@fabian_theis Fabian Theis described the steps his research teams have taken to build the capability to build this kind of model, and what it took to validate it in the lab.
Don’t sleep on this, Virtual Biology is heating up again.
3/
The latest generation of models need to control biology
With lab-in-the-loop as the bottleneck for both experimental biology and data generation, understanding how to design experiments that will produce the outcomes you want is increasingly important.
@fabian_theis
1/
@fabian_theis This enables an active-learning environment in which an AI steers the next step of an experiment to maximize learning or some objective function, and it can also help use design a campaign that can explore a landscape of possible biological states.
2/
Why 2D brain segmentation is hopeless in grey matter:
It is so densely packed with dendrites that an H&E mostly shows slices of wiring from cells near and far.
For gross tissue features, raw unsegmented RNA gives you more signal than forced cell boundaries.
4/ The cells branch.
One cell forks for 22 µm. In 2D, that fork looks like two separate profiles 16 µm apart with nothing connecting them. The model counts two cells.
All a 2D segmenter sees is a flat slice. (Data: Greiner et al. 2025)
Why standard nucleus-first segmenters fail on heart muscle:
Cardiomyocytes run up to 140 µm long. In a 5 µm section, you only catch a slice of each cell.
Most profiles have no nucleus at all.
If your model seeds from nuclei, it has nothing to start from. 1/
3/ Slice angle changes the apparent shape.
Along fibers, the cell is a ribbon; across, a polygon.
Turning the knife 60° drops ribbons from 63% to 9% and jumps profile count from 37 to 67. 57 µm deeper, fibers rotate 60°, so one section cuts along some cells and across others.
4/ The cells branch.
One cell forks for 22 µm. In 2D, that fork looks like two separate profiles 16 µm apart with nothing connecting them. The model counts two cells.
All a 2D segmenter sees is a flat slice. (Data: Greiner et al. 2025)
3/ Slice angle changes the apparent shape.
Along fibers, the cell is a ribbon; across, a polygon.
Turning the knife 60° drops ribbons from 63% to 9% and jumps profile count from 37 to 67. 57 µm deeper, fibers rotate 60°, so one section cuts along some cells and across others.
🧪🧭 What if we could ask a model which experiment gives us the cells we want?
Our new preprint, LabCompass, takes a step in this direction through generative inverse design. For me, this is one of the most exciting directions for the field.
Together with many others, we have been working on predicting how cells respond to perturbations. How can we use these models to design experiments that achieve a desired biological outcome?
🎛️ This inverse problem is difficult. Combinations and doses create spaces that cannot be exhaustively screened, and each experiment produces a distribution of cell states. Predictive models are also imperfect, particularly in unexplored regions where we want to discover something new.
LabCompass couples a generative model of experimental designs with a forward model of cellular responses. The desired phenotype guides generation in continuous design space, while a learned prior favors plausible designs. Selected candidates are tested experimentally, and the measurements inform the next round.
💡 The method is not tied to a particular virtual cell model. It can work with different differentiable forward models and cellular readouts, and the same trained models can address different phenotypic objectives.
🩸 We demonstrated this in primary human blood stem and progenitor cells. Across the study, we profiled 78 million cells and experimentally validated conditions enriching megakaryocytic, erythroid, monocytic and early lymphoid populations.
One result I particularly like: distinct culture recipes produced similar target enrichments. Several solutions to the same biological objective give us alternatives to investigate and develop further.
🔬 The opportunity extends beyond blood and cell culture. We also applied LabCompass to single-cell transcriptomic data from viral infection. I’m excited to explore its potential for other perturbation systems, drug combinations, differentiation and organoids.
This raises a question for virtual cell models: alongside prediction accuracy, how effectively can they guide experiments toward desired biological outcomes?
If this connects to a problem you are working on, I’d love to hear from you. We’re keen to see how LabCompass transfers to new systems and happy to explore ideas together.
A wonderful collaboration with Bertie Göttgens and Nicola Wilson at the Cambridge Stem Cell Institute. Huge credit to joint first authors Licyel Paulas-Condori, Alessandro Palma, Lorenzo Consoli and Juan Antonio Rubio-Lara, and the whole team across both labs!
📄 Preprint: https://t.co/EVieJpAFyY
💻 Code: https://t.co/oPCXkaSAiH
#InverseDesign #VirtualCell #SingleCell
I sat down with Latent Space to talk about the research preview of Omnii, our latest genome language model, and what becomes possible when AI can reason more directly across biology.
We covered a lot of ground, especially the potential across human health.
Much more to share in the coming months.
https://t.co/Eh5TYFoDED