
- ▶ 12:38
Organoids give multicellular readouts that cell lines can't, but nobody yet knows how to model them
Fabian is convinced the field must move to multicellular systems. Organoids include mini brains that connect and compute, and heart organoids that beat. They allow disease study where tissue is hard to get. Modeling is at the research frontier: cell types give a useful category language for dissociated cells, but there is no equivalent for local niches. For now organoids are treated as distributions of cells, though graphs could be more subtle.
- ▶ 17:41
With enough data, flexible models beat hand-built inductive bias, but perturbation data is not there yet
Fabian says the literature shows that building in strong priors can cost predictive power. His example is AlphaFold, where early versions hard-coded rotation invariance and later versions learned the symmetry from more data. Because perturbation space is still sparse, he thinks hybrid models remain right for now. Longer term, he expects flexible, data-driven models with interpretation added afterward, and he wants papers to show real scaling laws.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.