Scientist Stories: Richard Scheuermann, Machine Learning & Single Cell Genomics

Good afternoon, everyone. My name is Dan Wattendorf. I join you as a recent— to the Bill and Melinda Gates Foundation. I'm a director at a new team called the Innovative Technology Solutions in Global Health. I joined the Gates Foundation for some time at both DARPA and the NIH. I'm a little bit of a voyeur to your community. Human geneticist and, because of a particular lark of being in the U.S. military, ended up funding a lot of immunology and vaccine research while I was both at NIH and DARPA. This last session is, I think, particularly interesting because I think it represents some of the things that Gary has been trying to encourage, which is cross-disciplinary approaches to new technologies, and I think all of the speakers here are are exemplars of it. It was interesting seeing the earlier sessions today because I think all of them have particularly tools to address some of the things that came up. Our first speaker is Richard Shuerman, joining us from JCVI. He's the director of the J. Craig Venter Institute in La Jolla, an adjunct professor from UCSD. Richard's done a lot of work in computational methods and really been a key contributor to several of the big computational databases that have been used by many, both for flu viruses and immune databases. But his latest work, you know, has been on looking at machine learning applications to single-cell analysis, and particularly single-cell RNA and single-cell nuclei readout. So we look forward to hearing from him.

Richard, thank you. Thanks, Dan. So we're in the homestretch. You know, I think when Wayne started the conference today, he made an interesting comparison between the Human Genome Project and how it led to precision medicine and kind of the aspiration of the Human Vaccine Project, which is to lead to precision vaccinology. And so, you know, at the Venture Institute, we're still continuing to try to push the boundaries of the application of genomics technologies in a variety of different areas. And so what I wanted to do today is tell you about how we're trying to combine machine learning methods with single-cell genomics assays to try to get a better understanding for what are the cellular biomarkers that might predict vaccine responses. So let's see. How do I move this forward? The middle one. That big green arrow. And so Toby already kind of showed you this slide, but just to make the point that, you know, we have two flavors of vaccines that we're kind of interested in evaluating. Examples like HPV, where a single dose can lead to long-term protection in virtually everybody who gets vaccinated, versus something like hepatitis B, which shows quite variability in vaccine responses.. And so I think being able to compare the immune system and how it responds to these different types of vaccines is going to be very informative to us to kind of unravel the rules of immunity that we're very interested in exploring as part of the Human Vaccines Project. So as you all know, I'm sure, that different cell types play different physiological roles in the body and that the cell identity and its function The phenotype of the cell is basically dictated by the subset of genes or proteins that that cell expresses. And any kind of alteration in the normal expressed parts list gives rise to disease. And so what we wanted to do is— I guess our kind of underlying hypothesis in this project is that, you know, we propose that differences in the cellular response to vaccination are likely responsible to differences in vaccine efficacy and protection. I don't think that that's much of a stretch. Unfortunately, if you use transcriptional profiling of bulk populations, that can obscure the cellular complexity of these responses by focusing the analysis on previously defined cell types at a relatively kind of high level, low level of granularity. And so what we wanted to do is use single-cell genomics, single-cell RNA sequencing, to quantify cellular responses in an unbiased fashion. And the idea there is that that would allow us to evaluate both known cell types as well as novel cell types and what role they might be playing in the immune response to vaccination. So one of the computational techniques that we decided to apply in this area is machine learning. I'm going to go through a little bit of a background on machine learning in a bit, but typically people use machine learning as a way of solving a classification problem. They have different classes of things and you can develop a machine learning model that allows you to classify new things into these different buckets, into these different classes. But it turns out that machine learning can also be an excellent tool for identifying useful classification features. Which may end up being useful biomarkers in a variety of different settings. And these features that are useful for classification also frequently point to important underlying biological phenomena. And that's really important for us if we're trying to elucidate these rules of immunity. And so we're very interested in using machine learning methods like random forest that retains explainability, that highlights these useful classification features. And we think that those kinds of approaches have clear advantages over other kind of deep learning methods where some of the features get obscured by these different layers in the classification models. So this is, you know, this rule of— rules of immunity project that Toby described this morning. The idea here is that we used the hepatitis B vaccine as kind of our model vaccination, specifically because we expected to see variability in response to a single dose of the vaccine. And we wanted to use that variability for us to try to understand what are the determinants of that variability. This is a relatively small study. There's only 15 people in the cohort. And I think one of the take-home messages that I want to leave you with is that even with a study as small as 15 people, we still think that using techniques like machine learning and other kinds of integrative analysis approaches that we're able to tease out interesting phenomenon even from these small cohorts. So you also saw this. This is basically the study design, and, you know, basically what happens is that we collect a variety of different biological specimens throughout the course of this— this time course. And the samples that I'm going to focus on today were collected at day 0, so prior to vaccination, and then 1, 3, 7, and 14 days after the first dose of the hepatitis B vaccine. So this is— the idea here is that we want to see what's happening after a single dose. And then we basically took blood samples and then did single-cell RNA sequencing of selected immune cell subsets. And I'll go into that in a minute here. So we started out by focusing the analysis on 3 individuals that showed a response to a single dose of the hepatitis B vaccine and 3 non-responders. We collected samples, blood samples at these 5 different time points. And then Toby's group up in UBC sorted 5 different innate immune cell subsets from peripheral blood. So NK cells, monocytes, myeloid dendritic cells, plasmacytoid dendritic cells, and neutrophils. And this was no trivial undertaking because they had to do this with fresh material on all these days from 15 different participants, doing it, you know, in basically in real time doing the sorting. So it was a tremendous undertaking. So they sorted individual single cells into 96 or 384-well plates containing a detergent, and then they froze those down and then shipped those down to La Jolla for us to do the single-cell analysis. And so we use a single-cell RNA sequencing protocol called SmartSeq2. It's different from the more common method, 10x Genomics, and some of the other droplet techniques. We like SmartSeq2 for a couple of reasons. One is that it gives us full-length transcripts, and the second is that we tend to detect a larger number of genes expressed in any given sample than some of the droplet-based techniques. And I think that can be pretty helpful if some of the really relevant marker genes are expressed at relatively low levels. So in this particular study, we ended up generating cDNAs from over 14,000 sorted single cells. And we cherry-picked these based on a variety of quality control metrics and then basically took about 3,000 of those into single-cell sequencing, single-cell RNA-seq. And we ended up with over a terabyte of data that we end up playing around with. And this is our data processing workflow, but I'm not going to go into the details, but suffice it to say that We basically get the reads off of the Illumina instrument, and then we align those to the human genome, and then we use different computational methods to estimate both the transcript structure as well as the transcript abundance. And we use that information to do the downstream analysis. And so typically what everybody does with these single-cell transcriptomic experiments is the first thing they do is they do clustering. The idea here is that we want to group together individual single cells into groups that have similar transcriptional profiles. And there are a variety of different clustering techniques that you can use to do this grouping. And this just shows you the results using principal component analysis, tSNE, and this other method that kind of combines the results from multiple clustering methods. But in the end, we see very similar patterns in the data. So what I've done here is I've color-coded each one of these cell events, or these cell transcriptional profiles, based on the sorted population. And so, for example, here in purple, these were cells that were sorted based on expression of markers for plasmacytoid dendritic cells; in green, neutrophils; in red, NK cells. And first of all, you can see that The transcriptional profiles segregate very nicely with the sorted populations, which is not so surprising since these cells are actually quite different in terms of their physiology. So all that looked pretty good. But the one thing that we noticed is that there were some outlier populations that we thought suggested that there are subtypes within these parent populations. That may be interesting to explore. And the one that we ended up focusing on initially— sorry— is this— the myeloid dendritic cells had these two kind of major clusters that we saw that segregated pretty nicely apart from each other. And so we decided to kind of explore these two MDC subsets further, try to figure out what they are and what role they might be playing in vaccine responses. And so I'm going to tell you about a machine learning method that my group developed that we call NS-forest. And the goal of this particular method is not so much to solve the classification problem, but to pull out marker genes that— the minimum number of marker genes that are necessary and sufficient to be able to tag these different cell subsets. And so the idea is that we would— We would basically, you know, if we're interested in marker genes that are specific for this MDC outlier population, you know, can we use machine learning to help us pull out what are the best marker genes to solve that problem? Now, I decided that it might be interesting today to talk a little bit about machine learning in general. It seems to be a buzzword now in biomedical research. Everybody's talking about machine learning. Some people are actually doing it, but I'm not sure people really understand a lot of the details about what machine learning is and how these methods work. So I think for— to kind of interpret the results that I'm going to present, I thought it would be useful to kind of walk you through an example of machine learning and what it does. So if you're an expert in machine learning already, feel free to check your Facebook or your email for the next 5 minutes. But for those of you who are kind of curious about some of the details, I thought it would be worthwhile of going through it. So we use a method that's called random forest, or random forest of decision trees. So it's a type of supervised machine learning. And basically what you have to do is you have to provide some training data that is then used to build a classification model. So in our case, we've already done clustering of these single cells. So we have divided them into clusters. And so our training data is basically going to be, can you tell us the marker genes that are good at distinguishing between these different clusters? So the training data needs to consist of a set of data records, samples with known class labels and feature values. And so for us, the samples are single cells. The feature values are the expression of individual genes in the transcriptomic data. And the class labels will be the clusters that these different single cells are located in. One of the nice things about random forest is that the features can— actually, features can be continuous, discrete, or categorical. So in our case, our features are kind of continuous gene expression levels, but you could also feed it categorical data and it can still build decision trees based on those characteristics. And so then what you do is the training data has to have features and class labels determined. And then we construct a collection of decision trees, a forest of decision trees. And the big advantage of using random forest over just single decision trees is that this overcomes the problem of overfitting that can happen with single decision trees. And hopefully you'll get a sense of that when I go through the explanation. Okay. So this is a toy example of how these techniques work. So in this case, we have each one of these dots here will represent a single-cell transcriptional profile. They're divided up into classes, and in this case, the classes are color-coded, and they would represent the clusters that we see in t-SNE or PCA space. And then the features here would be the expression of two different genes in the dataset. So in this case, we have four different classes. And so what we want to do is we want to develop a decision tree that allows us to segregate these classes into different bins. And so you might start out by drawing a partition in this first feature vector space that would partition here. That gives pretty good segregation of the blue and red from the purple and orange. Then in the next step in the tree, we would then maybe draw a partition here. Now we've segregated these two guys. Over here, we get close, but we would actually need one more partition to be able to cleanly segregate the purple from the orange. And so ultimately what you want to do is by the time you get to the bottom of the decision tree, each one of these partitions, each one of these bins should be pure in terms of the class membership that you have in that partition. So this is the way a decision tree works. Now the question is, how do you know where to draw that partition? So the way that you do that is you basically draw all the partitions that you could imagine in any given space, and then we have a way of determining what's the best partition to draw. And the way that we do that is we calculate— first of all, we calculate a value called entropy. And entropy is essentially a measure of the purity of a partition in terms of a particular class. So if you— in a partition, you only have one class in there, there's no entropy. It's a pure population and you have an entropy value of 0. If you have a dispersion of all the different classes within that partition, then you have a relatively high entropy value. And so that might be, you know, 2 or 3 or 4, something like that. And so what we want to do at each step in the decision tree is to maximize what's called the information gain. And that's just the change in the entropy. So you want to maximize, you want to draw a partition so that you drop the entropy as much as you can in that one partition. And the way that we determine that is using this information gain metric. So basically what we do then is we can, we have computational methods that figure out where this optimal partition is and we keep going until we get to the end of the decision tree. So it's pretty straightforward. Now, the really important thing about random forest, and this is— I can't really overemphasize this— is that the reason that we do multiple decision trees is because if you just do a single decision tree, you can get trapped in local optimum— well, not really local optimum solutions. You can— there's the problem of overfitting the data. And the way that we overcome that problem is that we start out by instead of taking all of the data, all of the individual samples and all the features, we randomly select subsets of samples and subsets of features, and then we build a decision tree based on that random selection of subsets of the data. We do that, we generate one decision tree. Then we go back and we iterate that. We select out a different subset of samples and a different subset of features, and by doing that We don't get the problem of overfitting that we had in the first decision tree because now we're using different samples and different features. And so basically we just generate a forest of these decision trees. And in our case, we typically do between maybe 10,000 and 100,000 decision trees. And then it's this kind of collection of models that allows us to do classification without overfitting. But for our purposes, to look for marker genes, what we want to do is we want to understand which of the genes are being selected out at these different partitions in the tree that are really good at making— solving that classification problem. And so what you end up getting is you get a ranked list of features from the random forest based on how frequently they're used at different branch points in this collection of decision trees. And so this is an example here where— so for— in this particular classification, this is the HLA-DPB1 gene. It was available as a feature in a decision tree about 4,200 times. And when it was available, it was selected 4,100 times as that first split point. So what that says is that this is a really valuable feature in making that distinction. Between these two classes of data. If it's available, typically this gene will be picked first when you're figuring out which is the optimum partition. So we can— we have this beautiful rank list now of features that are useful in solving this classification problem. Okay, so what we wanted to do then is we wanted to take that as a starting point and we want to determine the minimum number of marker genes that would give us optimal classification accuracy. And so the way that we do that is we take this ranked list of useful features and then we build this decision tree using the top gene, then the top 2 genes, then the top 3 genes, et cetera. And then we measure classification accuracy. And we stop when we reach a plateau of classification accuracy or a peak in classification accuracy. And that tells us this is the minimum number of genes you need to mark this cell type. So we applied this approach to the clusters that I just showed, and so we basically selected out a set of marker genes for the— this kind of the main MDC cluster and this outlier cluster in comparison with all of the other clusters in the dataset. And we ended up finding that in both cases, 3 genes were sufficient to be able to uniquely distinguish between these clusters. And so for the main MDC cluster, it was the FC gamma receptor 3A gene and also the CDKN1C gene and this TCF gene. For the outlier MDC cluster, it was class II, NDRG2, and FC epsilon receptor 1A. Couple of things that you'll notice here is that some of these features are kind of quantitative differences. So for example, over here, the Fc gamma receptor 3A is expressed at relatively high levels in the main MDC cluster. It's also expressed in other cells like NK cells, but at much lower levels. And so some of these features are quantitative differences between cell types. But in other cases, it looks more binary. So the CDKN1C gene seems to be uniquely expressed in the main MDC cluster, but not in other cell types, whereas the NDRG1— or sorry, NDRG2 gene is expressed in the outlier MDC cluster and essentially not in the other cell types. So there's a— oops. There's several things that we do with these marker genes. So one thing that you can do is that instead of doing complete transcriptional profiling on single cells, now we can just use kind of a lower-dimensional, higher-throughput technique like quantitative PCR to identify these different cell subsets. Or if you want to look at spatial relationships in tissue, you can do multiplex FISH with these markers. We also use the marker genes to come up with semantic representations of cell types. So we're defining all these new cell type clusters; we have to give them names. So we can start conversing about them. And so we use these marker genes and other characteristics of the clusters to come up with semantic definitions that we can load into ontologies and be able to converse about cell types. And then the third thing that we're doing is we're developing reference databases of cell types with quantitative expression data that's focused on these marker genes. And so these reference databases of marker gene expression can be used to do comparative analysis. So you do your own clustering, you get your own cell type identified, then you can say, I wonder if this is similar to other cells that have been defined previously. You can basically use this reference database of marker gene expression to do that comparative analysis. Okay, so what we ended up doing then is we went back to the hepatitis C trial and we simply did quantitative PCR for these two marker genes. In the 3 responders and 7 non-responders. And we ended up with about 3,400 single-sorted MDCs that expressed actin as kind of our positive control. And a little over half of those were positive for at least one of those two markers. But when we looked at whether they expressed one or the other marker or both, what we found is that the expression of these markers are largely mutually exclusive. Which is kind of what our prediction was. So we had about, well, 1,779 single MDCs that expressed at least one marker, and only 17 expressed both. So they either expressed one or the other. So less than 1% expressed both markers. So these things really are mutually exclusive. And we think that these cells that express both are probably a sorting artifact, that we probably ended up sorting doublets in those wells. Okay, so the interesting thing now is that we can use this quantitative PCR data to look at the relative proportion of these different MDC subsets in these different samples coming from the vaccine study. And what we found was that the ratio of these two MDC subsets correlated with the vaccine response after a single dose of the hepatitis B vaccine. So there were basically 3 individuals that showed a response to a single dose. 2 of them showed a relatively high response. And both of these had a relatively high ratio of these 2 MDC subsets. So this is kind of provocative. It's pretty interesting. The other important point to point out here is that this is at day 0. So this ratio exists prior to vaccination. And so we think that the ratio is actually predictive of who is going to respond to the hepatitis B vaccine and who's not going to respond. So prior to vaccination, we can predict who's going to respond. So I think that's pretty exciting. Now, the caveat is this is just 15 patients. It's not a perfect correlation, so— and this is just one experiment. Maybe we just got lucky. But then we said, hey, I wonder if we go back and look in the bulk RNA-seq data that Toby's group had done with whole blood and looked for the expression of these two genes in bulk blood, would we see any kind of correlation? And, you know, interestingly enough, we do. So here are the two marker genes. And you see that this relative ratio— so here we have the expression of one marker gene, the expression of the other marker gene. And then this is serum antibody titers after the third dose of the hepatitis B vaccine. So we think that this ratio here of these two marker genes actually not only predicts the response to the first dose, but it actually predicts this response to all 3 doses. And in this study of 15 subjects, by the 3rd dose, almost everybody responded to the vaccine except for 1 person. That person, they gave 3, 4, 5 doses, I think, and they never ended up responding to the hepatitis B vaccine. Well, that guy is right up here. So, you know, there's still a lot of kind of follow-up studies that we have to do on this, but we think that this is this looks very promising in terms of identifying biomarkers that can potentially predict vaccine responses. And I think it really kind of emphasizes that kind of combining single-cell RNA sequencing with machine learning allows us, first of all, to understand the cellular complexity in these different samples in these different settings. But in this case, we think that we've identified multiple MDC subsets and PDC subsets. Subsets in these samples. And that using a very simple quantitative PCR technique, we can quantify the relative levels of these different subsets and then look to see how they correlate with vaccine responses. And even predict who's going to respond. Now, it seems that a relatively high level of an MDC subset that expresses high levels of MHC Class II is the one that predicts vaccine response. And I guess that's not too surprising, that if you have a high level of kind of preactivated MDCs, that you are kind of poised to be able to respond to the vaccine. And so that leads us to the hypothesis that maybe if we precondition patients in some way, by maybe by giving them an adjuvant before we give them a vaccine, and we boost the level of this MDC subset, maybe we can have everybody respond to the vaccine. So that's the idea. And I guess with that, I will just mention that this has been a really terrific collaboration between my group at the Venture Institute and Toby's group at UBC. And, you know, I think it's been a tremendous collaboration. We bring complementary expertise to the problem. And in the end, I think that's really the future of kind of the human vaccine project. So with that, I'll stop and take any questions.

Yeah.

[INAUDIBLE] Yeah. [INAUDIBLE] [INAUDIBLE] Yeah, you know, we haven't really had a chance to look at any of those kinds of correlations yet. I mean, this is all pretty new data, but I think, you know, it would be interesting to see if the variability that we're seeing in this initial level might be reflected in polymorphisms that are associated with the gene. But at this stage, we don't know. Do you know the function of the— CDK1 is a cyclin-dependent kinase inhibitor. And so we think that that is a marker of kind of a resting MDC subset, whereas the other subset expresses high levels of class II, and we think that maybe that's been activated in some way. But yeah.

It sort of strikes me that this is sort of person intrinsic, that you expect to see similar responses to other, what we would call, subunit vaccines. Yeah. But if you have their blood, so you could actually look to see whether there's a correlation between what you see with hepatitis B in response to the periodontal—

Yeah. If you had their records to look when they were vaccinated. Yeah, yeah. Well, now there's like all this data that's like in different databases, and we can go back now and see if we can see evidence for, you know, differential expression of these genes and responses to other vaccines. I thought you were going to ask about the HPV vaccine where everybody responds. So it doesn't matter what that ratio is for HPV apparently. HPV, the acute reactivate MDCs.

Hepatitis B I don't think does. Right. It's not needed to be preactivated because the immunogen itself is acutely activated.

Right. Yeah, yeah.

So do these two subtypes relate to the CD8-positive and CD4-positive DCs? So there are two sort of classically defined MDC subtypes that are developed in the developmental program. One is specialized to prime CD4 T cells, and the other one is specialized to cross-prime CD8 T cells.

Would it be interesting to look at that?

Yeah, I'm not sure about that. I'll have to look into that. And it might also be—

it would be nice to look in the tissues, in the lymph nodes, because the numbers in the blood tend to be a little more variable because they're just transiting.

Yeah, yeah. Well, and the other interesting thing is these things kind of disappear like by day 1 after vaccination. So they're moving around pretty rapidly.

We talked before just a little bit about ground truthing. —your classifier is predicting classification based on cell surface markers that are— we're saying this is truth of what these cells are. I mean, if you did this experiment the other way, where you're looking at just correlative protection based on genetic sequence as your primary, and then trying to determine which of the— which cell types they fall into, can you do it in reverse, and would you identify potentially new populations of cells that really aren't phenotyped by their cell surface marker?

[Speaker:Joshua Elkington] Yeah, I'm not sure about that. But, you know, the one thing was that, you know, we have done kind of a comprehensive analysis of all these different omics datasets that have been generated on the same samples. And based on the traditional way that you approach transcriptomic analysis, we didn't find anything that correlated with vaccine responses at day 0 in bulk RNA-seq data. So I do think that the direction that you're able to kind of validate these findings will make a difference. Okay, thank you.