Towards Virtual Tissue Models

Barbara Treutlein is pioneering ways to understand and model the human brain. As a professor at ETH Zurich and winner of the 100 Million Cell Challenge alongside collaborator Fabian Theis, she is combining innovative approaches to study brain development underlying mechanisms of disease with single cell analysis.

In this Innovation Spotlight, Treutlein shares her excitement of using organoids to model the human brain, harnessing AI to accelerate laboratory discoveries, and discusses how massive-scale single-cell experiments are bringing scientists closer to complete virtual models of human tissues. We use single cell genomic technologies to understand human organ development.

Could you tell us about your research?

We use single cell genomic technologies to understand human organ development. We focus on the brain, we use stem cell derived organoids to model human brain development. These are three-dimensional tissues grown in vitro that mimic the architecture of human brain tissue, but also functional aspects of the human brain. They are exciting systems because we can grow them at scale. We can grow many tissues for an individual. We can grow them from different individuals from whom we have induced pluripotent stem (iPS) cells. Then we can also genetically and environmentally perturb them. We use the single cell data from these organoids to learn how cells adopt their fate, which genes and gene regulatory programs are underlying cell fate specification, and how perturbations lead to disease.

Could you tell us about the project you're going to be running on QuantumScale?

Organoids are a relatively new model. In the past few years, we have tried to move them into really high throughput perturbation screens. So far, we have been perturbing 20-30 genes at a time either genetically or using drugs. But we are ready to increase the scale because we have established automated organoid culture and automated dissociation of organoids to do higher throughput screens . Our project relies on a high throughput pathway perturbation screen in human organoids. We are focusing on all major regions of the brain from forebrain to midbrain to hind brain. The idea is to perturb many signal transduction pathways with drugs and also perform a lot of combinatorial perturbations. The final goal is then to take all this data and train AI-based models that will eventually be able to predict unseen perturbations. AI could also potentially guide future perturbation screens to tell us which conditions to test, which ones are maybe the most powerful or promising.

Are you focused on understanding a particular disease or specific pathway as part of the screen?

We try to be as broad as possible, targeting all major pathways underlying, in principle, the function of any cell type, and then explore how these pathways are active in human brain cells and what's the crosstalk between pathways.

I think it's really an exciting time where we have the human models and can grow them at scale to run these kinds of perturbation screens in a relevant human context. We cannot use primary human brain tissue for this, and two-dimensional cultures of human cells do not mimic in vivo complexity. Organoids, in contrast, have a very large diversity of cell types that we can perturb together.

We can ask “what is the response of a given cell type?” and “why do different cell types react differentially?” The AI model can learn all this and, in the future, make predictions.

How many organoids can you produce at once?

We are growing multiple thousands of organoids.

We are now scaling up the number of perturbations by a factor of 10 compared to what we were previously doing. This includes individual perturbations but also combinations, which we haven't done before. The combinatorial space is huge, so we will not fully exhaust the combinatorial space. This is why we are excited to train models leaving out combinations we actually measured and then seeing whether we can predict them. I think we are still exploring what would be a good number of perturbations, number of cells, number of tissues to profile to be able to do predictive modeling.

We are aiming for 2,000-3,000 cells per sample, but this is driven by the cellular diversity of our organoids. Depending on how many cell types you find in a tissue, you may need more or fewer cells so you can profile enough cells per cell type and per perturbation. Because organoids are heterogeneous, we need to make sure to have, in this case, several thousands of cells per organoid and then replicates to also take into account and be able to control for technical batch effects.

When you talk to researchers doing cell line screening, they say that they do not need a lot of cells.

It is certainly very different. Sometimes these screens are in, let’s say, three cancer cell lines. Every one of these lines are very homogeneous, possibly containing just one cell type, so it's a whole different problem. In that situation, you can use 100 cells per sample. With our organoids, we are talking about complex human tissue that contains many different neuronal cell populations, glial cell populations, and other cell types. To me, that makes it exciting because we are going beyond cell lines that do not truly mimic human physiology.

What longer-term potential applications or discoveries might emerge from this work?

The big goal is really to learn the full phenotypic landscape of all human cell types. How many cell states exist in a human body, what genetic mechanisms exist to regulate these states, and what environmental cues move a cell from one state to another. Organoids are a good in vitro model of human organs that allow you to probe this landscape.

Eventually, if you put all this data into a big foundation model, the hope is you can predict how any perturbation would change a cell’s phenotype, which cell types are most affected, and what are the off-target effects of a drug. Ultimately, we are moving towards this huge goal of a virtual model of human tissues. But we need to generate these perturbation atlas data sets to move towards that.

Will there be a day where the AI model replaces the need to do an experiment in the lab?

We might be putting ourselves out of jobs eventually, but there are still many things to do. We are fine tuning the choice of drugs, and even there we are trying to go at it in a data-driven way. We infer gene regulatory networks from single-cell multiomic experiments, which allow us to do in silico perturbations to identify perturbations that may have the largest effects. Of course, once you train a model, this model can also further predict which future experiments you should do. I envision that these models would help us decide what to actually measure in the lab. Right now, there are lots of drugs you can test in principle, but which ones do you choose and which combinations? I would be happy if AI could help me make these choices to perform the most informative experiment.

It is becoming easier to do larger-scale screening with single cell technologies. How is this impacting your field?

In our case, we are hoping to move towards causality. We can finally start to explore the vast space of possible genetic and environmental perturbations that any cell in our body is exposed to and measure the response to these perturbations. We can do this for heterogeneous tissues with many different cell types, even if some cell types are very rare, because large-scale single cell experiments allow you to capture even rare cell types. Also, we can strongly increase the number of time-points that can be measured across cellular trajectories to learn more about cellular differentiation during development or as part of a response to perturbation. Overall, this gives us a much more detailed view on cellular dynamics that we could not capture previously with snapshot data coarsely sampled over time. I think it is super exciting that the field has opened up to these opportunities.

A version of this article was previously published here.