Scientist Stories: Fei Chen, Next Generation Tools for Spatial Genomics

We need to get information in space as well as in depth. And hopefully at some point in time as well. So I came up with the idea that we should call this Space Force because there's an enormous advantage of being able to understand not only the genes that are being expressed in a cell, but where those cells are and where they're positioned. Relative to one another. And of course, I just couldn't resist calling it Space Force. I was going to call it Space Balls, so it sounds like Space Force was better. So our first speaker is going to be Fei Chen, and Peter will introduce him. And I did want to also mention our next speaker. Our talks are given on the first Monday of each month. So our next speaker is going to be Amjad Askari, who is a postdoctoral scholar in Mike Elowitz's lab. And as you may know, Michael is going to be giving a talk this Thursday at the Salk, although he's not going to be speaking on MEMWAR, which is a way of doing lineage tracing at single-cell transcriptomic depth while preserving positional information. That's going to be given by Amjad, who developed that technology. So with that, I'll welcome you all to this continuation of our series in single-cell biology, Space Force. It's going to start with Fei Chen and continue with a series of really exciting speakers throughout the rest of the year. Peter, I'll turn it over to you.

Right. Thanks, Jeff. And thanks, Jeff and Brian, for putting this together. Yeah, so it's really fantastic to kick off this series with with Faye. I imagine most of you are already familiar with his work. So Faye is a core faculty member at the Broad Institute and assistant professor at the Harvard Stem Cell and Regenerative Biology Department. He obtained his PhD in biological engineering from MIT and actually only in 2016. Okay. So he's quite a clearly on a stellar trajectory. He was a Schmidt Fellow at the Broad Institute from 2017 to 2020. And as I think you know, and you know why we invited him to kick off this series, Faye has already made really fantastic contributions to technology development and application of those technologies in the area of spatial transcriptomics and genomics. Okay, so Fei is one of the co-developers of SlideSeq, which allows near single-cell high-resolution transcriptomic profiling in situ of intact tissues, and has also developed methods for in situ genome sequencing to allow spatial imaging of genome organization in single cells. So, um, you know, clearly some really important and exciting contributions to this area. And I'll—

that—

with that, I'll hand over to Fei and let him tell us more about it. Thanks very much for speaking, Fei.

Thank you. Um, thank you for that amazing introduction. Um, yeah, I'm really excited to kick off, uh, Space Force here today. And really today, yeah, I'll be sharing with you some of the tools and technologies that my lab is developing on spatial genomics. And I heard a little bit that you guys are also interested in maybe having a future seminar series on temporal genomics. And I am also very interested in that, but I won't have too much time to go into that aspect. So, you know, I like want to start off my talk with, you know, this video. And this is kind of like every cell in a developing Drosophila embryo from shortly after it's fertilized to like when it's practically walking off the microscope. And, you know, I think a lot of us are drawn to this video because it showcases the wonder, you know, of biology. It also showcases the immense spatial and temporal coordination that exists in biological processes. And so while in Drosophila this developmental process is fairly stereotyped, and we've had quite a lot of luck— not us, but developmental biologists have had quite a lot of luck in labeling cells and molecules one at a time or a couple at a time to build up an understanding of how this process works. But really, for mammalian biology and for disease processes, they're not really very stereotyped. And what we really need are tools that allow us to study these molecular mechanisms governing spatial temporal biology in that context. And so I think you guys had quite a great seminar series on spatial single-cell transcriptomics. And it goes without mentioning that this has been a recent technological advance that's really revolutionized our ability to study molecular mechanisms in cells. And that's because in single-cell transcriptomics, you can take a tissue and break it down to its constituent cells. And you can sample maybe this transcriptome or maybe the epigenome of each cell. And this is really kind of powerful because it's comprehensive. It lets you look at the cell types and states and the molecular programs and also maybe how those molecular programs change in the context of disease or other or development or other biological phenomena. But I think maybe, as was alluded to already in the introduction, fundamentally, we have to remember that what we're doing in these processes is we're taking complex tissues which have inherent structure, and we're dissociating them into this slurry of single cells. And in that process, you lose this context for cell-cell communication or maybe cell-cell interactions and differential expression as a function of where cells are localized in the tissue. And so we'd really like to know not only the states of cells, but maybe how they relate in relation to the tissue and how those states are an emergent phenomenon from the organization. Separately, I think, as was touched on a little bit, cells in tissues are pretty dynamic. And they might migrate, they differentiate. And so we kind of lose that temporal context when we dissociate as well. We get this snapshot. And that's something that we would also kind of need new tools for in the field. So today I'll really talk to you about approaches to allow us to look at— take the unbiased measurements of genomics. And by unbiased, I mean unbiased measurements of molecules, which are usually DNA and RNA and place that within the context of tissue architecture and structure. And this is really kind of driven by a couple of needs and motivations when we set out to develop these tools. One is just that— and I'm sure you all appreciate— the recent developments in single cell have enabled us to collect quite large atlases of cell types and states, but we have no idea, for example, how these cells are interacting or where they're placed in the tissue. For example, this is a couple hundred thousand cells from just a very small volume of the mouse bed nucleus. And what we really like to know are the interactions and relationships and communications between these cells. But at the moment, we have really good ways of collecting these high-dimensional datasets, but not very many great ways of placing those cells in context. And I think the second is something that we all, like, intuit, which is that we need ways to connect pathology to tissue. You know, we've had pathology for a long time, and just from simple H&E staining, we can recognize, like, pathological regions and the differences between maybe malignant and benign, or maybe disease and non-disease tissues. And— but we have like very few ways to connect those cell type and cellular architecture changes— sorry, like pathological architecture changes with cell types and cellular programs and molecular mechanisms. And so being able to connect pathology to gene expression is something that we'd really be quite interested in. Beyond gene expression, maybe we'd also be interested in connecting pathology, for example, here in the case of amyloid beta plaques, to the surrounding cell types and maybe how the gene expression in the surrounding cell types change. Right. OK. So to address this loss of spatial context when performing single-cell sequencing, we developed this tool called SlideSeq, which enables genome-wide expression profiling in tissues at about 10-micron resolution, which is about the size of single cells in mammalian tissues. And I want to mention that all of this work is a joint effort between my group and Evan McCosko's group. And we work quite closely. And we started our groups together at the Broad Institute around the same time. And so the big picture is that SlideSeq effectively lets you do untargeted single-cell sequencing and at the same time lets you know where those cells came from. In your tissue. And so how do we, how do we do that? In SlideSeq, we deposit individually barcoded 10-micron polystyrene beads in a 2D monolayer on a surface, and each one of these beads has maybe 10 million or so oligonucleotides on them. Each oligonucleotide has a PCR handle, what we call a spatial bead barcode, which is just that each bead has a clonal sequence that's unique to that bead but different between beads. On each oligo, there's a unique molecular identifier to allow for digital counting of transcript expression and then a poly-T tail, which allows for the capture of mRNA. And so we then deposit these beads in these 2D arrays. This is a coverslip that's 40 millimeters in diameter. And each one of these white circles is an array of about 100,000 beads in a 3-millimeter diameter. And so because the beads are deposited randomly, what we then do is we take a microscope and sequence the clonal barcodes for each bead in the array. I won't go into the details of exactly how we do this. But effectively, for each color on this microscope image, green, blue, red, and yellow that correspond to a different base, A, C, G, and T identity on one position in this barcode. And we go through and interrogate the different positions. And this is quite easy because the beads are quite large and quite fluorescent. And we can build up the sequence of colors and then the xy locations at each. Sequence location. We can do this up front. And, and so that one of the key strengths of actually SlideSeq is that once the arrays are made, it's very easy to then actually perform the experiment. We can take these arrays and take fresh frozen tissue and section a 10-micron section directly onto the array. Then you can take in a simple protocol, you can generate sequencing libraries, very similar to single-cell sequencing, where you perform reverse transcription and then some amplification of the cDNA library. We then take that cDNA library and we sequence it on an Illumina sequencer. And so you get— well, you get a result very similar to single-cell sequencing where you get for each bead the unique barcode that it's associated with. And you can imagine that that is in a matrix on the rows. Where there's a unique barcode— unique bead on each element of the row. And then across the columns are the digital counts across each gene in the transcriptome. And now if you remember, we had kind of these colors which represented the different bases in the bead barcode as well as their xy locations. We can match them up. And now we have an xy location for each of these beads. And at each xy location, then that means you have the counts of the genes across the transcriptome. Okay, so what does this data look like then? Actually, SiteSeq has kind of a similar data structure as single-cell data. And so naively, you can apply the same computational tools such as dimensionality reduction and clustering. Here, this is actually one of our first SiteSeq experiments where we ran it on the mouse. Hippocampus. And on the very left, what we've done is we've clustered the SlideSeq beads and annotated the clusters with the cell type identities. These are the principal cell types of the hippocampus. In the middle, what we've done is actually we've just taken the beads in the array and colored them by which cluster they belong to. And on the very right, you have what kind of this tissue looks like. For those of you who aren't familiar with the hippocampus, this is kind of the arc— this the cytoarchitecture of the hippocampus. You can see that we reconstruct basically the tissue architecture, but beyond that, we can now place kind of like the principal cell types like the granule cells, the hippocampus and the CA1, then this oligolayer separating the hippocampus from the cortex. And so the takeaway is that there's no microscope. We've actually turned the sequencer into a microscope. In fact, there's these 10-micron pixels, and they take images of genes across the transcriptome in a single experiment. And so I'll just quickly summarize what collecting this high-resolution transcriptome data allows for. One, it allows us to reconstruct the organization of tissues and connect it to molecular cell types. And because each bead is about the— the size of a cell and we get untargeted genome-wide expression, we can kind of reconstruct architecture of different tissues. For example, here you can see the reconstruction of the cerebellum with the principal cell types labeled. And you can see actually in orange this— the Purkinje single-cell layer Purkinje neurons corresponding to this histology here. But beyond that, we can use it to discover new patterns of spatial variation within cell types. For example, in our manuscript, we used it to discover spatial gene expression domains in these Purkinje neurons in the cerebellum. And so— and we've actually just discovered— like, in this— just, we developed kind of an unsupervised algorithm that will identify within cell types what are spatially nonrandom gene expression patterns. And using that, we've we define some subdomains of the hippocampus— sorry, of the cerebellar Purkinje neurons. And it also allows us to discover kind of and visualize spatial gene expression changes in pathology. For example, we previously demonstrated profiling the spatial temporal dynamics of a model of traumatic brain injury. And so this is a kind of a quick summary of some of the highlights from the first manuscript. For today, actually, I want to focus on some of the progress we've made since we originally published SlideSeq. And in particular, we focused on a couple of things. One is we focused on the ability of us to apply SlideSeq to a variety of different tissue types and organisms. Both in human early mouse development as well as organoids and in particular pathological specimens, mainly human oncology specimens. And second is that we focused on the capture, molecular capture of SlideSeq, or how many unique molecules each bead captures on the array. And just through kind of a series of molecular biology bead synthesis improvements, we've actually improved the yield of SlideSeq about an order of magnitude and to about 1,000 unique molecules per bead. And this yield is actually pretty similar to kind of first-generation droplet technologies like DropSeq. And we actually anticipate that just if you compare DropSeq to like 10xB3, there's another factor of 3 in there that we could potentially get out. And there's no reason practically that an array-based capture on the surface should be any less efficient than in droplets. So we're anticipating some more production there. And lastly, we've actually open sourced the array production and computational pipelines. And this has all been deposited on GitHub and protocols.io. And so these improvements to the SlideSeq technology have enabled us to develop many more analyses of the SlideSeq data. Here's just as a vignette, what we actually did was we profiled the developing mouse cortex at embryonic This is actually embryonic day 15.5. And here you can actually see that here's a hemisphere, and in this box region is the developing cortex where the cells develop from the left to the right, from the ventricular zone to kind of a structured cortex. And we've actually applied some single-cell algorithms for pseudotime, or actually RNA velocity, where we can look at the ratio of intronic to exonic transcripts to look at the rate of gene expression change. But we've actually projected now this RNA velocity onto the spatial coordinates. And what you can do is then you can construct kind of a manifold, a spatial manifold on both the rate and the directionality of gene expression change in this developmental process. And then what we can do is we can identify genes which vary along this manifold. We can then use that to learn kind of the relative enrichment of gene sets from human genetic studies of intellectual disability and autism and the relative ordering those genes lay on to basically this spatial developmental manifold. So we think that these kind of datasets and analyses will be very useful in understanding kind of in the future the spatial-temporal localization of genetic risk variants in CNS tissues. So, so far, I've shown you kind of a lot of data on unsupervised clustering of data from SlideSeq. That's what I mean by unsupervised is that we just use standard single-cell approaches where we act as if each bead is capturing RNA from a single cell. And that's mostly the case. We actually find that each bead captures RNA from a median of like 1.5 cells. And so there are definitely beads which are captured RNA from kind of a mixture of cells. And I can illustrate this kind of like pretty clearly with these two types of neurons— sorry, these two types of cells, the Bergmann glia and the Purkinje neuron, which occur kind of like interlaced in a single layer in the cerebellum. And if you look at kind of the unsupervised clustering assignments for beads which are assigned as Purkinje's or Bergmann, you have some singlets on the— on the diagonal— on the axes. But then in the middle, you have beads which are called one or the other but have some mixture of marker genes from both cell types. And really, if we want to learn about spatial gene expression changes within cell types or maybe cell-cell interactions, It's really important not to assign these mixtures as one cell type or the other, but to know what combination of the cell types they are and deconvolve them. And so that's exactly kind of an algorithm that we recently developed for spatial transcriptomic data in general, including SlideSeq. We've shown also that it works on Visium or MERFISH or other sorts of spatial transcriptomic technologies. And this is a supervised learning approach to decompose mixtures of cell types into single cells called robust cell typing decomposition. And the basic idea is that we might collect some spatial transcriptomic data with different pixels on it, and each pixel or point in the array have captured contributions from different cells. And you might also have a paired— and as I said, it's actually very easy to generate paired single-cell references. So you might have a paired single-cell reference where you've already annotated the different cell types. And so then the idea is that you want to estimate in each one of these pixels in what proportion and what combination of cell types best fit the observed gene expression profile. Using kind of just like a reference-based probabilistic model. And at the end of the day, then you can take each position in the array and assign that to a mixture of cell types. And so, so what we've actually done— we've done quite a lot of work to validate this computational approach with simulations with ground truth datasets. And what I mean by that is that we took actually single nucleus RNA-seq data and single cell RNA-seq data from the same tissue. And first we showed that we can use this algorithm to classify singlets from one data type to another. So this is on the bottom. This is the cell type as assigned in single cell RNA-seq data, and this is cell types as assigned in single nucleus RNA-seq data. And we show that we can get basically very accurate recon— classification of cell types from one data type to another, with the exception of maybe some similar cell types. There's some bleed-through. And then what we did is that we took cells from here and artificially constructed doublet pairs where we, by mixing, you know, one, for example, Bergmann cell and one Purkinje cell computationally, And then first we try to identify when cells are mixed into a doublet, and we find that we do decently well at that task. And then once we've identified the doublets, we can actually accurately decompose those mixtures into the fractional content of the constituent cell types. And so this is the ground truth, and this is the reconstruction in the decomposition. So what this allows us to do, if we go back to our example with these two intermixed cell types, is that we can use single-cell RNA-seq reference to both assign the cell type and decompose it. So now we have kind of the predicted singlets on the axes. And then for the mixtures, we have actually the proportion. And then we can decompose those proportions. Into the fractional content. And so what this allows us to do is it allows us to perform— not only does it allow us to, like, deconvolve the doublets, it also allows us to perform really high-resolution cell type mapping, where in your reference, in your single-cell reference, you often have the ability to finally subdivide, say, inhibitory neurons here as an example, This is 27 different subtypes of inhibitory neurons that live in the hippocampus. Here is the spatial data. The gray cells are actually like the excitatory neurons, and then these are the assignments for these 27 different cell types. And we have independent validation that these assignments are quite accurate. And here you can see that, for example, these are excitatory neuron clusters that form the different layers of the cortex. And so you can see this reconstruction of a cortical column. So in this sense, you're kind of leveraging the higher— both the higher sensitivity and also your ability to more finely grain cell types in the single-cell reference. But beyond that, I think it's important to also recognize that philosophically there's been a lot of work in trying to detect spatially varying genes from spatial transcriptomic data. And in general, these approaches have not been thinking about it from the abstraction layer of cell types. And if you think about it, the principal component of gene expression variation in the transcriptome is cell types, and cell types are not spatially randomly distributed in the transcriptome. And so if you want to look for genes which have like interesting patterns in your tissue, uh, the first things that you're going to find are cell type markers. Um, but what we actually are interested in is not like cell— like discovering cell type markers through spatial methods. We're interested in discovering kind of cell type specific spatial variation within tissues. For example, here are several genes which are smoothly varying in CA3 cell type. This is an excitatory neuron cell type within the spatial context. And beyond that, we'd actually want to find out— so that's how gene expression is changing within a cell type as a function of tissue organization. But I think, you know, just as interesting of a question is gene expression change as a function of who your neighbors are. And, you know, so here we did a simple test where we looked at astrocytic gene expression, so like gene expression within astrocytes as a function of their neighbors. And so these are actually genes which, which have differential expression in astrocytes when their neighbors are, for example, excitatory neurons. And like, for example, like SLC6A11 is a GABA transporter, and this is actually known to be a gene associated with parasympathetic astrocytes. But in general, I mean, this is just one example, but in general, we want to condition on cell types to actually like deconvolve the contributions of cellular environments on cell states. And that's like the thing that's like the most exciting thing to me about spatial data. OK, so kind of a quick summary of this first section on SlideSeq is that I've shown you how we've developed some experimental and computational tools for high-resolution spatial transcriptomics. One thing is that we've really increased the yield in SlideSeq v2. We've open-sourced the array production. And also, one thing is that for these arrays, We've really scaled our ability to construct them. And so we're starting to offer arrays to folks through our core facility at the Broad. And that's internal right now, but it should be opening up quite soon to external collaborators. And I've also shown you how we can leverage single-cell atlases to assign cell types to spatial data. And I think this is where really next level of analyses really begin to take off, where we're going to be able to get spatial differential expression within cell types and cell-cell interactions. And these are kind of really the tools that— computational side tools that we're really excited to develop. And at the same time, we're collecting some tissue atlases One, just as a platform for developing these computational algorithms and two, as resources for the community. And so we've been working on some atlas for the mouse brain and some developing mouse atlases. And so that's kind of exciting. So hopefully we can share some of that data going forward. So for now, for the last— maybe, let's see, 15 minutes of the talk, I want to transition and discuss some other work at a different spatial resolution, but, you know, and a different modality of biomolecule, but still the same fundamental problem, which is like, how do we address recovery of both the context of sequence and the context of space. And this is work that we developed on in situ genome sequencing. It's actually the first approach to directly sequence the genome within cells and tissues. And this is really work led by 3 graduate students over the last couple of years between my lab, Ed Boyden's lab, and Jason Buenrostro's lab. And so again, you know, like, as I briefly touched on in the beginning, but imaging approaches are actually really powerful because they give you this really high spatial context. You know, for example, like, these are like dendrites, and you can look at like RNA within that single dendrite, right? And But sequencing approaches in general allow you to look at sequence context, like for example, where on the genome a transcription factor binds. Right? But you lose the spatial context that you can get from imaging. And actually, this has been well recognized in the field. There's been a couple of approaches that have tried to bridge that. We've had Hi-C, which is an approach to kind of reconstruct the organization of the genome through statistical interactions between neighboring pieces of DNA. Now, this is challenging to do in single cells, but also it is a statistical reconstruction, so it's not an actual direct visualization of the architecture. And at the other side, we have— there's been really quite a lot of brilliant advances in DNA FISH-based imaging where you can kind of start tracing out the chromosome. And these sorts of approaches in general, they've been focused on really high-resolution tracing of single chromosomes, but also you lose like the sequence context, like what is like the novel mutations or SNPs or translocations. So you lose the sequence context. This is because it's a hybridization-based approach. And so we're in general interested in, when we set out, how can we sequence and image DNA in the same single cell, kind of like somewhere at the intersection of Hi-C and imaging. And so how does this approach work? Well, it's kind of— we can take cells and Normally, in Illumina-based sequencing, you might extract the DNA. You might amplify it and put it on that sequencer. And then you might align it to the genome. Here in in situ DNA sequencing, we fix the DNA. So we fix the cells with just standard paraformaldehyde. We construct sequencing libraries directly in situ. By that, I mean we amplify— genomic loci, and we use fluorescence-based sequencing similar to Illumina directly within the cells. And then we align it to the genome, but now we also know where those reads are in x and y and z. And so I'll just quickly go over how we do this. And this has really been quite a lot of amazing tech dev on the— front of the graduate students solve a couple of key technological questions. So in C2 sequencing, genome sequencing, we take fixed cells. And we actually use Tn5 transposase adapters to paste in sequence directly into the fixed genome. And this can either be accessibility sensitive or not. And to make it not accessibility sensitive, you actually just have to denature the histones a little bit, maybe with SDS or HCl. And this is something that people do quite often for DNA FISH anyway. So we took a standard DNA FISH protocol, and we found that you lose the accessibility sensitivity after you apply that and you put in TnF5 transposase. So now you paste in some sequencing adapters into your genome. Then we ligate these hairpins onto the ends of the sequencing adapters. That circularizes them, allowing us to do rolling circle amplification. And what that does is it generates clonal amplicons, where each amplicon is some amplified piece of your genome. And inside of each of these amplicons are these repeats, where it's the insert of your genomic DNA. And then some priming sites. And kind of importantly, we included in each one of these hairpins a unique molecular identifier. And that's actually important because it's rather difficult to do long— and by long, I mean greater than like 15 to 20 bases in C2. And I'm sure all of us who have done sequencing a lot are used to even in short-read sequencing on the Illumina sequencer or something like paired-end 150 bases, right? And but we actually kind of use a similar trick as what we're doing inside SEEP. We can sequence this UMI. Actually, we sequence it 18 bases. And so here, you know, it's one amplicon. You can actually see— you can actually call this basically by your eye. This is like, you know, we did a lot of optimization on actually the sequencing chemistry and microscopy to to generate such high-quality in situ sequencing reads. But you can take this UMI. You sequence it in situ. You can recover the amplicons and just PCR them back out, put them on an Illumina sequencer, and do paired-end short-read sequencing. By short-read, I mean paired-end 200 bases. Now you can sequence for SNPs or mutations or other aspects of sequence context. And now you match— at the same time, you sequence that unique molecular identifier on the Illumina sequencer as well. And now you match it up. This UMI on the Illumina and in situ. And now you have the XYZ position of your reads. And for example, in this single cell here, you can see the reads colored by chromosome. And you can immediately see that you can resolve chromosome territories, kind of like the zeroth-order organization of the genome. And so in our paper, we performed some studies in PGP1 fibroblast cell culture and mouse early embryos. For the interest of time, let me just talk a little bit about what we did in the embryos, which is that You know, we were pretty interested in profiling the genome organization of these early embryos because there's a growing body of work, obviously, which suggests that the chromatin architecture is both dynamic and changing rapidly in these early stages of development, and especially during, like, zygotic genome activation. And— but at the end of the day, you know, these embryos are really not accessible to existing techniques. Single-cell hi-C really can really only be performed on zygotes by microdissection of the pronuclei, and it doesn't really scale to 2-cell or 4-cell stages. And also, these embryos are quite large, actually, when you get to the 4-cell stage for DNA FISH. And so we thought this was kind of like a pretty unique an interesting opportunity for in situ genome sequencing. And the other thing that's really cool is that within a single 4-cell embryo, we can actually look at how the nuclear architecture vary within daughter or sister cells or cousin cells of the same embryo, which is particularly cool. And so we actually then perform from the study measuring the genome structure at 3 distinct stages of development at the zygote shown here, and then 2-cell and early 4-cell embryos. And here is the median reads per cell. And you could actually see that it goes down as you go up through the stages. And that's actually because the number of reads that we get per genome is entirely dependent on the genome volume. And it turns out in that— in these early embryos, there's a little bit of conservation of volume. And so we actually get more data per nucleoid in the early zygotes. We also collected at the same time as the genome sequencing data, the in situ sequencing data, we collected antibody staining, which is one of the beauties of actually doing kind of this in situ work is that you can layer on kind of multi-omics. Here we labeled kind of lamin, DAPI, which allowed us to see like the pronuclei as well as— excuse me— nuclear density, and then SEMPA staining. And so for example, here's kind of one— when you put that all together, here's kind of like data from one pronuclei. This is a zygote and— sorry, one zygote, and the maternal and paternal pronuclei are actually separated at this stage of the zygote. And you can see, basically, you could assign genomic reads, and at the same time, you can look at their distance to protein landmarks. And because we also get the sequence information, we can actually assign— this is— actually kind of an inbred cross, so the SNP density wasn't that high, but we can still see, just based on the SNPs, the assignment of chromosomes to maternal versus paternal lineages. So here you see that the pronuclei are separated, so it's It's clear that one pronuclei has mostly maternal and one pronuclei has mostly paternal. Here you can actually see that the 3D architecture of these territories is along with the assignments. This is kind of a simple connect-the-dot model where we've just traced out the fold of the chromosome by connecting nearby loci. It turns out that you can use this— this is like the simplest thing. This is just visualization, but you can actually layer Hi-C data on top of this to construct kind of like a constrained nuclear architecture model that has the folds. So what we did was we actually can use these stages to characterize kind of like the organization and the changes in the genome across these couple early stages. For example, in— this is one cell of a two-cell embryo. You can actually see that the haplotypes of the different chromosomes after that first cell division, they're in the same nuclei now, but they're actually completely polarized. There's no mixing, very little mixing between the paternal and maternal homologues. And actually all the chromosomes are aligned on one axis, the centromeres over here and the telomeres on that side. These are mouse, The mouse chromosomes actually only have really one arm. The centromere is really on one end and the telomere is on the other. So that's why we're using the centromere to telomere as a measure of one end of the chromosome to the other. And so what we can do is then over developmental stages, we can, for example, measure the mixing between different SNPs and calculate the intermixing between the maternal and paternal haplotypes, as well as the distribution of mixing over time to look at different models of this genome mixing. And we can also look at the polarization of basically of these chromosomes along this dimension. This is actually a well-known thing called the RABO. Well, I don't know how well known it is, but in cell culture, it has been observed. And this is called the Rabble configuration. And we think that this is because the cell division times in these early embryos are really, really short. And this polarization of the centromere to telomere is actually a function of that. But if you basically look at kind of the alignment of centromeres and telomeres between neighboring chromosomes. They become very synchronized, actually, as soon as you go from the zygote to the 2N, 2N4 cell. So for the interest of time, maybe I'll just highlight just one really cool aspect of this data is that we can use the fact that we are sequencing the genome structure of cells within the same embryo to look at how genome structure changes or doesn't change as a function of inheritance. So here, if you look at basically two embryos— this is one embryo up here and one two-cell embryo down here, and this is just two sets of chromosomes highlighted— you can actually see that these two sets of chromosomes are actually just by eye, have similar positioning within the nucleus relative to each other and relative to their homologs. And since these are 2-cell embryos, these are by definition sisters. We can look at the 4-cell embryos actually, and we can look at now across all of the chromosomes inside the nucleus and look at the correlation of their positioning. So like how similar is one cell, the organization of one cell's genome to the organization of another cell's genome. And in a 4-cell embryo, there's a pair of sister cells and a pair of cousin cells. And so if we assign, if we look at the distribution of the 2 highest most correlated guys, they're very correlated. And they're similar to what you would see from the two-cell embryos, as in like, basically like sister cells have very well, very correlated genome structures in chromosome positioning. And then if you look at the cousin cells, which is like the least two correlated pairs, they actually are more correlated than just random pairs of cells between embryos, suggesting there's still inheritance of of this genome structure across two cellular divisions. And this is, I think, particularly maybe interesting as we think about early genome activation and the really early choice of cell fate as it goes from 4-cell to 8-cell. That's when some of these things happen. And we might want to consider basically how— we looked at the positioning of the chromosome. But if you look at the folds of the chromosome, there's also some inheritance of genome structure at the folding scale. And so we're pretty interested in how basically this inheritance of genome structure leads to changes maybe in inheritance of fate in these early embryos. And so maybe I just want to summarize on the in situ genome sequencing side. We developed a technology that allows us, for the first time, to sequence DNA sequence directly within cells and tissues. We think that this is really powerful for looking at the organization of genomes. And we did it in early mouse embryos. And we get about 1 megabase resolution on the chromosome and genomic structure. And that allows us to kind of make comparisons of genome structure within intact embryos, but also look at kind of like the changes in organization of genome structure as a function of these early developmental stages. Now there's a lot of other, you know, uses potentially for the technology. Maybe, for example, like in tissues and such as tumors, you might be able to discover de novo changes in ploidy or translocations of chromosomes, we actually demonstrate that we can see that in the embryos. And maybe you can also imagine using this to do some long-range haplotyping because of the ability of you to spatially localize nearby reads and perhaps even assembly of de novo genomes with this approach. And so I want to say that in summary, I told you a couple of different strategies we're using to measure different types of genomic molecules in the context of tissues at high resolution and high throughput slide-seq and at very high— sorry, at cellular resolution and high throughput slide-seq, and then at very high subcellular resolutions in C2 sequencing. And so as I said, my lab is also pretty interested in developing synthetic biology recording tools, which I don't really have the time to go into today, but I'd be happy to talk about that offline. And so thank you very much, and I'm happy to take questions. But before I do, I just want to, you know, thank all the folks who did this work with me. In particular, Evan and I have worked very closely on SlideSeq with the team of Sam, Bob, who led the original work, and now Dylan and Evan Murray have joined in. And then as I said, Andrew, Paul, and Zach have really led the genome sequencing work. Thanks.

Great.

Thanks very much, Fei. That's really, really fantastic. Really enjoyed it.

Thanks.

We have a few questions in the chat. I'll start out by reading off some of them, and then maybe we can just open it up and people can ask direct. So first question is, so how does SlideSeq perform in— so this is from Danielle Engel. How does SlideSeq perform in tissues with high RNAs levels, such as normal pancreas?

Yeah, that's a great, great question. I have to say that we've never tried pancreas. We've tried some other tissues where, you know, if you look at this, there's this chart with like RNA level as a function of tissue type. And the main takeaway of that chart is that the pancreas is like way off the chart and there's other tissues that are, you know, high, much higher than brain, but not nearly at the level of pancreas. And those tissues like liver, kidney, like work fine. I'm not sure. I would be worried though about the extremely high RNase contents of the pancreas.

Right.

And Jeff has a question. Jeff, do you want to ask direct?

Sure, thanks, Faye. So first, I just wanted to say that was really a beautiful talk, literally and figuratively. But, you know, I had a conversation with Evan McCosko, and I would really like to hear from you how SlideSeq v2 compares to other methods such as SeqFish or MERFISH. In terms of cost, read depth, and precision of localization. And the other thing is, it seems to me that what's really limiting for implementation broadly are the availability of the pucks that you have with the beads on them. And you said that these are being made in the core at Broad, But first, when would they be— when would others be able to access them? Or can other people make them simply using the protocols that you've made available in GitHub?

Right. Those are great questions. I guess that's 3 questions. Maybe I'll start backwards from the last one, which is that some groups are trying to make them with the protocol from GitHub. And I think one— We've helped a couple of them, and usually they tend to be microscopy cores, and one group has succeeded. I would say that, you know, the ability— the core at the Broad that's making them, they're already just like basically making them for my lab and like basically everybody at the Broad, and so they should be ready to, you try to transition to outward collaborations very soon. And the question about the trade-off between MERFISH— so obviously MERFISH and seqFISH and these sorts of approaches are, I mean, microscopy-based. And they are very good in the sense that they allow you for a targeted set of genes, hundreds maybe, allow you to localize and count the mRNAs for those genes. Now, they do require a decent amount of microscopy expertise to set up in your own group. And maybe those sorts of things will be commercialized. But that's one of the main things is there's some upfront cost. And the second thing would be that you have to choose your targets versus an unbiased, untargeted genomics-based approach.

Well, SeqFish+ according to Long Cai goes full genome. He can do the full genome resolution.

Right. Yes. Yeah, I understand. But I think that there's an inherent trade-off if you want in the amount of time it takes to do that experiment.

Yes.

Both the time and capital cost to do that experiment over 100 genes. I think 100 or a couple hundred targets is a very feasible experiment with MERFISH. And then very few labs are doing a whole genome-wide experiment. And so that's, I think, the main trade-off. And I do think that to answer your question of availability, there's going to be quite a lot of companies commercializing array-based capture. People are looking into licensing Sightseek, but I think 10X is going to come out with— has come out with Visium, and they're going to come out with some higher resolution versions of Visium as well. And I think that the commercialization is going to play play a big role in people's ability to use the technologies.

Can I ask just one other question that had to do with your deconvolution method? Yeah. Where you're trying to separate potential doublets from singlets of different states. In cancer biology, cancer cells often transition or reprogram into different states and they acquire hybrid states. So how do you distinguish something that you might call doublets from something that's a true hybrid state cell?

Right. I guess the question is whether you see those hybrid states in like a single cell reference, if you see those things in a sampling of single cells. The other thing I would say is that you don't have to do that sort of deconvolution with cell types. You can also do it with like factors like NMF or like cellular programs. And that may help with that particular type of definition. Right, because you can say that this cell is a mixture of these programs. Now whether you think that that is a doublet or you think that that is a real cell that has the mixture of those programs, that's kind of— that probably has to be determined independently.

Okay, thanks, Faye. And a question from Ulysse Uzcarelli.

Are you there, Ulysse?

Can you hear me?

Sorry.

Go ahead.

Okay, very nice talk. I'd like to ask, with this genome sequencing, what percentage of the total, let's say, mouse genome can you sequence? Or when you map, what percentage of the— Right, that's a great question. Now per cell, I think—

Okay, I think we've lost Faye temporarily.

Percentage of the mouse genome, but over all the cells, then you— Hello? Oh.

Go ahead. Maybe you want to— at least at my end, you cut out for a bit there.

Yeah, I think my internet— connection is unstable. Can you hear me?

Yeah, we can now.

Yeah. Apologies. Oh, I just said that in each single cell, the coverage is rather low, maybe like 0.01% or 0.1% or something at best. But across many cells, then obviously your genome— your coverage is— then becomes like you can cover the whole genome. Maybe not even that many cells. And thank you very much. And how about the SlideSeq? What percentage of the genome can you cover with mRNA transcripts from a single cell? Similar to like DropSeq, it's like 1,000, a couple thousand, like 1,000 UMiS or so. So it's like similar to like a DropSeq experiment. Okay, thank you. That provides the right context.

And a question from Palav Kusuri.

Yeah, hi, wonderful talk, Faye, and beautiful technology. I really love the creativity. I have two questions, one about SlideSeq, some of the assumptions that go into the way you do this. So you identify where the beads are relative to each other, construct this Voronoi diagram that basically segments the area into the different regions that correspond to the different beads. And then when you assign these transcripts, you base that on the assumption that each transcript goes to the closest bead. Is that correct?

Well, we actually don't do this Voronoi thing. We— yes, we do make this assumption that each transcript goes to the— We make— it's the opposite way around. We make the assumption that each bead, each sequenced transcript from the bead came from something nearby. Because like when we sequence it, we know which beads go with which transcripts, just like when you sequence like, you know, because like you sequence the bead barcode and know what the transcript identity is.

Yeah. And so do you— I mean, okay, so to put it another way, For each bead, do you assume that the transcripts all came from a cell that is overlapping in space with that bead?

Yes, we do make that assumption.

Is that— how sure are you that that's true, that it's not possible?

We made some pretty careful measurements of diffusion, which I didn't show in this talk, that says that the diffusion length scales are pretty low, like less than, less than the size of the bead, and so that's why we're basing that.

Okay, okay, yeah, I'm just wondering where that came from. That should be easy to test. Yeah, okay, and the second question, uh, second question is more sort of a, a curious, uh, exploration of possibilities. I'm wondering if, would it be possible if you did immunohisto staining, like antibody staining, Nowadays you have all these antibodies with oligos on them.

Yes.

Could you use this approach and then release the oligo and use that as a combined proteomics/transcriptomics approach?

Yeah, I think so. I think, you know, there are some folks who are in other types of technologies who are doing that already, and it just requires the molecular biology being worked out. But yeah, that's a good idea.

Yeah, because you can imagine like this You have these antibodies with oligos, you could just make a duplex and you just need to release it somehow.

Yeah, I think that 10x, there's like some work where some other folks have published a little bit on that.

Okay, sounds like it's in the works.

Yeah.

Cool. Okay, great. Well, thank you.

Thank you.

Any more questions for Faye?

Actually, I would just like to follow up on that.

Go ahead.

What's your association with 10x and how involved are you with their product development?

Oh, I am not— I'm not developing their products. I don't have any disclosures with my affiliation with 10x. I mean, I think that in some ways we're like developing some— we're like kind of pioneering in like what sorts of tools can be done and what sorts of analyses can be done. And, you know, it would be great if they could commercialize some, some of the things so that more people have access to it.

Yeah, yeah. Okay.

But I have no, you know, financial or other conflict disclosure.

I guess for, for us as sort of, uh, you know, potential consumers or, uh, users of your product of of the 10x products, it would be good to know if they have sort of the Chen stamp of approval.

Oh yeah, for sure. I mean, I think 10x does a great job. I think the main thing is, is that their arrays are a little bit low resolution right now. They're at like 50 microns, or I don't know how you want to count it. It's 50 micron, 50 micron spacing. So, but they are definitely working on a high resolution, like something similar to SlideSeq. That is slated to come out next year.

And where does the NanoString GeoMx fit in with this? That's kind of substantially lower resolution, isn't it?

Yeah, substantially lower.

Does that have any advantages?

I'm not— I think the main advantage is that they've highly optimized those panels to work with FFPE, and there's a lot of clinicians who've really bought in and we collect a lot of data and the data looks great. So I personally don't have a lot of experience, but I think it's really valuable to get that clinical perspective on the technology. So I've definitely collected a lot of clinical data.

So if you've got archived FFPE sections, then the nanoString would be the way to go.

Yeah, I think so. Okay. At least currently. But yeah. Thank you.

Yeah. Any other questions for Faye? If not, then thanks, thanks very much, Faye. It's a really wonderful talk and Great turnout. We were up to over 150 attendees, I think. So great job.

Thank you.

Thanks, Faye.

I had fun. Yeah, great day.

Yeah, thanks. And next meeting is March 5th, everybody.

April 5th.

Oh, sorry, April 5th. Yeah, thanks, Jeff. We'll let you know by email.

Bye.

Thanks, y'all. Bye.