Leskovec explains that Stanford's AI virtual cell models learn cell types, states, and relationships in a fully unsupervised way, with no human biological knowledge injected into the model. ✦ AI generated
Jure Leskovec · The TWIML AI Podcast · 2026-05-21 · original ↗
starts at this moment · 4:21
“And can this kind are we able to kind of collect this information over these orders of magnitude different scales to get a strong datadriven representation of the let's say underlying uh patient in this example.”
So you don't need to insert any human bias any human knowledge of biology. The biology emerges from the data itself, right? Like cell types, cell states, relationships between them. Um that kind of human biology, how we describe it, actually emerges directly from the data.
verbatim transcript · starts at 4:21
4:21kind of collect this information over these orders of magnitude different scales to get a strong datadriven representation of the let's say underlying uh patient in this example. And the interesting thing is that this is purely doable and it's uh trained purely in an unsupervised self-supervised way. Right? So you don't need to insert any human bias any human knowledge of biology. The biology emerges from the data itself, right?
4:52Like cell types, cell states, relationships between them. Um that that kind of human biology, how we describe it, actually emerges directly from the data, right? So the model learns how to best describe the underlying processes and phenomena without us pushing it on it from from the top. Um that's kind of the exciting uh interesting kind of emergent capability there. >> And tell me if this question makes
5:20sense. I think it's related to um the the way you're describing the the training process. But is the data set that you're training on mechanistic in nature or behavioral in nature in the sense of like are you observing some behaviors of cells and then training on that data and they're you know some kind of faithful representation of mechanisms are emergent or is the does the data have mechanistic properties to it?
5:49>> The data we are using in this case case is called single cell RNA seek data. uh this is data that uh large international consortia are are are collecting but basically what it says is that you can take some sample from some let's say some tissue uh and then for every cell in that sample you measure the the number of different protein molecules inside that cell. So every cell is now
6:15represented by a 20,000 dimensional vector that tells me the abundance of that specific protein in that specific cell. Right? And every cell has different uh let's say ratios of these uh proteins depending on its uh on its type depending on its state uh and things like that. So that's the that's the raw input data and then of course because we know what the protein is we can actually bring the protein