Atlasing Organs, Aging, and Development with Genomics

All right, welcome to the atlasing session. Very thrilled to have as our speaker here Professor Goji Goh, a former colleague of mine actually from GIS, and he's a deputy director of the Stem Cell Center for Stem Cell and Regenerative Medicine at Zhejiang University School of Medicine, Deputy Chair of the Stem Cell Society at Zhejiang University, got his PhD in Singapore at NUS in 2010, moved to Harvard for his postdoc, and then 2014 he went to start as a PI in Zhejiang. Uh, got the NSF, uh, National Science Fund for Excellent Young Scholars Award in 2017. And I think many of you know him, or most of you know him already, really a pioneer of large-scale single-cell analysis and atlasing. So who better to start this session than Goji? Welcome, Goji, over to you.

Thanks. Thanks for Sriyank's introduction. And thanks for HCA as well as Australian colleagues for organizing this wonderful event. And thanks for the invitation. And I would like to share some of our recent work on inferring genetic models from cell landscapes. There is one thing that is always puzzling me since high school. As we know that in biology, we're still quite descriptive and we do not have any predictive model or equations like what we have in physics and chemistry, like the periodic table. You can easily predict unknown chemicals using the periodic law. Indeed, we know that there are connections among DNA, RNA, and protein, but we do not know how quantitatively they're connected and how a single genome ended up coding different types of transcriptome to make cell types. Apparently, there are many variables in biology. You need lots, lots of measurements until we have enough data for modeling. And this is the challenge. And so how does a genome encode the cell landscape? Back in 1950, Waddington proposed this epigenetic landscape, and although it's a concept, but it beautifully describes that every cell type corresponds to an attractor or an equilibrium state that is self-reinforced and that really attracts unstable stem cells into these valleys to make different cell types. And why not? Can we simply measure the whole landscape at single-cell level and then try to see if there is a model? And this is what HCA has been dreaming of. And indeed, because of lots of advances in single-cell genomics, we can now measure transcriptome, genome, proteome, and epigenome at very high throughput. And this not only solve the problem of heterogeneity, it also solves the problem of consistency in biological measurements. And because individual cell is the fundamental unit of genetic regulation, if you want to model biology, you have to use data that is from this individual units. And here we go, the single-cell genomic data really solve this problem, and it doesn't matter where you're from. And what tissue you're looking at. You don't have to cherry-pick marker panels. You just do a single-cell genome mode transcriptome analysis, and the data is considered to be comparable across different labs. And this is the whole concept. In Zhejiang University, we established this microwell-sea homemade methodology. And different from the droplet-based method, we use agarose microwell plate to capture individual cells.

Ourselves.

We synthesize our own barcoding beads, and the beads will block the cells below. And then after lysis, every cell gets barcoded with cell barcode and molecular barcodes. Then we pull them for sequencing. The advantage of micro-well seq is that there is a quality checking step, and if you do see very high cell concentration, sometimes in some region you have that, you can just pipette it to reduce the local concentration, to reduce downplay rate, we get a very nice, uh, species mixing experiment. And then since we synthesize the microwell array using only agarose, it is very cheap and you can easily scale up the experiments with minimal cost. And we have performed some optimization. First of all, by using random primers rather than template suite, we can further prove the sensitivity of the assay. And this is quite similar to the SeqWell S3 from Alex Shelek lab. And then by coupling the pre-index technology with the microwell method, and basically we fix the cell, we do pre-index using the RT. And previously, you know that from every 10 well, you can capture only one cell for Poisson distribution. This really helps to capture single cell. However, because now we have a barcode for each cell, you can load up to, you know, 45 individual cells in each well. And then the throughput of the analysis increased by, as you can see, 40-fold. So you can analyze, you know, up to 1 million single cells in one experiment run. And then we actually use the BGI T7 sequencing machine, and this is a Chinese homemade machine, and so the cost is cheaper when compared to Illumina sequencing. As a proof of principle, we use this method to run the drug screening experiments, and it really helps since you can provide the pre-index for each of these chemical combinations. After the reverse transcription, you pull them and then you run the microwell. In one experiment, you can test lots, lots of combinations or parameters. I think a similar type of experiment can be used to run the different donors, and using one experiment without batch effect, you can do population genetics or population atlas genetics. I think the throughput is quite important for atlas experiments. Using the microwell-seq platform, we have actually measured any cell types that we can get. We, in principle, we build this cell landscape for human, mouse, zebrafish, Drosophila, Xenopus, and earthworm. So far, we have profiled more than 6 billion cells, and some of them are published, some of them are not published. And the good thing about this dataset is that it's all microwell-seq. It's highly comparable. And then the aim of this project is that we can generate enough biological data for the AI to interpret genetic models in the cell landscape in nature. In principle, we want to ask two questions: how many cell types are there, and then how are cell fates regulated? As an example, in our human cell landscape data, we collected 60 human organs and tissues, half of them from half of them from adult state. In total, we have 600,000 single cells. And to answer the question, what is the cell type? We think it's a complex hierarchy. So we do the Seurat clustering to generate the major clusters. Here we have 102 major clusters. And then for each of these clusters, we extract the data and do the Seurat clustering again. This helps to predict a total of 843 subclusters in this human cell landscape hierarchy. Then for the mouse cell landscape, we supplemented more data from our previous NCA work. We collected the neonatal, 10 days, 3 weeks, 1 year, 18 months, and 2-year data sets. We found that at neonatal and 10-day stage, And there are lots of novel cell clusters or cellular states that is different from either fetal or adult state. We think that the tissue or cellular differentiation persists after birth, and particularly for neonatal and 10 days, we have many interesting cell populations. In total, we have more than 1 million single-cell data from the mouse at different stages, and we predict a total of 2,000 5,500 cellular states. And is there any novel cell types? This is a question that I've been asked for most of the times. I think the answer is yes. We extracted stromal cells from different tissues and to combine them and make this stromal cell landscape as well as endothelial cell landscapes. And since the stromal cells are really poorly studied, and before single-cell assay and the FACS sorting for different tissues, and people follow different protocols with different marker panels, therefore you can never combine different stromal cells to look at the whole picture. But for single-cell transcriptome data, you can do that. And after the clustering, we consistently see several clusters expressing immune-related markers. Some of them express IL family proteins. Some of them express CXCL family proteins. Some of them express antigen presentation related. And in endothelial cells, we discovered two clusters expressing immune-related markers as well, particularly the antigen presentation related genes. And using the immunostaining, we can see the co-staining of endothelial cell marker and the HLA-DR, which is MHC II. Family genes in the human bladder sample. And they're co-stained here. And this immune-related structural cells not only seen for human, we have seen them in mouse, in zebrafish, and in Drosophila. Particularly in Drosophila, the Malheppen tube, they do have a cluster of cell types, epithelial cell types, express epithelial markers as well as immune-related markers. We think that during evolution, structural immunity is the key mechanism for immunity before the evolution of actually immune cells. The funny thing is that during aging, many of these immune markers get alleviated. And this is strongly correlated with lifespan time course. And for endothelial cells, epithelial cells, and stromal cells, you can see immunoglobulin genes activated in lots of tissues. And also, in endothelial, epithelial, and stromal cells, you can see antigen presentation-related markers get gradually increased through age. Um, therefore, to answer the question how many cell types are there, I don't think our answer is perfect, but at the current stage, we have at least detected 2,500 cellular states for mammals through the lifespan. And the cellular hierarchy is more complex than we think. There are transitional cell types that is half epithelial, half immune, and there are inflamed mesenchymal cells and endothelial cells. During development, there are lots of cells that is under epithelial-mesenchymal transition, and there are progenitor cells that co-express markers of myocyte, endothelial, or myocyte-neuronal markers. We believe that a cellular hierarchy is better to present the actual cell types in mammals. To answer the second question, how are cell types regulated? And my lab are divided into two small groups, two small groups, and we try to answer the question in two formats. One of the groups tries to find whether or not there is any common regulator that controls sulfate distribution in multiple cell lineages. And the other group takes the approach of a holistic approach. And they hope to generate a common model to interpret the trajectory from the genome to the cell landscape. And for the first approach, we combine the data from fetal stage to the neonatal, P10, and adult stage. We build the trajectory for different lineage. One thing in common is the drop of transcriptional entropy. It doesn't matter what single-cell entropy estimation you use, such as STEM-ID, CCAT, or SLICE, or Cytotrace. You can always see the reduction of transcriptional entropy. And in other words, transcript entropy corresponds to heterogeneity of the gene expression or transcriptional activation. We then look for what genes are commonly upregulated in mature cell types. And as you can see here, there are markers that is commonly downregulated from fetal to adult stage, such as CFP family proteins and HOX family proteins. And some of these transcription factors are highly lineage-specific, such as oocyte-specific or erythroid-specific modules. And some of the transcription factors are are commonly activated in multiple lineages during mouse cell development and cell type maturation. And these common markers include the clock, AP1 family proteins, and XBP1. We then take the cross-species landscape analysis approach. And we try to look for what common regulators are conserved through the evolution since we have so many cell landscape data. Here we go. And there are genes that is commonly downregulated. These are conserved markers. These all relate to ribosomal proteins. And most of them are downregulated in multiple lineages and multiple species during development and cellular differentiation. And then there is only one transcription factor that is upregulated in multiple lineages and multiple species. This gene is called XBP1. And of course, in vertebrates, there are more conserved hits. XBP1 is a transcription factor that is known to be regulating UPR. And here we think that what we are looking at is transcription activity of this particular transcription factor. We use CRISPR-Cas9 to generate knockouts, and we look for single-cell atlas pattern during development. And since embryo dies at embryonic day 13.5, analyzed at 12.5, we see that Nearly all of the differentiated mature lineages are downregulated. We have less cell number. However, we do see an increase in progenitor cell types in most of the lineages, and we see an upregulation of single-cell entropy in the XBP1 knockout in nearly all of the lineages that we have tested. Another approach we take is to generate a model to interpret the trajectory from genome to cellular landscape. We know that transcription activation and transcription factor syntax are important for this cell lineage regulation. We use the deep learning approach to predict gene expression from DNA. And some of this work has been done in cell lines, but it hasn't been done at single-cell level and landscape. Scale, we build a new approach to interpret this, and the input is 13-kb DNA sequence around each TSS. The output is expression probability of each gene in each cell. So each— it's really a mega task neural network since we have to predict the expression probability of an individual gene in all of the single cells on the cell landscape. And after hyperparameter optimization, we train with the imputed single-cell landscape data, and we get a pretty nice AURC for nearly all of the species that we have tested. And as you can see, the MAGIC imputed single cell generates the best prediction rate. And at single cell level, we have good prediction or good model for some of the cells, but some of the cells we have poor level, the variation is big, and the MAGIC's performing imputed data perform better than the pseudoball data. The funny thing is that for all of the species, the germ cells always have the best prediction rate. Perhaps it's less dependent on epigenetic regulation and more dependent on transcriptional regulation. That can be one possibility. And metagenome with related species really helps to improve the neural model Using the single-cell data, our neural model performed better than many of these other cell population-based deep learning models for gene expression prediction. Then we exclude some of the chromosomes and train the model, and then we predict what genes are on those excluded chromosomes, and then we predict gene expression. Then we use those predicted gene expression on the leftover chromosome to run the cell landscape data. Then we compare the predicted cell landscape with the actually measured single-cell transcriptome landscape. We see a pretty nice overlap from the prediction with the actual single-cell gene expression experiment data. And so what Neuma has learned since this is a deep learning model, and you can do perturbation of regulatory elements in any region of the genome and then find its consequence of that mutation, synthetic mutation perturbation. And then we found that most of the, um, NuWa recognized regulatory elements correspond pretty well with epigenetic data or ENCODE data from the published works. And some of them overlap, some of them doesn't overlap. So we look and try to investigate why some of the reasons does not overlap. In this case, we compare the predicted regulatory region with the single-cell ATAC experimental data and ENCODE data. We found that there is a very strong regulatory signal at this region, which is corresponding to the 5S rRNA region. Because these regions are highly GC-rich, we think that in many of these PCR-based epigenetic experimental approaches, they missed this information. Therefore, they don't have the signal. But for the AI model, it doesn't have any problem, you know, similar to PCR bias. So it detects strong regulatory signals because 5S RNA should be expressed anyway. And it also detects the regulatory regions such as CCRE regions. And for many of these ATAC peaks, actually, they they can find overlap with the new predicted transcriptional activation signals. And from this model, we can test what filters are important for coding cellular landscape of different lineages. We can generate filters that is for, for elements that is important for human, mouse, zebrafish, and different species. Of course, we noticed that Most of these filters or NuWA elements correspond to known transcription factor binding sites. We see moderate lineage specificity of these NuWA elements with the lineage-specific expression. But we do see stronger lineage specificity when we combine the filters or NuWA elements, suggesting that transcription factor combination is the key for ensuring lineage specificity. We did a pretty general analysis. We think the lineage-specific elements or deep learning-related filters or elements are more conserved across species when compared to the actual lineage-specific transcription factors. And this might exert some of the evolution constraint on animal— in the animal kingdom. So, to summarize, to answer how many cell types are there, we think we need a complex cellular hierarchy with thousands of cellular states to, you know, make a finalized mammalian atlas. And then for the question, how are cell types regulated? And first of all, we found a pretty common regulator, and this is really not analyzed before, and this gene seems to be quite conserved across species, and it controls cell fate, cell attractor, or lineage stabilization for many species in many lineages. And then we generate a complex deep learning model to explain the trajectory from the genome to cell landscape. We have to say that perhaps in biology we can never get a simple model, and then finally you will realize complex This model is the only way. I will stop here, and I would like to thank two collaborators from my Zhejiang University, Liu Guanghui and Xu Jing, and Shi Qiang from IOZ. And most of this work are done by Professor Xiaoping Han and PhD student Renying Lijia Qi, as well as Jingjing Wang and Zhang Peijing.

Thanks very much, Goji. That was very inspiring. A couple of— okay, while I wait for other questions to trickle in, just quickly. Yeah, so coming to the first part of your talk, you mentioned immune-like, I think, epithelial cells and endothelial cells that express non-endothelial markers and so on. And I guess this goes to the concept of cells as bags of programs rather than sitting in boxes isolated from each other as discrete and completely distinct cell types, right? So an epithelial cell can express some immune programs, an immune cell can express some other mesenchymal programs that immune cells don't normally express, and so on. So then the question is, you know, and then ultimately because markers are never completely unique to one cell type, right? So how the immune-like epithelial cell, how immune-like is it?

Well, it's definitely a great question. It's a great question. It's different. I mean, it's different from immune cells. Indeed, it expresses some of the immune genes, I think for the IL family genes, IL-6, I think they may be the main contributor of IL-6 even for COVID-19. That's their responsibility. But for antigen presentation and immunoglobulin, I have to say I hesitate to say that they have exactly the same function as immune cells, probably no, because we have tried quite a lot of functional experiments, and their ability to activate T cells is very limited. And then the VDJ recombination, you know, you always detect some, some signal, but you always worry about contamination. But one thing for sure is that the diversity of VDJ recombination is way, way less complex than B cells and T cells. So Even if they are inflamed epithelials, I don't think they're the same as, you know, the actual T cells and B cells. But IL-6, IL family genes are their responsibility. I think that's the only thing that is confirmed. Yep. Thanks.

Got it. Thanks, Goji. There's a question from Chun-Cha Han. Does the training of NeUA rely only on gene expression data? Or does it require external, say, epigenome data, histone marks, and so on?

It's a great question. I mean, the reason why I do it is because I think only training by expression data is quite amazing, and that's the approach we take. I'm pretty sure if you include more epigenetic data such as single-cell ATAC data, you can increase the prediction rate or accuracy. I'm pretty sure about that. But the most exciting thing in my project, I think, is that we have never trained with any epigenetic data, only gene expression data, but still they can recognize how the genetic network is wired and what regions are actually epigenetic, should be epigenetic marked. This is how I get confidence with CNN or deep learning and technology. And so thanks.

I'm sure adding more data would help. That's an excellent point. And I guess you're working with some very, so, you know, distant or more basic model organisms as well as human and mammalian data and so on. And at least for, certainly for humans and for mouse, actually maybe more so for humans, people have tried for a long time to go from sequence to cis-regulatory function, right, to predict cis-regulatory function purely from sequence, and with the ultimate goal of predicting the effect of SNPs. Have you compared your newer predictions of the effect of SNPs to, say, eQTL analysis or any other standard ways of evaluating the, you know, human genetics predictions? Just a short answer, we're almost done.

Yeah, we have another project we haven't and talked about it because that project is fully, you know, focused on human genetics like those SNP sites. And the good thing is that there are overlaps. First of all, the eQTL and single-cell prediction model recognized elements, they overlap, and it has a significance. But the top hits of single-cell, you know, deep learning models are not those top hits in GWAS data. And the answer is very simple, because for these gene expression-based genetic models, the top hits will kill the embryo during development. Am I clear? Because the most important thing is that genetic model, this type of model predicts must be very strong phenotype and they don't even develop into embryo. However, for GWAS data, it's usually mild effect because at least you have to grow into mature cell types and then through the, you know, after 10 years or 20 years, you generate a disease phenotype. So in the end, it becomes a hot question to be answered because— so why they don't overlap that well? So we still have to investigate further. Thanks.

Excellent point. Yeah, effect size is a big issue. We are out of time, unfortunately. Fascinating talk, Goji. Always a pleasure to have you with us. Thank you.

Thanks.

All right, on to our next speaker, my co-chair for this session, Aki Minoda from the RIKEN Center for Integrative Medical Sciences. Aki, over to you.

Thanks so much, Shyam. Let me put this up.

Are you seeing it okay?

We can see it.

All right, great.

All right, so thanks for the introduction. So my name is Aki Minoda from RIKEN in Yokohama in Japan, where Jay and Piero are also. So today I'd like to share some information we have on the analysis from our efforts on mouse aging promoter atlas, as Jay kindly introduced. Using the 5' technology. And also, we have the model from the JUMP3 to be able to gain insights into the effect of the microbiota on aging. Right, so we have several different tissues. We have 10 different tissues that we have generated the 5' datasets on using the 10x Genomics. And we have the model of the SPF, which harbours normal standard microbiome as a control, and germ-free, which lacks completely the microbiota. And we've taken tissues from young and old, young from 9 weeks and 19 months for the old. And we believe this is going to be an excellent resource for the community of the aging research, as well as this— I believe this will be the first promoter mapping of the mouse tissues. So I believe— we believe this will be a valuable resource as well. So of course, there are many questions we can address, but fundamentally, we are interested in determining what are the effects effects of the microbiome on the rate of aging. Does it actually drive aging or not? What are the common aging signatures? Could we find across tissues or cell types? We're also generating single-cell ATAC-seq from the same tissues. So what are the chromatin aging signatures? And we'd like to integrate the datasets. This effort is a huge collaborative effort, mainly from Rican on the same campus as us, and the headquarters is basically comprised of myself, Tommy Toruatea, Renchen Chang, and the researchers who are really analyzing the datasets in depth. The data have been generated by many members from my lab here, but we have so many different collaborators who have kindly provided the tissue samples, which was a huge amount of effort we put together, many technicians also from their labs. So they kindly gave our samples, but not only that, of course, we are obtaining their expertise on the biology of each tissue as well. So here is the setup that we did. We took, as I mentioned, two different ages, SPF and jump-free. Our data is only from female, at least of that we are generating single-cell datasets on. And we have generated three mice worth of data, at least. And these are the list of the tissues depicted here. And as I mentioned, this is all single-cell 5' RNA-seq that we're calling Phase 1. In Phase 2, we also have— we'll be generating flow-through 5' RNA as well as single-cell ATAC-seq for these ones, as well as other tissues that we have flash frozen that we did not have the single-cell dissociation protocols at the time in our hands. I'm also very excited to say that even though this is not the single-cell level, we have tissue-level bulk metabolome datasets. So, standard sort of soluble metabolome dataset, as well as the lipidomics. This is something that is not so standardly carried out. So, very, very looking forward to being able to integrate such datasets. Of course, we have generated a microbiome dataset, and we have some tissues to be able to do some imaging. So far, we've generated easily over 200 single-cell libraries, and really, it's a— as Goji, it's a huge effort to integrate all of these together. So this is just a one snapshot of all of the datasets we have in our hands put together. And as you can see, the cell types tend to cluster together. And nicely separating. You may be aware that there has been a Tabula Muris Sinus that is very similar idea to what we're doing. Of course, we didn't know that we were doing each other at the time, but actually it's working out to be quite complementary, we believe. So the Tabula Muris Sinus generated both the full-length and 3' single-cell datasets, RNA datasets, and have deposited into database. And we are in active talk with them and to collaborate with them to actually share our data that will be 5' RNA coming from both SPF and GEM3, as well as the single-cell ATAC-seq data that we're generating right now. So what we want to know is how How do cells age? And how— what is the effect of the microbiota on that? So, these are some of the things that Ren Chen and Tommy are analysing at the moment. So, of course, we can do the cell type proportion changes during ageing and differential pathway activation amongst the cell types. Cell-to-cell communication is something we're interested in analysing as well. And of course, we have the transcription start site level resolution of the single-cell datasets to also determine changes at the enhancer regulation level. And then later on, we would like to integrate the single-cell epigenomics data, that is specifically the single-cell ATAC-seq data with the 5' data, as well as some bulk metabolome data as well as the microbiome, which will be a bit of a challenge also, and we need to— but we're working on this. So, if you look at the cell type proportion changes, so we have 4 different types of models, basically. So, here we have the standard microbiota-bearing SPF, young and old. And also, we have the germ-free that don't have the microbiota, germ-free completely. And so, we can do what is the differences between the old mice, between the SPF and germ-free, to gain insights into the effect of the microbiota on aging. Of course, as a sort of a filtering, pre-filter control, we also look at the the changes, the differences between SPF and germ-free at a young age. So, if there's already some differences here, then we cannot— we'll actually filter them out. So, this is an example from the bone marrow. So, you can see we just depicted some of the selected cell types. As you can see, there are some changes with some cell types that are more obvious. For example, the T cells, the T cells— actually, you see the increase in T cells with old mice, but that increase is actually not observed in the germ-free old mice, suggesting at least the microbiota is actually driving this increase in the T cells with age. So we do all of these comparisons systematically. So there are some— Ren Chen is also working on to best summarize all of these, a lot of data. This is one example. So you can see that here we can actually see this, the changes in the T cells here, as I mentioned before. And you can see in this summary plot, basically, that between the SPF, young and old, there is an increase, but that increase is not observed in the germ-free. And so, we can do this in a systematic way. So, we can also see an example where there's an increase in old, but it appears to be microbiota independence, since we don't see the green color here. And that if you— we see the square, it means the actual comparison, the significance is not so clear because there was some difference in the young John Free and SPF. So I think many of you are probably aware that this proportion actually analysis is proportionate, it's not the absolute cell number, which may also change. And we're really working hard on getting the annotation right, because obviously the biology can— and the interpretation can change greatly if we get the annotations wrong. So one of the things that we're also interested is inflammation, because this is one phenotype that is known to occur during aging. The sort of low-grade but chronic inflammation. And so we have used some genes that are known to play a role, have been shown to have link in inflammation, and looked at it by scoring inflammation score, so-called, using AUCell, this score, gene set activity. So what we see is actually in general, we do tend to see this overall inflammation signature with age, and that some of them do, they are rescued by the germ-free. So here is another sort of summary plot. In general, you see the purple on the left. And then the green on the right suggesting that in the normal SPF aging old age mice, there is an increase in inflammation across many different cell types. But that increase is not observed in the germ-free in the same extent. It also is, of course, important to see, to break it down to tissue level, because these are what we are showing here is actually, or cell types of all tissues combined. So whereas some of them tend to be consistent in increasing inflammation gene expression in the old, it's not always the case. So some of them could be more tissue-specific. Some may have some outliers, or exceptions, I should say. But in general, the trend appears to be that there is increased inflammation in the aged with the microbiota, but not as much in the absence of the microbiota, suggesting inflammation, a lot of the inflammation may be driven by the microbiota itself. And this is another example. These are non-immune cells. Also, we see actually a lot of the inflammation genes increased with age, and some of them are sort of rescued by the absence of the microbiota. As I mentioned, we're also interested in looking at how cell-to-cell communication changes with aging, and this is depicted by this. So this is increased— you can use Cell Phone DB or NATME. NATME is from Alistair's lab, and we can see that the there is an increase in cell-to-cell communication, which is often seen upon inflammation or immune state activation, whereas it is not really observed in the old. I'm running out of time, so I'm just going to go through, just to mention one thing, which is that we are also using ScaFE, as Jay mentioned, to look at the transcriber regulatory elements. And for the 5' datasets, we actually, Renchen and Tommy actually run Cicero, which is a tool that was originally developed for the chromatin accessibility, but actually we tried it on the 5' dataset and it seems to work well. And actually the prediction of the enhancer usage that appears to be changing with age, but also with the jump rate. So with that, I'd like to thank everybody who's been involved. A huge, huge, massive thanks to everybody who has been helping to make this possible. Thank you so much, and I'll take questions.

Thanks very much, Aki. We have time for, I think, one question from Al Forrest. Are you looking at differences in TCR repertoire between SPF and germ-free mice?

Yes, we're looking at the TCR. So we're running TCR-BCR on all of our tissue datasets actually. So yes, and I hope we can use it. Okay, fantastic. We haven't really been able to dive into the sort of the results yet, but we're doing it.

Fantastic. Okay, then maybe a short answer to one more question and then we'll close this round. Really intrigued by the fact that if you have— you don't have the microbiota, then you don't have that— you have less inflammaging. Do you think that could translate to human also? I mean, not that you'll have germ-free humans, but modern civilization is about reducing the amount of germs, right?

I know. I don't want to give a wrong message or impression that, you know, the microbiota is bad for you. Certainly, I wouldn't say that. But it appears that they are definitely playing a role, and I would bet that it is also the case in humans. Thank you very much. Hopefully we'll find out in some way.

Yes. Looking forward to it. Thanks again, Aki.

Thanks, Sheryl.

All right. Our next speaker is John Chambers from the Lee Kong Chian School of Medicine talking about the SGH 100K cohort study.

John?

I think you're still on mute, John.

Sorry, not on now.

Yep, all good now. All good. Great, thank you. Shan, thanks very much for the introduction, the introduction and the invitation to speak. My talk is going to be a lot more prosaic than the fantastic talks I've just listened to. I think the reason why Shan wanted me to present on the SG100K programme is just to inform the community about this resource that we're putting together that we hope will be of tremendous value to colleagues in Singapore and outside of Singapore. For scientific collaboration and exploration. So I'm John Chambers. I'm the Chief Scientific Officer for the National Precision Medicine Program and also the PI for the SG100K cohort study. So what is— what's this all about? Well, this slide summarizes the motivation for what we're up to. Singapore, like many other Asian countries, is undergoing an epidemiological transition. Life expectancy is high, population is becoming elderly, and there's a rising burden of chronic disease. And so to mitigate against this, we're keen to improve our understanding of disease etiology, risk stratification, and treatment selection for the population—so-called precision medicine journey. But we face the problem that most cohort studies, most epidemiological resources, are very Eurocentric. And so there's a need for us to put together an Asian-centric longitudinal population cohort to address the needs of the region. This slide gives a brief overview of the whole National Precision Medicine Programme. It's a 10-year vision. Phase 1 is complete and involves sequencing and data sharing around 10,000 individuals. We're now in what we call Phase 2, which will focus on the SG100K cohort that I'm about to describe and will bring whole genome sequencing, deep research phenotypic dataset and rich biological samples for exploration. The ultimate goal will be phase 3, in which we scale up the learnings from phase 2 into the healthcare system and start dealing with precision medicine at scale. Right, so what's the design of the SG-100K study? It's a It's done in a two-stage process. The ultimate goal is 100,000 Singaporeans aged 30 to 84 years, with oversampling of the Malay and the Indian communities to give an overall ethnic mix of Chinese 60%, Malay 20%, Indian 20%. As I said, it's a two-phase— putting this together is a two-phase effort. Phase 1 Phase 1 is built around sharing and integration of data and samples from 50,000 people who were initially recruited to the MEC and SEED studies between 2004 and 2011. The advantage of using these legacy samples is that they've got long-term follow-up, which enables identification of instant clinical events. It's great for biomarker discovery strategies. Phase 2, which is more exciting component of the study is new recruitment of 50,000 people in which we're able now to implement state-of-the-art comprehensive clinical and behavioral phenotyping as well as deep biological and imaging assessment. And the depth and breadth of the information available for phase 2 participants will greatly strengthen our ability to advance the precision medicine agenda. I'll briefly talk through— I'm going to focus probably on what we're doing in Phase 2. So as mentioned, people are recruited from the general population with a very active team of community champions. This helps us get a representative— it's a representative population study. We take broad consent from individuals to allow us to do a wide range of health-related research, to share data with academic and commercial collaborators, including collaborators outside of Singapore. I emphasize this because we are here to work collaboratively with the international community. We also have permission for deep molecular phenotyping, linkage to health-related records, and interest in the recontact by genotype and phenotype. So there happens to be an interesting question around a specific gene, set of gene variants, we're in a position we can recall people and do deep phenotyping studies. Participants attend a dedicated SG100K assessment center for an initial first visit that lasts about 5 hours. We collect information of 4 domains: physical questionnaires, physical measurements, imaging, biological samples. The questionnaire set is summarized here. There's information on health, lifestyle, cognition, diet through FFQ, using tools that are internationally standardized and also locally validated. Physical measurements summarized here: height, weight, blood pressure, 12-lead ECG, handgrip strength, spirometry, arterial stiffness, and treadmill test. Pretty standard stuff, population studies, not, not where the real excitement is at. This is where it starts to become much more state-of-the-art and much more exciting. These are some of the imaging evaluations we do. There's 2- and 3-dimensional carotid ultrasounds. We get early atherosclerosis. We have DEXA scans for bone mineral density, body fat, muscle composition. We have extensive retinal imaging which includes a fairly standard color fundus, but more excitingly optical coherence tomography and optical coherence tomographic angiography. And for those of you not in the know, that's a structural assessment of microvascular system which traditionally has been very difficult to get objective measures on. Stepping it up again, this is our biological sampling strategy. We go well beyond what's being collected in any other population cohort internationally. We've got the usual blood serum, Axgene tubes. We've also got skin tapes, stool samples, urine, and saliva for participants, allowing us deep assessments of host physiology and systems biology, as well as the interaction of host with microbiome. We've got a state-of-the-art cloud-based fantastic database for collecting the data in a high-quality way. We've with direct data capture from machines, of course. And just to illustrate, cohorts are often seen as being, as taking from the community. It's very important for the longevity and just good practice to give back to the community too. So we give a comprehensive health report to all participants, including reporting of incidental findings. And that's really been well received by the participants and has been a major driver to people wanting to take part. In fact, you know, we've got a 6-month, 10,000-person waiting list of people who want to join the study. I think that illustrates how robust our community engagement strategies have been. The initial strategy for generating molecular data is summarized on this slide. We're funded for whole genome sequencing at 30x depth and for generation of DNA methylation profiles by Epigrade for all 100,000 participants. This will be completed by the end of 2024. A phenomenal resource for looking at functional genomics, structural and functional genomics in a multi-Asian population. We're linking up to the electronic healthcare records. Singapore's got high-quality electronic records with high coverage of population. This will be a tremendous asset for strengthening clinical phenotypes and determining health outcomes. We've got a lot of data. Too much data, some people say. Not enough data, some others say. But the data are— we can show high quality, which we can confirm high quality samples do by demonstrating this. We expect some of the expected pilots of the proof of principles, proof of concept studies, the expected relationships. And turning to what the data look like, here's some of the preliminary whole genome sequence data. Shows the beautiful complex population structure of Singapore— 3 main ethnic groups: Malays, Indians, Chinese, with lots of admixture going back over centuries, probably. Understanding how we can use these data for clinical and molecular epidemiological studies is really exciting. There is going to be a lot of method development needed, for example, to unravel how, how this genetic structure, for example, also translates into social structure and how those social structures in turn impact on phenotypic variation and disease susceptibility. We've got a lot of experiments on the way. Like many others, we're doing polygenic risk. This is terribly important for the precision medicine journey. How can we use— this is one use case for how we use genetic information to strengthen healthcare in the population. My own personal interest is much is around the DNA methylation side and how we can use the genetics of DNA methylation to help walk the pathway, a functional molecular pathway between the genetic exposure and the phenotypic outcome. And this cohort study is going to be phenomenal for that with the depth and the size and also the breadth of phenotypes available. So this is our our strategic vision. We've got deep research phenotypic data, we've got deep molecular data, we've got healthcare records together making a unique data repository that will allow us and others to explore the biology of disease in Asian populations and identify new approaches to risk stratification, therapeutic intervention. We are committed to making the data available to colleagues. We have already got a national data access platform for our pilot data, and once the SG-100K dataset's available in a couple of years, we'll be extending this to the whole SG-100K dataset. So this is a summary. We've put together a unique resource for population precision medicine research. We're closely aligned to the national research agenda. We're filling in a globally, a really important globally unmet need for research for Asian populations. We've got high-quality data and samples We've got strong pathways to molecular and clinical phenotypes in place, and we've got a strong commitment to collaboration and partnership. We are— I'm one person, we're a very big group, you know, there's probably 1 or 200 people directly involved in data collection and initial data triage at the moment, but it's reaching out into the ecosystem and there are now probably 500 or 1,000 researchers working on the data. It's fantastic. I also need to be very appreciative of my colleagues in the team who've had a tough time over the last 18 months keeping this project going, and I'm very grateful to them and the participants who've taken part. Thank you for your attention.

Thank you very much, John. Thanks, John. Very exciting, and I hope this gives our audience a flavor of the context in which we do our work, especially cohort-scale single-cell science. And your talk is very timely because the single-cell field is moving from just datasets from a handful of subjects to datasets across cohorts. And so there's a lot of synergy, I think, for single-cell folks to be working closely with epidemiologists and people managing cohorts of this sort. I would say for background also, for example, Baradom and Ponpan in Mahidol University also doing their either single-cell work in collaboration with their National Precision Medicine Program. And for the audience, the IDA data that I showed earlier today was with samples from John's cohort in Singapore. So I think people want to know, where do you see single cell adding value in epidemiology?

Oh well, it's all about data quality. If you measure things properly, you discover things. If you measure things badly, you don't. So there's a beautiful illustration of that from measurement of physical activity. If you measure physical activity using a questionnaire, you get very weak relationships. Measure it using the accelerometer, you get beautiful strong relationships. And so much of our molecular epidemiological work is limited by to the use of mixed mixed cells collected from whole blood. We need to broaden this, we need to broaden it by collecting cells from more than just whole blood, but also when we work with whole blood we also need to understand what's going on in individual cell subsets. And I think none of the large-scale studies have addressed this important need, and I think once we do over the next— over the coming years, we're going to see an explosion of discovery. Seperate. It would be very meaningful. And thanks, Sham, for highlighting the relationship with IDA. I apologize for not mentioning that. Yes, we are using this cohort to drive the IDA data sample collection. And yes, we do have separated lymphocyte subsets for a fraction of the research participants. We'd like to make it more, but it's a bit expensive. It is.

It is. But the costs are getting lower and lower, I promise you. So, uh, slightly different question again. When we do this multinational studies in the single cell field— oh, we're already out of time, so okay, maybe two-sentence response. How do you address the challenge of returning results to communities and giving them some benefit? Like, for the single cell community, what can we do?

Uh, I, I— you probably can't. Because I think what research participants want is they want stuff they can— that's actionable, that they can work with. And so that's mostly things like whether they've got diabetes, high blood pressure, cardiovascular disease, you know, clear risk or disease. Now, if in the future you find something that is actionable, then great, we can return that. But at the moment, I think you can let sort of the clinician scientists do what they can to return value to participants. In the future, you'll create discoveries that bring social benefit, which would be fantastic. Excellent point.

One step at a time. Thank you very much, John.

That was extremely illuminating.

Thanks. Okay, our next speaker for this session is Jing Xu from the Chinese Academy of Science State Key Lab of Stem Cell and Reproductive Biology, Institute of Zoology.

Jing, look forward to your talk. Yes. Accident here. Sorry, can anybody see me and hear me? I just had an accident with my Zoom. APP.

No problem, all good.

We see and hear you. Yeah, okay, let me share my screen. Oops, I think the problems comes from the sharing screen. If I use the PowerPoint, sorry, here I'd like to exist my WeChat. It may consume too much of my Does it work?

We're seeing the presenter view right now. Ah, okay.

Now we see it. Okay.

Sorry.

Really, I apologize for the accident. And many thanks for Shiyam's And Aki's invitation is really an honor to share our work in the HCA Asia meeting. And what I'm interested in is the aging studies. And my topic is about the gene therapy alleviates aging defects by preventing cellular senescence. And I'm from the Chinese Academy of Science. And aging is very replicated in the heterogeneous. Our organs and different cell types, they are affected differently by aging. How to interfere in aging and how to rejuvenate the degenerative tissues and cells are what we have been interested in. And to explore the regulatory mechanism underlying aging, we need to figure out a way to distinguish effective factors from the tons of noises. If we regard aging as this sphere, aging intervention would be the shield that can resist aging. To systematically analyze the two opposite processes may help us to reduce the complexity of aging research and to capture the key information. Closely related to aging. And among all the ways to retard aging, calorie restriction is the most recognized one to extend lifespan in diverse species, from worm, mice, to the nonhuman primate, without causing obvious side effects. Therefore, we selected 3 groups of animals for aging studies, including young rats, old rats, and old rats with 9-month diet intervention. That's 70% calorie restriction. The weight of CR animals was significantly reduced, while the blood sugar level was not affected, indicating that these mice were not malnourished.. And consistent with previous reports, the lifespan of the CR rats was significantly extended. We then selected 9 tissues involved in the body metabolism, including brown adipocyte, aorta, kidney, liver, skin, bone marrow, muscle, and brain for the single-cell analysis. In total, we identified 109 cell clusters that could be divided into 42 major cell types across the 70 cells. And we also have the single nucleus atlas. Now, among all the cell states we identified, more than 80% of them have changed during aging. And CR intervention can reverse the aging-related changes by more than 60%. We found that the main cell types that undergo proportion change in somatic cells— that's the upper part— are tissue stem cells and vascular cells that decline with aging and recover with CR. In opposite, among the immune cells, for example, the neutrophil cell, the neutrophil plasma, cells, they increased during aging and significantly decreased after CR. So, our data suggested that the common changes in multi-organ aging are manifested in chronic inflammation of these tissues, reduction of stem cells, and vascular degeneration. And taking all the CD45-positive cells out, we can we constructed a multi-organ immune map for aging and CR intervention. Compared to the blue young group, and we can see that the yellow older group has increased immune cell types and numbers, which can be reversed back to the young state as the green CR group shown by the CR. And the numbers of immune cells in multi-tissues, except for marrow, increase during aging. And that increase can be prevented by CR. And the numbers of neutrophils and the plasma cells will predominantly increase during aging. And we then analyzed the differentially expressed genes and the aging DEG that were partially rescued by CR are referred to as rescue DEGs. We can see here CR rescue genes account for almost 1/3 of aging DEGs. And each organ shows the similar trend. And the most dominant tissue is the brown adipocyte. And the commonly downregulated genes across tissues during aging were enriched in the muscular development, the regeneration, and their downregulation was antagonized by CR. And aging also led to enhancement of multiple pathways, including immune system and inflammatory pathways. And this enhancement was attenuated by CR. Then the functional comparative analysis of aging and rescue the ages among the tissues further supports the role of CR in anti-inflammatory and tissue regeneration and the pro-angiogenic pathways, all of which are pathways potentially contributing to general protection. So, I won't go too much into the details. And if you are interested, you can download the paper and read. And also, So, you can check our website, the Aging Atlas, and to check your interest genes to see their expression changes. So, our question is, what is the cause for the accumulation of the immune cells? If we check the organs from both young and old animals, one of the most common phenotypes is the increase of senescent cells. For example, here in the mouse, joint, or in the monkey spinal cord, or the human heart, we can see the senescent cells accumulated with age. And consistently, the accumulation of senescent cells can be observed in the aged rats, as I just mentioned, and the CR can reduce the accumulation. So, what can senescent cells already do? Synanxin cells can secrete a number of pro-inflammatory factors, chemokines, and proteins. This is called the SASP, the synanxin-associated secretory phenotype. And these factors can cause chronic inflammation inside the organs. And at the same time, if those cells are stem cells, these synanxin stem cells lose their potential for self-renewal and differentiation, which leads to a decline in organ regeneration. Together, they can make the aging defects of these organs. And therefore, depleting the accumulated senescent cells would be a way to attenuate the chronic inflammation and alleviate aging. And senolytic drug specifically targeting the synthesin cells can effectively interfere in various aging-related diseases here. In principle, these drugs are expected to inhibit the chronic inflammation of aging organs, but they may have defects. For example, they may result in the stem cell exhaustion and unexpected structure damage in the tissues. Therefore, it may be better to rejuvenate these senescent cells to their younger state rather than eliminating them. And human mesenchymal stem cells are the cells we're interested in because they are involved in the maintenance of many tissues and organs. And we know some of the regulatory genes for these cells' senescence. However, our understanding is just the tip of the iceberg, and there are many more for us to explore. Therefore, we adopted the CRISPR-Cas9-dependent genome-wide screening method to discover the potential senescence regulatory genes. And the stem cell aging model we studied here are the HTBS and the Wernicke syndrome. Both are the premature aging diseases caused by gene mutations. These pathological aging models are recognized to reflect the physiologically aging phenotypes to some extent. And the library for us to use contains 2 to 3 guide RNA for each gene in the genome. And it can target gene with the nucleus Cas9 to cut the gene and silence it. So, by diluting the virus concentration, we can make only one guide RNA introduced to each cell. And by infecting a large amount of cells at the same time, we can obtain tens of thousands of different cells. Each of the cells has a specific gene that has been knocked out. So in this way, senescent cells can be eliminated during passages. And the final enriched cells other cells whose senescence has been retarded. As shown in this figure, we collect these cells at 4 weeks, 6 weeks, and 8 weeks after infection for the genomic sequencing. The higher the sRNA sequencing abundance, the stronger the ability of the gene after knockout to retard senescence. That is to say, the stronger the ability of the gene to promote senescence. So, we found a lot of candidate genes. And among them, there are only 11 we have known with a known function of the aging regulation. However, based on the GO analysis, we found that these genes are related to the cytokine-mediated signalling pathway, or the insulin response, and etc. These are consistent with the well-known aging regulatory pathways.. And starting from the 6 weeks, a gene encoding histone acetyltransferase, KAT7, reached to the top 1 candidate position and also ranking top in the other stem cell aging model, both in the Burnell syndrome and HCPS. And the further verification at the cellular level also proved the silencing of KAT7 can effectively delay the cell senescence. And it is also manifested by the decrease of SAST factors and the increased differentiation potential and increase of the cell retention ability after transplantation. And in addition to the premature aging models, we also try the other aging models. For example, the RAS oncogene-induced senescence. And the UV hydrogen peroxide-induced senescence. And they can all be delayed by silencing of KAT7, indicating KAT7 is a ubiquitous geroprotective target. And how does KAT7 function in promoting aging? And we all know KAT7 is a member of the MIST acetyltransferase family, and its substrate substrate is H3K14 acetylation. And in the human mesenchymal stem cells, we also confirmed that the silencing of KT7 can make their substrate diminish. For example, here, the Western blot, immunofluorescence, and ChIP-seq. And we can— we've already confirmed that the knockout of KT7 can retard the senescence, and overexpression of KD7 can accelerate senescence. Therefore, we perform the integrative analysis of the H3K14 acetylation ChIP-seq and the RNA-seq of the Werner cells, the Werner syndrome cells, with or without silencing KD7. And we also have the RNA-seq data from the YCAP cells, which is always without expression of KAT7. And we put all the data together, and we noticed there are 21 genes have the similar codon patterns. And we noticed the P15 is a cyclin-dependent kinase inhibitor. And we knock out KAT7 in the cells that can retard senescence. If we addition we add the p15 inside the cells, it can promote the cells into the senescence programme again. And also, if we overexpress the KT7 in the young cells, it can make the cells senescent. And if we silence the p15, it can reverse the aging-promoting effects of KT7. So this Through these two sets of experiments, we proved that KAT7 plays a key regulatory role in cellular senescence via the KAT7-H3K14 acetylation and p15 axis. And in addition to the premature aging models, the replicative senescence models, and also the primary mesenchymal stem cells from the physiological age-dependent. And also in the mice, we notice in the liver of the mice, there are increased expression of KAT7 and also its downstream target, the p15. Therefore, by injecting the luciferase reporter through the tail vein, we notice that directly target the liver of the mice. And we then use the same injection method we introduced the CRISPR system to target KAT7 in vivo. As shown here, only liver had the KAT7 gene edited among all the organs we detected. Jing, will you be wrapping up? Okay, almost there. So, through this method, we injected the 20-month-old mice and also the pro-mature aging mice, and they do show their lifespan can extended, and their behavior can be promoted after the gene therapy. And these are all the data from the mice liver. We can see the ablation of KDMF alleviates liver senescence and inflammation. And then we also tried the human hepatocytes, and it shows the same effects. And also, the gene expression pattern are quite similar. So, I'll just stop here. And what we are doing is to decode the aging by genome-wide screening and the construction of those atlases. And based on the single-cell atlas, we noticed that— we confirmed that aging can lead to chronic inflammation, and those senescent cells may be the cause. And we try to find a way to reverse aging, and it is potentially druggable. So, I'm sorry, I cannot use the background. Of the picture here. And so I used it for my last acknowledgement slide. And many thanks to my collaborators from the LOZ-CAS, the BIC-CAS, and also at Zhejiang University, Guoziguo. And thanks to all the audience.

And I'll stop here for questions. Thanks very much, Jing. Very great to see the translation. We're going to have to skip Q&A, unfortunately. For lack of time. I'd like to suggest to the succeeding speakers, please pin the Nextech KC window so you can see the timer during your talk.

Can you help me to stop that?

I cannot find the button. No problem.

Yeah.

Okay, I'm going to hand over to Aki Minoda, my co-chair, who will chair the remaining half of this session, and then we'll go over to the panel discussion.

Look forward to seeing you all there. Thank you. Thanks, Jayam. So I will take over from here. So the next speaker I'd like to introduce is Yonghyun Park. He's an assistant professor at Korea Advanced Institute of Science and Technology.

Take it away. Thank you, Aki, for a nice introduction. And can you all see my slide and hear me well? Yes. Okay, great. I'll start then. So thank you very much for inviting me to this great meeting for the Human Cell Atlas Asia. And today I want to spend some time to discuss how we can build a strategy to integrate the big data that we are generating at single-cell level to connect all the development and aging and disease across multiple organs. And if you think about the scale of the Human Cell Atlas, you can imagine that there are multiple axes. For example, we are dealing with the development and aging in normal physiology across multiple organs. And then we are also now starting to look at the ethnic diversity. And then if you think about all the pathological situations that we can develop through our lifetime, you can see how enormous our task is. So for Human Cell Atlas, we took this approach of divide and conquer so that we divide all the data generation steps types by separating by organs or developmental stages or disease status. So I think we have been quite successful in that, so that now there's already 164 datasets that have been deposited in the TCP in Human Cell Atlas. So now I think it's time to think about how we can then come back to our original quest of making a coherent Human Cell Atlas by integrating all these kinds of separated datasets. So today I'm going to talk about what we are going to need for the good data integration and one of the strategies that we took to do this. So let's first look at our ingredients. So what kind of datasets that we have in single-cell data formats. So first we start with the sequence files like FASTQ and BAM file, and then we turn them into the cell-by-gene matrix, the count matrix, which is in the format of Loom or Matrix. And finally, by the study, we merge all these kind of count metrics and we annotate them with the different set of annotations, and we also include metadata. So for like a data scientist like me, the main interesting starting point will be annotated metrics, which is already like merged and coherent and has all the metadata. But actually there's a lot of orders when we try to like integrate multiple datasets that has been generated from the multiple gaps. For example, if you think about the gene, the structure of the gene and cell matrix, ideally we would want to have our matrix from different studies to have an identical sort of the list of genes and gene annotations. But in reality, all different people are using different versions of the gene annotation, and some of genes get omitted because they have been cut off due to the low expression. So we never have an identical gene expression matrix. And I personally think genes with low expression shouldn't be removed at the final data deposition step because that kind of hinders the data integration process. And also for the metadata, people are using different nomenclatures, which makes it quite difficult to combine all these into together. But then for this part, I think Data Coordination Portal is doing really, really a great job on this. So if you go to the DCP homepage, then now you can see that most of the datasets that have been deposited are now being remapped using the UniFOM pipeline so that all the same structure of the gene and cell count matrix are getting produced. And we can see always the same metric for the QC. So this problem, I think, I believe it will be solved soon, like by the beginning of the next year maybe. But then let's think about the cell rotations, because still our pipeline doesn't offer us a chance of shared and unified cell rotation. And to achieve the harmonized cell rotation, I believe we should at least at once, at least one time, bring all the cells from the different datasets into the same space so that we can name them with the same logic and same categorizing algorithms. And if you have tried to do that, most of you I think has encountered this problem of batch effect, which is a very long-lasting problem in the single-cell biology. So personally, I have encountered this batch effect when I was participating in the project of making an atlas of human thymus, which was a very interesting organ because it is one of the fastest degenerating organ in our body. So we wanted to connect the thymic biology from its development to the ageing. So we collected thymus from 23 different donors, which, and from different stages, including fetal stage and pediatric and adult stage with ageing going on. And then what we noticed is is instead of our dataset being separated by biology, it's largely separated by batch. For example, different methods of 10x that we used or different genotypes because we have 23 different donors here. And so in general, nowadays there are many tools that solve this problem of batch effect. And I think there are 3 aspects that we need to think about to assess what is the good batch correction. So first criteria is whether it's doing the good batch removal, so good harmonization across batches. But then the secondly, it should also conserve biology well, so it should maintain the biological integrity, not doing the over-correction between the same different cell types and no distortion in the global structure. And for the human cell atlas purpose, I think the third characteristic, the scalability, is the most important. Because nowadays we are dealing with thousands of different datasets with tens of millions cells in it. So when I thought about what would be the most efficient way of batch correction, if you think about the general procedure of single-cell data analysis, you can see that most of our useful data analysis pipeline stems from the neighborhood graph. So if we can correct the batch effect at the neighborhood graph level,, then that could be potentially the most efficient and scalable tool. So then we start to think about what would the batch effect look like at the graph building step. So this is the situation where I have supposed 3 different cell types in a single batch that has been embedded in gene expression space. And then you can imagine if you have a batch effect, it will sort of shift the gene expression program in certain direction.. So in this gene expression space, they will move into certain direction, and then you have another batch and it will move into another direction too. Then when we try to model this by making the k-nearest neighbor graph, you can see that because single-cell data is very dense, most of the connections will be only made inside a batch. So the within-batch connections will be made, but actually batch effect is usually not that significant. Like, it is not as significant as you see in the UMAP. It's just slight distortion, but because of the density of the single-cell data, the across-batch cells never get to see each other. And another thing about that we can think about the batch effect is this. So if you now, like, think of this diagram as a cylinder, and then you put it— you, you look at it from this direction, So you just focus on the batch effect. And then you can think of the two scenarios. So first scenario is you have a random batch effect that the batches move in random direction. And this happens when you have a batch effect from the genotype or experimental batches. But then in second scenario, you can have a systematic and recurring batch. For example, when you compare different technologies in the single-cell technique, or if you have a situation like development and disease, you will have batches clustered and get separated in a certain way, in certain and predictable direction. And when we looked at this, we thought that just by connecting the graph and by allowing the, the cells from across batches to look at each other and measure the distance from each other, this kind of the random batch effect will be nulled out and the cells will be brought into the core of this kind of cylinder.. So that's how we sort of designed the algorithm called BBKNN. And indeed, the BBKNN was really fast and memory efficient because it only works at the graph level. Then we could see that we can really integrate, like, big-scale single-cell datasets. So here we have integrated 9 different big-scale mouse single-cell atlas datasets, and then we can see that this structure shows intuitive trajectory where the embryonic stem cells are located in the center. And then you can see the differentiation of all different cell types, including the hematopoietic lineage down here, gets stemmed out from this, the core. And using this kind of structure, we can really see what kind of transcription factors are lineage-specific in this pan-murine single-cell atlas. So PBKNO was quite useful by efficiently connecting the random batches. But I haven't mentioned about the systematic patches. And actually, when you have systematic patches in the BBKNN process, where you just connect the things at the graph and then keep the distances, you can see that these patch differences will sort of reinforce each other as you put in more and more of these kind of patches. And actually, what will happen is they will be aligned, but they will be sort of located in parallel to each other. So BBKNN aligns but not correct for these systematic batches. And if you see this batch effect benchmarking paper, this is the totally expected behavior of BBkNN where you have two different single-cell techniques that are sort of separated but parallelly aligned. So the same cell types are located near to each other. So we know that BBkNN is good at keeping the biology, but also it keeps a systematic batch structure. So we usually use this for the exploratory data analysis when you want to observe the overall structure of the technical batch effect or biological batch effect. But then sometimes you want to solve this problem and solve the systematic batch effect. So how can we do that? So to solve this problem, we have come up with this type of the batch correction, so the tandem batch correction approach, where before, at front of the BBKNN process, we had added another function to regress out only the systematic batch effect. And this is the simple and traditional linear model. But you know that when you try a linear model for the batch effect remover, you have a sort of the risk of overcorrecting and removing some known biology, right? So the way that we solve the problem is that these days, nowadays, there are many models available that we can use to predict the cell type. So we use the model to predict the cell type even before batch correction, and then keeping that, we put that as a variable in our linear model. And then we only ask the regression to correct for the systemic batch effect that is recurring. And we also do the regularization here so that we don't do the aggressive batch correction here. We just slightly bring the different systematic batches into more closer proximity. And that's how we came up with— 30 more seconds. Yes, I can wrap up now. So this is how we achieve this kind of coherent data set integration, which actually allowed us to look at what kind of developmental changes and aging-related changes are preserved there. And that is possible because we are not regressing out the systematic batch effect for the development and aging.. And this is just one other thing that I want to highlight, that we can then do the multi— combine the multi-organ data to find the rare population that is jumping from organ to organ. So that's the take-home message from here. So we apply the tandem batch correction with the first one to regress out the systematic batch effect and second one to regress out the— to align for the random batch effect. And this one is available in our Zenodo repository. So if you want to test this code, you can find it from here. And nowadays there are many model-based stereotype prediction tools, and I think it's very important to share the well-annotated dataset to build a better model, which will help us to integrate the different datasets. So I'd like to thank our lab members when I was doing this work in Sara Takematsu's lab, and especially Christoph, who worked with me on BBKNN. And our lab members who are studying together in the CAISS.

Thank you very much. Thanks, Yongmin, for a great talk. It's really appropriate to be able to integrate all of these datasets that's really accumulating in a DCP. That's really great. So I'm also curious, other than the batch effects, when you're integrating all of these datasets from different labs, Are there any other particular obstacles that we should be looking out for?

Usually, I think that's still the main one, that people are sharing the count matrix, but then the gene annotations are quite different. I think that's one big main obstacle that is making a lot of batch effects, which is unnecessary. So I think from this kind of ZERATRAS meeting, we can discuss more about that.

But still, TCP is solving the problem. Problem. So I'm also waiting for that. Right, right. So do you know how DCP is actually doing this?

What the process of the annotation is? So I think DCP is not yet working on the cell type annotations yet. So they are making this coherent gene and cell count table. So it's still up to the researchers. They need to upload the well-annotated cell type annotation.

So that we can just take from there and build a better model and make a coherent model. Right. I think there are various workgroups by tissues that are working together worldwide and to sort of come to consensus of which ones should be annotated, how. So hopefully, I mean, biology, I think, can really vary the outputs and the interpretations can really change if you don't get that right. So we really need to get together as a community, really nail this, because I think everybody's, so many people around the world, including us, spending so much time on annotation. And if that's streamlined, I think it can save us a lot of time. And time is money. I totally agree with that. I have just one quick question. So from Letu, nice talk.

How is PBKN performance affected by graph generating and cell type prediction methods? So I think for the cell type prediction method, for the regression, when you try to combine the regression-based approach. I think it really doesn't matter. So if the model is predicting the cell type in a coherent way, then that's not producing any more batch effect. And if the model is good enough, then you can have a rough idea about the biology. Then when you try to regress things out, then it will try to look at the consistent batch effect across those cell types.

I think that's what matters. So yeah, that's my key question about that. Thank you so much. We could also discuss this later on in the final session. Right, thank you so much. Next up, we have Sean Wilson, who is a graduate student from Murdoch Children's Research Institute in Australia. The title of his talk is The Challenges of Annotating Cells in Organoids.

Carry on with our topic. All right, thank you. Thank you for that introduction. It's a great honour to be here presenting at the Human Cell Atlas meeting today on the topic of challenging— the challenges of annotating cells in kidney organoids. So the kidneys are a pair of organs with the primary role of filtering the body's blood supply. And they do this through the functional unit, which is the nephron, of which each human kidney has on average about 1 million of these nephrons. So the nephron is a specialized epithelial structure at which, at one end, the glomerulus, filtration of the blood occurs, and that filtrate then passes through this specialized epithelial structure from the proximal tubules, loop of Henle, distal tubules, and then is passed to the bladder by the collecting duct. So globally, chronic kidney disease was the cause of 1.2 million deaths in 2017, with nearly 700 million cases reported. And with more than 2.5 million people currently receiving renal replacement therapy of some kind, which is estimated to double by the end of this decade, understanding genetic causes and the diseases of kidney disease is a really important research area to ameliorate this global healthcare burden. So the way in which we approach this area of research in our lab is the generation and application of stem cell-derived kidney tissue, which we call kidney organoids. We generate kidney organoids by following a protocol first published in 2015 from our lab, in which we take pluripotent stem cells and then guide them through in a differentiated direction towards kidney cell types, of which we sort of follow the in vivo process of that developmental sort of trajectory. So what we have after a month of culture is a 3D micromass of tissue, which contains tens to hundreds of connected structures that we can recognize nephrons. And we use this model system to research various themes of human developmental biology, toxicology, and drug screening and discovery and disease modeling. But what we hope is that organoids and kidney organoids will be able to be utilized optimally in personalized medicine for things such as cellular therapy and renal replacement. But for this dream to be realized, we have to have a really solid understanding of the cell types and structures that we generate within kidney organoids. And for this, we turn to generating transcriptional data to tell us about what we have in organoids. So the questions that we ask our transcriptional data is what cells are actually generating in vitro, how similar these cell types are to their in vivo kidney cell type counterparts, how mature these cells get within our organoids, and how we can compare between multiple protocols and datasets.. And so we've, we've encountered a number of challenges in the accurate classification of the cell types within kidney organoids. One of those is that organoids are actually developing tissue, so they're the equivalent of a late trimester 1, early trimester 2 human fetal kidney. So what this means is that kidney organoids contain cell types at a range of developmental stages, and the developing structures within the kidney is a really dynamic process within the three lineages the nephron, the ureteric epithelium, or collecting duct, and the stroma. And so we have this sort of range of cell types present within any kidney organoid at these later stages, which is really hard to delineate into distinct cell types. The other problem we have is that many of the classical markers that we've previously used that have been really clearly identifying cell types in histological research aren't specifically expressed when we we look at the transcriptional profile. And one really prominent example of that is that when we previously used GATA3 and HOXB7 to delineate the collecting duct in the kidneys, when we look at transcriptional profile, the distal nephron, which is adjacent to the collecting duct, also expresses these genes, even if it is at a lower level. But that means when we're looking at this transcriptional data, it's much harder for us to be able to delineate between these two populations.. And so to overcome this challenge, we've been trying to utilise reference atlases in our work to improve the quality, accuracy, and robustness of cell classification in organoids. One way in which we've approached doing this is to utilise a human fetal kidney reference atlas, of which we were provided a really high-quality sample dataset by our great collaboration with Jason Spencer's group. And once we had performed that accurate— what we thought was a highly accurate annotation using using sort of standard methods, then looked for and isolated enriched gene sets within this, within the populations we identified in the human fetal kidney, and were able to use these to score the transcriptional profiles of the datasets that we were generating. And one way in which we utilise this was to try and answer that question of what the GATA3-expressing epithelium in organoids was. And when we isolated this population using organoid-derived from fluorescent reporter lines that we generate in our lab and score the transcriptional profile, we found that the GATA3-expressing cells are much more like the distal nephron than they were the collecting duct. However, if we isolated and cultured these cells in a specific way in collecting duct sort of maintenance media, they took on a more collecting duct-like identity. And we could again see that reflected in their transcriptional profile and the expression of these gene sets. So this gave us some really interesting biological data. It showed that we were generating a plastic cell type that we're able to control the identity of. But it also provided us a way in which we could enrich for this population and then slightly alter the media and mature that population of cells to become a more mature-like collecting duct. And we are now able to utilize this in our lab to model diseases of the collecting duct, such as autosomal recessive polycystic kidney disease. This is a really great outcome. But moving into trying to ask questions around single-cell profiling, we wanted to see if we could utilise the same sort of concept to robustly specify unique populations within single-cell datasets. And so this is really complex and it's sort of been touched on already in a number of talks in this meeting. But with kidney organoids, we've got a really complex tissue with many different transitioning cell states and within the dataset itself, this disproportionate population dynamics, the different sort of percentages of cell types are different that are present, as well as when we're looking at an organoid culture, we have in vitro artifacts and unwanted off-target cell types that complicate classification of these data points. So the traditional method has always been to perform sort of a clustering analysis, which is rather user-guided. Sort of setting the resolution at which you decide your clustering is going to work, and then what populations you have based on that, it's fairly arbitrary. And so what we wanted to do was try and look to see if we could make that less arbitrary. And we turned to looking to classify, generate, use machine learning to train classifiers, and using the reference Human Fetal Kidney Atlas as a sort of ground truth that we could then transfer into our organoids. And so to achieve this, we integrated a number of high-quality reference datasets that were publicly available across multiple samples and time points. And once these are annotated as optimally as possible, we use these as a reference dataset to train a classification tool. And so we utilised a tool called SCPred, which is developed by our great collaborators in Joseph Powell's group, and we look to try and utilise this to generate models that could be applied in a robust and reproducible way. So to start with, we just looked at classifying the lineages, the general lineages within the kidney system, which is the nephron progenitors, the nephrons themselves, the collecting ducts, the stroma, and the endothelial cells that also arise within kidney. So scPred essentially trains a set of binary classifiers on the annotation of a reference reference dataset within a single-cell reference dataset, and then you can apply those to a test dataset. When we applied the scPred model based on this annotation to held-out human fetal kidney data, it seemed to generate really high accurate assessment of cellular identity, really high probabilities, and, and quite strong signal there.. And then when we transferred that to the existing field of organoids, of which there are, I think, around about 300,000 cells across maybe 60 individual sort of organoid samples, it performed similarly well. Of course, there was a greater distribution of some of these probability scores because it's an organoid in vitro system, but we were impressed enough that we looked to generate more specific models of the lineages within the kidney, so the nephron, collecting duct and the stroma, with sort of further models that would identify more specific cellular identity within our kidney organoids. And we pulled this all together into a hierarchy of models, and which, which we term DevKidz-CC, which is now an R-based software package in which you can use this model system to robustly classify a novel human— sorry, a novel organoid data set. And what we use this to do when we, when we, when we had this classification method was that we could, we felt that we could now directly compare all of the available data sets that were in the published sort of literature space in a way we hadn't been able to before because of many of the challenges around data integration, batch effects, and things like that. And so when we looked at directly comparing organoids generated from the three sort of main protocols that you use to generate kidney organoids that were published around the same time. We can now see that there is clear— there is variability, not just between the different protocols which generate organoids in slightly different ways, but also within a protocol. We can see, as I should have mentioned, these are all end-stage organoids that are sort of the most mature organoids we have. And so we can see there's variability that's present between line and batch and even the labs that are generating these datasets. But we can start to now quantify that in a direct way and answer some of these questions that we have around the different protocols and the usages. Because we have this direct, this way of classifying organoid datasets in a directly comparable way, We also pull this data into an interactive, interactive database, which is available through our GitHub as a, as a Shiny app you can use in the web browser. And you can now investigate any gene of interest that's expressed in any of these datasets and look to see where it's expressed, in what sort of kidney identity, what population. And this compares in comparing these datasets. And so in summary, I think I've shown that there is a utility of organoids relies on understanding the fundamental biology of the organoid systems that we're generating. And because cell classification of developing tissues is really complex and organoids even more so, we need these really high-quality reference atlases to compare our cultures with. I think I don't need to impress upon this audience the importance of integration of molecular and computational biologists as crucial to overcome these challenges.. And I'd like to think that we've provided an example of how we can integrate some of these different datasets in a way using this DevKidCC database. And with that, I'd like to conclude by thanking and acknowledging my two supervisors, Melissa Liddle and Sarah Howden, in the Kidney Regeneration Lab at MCRI in Melbourne, and our fantastic collaborators, including the Powell Lab and the Spence Lab, and our funding bodies.

And with that, I'd like to conclude my presentation. Thank you so much, Sean, for a great talk and keeping the time. It was perfect timing. So while we wait for possible questions, I have one question quickly. Short on time, so if you could just make it quick. So if you have— if the reference is the atlas, let's say the fundamental or healthy, and you're comparing to other states other models, how does it fit? How do the cell types or cell states that weren't in the original reference would appear?

Would it be designated as one of the cell types? Great question. Great question. It's a difficult challenge to overcome. So the key thing with the reference that we use is that it's all the cells in the human fetal kidney. But I think I mentioned in organoids we generate off-target cell types. And there's also— they're not exactly the same as what we have in the human fetal kidney just because they're an in vitro system. And that's why we wanted to use the human fetal kidney as a reference, because it gives us the best ground truth state, which we don't really have if we're asking for an in vitro system. And so any cells that don't have a high enough— that are not scored high enough on a probability within the model remain unclassified and are subsetted out of that. And then you can investigate those cells separately. And what we're finding is that we get a lot of muscle and neural cell types that are generally off target and aren't classified within our model.

But it is a challenge that we have to overcome. Sounds great. Thanks so much. All right. In the interest of time, I'd like to move on to the next speaker, who is Kylie James. From Garvan Institute Medical Research, who is a group leader there with a title of Gut-associated lymphoid tissue developmental programs readapted during Crohn's disease.

Thank you, Kylie.

Hey, thank you very much, Shiki. I'll just share my screen. Okay, you should be able to see that now. Okay, so thank you again for the introduction, and thank you to the organisers for the opportunity to share my work today. So as Aki said, I'm at Garvan, but I've only just joined, so the work I'm going to share with you actually comes from my time as a postdoc fellow at the Wellcome Sanger Institute in the UK. And there, our project was— or our work was very much focused on understanding the human intestinal tract. And I'm going to share with you an atlas and in particular focus on just one little vignette and where we use this atlas to look at gut-associated lymphoid tissue development and how these programs are readapted during Crohn's disease. So firstly, just a moment on gut-associated lymphoid tissues or GALTs. So these are hubs of localized immune activity in the intestinal tract. And so they're key sites of antigen presentation or activation of lymphocytes. And they include the Peyer's patches of the small intestines, the mesenteric lymph nodes, and the cryptopatches. And these structures are known to begin their development during in utero fetal development. But later on in life, we also see structures that resemble these that we call ectopic lymphoid structures or tertiary lymphoid structures that appear in situations of chronic inflammation and fulfill similar roles in supporting the requirement for additional kind of immune responses. And these structures have been reported in inflammatory bowel disease but also other chronic inflammatory or disease situations such as arthritis and also cancers. So understanding the kind of the process of development of these structures is really kind of key in understanding how we might be able to switch off that inflammation. So what we know about the genesis of GALTs, particularly from in utero development, mostly comes from model experiments in mice. And what we know from this body of work is that there seems to be 3 main key players So I'll just take a moment to explain these cells and their interactions. So the first is the lymphoid tissue initiating cell, or the LTI, which shares an origin with innate lymphoid cells. And these come in from circulation and are drawn into the tissue through signals they receive from the second cell player, which is the endothelial lymphoid tissue organising cell. Cell or ELTO cell. And these ELTO cells sit at lymphoid— at the intersection with lymphoid vessels and these blood vessels. And then when inside the tissue, these LTI cells can then interact with the third cell player, which is our mesenchymal lymphoid tissue organizing cell or MLTO cell. And it's really the interactions then between these two cells that causes the activation of these MLTO cells, who then can produce chemokines to draw in lymphocytes from the environment, from the circulation into the developing GALT. So as I said, this mostly comes from studies in mice. And so what we didn't know whether the same process occurs during GALT formation in the human intestine. And so we developed an atlas, and this is one of the questions that we were able to address. So a little bit about the gut atlas that we built, and this was co-led between myself and a PhD student, Rasa Elmentaiti, and together we did quite extensive sampling of the intestinal tract throughout human development. So first beginning in first trimester development, so 6 weeks post-conception, through to adulthood, which covers 20 to 75 years of age. And within this, we also included ileal biopsies from children undergoing endoscopies to see whether they had Crohn's disease. And so this has both control and Crohn's disease samples. And for most of these samples we did, we sampled multiple anatomical regions and mesenteric lymph nodes where possible. So from these samples, we did then did single-cell RNA sequencing using the 10x Chromium platform, which was the 5 prime. So here I'm showing all of our data together, so pooled, and after quite stringent quality control we have together 428,000 cells or thereabouts. And we first annotated these to the major cell lineages which I'm showing you here, and then we went through a process of subsetting these out and performing much more fine-grained annotation so that together we annotated a total of 132 cell types and states. And of course, I'm not going to go into each of those today. But if you are interested, please do check out our recent publication of this work in Nature. And you can also access the dataset at cellatlas.org. So to look at GALT formation, we went after those three main players that I spoke about before. So first, beginning with the MLTO cell. So we subsetted out the mesenchymal cells, which I'm showing here.. And amongst these, we identified these populations, 3 populations in particular here, which had the key signatures indicative of what's been described in mice. So that's specifically with expression of CCL21, 19, and CXCL13, which are really important for drawing those lymphocytes out of circulation. And in particular, one population which we labeled MLTO cells, was only found from 12 postconception weeks. So this is when we start to see mesenteric lymph nodes form in the fetal gut. And these cells were also enriched thereafter in the mesenteric lymph nodes. So then looking at for that second population, we then subsetted out our endothelial cells, and that's what we're showing here. And you can see quite big populations of arterial— and venous endothelial cells, but also these populations of Lyve1-positive lymphatic endothelial cells. So these are the cells that will surround the lymphatic vessels in the gut. And in particular, we found one very small population, which is about here, which was very specifically the highest expressors of ICAM-1, VCAM, and MadCAM-1. So very important molecules for recruitment of lymphocytes. And again, highlighting them as having the required signature to carry out the role of an ELTO cell. And again, this LEC2 population or the ELTO cells were found from second trimester development and highly enriched in the mesenteric lymph nodes.. So then on to our LTI cell, which, as you remember, I said has a common origin with innate lymphoid cells. So to identify these, we looked in our innate lymphoid cells and T-cell populations. And we found this arm of cells that had a signature indicative of an innate lymphoid cell. So down here we have ILC-3s from— that are specifically found postnatally., but also 3 populations that were exclusively found in the fetal gut. And I've just highlighted those and the specific molecules that we attribute to those here. So, one population that had more of a precursor signature with a high CCR7 expression, and then 2 other LTI-like populations that differed by their expression of NTR2. And we hypothesized that all three of these populations fulfill the role of LTI cells in GALT formation in the fetal gut, but perhaps at different stages during that process. And furthermore, we used RNA scope in collaboration with Omer Bayraktar and Kenny Roberts to identify these three populations of fetal LTI-like cells in the fetal gut. So next, As I mentioned, we had Crohn's disease, pediatric Crohn's disease samples within this dataset. And so, we looked to see whether there were equivalent cell types that we'd identified in the fetal gut also present in the Crohn's disease samples. And so, to do this, we did a logistic regression. I'm comparing the fetal cell type signatures with those in the Crohn's disease samples. And here, a high probability of similarity is depicted in the red colouring. And what we saw was there is a high degree of similarity between the ILC-3s in the paediatric Crohn's disease samples with our populations of LTI-like cells. And there was also a high probability of, or equivalent signatures with our MLTO cell in the fetal gut with populations of stromal cells in the Crohn's disease samples. And in particular, I just wanted to highlight the stromal 4 population, which we found was highly enriched in the Crohn's disease samples compared to the controls. And other papers by Allison Simmons in particular have also shown equivalent populations enriched in their studies of inflammatory bowel disease. These with single-cell RNA sequencing. So next, we wanted to see what signalling profiles were conserved in development and differed or differed in development versus during inflammatory bowel disease. So we screened for the expression of a whole range of chemokine ligands and receptors. And between our fetal cell types and our paediatric pediatric Crohn's disease cell types. And what we saw was that the chemokines that were key or were used to identify these different cell types were, of course, conserved between pediatric and fetal development. But just to highlight, there was also the additional subset of chemokine ligands in particular that were additionally expressed by stromal cells in the paediatric Crohn's disease that were not seen in development. And in particular, these ones are very important for recruitment of inflammatory monocytes. And it just highlights that there's this additional kind of requirement for these cell types in the context of disease to recruit inflammatory cell types. So this actually brings me to end of my summary. So here what I've shown you that by building a gut cell atlas of the human— throughout human development, we've been able to show that as in mice, there is this 3-cell model of lymphoid structure formation also in play in the fetal human mucosa. And in addition, kind of using the power of single-cell RNA sequencing, we've been able to tease out these three subpopulations of LTI cells that haven't previously been described and could represent a species difference. And then furthermore, by then comparing these cell types to those that we see in Crohn's disease, we've seen that there's equivalent populations in both settings, which could suggest the adaptation of these developmental programs during during disease. And this could represent, you know, a way that we can kind of block that inflammation and is something that we are following up with spatially resolved single-cell RNA sequencing. So I just want to thank Sarah Tigeman, whose lab I was in for this work, my co-lead Rasta Amai-Taitai, a whole range of colleagues and collaborators, donors and their families, our funding bodies, And thank you again, and I welcome any questions.

Thank you, Kylie, for a great talk. It's so great you're looking at such subpopulations. There's no questions from the audience so far, so maybe you want to put question for me. So in terms of like differentiating tissue resident cells versus those that have been coming from the periphery, is it Is there any— is there a way of differentiating from transcriptional profile? Or do they— I'm actually curious, do they change their identity or expression profile when they reach the—

basically a different environment, which is a new tissue? Yeah, no, it's a really good question. And certainly they're based on a whole like a lot of previous research, we do know that there are markers that identify tissue residence. CD69, for instance, would be one for tissue resident T cells. And there are a lot of chemokines that specifically direct migration of cells into peripheral tissues such as the gut. So we can rely on, on previous knowledge and to annotate cells as either overall. But also, using single-cell datasets from peripheral blood can also give us an indication if we look for, you know, paired, you know, equivalence of signatures between periphery and gut would be a way of identifying novel markers, for instance, of tissue resonance.

But it is a tricky thing. Yeah. Great. Thank you so much. Thank you. Yeah. We'd like to We'll move on to the next talk. Thank you, Kylie, which is the last talk for this session before we go into the discussion. And the last talk is by Yingxin Lin. Hopefully I'm pronouncing it not so far off. She's a graduate student at the University of Sydney.

She will talk about SC joint transfer learning for data integration of atlas-scale single-cell RNA-seq and ATAC-seq. Can you see my screen? Hang on. Yes. Yes. Thanks to the organisers for giving me the opportunity to present. Today I will talk about transfer learning for data integration of atlas-scale single-cell RNA-seq and single-cell as text-seq data. Recent development of single-cell technologies enables measurement of different modalities of individual cells, such as genomic chromatin accessibility, gene expression, protein abundance, and spatial transcriptomics. However, each technology measures only one aspect of cell identity and has its unique strengths and weaknesses. Therefore, integrating multiple measurements of single-cell omics is important for not only cell type classification, but also deeper understanding and characterization of cell identity and functions from different angles. There are two approaches to integrating multiple omics. One is from the biotechnology approach to measure multiple modalities of one cell at a time, such as CySeq and SNARE-seq. And the other approach is using computational methods. So today we will focus on the computational method, and more specifically is the integration of chromatin accessibility and gene expression on the atlas-scale data. So why do people want to integrate single-cell ATAC-seq with single-cell RNA-seq? So first of all, single-cell RNA-seq and ATAC-seq characterize the cells from different aspects. Integrating two data will allow us to have a better description of cell identity. And secondly, it is well known that single-cell RNA-seq is sparse, but with around 80% of zeros. But single-cell ATAC-seq is even sparser with 99% of zeros, which actually makes the data nearly binary. Therefore, single-cell ATAC-seq alone may not have enough power for cell type identification. And lastly, since single-cell RNA-seq technology is relatively mature and widely used, there are large amounts of public single-cell RNA-seq data with well-annotated cell types that can be used as reference to transfer labels to single-cell ATAC-seq. There are not many existing methods for integration of single-cell RNA-seq and single-cell ATAC-seq. And most of these methods have several limitations. For example, these methods only tested on simple scenarios and they are very sensitive to the features that are being selected for the integration. And also, most of these methods require long computational time and large memory, which is not practical. And there are also several challenges to integrate at this level, single-cell RNA-seq with single-cell ATAC-seq data. Such as there are varying resolution levels between two modalities and they have different scales in measurements. And also the cell types are very diverse across different tissues. To address these challenges, we have developed an end-to-end training method with joint update between dimension reduction and label transfer named SCJoint. The input of our method includes gene activity scores matrix from single-cell ATAC-seq data and gene expression and cell type labels from single-cell RNA-seq. The methods consist of 3 main steps. Step 1 is a joint neural network-based dimension reduction and a semi-supervised transfer learning. So this is based on a neural network that contains a joint embedding layer and a prediction layer. And the output of the first layer is a joint embedding for cells from single-cell RNA-seq and single-cell lexic. The loss function of the joint embedding layer consists of two main parts: an NDR loss and a cosine loss. The NDR loss encourages the features to be more discriminative across samples and at the same time encourages the projection to be sparser and minimize the correlation between the different dimensions. And the cosine is trying to let the cells that are neighbors in the other omics closer to each other. Then in the second step, we transfer the labels by k-nearest neighbors based on the joint embedding. And in the last step, we will use the transfer label to refine our— to rebuild the joint NDR models and then to refine the transfer labels. Next, I will show 3 case studies of the SC-JOIN. The first one is mouse cell atlas studies, and the second one is the integration of multimodal data. And the third one is the analysis of paired data. So first, mouse cell atlas. We apply our method on two mouse atlas data from single-cell RNA-seq and single-cell ATAC-seq. The mouse cell atlas of single-cell RNA-seq is tabular from 20 organs and tissues. Tissues by two technologies— generated by two technologies, SmartSeq2 and 10x Genomics. And these single-cell ATAC-seq data are sequenced from 13 tissues with 81,000 cells and 29 cell types. Overall, there are only 19 overlapping common cell types between the two atlas data. Here are the results of the overlap cell types integration of the two atlas. The first row is colored by the cell types and the second row is colored by the technologies. We can see that our algorithm is able to remove not only the batch effects from the single-cell RNA-seq technologies, but also the modality effects between the single-cell ATAC-seq and single-cell RNA-seq. And at the same time, scJohn preserves the cell type signals. And we also achieve a higher accuracy rate than other methods. This observation can be confirmed by the evaluation matrix. Here we use silhouette coefficient to evaluate how good the joint embedding of different methods in preserving the cell type signals and the batch effect removal. For the x-axis here, a higher coefficient indicates the cells of the same cell types are staying closer to each other. While a higher y-axis here indicates how batch are missing. Therefore, the best method should be on the top right corner, which is scJoint, as it achieved best performance in balancing the modality, effect removal, and cell type signal preservation. We also perform a stability analysis on the atlas data by downsampling cells from single-cell RNA-seq as training data. Compared to SURET, scJONs maintain a high and stable performance even when only 20% of the cells from single-cell RNA-seq datasets are used. This suggests that scJONs can be applied in practice where only a subset of single-cell RNA-seq data are annotated. So this comes to our first conclusion that scJONs shows robust performance on highly heterogeneous atlas data. Next, we will show scJohn's ability to incorporate additional modality information in the integration of RNA-seq and ATAC-seq data across different conditions. To illustrate this, we perform an integrated analysis of size-seq and seq-seq data from PBMC. Size-seq is one of the recent multimodal technologies that sequence gene expression and measure protein abundance in the same cell. Similar to SciSeq, a seqseq published in this year is able to sequence open chromatin and protein abundance in the same cell. Here we use a PBMC data from two conditions. One is control and another one is stimulated by CD3 and CD28 activation. Overall, scGene generated a better gene visualization of two technologies. Especially for subtypes of T cells. So, let's take this naive CD4+ T cell example. They are clearly separated from other T cell subtypes. And from the t-SNE, we can see that the two technologies are well mixed. Moreover, scRNA is the only method that's able to reveal the difference between two conditions. We can see that compared to B cells, NK cells,. And the naive CD4+ T cells from the two conditions actually clearly show difference between the conditions of control and stimulation. Again, looking at the transfer accuracy rates, aciDRAW outperforms the other two methods by a large margin, where we have 87% accuracy rate, significantly greater than the two other existing methods. Methods. So this is our second conclusion: scJoint enables accurate integration of single-cell multi-omics data. And finally, I will use SNASEq data as a gold standard to evaluate the performance of scJoint. We compare scJoint with two more methods that are designed for paired data. We can see that the joint embedding of scJoint is clearly grouped by cell type. And looking at the evaluation matrix here, we find that SC-Joint output— SC-Joint has very comparable performance with the paired methods while outperforming the other unpaired methods. So in summary, SC-Joint has comparable performance with paired method in the paired data. SC-Joint is also very computational efficient We use scJohn to integrate millions of cells from the Human Fetal Atlas generated by RNA-seq and ATAC-seq data. Due to memory limitation, most of the methods are not able to handle the data with more than millions of cells. And we found that scJohn is the only method that can integrate millions of cells in a short period of time, and at the same time achieve a very high accuracy. Accuracy rate. So scJohn is computationally efficient and consistently achieves high accuracy rate. In summary, scJohn is able to integrate at a scale single-cell RNA-seq and single-cell ATAC-seq data and construct John embedding from two modalities that removes batch effects and at the same time preserve the cell type signals. And also the label transfer by scJohn reach a high classification accuracy rate compared to existing methods. Lastly, I would like to acknowledge this project is with close collaboration with Tongyu and Sean from Wing Wong's group in Stanford. And I would also like to thank my two supervisors, Rachel and Jing, and all the members from Sydney Precision Bioinformatics Research Alliance.

And if you are interested in more details of SCGEM, You can check our manuscript in bioRxiv and also try our method, which is available in GitHub. Thank you. Thanks so much, Yingsi. We have a question from the audience. Li Tu, thanks for the talk.

How does casting binary nature data of single-cell ATAC-seq data into gene activity matrix affect the interpretation of the joint learning? Yeah, so one trick actually we did in the SC joint training is that we binarized both the single-cell RNA-seq and single-cell ATAC-seq.

We found that this preprocessing step actually with the transfer learning between the single-cell RNA-seq and single-cell ATAC-seq, which is probably due to the binary nature of single-cell ATAC-seq. Okay, I hope that answers the question. So in the interest of time, we are actually 15 minutes over, running over. So I'd like to thank Yingsen for this talk.

And move on to the last session, the last part of the session, which is a discussion session.