Scientist Stories: Oana Ursu, Phenotyping Coding Variants in Cancer using Perturb-seq

Cool, so maybe we can just go ahead and get started. There's a couple minutes of intro stuff, which will give people some time to file in. So welcome everyone and thank you for virtually joining us for the latest version of the Variant Effects Seminar Series, or VESS. My name is Diego. And I'm joined by Stephen Irwin, who is the other member of the VESS organizing committee. These seminars are made possible with support and a lot of help from Laura Muffley and Alex Hopkins, among several others, and our sponsor, the Atlas of Variant Effects Alliance. So a big thank you to them. Before jumping in, I have a few housekeeping items to mention. Each talk is about 20 to 25 minutes, so there should be time for a few questions. Today we just have one speaker, so there'll be a little bit of wiggle room with these times. Please post questions to the Zoom Q&A or chat, and we'll read them at the end for the speaker to answer the questions. And if you have any additional questions after the end of the seminar and you're an AAVE member, feel free to visit our seminar series Slack channel for more discussion. You can catch the seminar today and previous presentations that have been recorded on our CMAP YouTube channel. And finally, we're now posting updates on our Twitter account @VarianceEffects, so be sure to follow us there. This has never been an issue, but as a reminder, please be considerate when posting comments or questions. We expect everyone to be respectful and use welcoming and inclusive language. All the details of this seminar series can be found on our website. And we should be updating the future speakers' schedule fairly soon. And without further ado, I'm very excited for today's presentation. Ioana Ursu is a postdoc working with Aviv Regev at Genentech. She studies the molecular basis of cellular decisions and the resulting cell states using pooled high-content screening. Her latest work is applying this technology to profile the impact of coding variation in cancer. And this was a collaboration with J.T. Neal and Jesse Boehm at the Broad Institute. The title of her talk today is Phenotyping Coding Variants in Cancer Using Perturb-Seq. So Ioana, feel free to start sharing your screen and take it away when you're ready.

Cool. Thank you so much for this wonderful introduction. All right, so I'm really excited to be here and tell you about our work basically studying coding variants with single-cell screening technologies. And before I kind of dive into the science, I just want to mention that I'm one of many in a team that has been a really wonderful collaboration. And in particular, JT Neal has been leading all the experiments that I'm going to talk about, and I worked on the analysis. And this has been really a collaboration between the Regev lab and Jesse Bone's group at the Broad Institute. And if you have more questions or just want more references, the paper is now online, so you can also look at that. So basically what motivates us is to try to understand the roles of coding variants, and in particular in cancer, that's important. Because what happens when a patient comes in is that doctors will look at the different mutations that a tumor might have and make decisions about the treatments that are going to be used. And so for that reason, mutations in important cancer genes are sort of actually having consequential effects. But what happens a lot of times is that a patient will have a mutation in a gene that is sort of actionable, but that mutation is just quite rare. So it hasn't been seen in the population enough for us to understand whether it's a mutation that impacts that gene or not. And in that sort of case, it's basically there's not enough information to actually make the best decision for that patient. And that's where we, and obviously the field, and especially this group of studying variant effects want to really understand the role of each coding variant, such that then in such a situation one could look up the expected role of a variant. So here's an example. It's the PI3 kinase gene, and I'm showing you basically along the protein how many times each mutation has been observed in cancer patients. And you can see that there are some hotspots of mutations that have been observed really frequently and that are known to be sort of driver events, but there's all these other mutations that occur in 1 or 2 patients that we would want to understand. And really what in an ideal world we'd like to know how these variants affect the function of the gene, like is it a gain of function or loss of function and so on, how this translates to effect on tumor fitness, and then especially how all of this translates then to drug response in order to prioritize the most efficacious drugs for each patient. And this is, of course, challenging to get such an atlas because there's a large space of possible variants, and because depending on the gene of study, usually you have to develop a separate sort of functional readout of whether a variant changes the function of the gene. So it's not straightforward to to study variants across multiple genes without investing this type of effort. And really what I'm going to talk about today and what the field is thinking about is how can we annotate the effect of many coding variants in high-throughput with a single assay such that we wouldn't have to develop a new type of assay every time we want to study variants in a different gene. And the key idea here is that when we want to get from variants to their effect on gene function and tumor fitness and drug response, one thing that we could do is just measure something about those variants that allows us to embed them in a phenotypic space where variants with similar functionality are sort of clustering together. And then all we need to do is understand how this phenotypic space then maps to what we care about the variants. In particular these three axes. And so today I'll talk about using gene expression to do this type of embedding, but you can of course think of other modalities as well that might shed light on variants, such as imaging-based approaches or chromatin accessibility or anything else really where variants could distinguish themselves based on what they do to cells. So to understand what variants do, the most intuitive thing to do would be to take some cells and engineer the specific variant at its endogenous locus using something like base editing and then see what happens to the cells. But since, especially when we were working on this, this was not a very efficient process, we actually did something that is a little bit different. And so the idea is that we want to measure the function of the variant by looking at how do cells react when we overexpress the variant. So in other words, the cell will be a sort of a biosensor that tells us whether the variant is doing something or not. So let me give you an example of this. So we start with a cell and then we overexpress the wild type version of a given gene. And that just overexpressing the wild type gene might lead to some change in cellular state. In this particular case, it's becoming a square rather than this circle. And so then what we can do is do the same, overexpress the variant, and ask ourselves whether we see the same changes we did when we overexpress the wild type. Because if we do, that means the variant cannot really be distinguished from the wild type. So probably it's a neutral variant, it doesn't really do anything to the gene. If on the other hand, when we overexpress the variant, it just looks like nothing really happened to the cell, that might indicate it's a loss of function. It will just look as if the overexpressed protein is not even there. And the third possibility is that the variant, it doesn't look like the wild type, and it also doesn't look like cells without a variant. And in that case, that would be a change of function, something like a gain of function or dominant negative or other sorts of interesting things. So this type of approach has been used successfully before in an arrayed format using gene expression as a readout of the similarity in the cell states. So what you would do is that you would get the gene expression profile of cells that overexpress the wild type and then compare the profile of overexpressed variants and see if it's the same. You're essentially asking, like, does the variant deviate from what you would expect when you overexpress the wild type? Then it's impactful, or it doesn't. So this was called eVIP for expression-based variant impact, and indeed it was used across multiple different cancer genes and variants in them quite successfully. However, because it was sort of requiring an arrayed format, it was difficult to sort of scale up. And one of the main contributions of the work that I'm presenting now is doing this type of scale-up using pooled single-cell screens, which in expression readout. And so for this, we use the technology that's similar to to perturb-seq, but with the caveat that in perturb-seq you usually knock out a gene and then see what happens to cells. Here we are overexpressing. So what we do is we start with the synthesized variant sequences, and each different basically coding sequence of this gene has a variant in it, and each variant is associated with a different barcode. At the end. And so then we transduce these into cells such that each gets one. We select for the cells that got a variant and then do single-cell RNA sequencing. So that gives us for each cell its expression profile as well as the identity of which variant was present in which cell. And actually the variants are expressed at large enough amounts that you don't even need to enrich for them in this process. You just kind of quantify them as if they were any other gene. This addresses some of the considerations I mentioned in the beginning. In particular, because you can do this in a pool, you can have a pool of different variant coding sequences, it's much easier to scale up while maintaining really good like really low batch effects and things like that. It's also gene agnostic, so as long as you can measure any difference in gene expression between the wild type overexpressed and the variant overexpressed, this should work. And the third point that we're going to get into more detail is that actually it's a very information-rich readout because we're not only finding out whether the variant looks the same as the wild type. But actually, if it looks different, we can look at exactly the genes that are changing and further group the variants that seem to deviate from the wild type into different categories of impact. For example, you could have a gain of function, you could have other types of changes of function, and so on. Cool. So in this In this work, we started with a proof-of-concept experiment looking at 200 coding variants in the p53 and KRAS genes. And we chose these genes because they were genes that, first of all, were important for cancer, and second of all, they had other types of functional assays associated with them such that we could benchmark our technology. And for each of these genes, we We chose the 75 most frequently occurring missense mutations in cancer. So this is across basically cancer cohorts. And we chose the most frequent ones because being enriched in a cancer cohort is indicative a lot of times of sort of impactful genetic variation. And then we had 25 controls. 15 were synonymous variants, so those shouldn't change the the sequence of the protein. And then 10 of them were missense mutations that were not observed in cancer, but they were observed in just sort of healthy people in the ExAC Consortium. And this was done in the A549 cell line. This is a lung cancer cell line, which again had also been previously shown to be a good biosensor in terms of p53 and KRAS. And of course, future plans actually include looking at these variants across multiple different contexts to look for any effects that depend on specific genetic backgrounds or tissues and so on. And because we didn't know the effect sizes of these variants, we actually had quite a few cells per variant, 1,000 cells for each variant, and we could then do subsampling and see how many cells you actually need to to accurately measure effects. So let me tell you what we found about p53. First, I'll walk you through some of the quality control metrics. One important one was to check whether the expression of the variant deviates a lot from the overexpression of the wild type, because then if that were the case, then it would be hard to basically distinguish whether the effects we're seeing are due to just expression of the construct versus actually the effect of the variant. And for P53, for most of them, except for 2 variants, so most of the variants were within a 1.5-fold change from the wild type. So there was quite uniform expression of the variant. And moreover, then we looked at how good are our assessments of which cell got which variant. So we looked in each cell how many UMIs were supporting its assignment to a particular variant, and in more than 70% of cases we had at least 2 UMIs that supported the same variant. So we're confident in our assignment of cells to variants. Cool. So then the next question was how to actually measure this distance between a variant and the wild type overexpression, because this is really our measure of the impact of the variant, of how much it's actually doing something else. And to do this, we basically took the cells that were overexpressing the wild type version of the gene and the cells that were overexpressing the mutant, and define basically a distance metric between these two groups that would be proportional to the effect size of the variant. And we call this the single-cell EVIP score. And we tried many things including correlate, just simple correlations. What worked most robustly was this Hotelling t-square statistic on the principal component space. So essentially this is like a t-test but using multiple dimensions. So it would be the same as basically testing each individual gene, whether it's differentially expressed between the two groups, but now we're testing each principal component together into one score. And then we can shuffle the assignments of cells to variants and compute an empirical FDR to identify a threshold score for a 5% FDR. And so when we look at these scores, we can compare them with our controls, which we expect to not have a large effect. And indeed, all the synonymous variants fall below this 5% FDR line. And also for p53, all of these exact variants that occur in just general healthy cohorts also fell below, so they looked wild-type-like. So we think they're neutral. On the other hand, we had some positive controls that were hotspots, that were occurring at hotspots in p53 that tend to be particularly enriched in cancer patients, and those had high and significant scores. And then all the other variants in our dataset also, in the majority of cases, had significant scores. And that makes sense because in p53 most variants would be expected to be loss of function, and I'll show you that that's indeed what we see. So this is on a global level to just estimate whether a variant is at all different from the wild type version. However, as I mentioned, we actually have these rich gene expression profiles for each variant, so we can actually cluster them and see how these variants group together. So this is a Spearman correlation matrix between the average expression profiles for each variant. So each row is a variant, each column is a variant, and we're essentially looking at the correlation between variants. And there's two large groups here. The first one contains basically all of our controls. And we annotate it as this sort of black-colored cluster. And then the rest of the variants kind of fall into the other cluster, which also further subclusters into two, and those would be the impactful ones that are likely to be loss of function, which we call impactful 1 and impactful 2. And we can compare these to signatures, so to basically to genes that are known to be induced when you overexpress p53. And indeed, in the controls where you're overexpressing a sort of functioning version of p53, you see induction of these genes like CDKN1A, and you actually see either no change or a decrease in this gene for the rest, and similar for other such signatures.. And so what actually happens at the single-cell level for these variants? So we can look, we can do a low-dimensional representation of genes in this UMAP space where cells that are more similar to each other are going to be closer together in this space. And then we can, so basically we look at the density of cells in this space. And in gray we have the wild-type-like overexpressing cells and in blue the the ones with these impactful p53 variants. When we compare this to the distribution of cell cycle phases across these cells, we actually see that where these two groups differ is in the proportion of non-cycling cells. That makes a lot of sense because overexpressing p53, like a functioning version of it, like in the wild-type, would lead to cell cycle arrest. You would expect to have an increased amount of non-cycling cells, which is what we see. So even at the single-cell level, we can sort of start to understand the types of molecular mechanisms that account for the differences we see between variants in p53. And finally, we also looked at how our scores compare to specialized p53 functional assays, which are done on cells treated with different DNA damaging agents and looking at survival of cells across these different conditions. And we find a good agreement with those. There's like 3 different ones. And actually, we find the largest agreement with one of them, with treatment with NUTLIN3 in a p53 wild-type background, which measures dominant negative alleles. So those are variants in p53 that not only are loss of function, but actually they sort of interrupt the function of even the endogenous p53 that was expressed in the cell. And that makes sense because actually our cell line has the wild-type p53, which means it would allow us to measure these types of effects. So that's about p53. Now I'll go to KRAS, which turned out to be a lot more interesting because we ended up studying a lot of variants that are actually quite rare and have not been studied as extensively as the 53 ones before. And again, we did all these quality controls. And again, there were 2 variants that were expressed at a little bit lower levels than the wild type overexpressed. But altogether, we were able to distinguish the control variants from known impactful ones at hotspot locations, and we had good confidence in our variant-to-cell assignments. So now I want to actually show you what the gene expression profiles look like for each of these variants. So each row in this matrix is one of the KRAS variants in our dataset, and each column is a gene that we measured. And what I'm showing you in this matrix is the average expression of each gene in each variant. And so based on this, we can essentially cluster the variants across the genes. And again, we end up with a few different categories of variants. So the black cluster is the wild type-like, that also includes just the overexpression of the wild type. And then there's 4 other groups. And the red one, the impactful 4, contains most of the sort of known gain-of-function KRAS variants. But there are also these other groups in between that we kind of studied further. And again, we compared with cellular assays that measure KRAS functionality, in particular growth and low attachment, which is a proxy for KRAS gain-of-function. Indeed, the largest effects were in the gain-of-function group based on our data. So we found good agreement with functional assays. The really interesting thing in this dataset was that these categories that we defined ended up being much more of a continuum rather than just discrete categories. You can see it here a little bit. You can see it even better if you look at, for example, principal component 3. So you do principal components on the cells and then look at the distribution of principal component 3 scores across cells of each variant, and there looks to be this sort of continuum. And actually, this variation in PC3 quantitatively correlated with this growth in low attachment assay that was measuring KRAS functionality. Suggesting that we're seeing sort of a gradient of how much gain-of-function there is in KRAS rather than just a discrete— it's gain-of-function or not gain-of-function. And this couldn't be explained by the— just sort of the expression of the construct in each cell. And we tried to think about how this works at the single-cell level. So you could think of the cell state space kind of going from cells that are very wild-type-like to very gain-of-function. And then wild-type cells are going to be enriched in those wild-type-like sort of black-gray cell states. And then you might have like a weak gain-of-function that has more pink but also some black and some red, and then sort of increase the proportion of cells that are strongly gain-of-function. And indeed, that's actually what we see happening in our data. So if again, if we take the cells and embed them into this lower-dimensional representation, there is this cell state 0, one cluster of cells that is particularly enriched in gain-of-function KRAS variants. And we can really observe a continuum of variants from from sort of the left of this UMAP to the right. So I'll show you that here. So in black, there's an example of a wild type overexpressing set of cells, and in dark red is a known gain of function. And then this particular variant in pink is a synonymous variant, so it doesn't overlap with the gain of function very much. But as we go along our continuum, there's cells that are more and more concentrated on the gain-of-function, and they sort of span this space quite continuously. And indeed, the proportion of cells in this cluster is what correlates with the growth and low attachment. So this suggests basically that, at least in this cellular system, KRAS variants might be falling more along a continuum rather than just the impactful and not impactful variants. And of course, much more work needs to be done to study how much this is actually happening in patient samples and so on. It's kind of fun to think about. And finally, where this type of approach like I described here would be most relevant is for variants that are rare, as I mentioned in the beginning. And that's because traditionally the way variants are sort of assigned driver and passenger roles is based on how often they appear in cancer cohorts. And that makes a lot of sense because if a variant is enriched in patients, it's probably because it has an effect. However, there are some variants that occur rarely in patients, and those are the ones where knowing how often they occur is not particularly informative and where doing an assay like this is particularly helpful. And so I'm showing you some examples here. All these red variants that have large effects with our approach but just occur rarely in patients would have been missed if just looking at their frequency. So some of them are rare because they involve multiple substitutions. So it's just harder to create these mutations in a patient in the first place, regardless of their effect. And then there's other ones that are rare for unknown reasons, but are deemed to be impactful in our assay. And some of them occur at the same positions as known gain-of-function variants. So we're quite confident that they're actually gain-of-function. And then, of course, we did subsampling experiments to kind of study how many cells you need per variant. And I'll make this quick, but for the highest impact variants, you need about 20 cells. And then for the lowest ones, you need from 100 to 400, depending on the variants. So for example, with a million cells, we could have profiled all of the possible PTEN53 or all of the possible KRAS variants. And with 10 million cells, we could profile all the 1KB cancer genes, those 57 genes that are actionable. So in summary, I've told you about measuring variant impact using a single-cell gene expression and how this allows us to scale to many variants across multiple genes. I've told you about proof of concept experiments in KRAS and p53. And how this allowed us to group variants into different types of groups of functionality, and these agree well with labor-intensive gene-specific functional assays. And our approach allowed us to learn about shifts in cell states that might be underlying the functionality of these variants. And of course, there's a lot of exciting future directions. One of them is to look across multiple contexts, to look at variants whose role depends on the tissue or on the genetic makeup. I've also mentioned other modalities, and the most exciting part is to try to engineer these variants at their endogenous locus with base editing. So with that, I'll acknowledge again this wonderful team that I've had the pleasure to work with and I'll take more questions.

Hey, great. That was a super cool talk, and we already have a couple questions that have begun to accumulate. So if anyone has a question now, you can feel free to type it into the Q&A or the chat, and we'll ask it to Oana on your behalf. So we'll start with the first question, where Jason Ian Commander has asked, um, genes that halt cell cycle will have relatively large effects on expression, and he's wondering if you've tried genes that cause more subtle phenotypes.

Yeah, so we— I mean, obviously in this example we just, we studied p53 and KRAS, and so in KRAS the cell cycle changes were not as good at distinguishing between the variants and the wild type-like. In the previous work where this was done in an arrayed format, they did more genes and it worked well there. But yeah, I think it will depend on how transient the effect is and sort of how much it translates into a transcriptional output in the end. But because, for example, KRAS wasn't just a cell cycle effect like p53, We think this might work quite well, but obviously we need to study more genes and see.

Sure. And then they've also asked, if you were to start with a new gene, how would you check if the control variant perturbation is big enough to give a robust enough transcriptional signal? And maybe you can speak a bit— they've asked if you have any practical tips if someone's trying to start to boot up this assay on a new gene?

Yeah, so I think basically you need two things. You need to have like the wild type sequence and you need to overexpress a mutant one that you just know should be different from wild type. Those will be the controls. So then a pilot experiment would be just doing those two conditions and seeing if there's a significant difference transcriptionally between them. That's where I would start. I think, I guess the practical tips would be that there's still sort of a size limit right now on how much you can put in the construct, around like 300— 3 KB. So it would work much better on shorter genes. Yeah, and then maybe ask JT for more tips.

Yeah. Another question that I might have is, so you had mentioned that there's obvious— like, it might be improved to start to edit the endogenous locus rather than using these exogenous overexpression systems. Did you see anything in your data that suggested that some of your variants you might have been interpreting incorrectly as an artifact of overexpression? Like, how valuable do you think it would be to switch to the endogenous locus? That might address the size constraint you mentioned, but did you see anything that seemed to actually be an issue with the overexpression?

So we didn't see anything that would be an issue, but I think doing it at the endogenous locus would make it so that the actual gene expression changes you see would be meaningful. Because right now, this is an artificial system where you're overexpressing the variant and whatever gene expression changes that occur, they're not necessarily what the variant would do at its endogenous levels. I think you would just have much better interpretation of what happens biologically first. Second, I wonder how much of what we're seeing in our KRAS experiments is actually also a interaction effect with the endogenous gene because the cell line that we used, A549, actually already has a gain-of-function KRAS mutation in it. It's quite magical that we can read out these additional gain-of-function variants. We've been wondering about whether it's those two different variants in the cell competing with each other or basically what you would get in a KRAS wild-type background.

That was surprising when you had It didn't even dawn upon me that since you're overexpressing these in another cell line, you would have— you could theoretically have a wild-type allele still expressed there. I was wondering, when you talk about a loss-of-function variant, do you not anticipate having a problem where you might want to maybe knock out that gene in the endogenous locus before running these assays?

Yeah, you might want to do that. I think that's why when we were comparing our things with the p53 functional assays, some were done in the p53 wild type and some in the p53 null, exactly because those assays would be able to distinguish between just loss of function or loss of function and messing with the endogenous p53. And yeah, we correlated the best with the ones that were done in the p53 wild type because that's what our cell line was. So we're just better at picking up that. So yeah, it sounds like that might be a good idea to try to delete the endogenous gene. I just don't know how stable that would be, things like that. You might just get some cells that didn't have a deletion and outgrow the other one.

Right. And then you focused almost exclusively on missense variants. You didn't include any nonsense variances that because of the system you were using, so like the overexpression of a nonsense variant will obviously be different than the impact of a nonsense variant in its endogenous locus. Was there any reasoning why you focused— or like you just— the assumption is that it's going to be deleterious if it's a nonsense?

Yeah, I think so. I think it was more of maybe a priority of the variants to study, and maybe you're more confident that things are a loss when they're nonsense. But yeah, I mean, that would be another interesting thing to do.

Great.

Okay.

So I don't see any more questions coming up. So I think that we can end it there. So thank you again for taking your time. It was a really interesting talk, and I think a lot of people are going to have a lot more thoughts about it after they've maybe read the paper to go along with it. And to remind everybody, the talk has been recorded and it will be posted on our YouTube channel later, so if you want to follow up, you can watch it there. And then as a reminder, we will be having our next seminar series on March 1st, so in about a month from today. So thank you again.

Yeah, thank you for having me.

Great talk.

Thank you.