In this talk, Amos Tanay (Weizmann Institute) presents a framework that leverages single‑cell transcriptional atlases, modeled as quantitative manifolds, together with temporal flow models to dissect gene function during mouse gastrulation. He describes how the team constructs a high‑resolution cell‑state manifold of the whole embryo, visualizes developmental trajectories as flow diagrams, and uses these tools to interrogate the DNA demethylation TET system. By comparing whole‑embryo triple‑knockout embryos (mutant embryonic mass with wild‑type extra‑embryonic tissue) to chimeric embryos containing both mutant and control cells, they reveal starkly different phenotypes: the former shows loss of multiple lineages and a shift toward extra‑embryonic mesoderm, while the latter appears largely normal with only specific lineage deficits (e.g., blood, notochord). High‑resolution ATAC‑seq and DNA methylation profiling pinpoint disrupted transcription factors and signaling molecules, notably loss of Lefty2 expression in mesoderm, which impairs epiblast‑to‑ectoderm conversion. Methylation analysis shows widespread enhancer hyper‑methylation in knockouts, yet many promoters remain protected, indicating selective vulnerability. A lineage‑specific up‑regulation of TET3 in the notochord illustrates direct regulatory roles. Tanay emphasizes that gene function in development is highly quantitative, context‑dependent, and cannot be reduced to single‑gene master‑regulator models. The talk concludes with a discussion on the combinatorial nature of developmental regulation and the challenges of translating atlas‑based insights to disease mechanisms.
Chapters
our next speaker, who is Amos Tanay from the Weizmann Institute in Israel. So I'd like to welcome Amos to tell us about single-cell flows over developmental manifolds, or I guess, what is the developmental function of my gene?
So I want to give this brief discussion some ideas we are looking at With respect to the very classical question of how to use a cell atlas in order to understand gene functions, think about it in the developmental context, of course, it's very natural to replace gene by a gene variant. So in essence, we are all busy generating atlases, we think of those as manifold, and here is one manifold we were generating together with Jonathan Sternzel group at Weizmann. Syphon Chang and Marcus Newhart are the leaders here. And for us, this manifold is actually a snapshot of a quantitative, in that case transcriptional, cell state, which— or cell state space. So each position here on the atlas actually represents the distribution of gene expression. And our aim is to try and convert this into something that is as physical or biophysical as possible in the sense that we interpret possible movements along of cells in development along this atlas as something that have an actual mechanistic and regulatory sense. So, this mouse restoration embryo manifold is something we want to use as a basis for doing much more. So in practice, by the way, highlighting some of the tools that were initially internal in the group are now out in BioRxiv, that allow us to construct those manifold representations. Again, this basic representation of the cell state space is a set of transcriptional, quantitative transcriptional distribution. But given those representations, we want to go more and understand development. And for us, the major challenge here is transforming what is initially a picture that is giving us to add the resolution of a single cell to something that is actually capturing a tissue or an evolving tissue. So for mouse gastrulation, it's very natural to think about capturing the whole thing. Here are single mouse embryos that we can map on our manifold, and we did it systematically. And in a recent paper, we generate those huge and frightening flow diagrams that you can simplify to generate those vein diagrams. That actually tells you how the embryo is gastrulating over time. Time here is the y-axis, and the different colors mean different cell types that you can follow. So that's the basis. Okay, so we have a manifold representation. Let's assume that we even have a temporal flow model on top of it. And now let's go back to the question of how we understand gene function. Again, a very, very classical question. And the problem is that when you're looking at gene function, you in the context of such a complicated bifurcating or multifurcating process, there is no one single answer you can come up with if you really want to understand what happens. The classical idiom is of course to perturb the gene and look what happened. We need to do something causal. But when you perturb the gene, of course, maybe something completely new is going to come out, like a state that was never there before. But more likely, there are going to be all kinds of perturbations like intrinsic effects can prevent differentiation and prevent the appearance of a cell type altogether. Other types of perturbation can slow down or accelerate the process, which will generate huge changes or huge, at least confounders if you just compare one system to another. And of course, there are going to be an accumulation of many indirect effects because if you are perturbing even mildly the signaling in one cell type, this is going to propagate as a shockwave throughout the system, and you will see many, many effects that are not a direct impact of your own beloved gene. So this is what we want to understand. And for this talk, very, very quickly, we are going to talk about the function of a particular system of interest, which is the TED system. So replace the general question by a question of what the TED system is doing in plasturation. So in a sense, the TED system is a demethylation or DNA demethylation system that works together with the DNMT system, which is the de novo the DNA methylation system to somehow stabilize the DNA methylation of the developing embryo, and later also of other types of tissues. So we think of them as a demethylation agent, and the literature is not completely clear about what these guys are doing. It is clear that if you knock out all three of the genes, you get embryonic lethality. So things look bad. But we want to do more, and we want to use our ATLAS for that. So with Jonathan, we designed a series of experiments that leverage our ability to try and separate direct intrinsic and indirect effect. So the first of those is a whole embryo mutant where you replace only the embryonic mass, not the extraembryonic mass, with mutants, and you compare it to controls. This is the kind of scheme you are running here. And when you do it and you replace triple knockout test cells and generate embryos that have normal extraembryonic but a completely mutated embryonic part, you see major effects that you can project on your manifold and on your flow model. So here are controls at the left and the mutants on the right here. And you see the first type of effect that we will call major compositional differences. So there are kind of several lineages that are completely missing, okay? In particular, the embryo, and you also see it under the microscope, is channeling itself toward the extraembryonic mesoderm lineage and fill up the cavity that is usually going to be the part where the placenta and the allantois is going to grow. So now it's a very aberrant process. And it's— this is what's very similar to the germline knockout observation that people have made before. So this is telling us that that is a terrible thing, and it is maybe blocking all possible cell types. But the picture looks very different if you are using an alternative system where you are injecting and generating chimeric embryo where you are putting inside the inner cell mass both controls and mutant cells and grow them together in green and in red. And then the picture looks very different. And suddenly the knockout, those genes, seems to have almost no effect. Okay. So what you see here are the distribution of the composition differences. There are differences. For example, there is no real blood and there is no notochord. In the chimera, but the other type, other lineage seems pretty normal, which is kind of shocking. And again, there are lots of additional intricacies, for example, in the complete tetraploid embryos, the whole embryo mutants, we did have blood. Okay, so you understand that the process is complicated, and each gene is functioning in a highly context-dependent fashion. And And we want to resolve this. So luckily, we have the atlas and we have all the temporal flows, we can follow what happens at very high resolution. Let's start from the epiblast. So these are those Venn diagrams that show you how things are evolving in time. So from the epiblast, you should transform into the primitive streak and later into the early neuroectoderm lineages. And first of all, you want to check if the mutants are generating a proper epiblast. And if you just look at the transcription programs that are supposed to run in the epiblast, these are the distributions here, it seems pretty normal, or almost completely normal. Which is striking. You are knocking out the TET system, but the epiblast, by and large, the program is almost normal. You are continuing from the primitive streak, which is a transient state, and generate the nascent mesoderm, which is a hub that will later generate many different lineages in the mesoderm. And in the mesoderm, you can also make those comparisons and look at the nascent mesoderm program that the knockouts are running, and it also seems quite normal. And by the way, it's very similar if you're looking at the whole embryo, So this is quite striking. So the knockout, in particular the whole embryo, is taking only one path, okay, but when it is running this path, it is doing it pretty normally. And we want to understand what is initiating this blockage of conversion of the ectoderm or the caudal epiblast or parts of the rostral mesoderm that are not present in the whole embryo. So we can now do very high-resolution analysis which those ATAC-Seqs are giving us and control everything at the maximum level. We control time, we control lineage, and then we do differential expression analysis per embryo. And each dot here represents a single embryo that is completely controlled and find genes that are still biased compared to what the wild type should have run. And you find your two major transcription factors, or a transcription factor on a signal molecule in the epiblast that are not normal and supposedly are causing slowdown of the epiblast conversion. And very, very interestingly, inside the mesoderm, massive perturbation of some major signaling molecules, in particular the famous Lefty2 that was implicated before with the whole effect of the TET mutant. So now we kind of understand what happens, because if you are not secreting Lefty2 in the mesoderm, okay, you are going to prevent the continuous conversion of the epiblast toward the ectoderm here. And you're also going to prevent the continuous differentiation of the nascent mesoderm to the rostral part, and everything will be channeled toward the extraembryonic mesoderm. On the other hand, if you have proper signalling support from normal host or normal controls, this will be prevented. So we can really try and continue and understand how this whole thing works. Now let's go back to the actual mechanism of action of those TETs. And we think it should go through methylation. So here are very quickly comparison of DNA methylation in control and in knockouts. And you see in the y-axis, massive accumulation of methylation in the TKO when you sort them for embryos. This is, again, using this very, very well-controlled system. Turn out that when you are losing the TETs, you are still protecting from methylation many different loci, in particular promoter CTCF sites and regions that are protected by, you know, the core epigenetic machinery that maintain activity. On the other hand, as you see in this graph here, the lower right one, almost all of the putative enhancers are gaining dramatic methylation. So it means that this knockout that we are looking at have an effect on thousands of different loci in the genome. It means that these thousands of effects are usually well tolerated, but in some cases, they will generate transcriptional perturbation. And this would spread in the developmental hierarchy and create those dramatic effects that we are seeing. There is tons of additional data. This is highly unpublished at the moment, hoping to submit it soon. Additional data on double knockouts, and in parallel, a story that is doing exactly the same to DNMTs with very similar results. Okay, so in general, we think that TETs, or generally the DNA methylation machinery, is tinkering expression. Okay, in some case, this tinkering is fatal or have dramatic consequences. But most of the changes you'll see in the embryo are going to be a consequence of very mild perturbation that intrinsically happened. And in essence, you need to define the function quantitatively in each lineage time, you cannot just say this gene is responsible for that. Won't work, especially when we have such beautiful quantitative model to support this. So just a final word about things, how things get even more complicated in development. So we looked at a particular very early developmental time window, which is early, early gastrulation. Here is differentiation of the notochord that is happening immediately during gastrulation. And somehow those TKO cells, the tetrapods, failed to generate this notochord in chimera or whole embryo context. Okay. So every gene you are going to look at is going to come up again and again and again, and we have additional function. In that case, it is particularly interesting, we can use our temporal atlas to trace back the kinetics of the TET machinery. And we see that only in the notochord TET3 is going to climb up and get induced. And this is an example that where an enzyme or a gene is actually directly regulating lineage, okay. So You also have these types of effects in any developmental system you're going to. I'm going to finish up now. I want to thank the group, Markus, Niten, Svald, Zohar, Mukamel, and the beautiful people in my group. Of course, this major collaboration with our neighboring lab, Jonathan and his gang, in particular, Saifan Cheng, amazing work. In all aspects. So thank you very much.
Thank you, Anas. We might have time for one question. Maybe I'll just mention one if people are— we don't have any coming in from the question platform, but I was just wondering about your thoughts about, you know, the reuse of genes or maybe more generally molecules in development, you know, is it, you know, how should we think about that? Is it, should we think about it in terms of combinatorics, or are there specific control molecules that are like, you know, is it always combinations, or are there, you know, are we going to be able to frequently see, you know, a specific transcription factor that really is a master regulator for individual processes in some kind of sequence?
So that's definitely— it's a great question. And in particular, I want to highlight the point that in any system we are looking at, like, just giving one gene one function is not working, okay? It's a system that seems to be evolving to robustness through the availability of multiple factors that regulate any process you can think of. I think the challenge today is going to transform our quite unbiased picture of the system through generation of atlases, okay, into mechanistic models that we can work with. Because the phenomenology of that, like identifying the cell types or even doing the 3D modeling in the coming years, that's— I think we all understand that it's kind of doable. But understanding now how to take a kid with a certain gene variant and explain a disease, okay, or even for our own curiosity to understand how the whole thing is working, this is still a long way from us. And I think we have a lot of clues coming up, especially from the simultaneous modeling and recovery of those lineages and the temporal dimension, which is being discovered now gradually. And I think it will be amazing to continue and look at it at higher resolution.
Great, that— thanks, that obviously has big implications for disease research, as you mentioned. Okay, thank you, Amos.
