Scientist Stories: Lea Starita, Understanding the functional effects of coding variation, at scale

Thanks, Malta. Thanks for inviting me to be here today. And as Malta just told you, I'm from the Genome Sciences Department in the University of Washington, but I'm also part of this new venture in Seattle, the Broadman Beatty Institute for Precision Medicine. And as you all know, beyond the absolute crap American healthcare system, one of the main obstacles to actually practicing precision medicine is the fact that we are much better at reading the human genome than than actually understanding what we find after we sequence it. And nowhere is this more true than for missense variation. And so to drive this point home, I've made this little plot of the ACMG 59. Now these 59 genes are special because they've been deemed actionable, right? So if a pathogenic variant is found in one of these genes, a physician can actually do something about it to mitigate the disease risks. So for example, BRCA1 or 2 mutation and prophylactic mastectomy and oophorectomy. And the more we sequence, the more we find, right? So in the United States, it's actually all genetic testing pretty much is done by private companies, so it's kind of hard to figure out how much testing is going on for each gene. But each one of those companies has to register the genetic tests with the NIH. So I just used the number of tests registered with the NIH to figure out kind of how much testing was going on. And that's here on the— where did it go— on the x-axis. And on the y-axis, what I did was I just scraped up ClinVar. Okay. Okay. So I scraped ClinVar in order to find the total missense variants that were— that had been found by clinical sequencing. And ClinVar is just a kind of a repository for variants that have been found by genetic testing. And what I did was I just— and then the size of these dots are the number of BUS. And I've animated this to show what happens over time, right? So, the more we sequence, and I think the thought was that the more we sequence, we would really start to understand the effects of these mutations, right? We would have big cohorts where we would have cases and controls, and therefore we would understand, you know, which variant was associated with disease and which ones weren't. But that's really not what's happening, right? As there's been this explosion in genetic testing because sequencing as sequencing has gotten cheaper, we're actually finding more variants of uncertain significance. And if I stop this at 2019 and I put labels on some of these, you'll notice that these are pretty famous genes, right? These are extremely well studied. So BRCA1 and 2, we've got the mismatch repair genes, APC. You'll notice, you know, these are extremely well studied. They have a lot of different structures and functions. But they still contain all of these variants of uncertain significance, right? So these are not genes for which we don't know the effect, but these variants of uncertain significance means that we have no idea the effect of the biochemistry of the protein or splicing for most of these extraordinarily well-studied genes, right? So if we can't get it right for these guys, you know, how are we going to get it right for genes that we really don't know as much about? And there are many strategies for gathering evidence to reclassify these genes as either pathogenic or— these variants as either pathogenic or benign. The gold standard is co-segregation in families, right? Does the disease co-segregate with the mutation in families? But, you know, these are accumulating at such a rapid rate that it's really difficult to do this, even though, you know, the There's been kind of modern twists on gathering these pedigrees through social media and direct-to-consumer genetics. And we also live in this amazing era where hundreds of thousands, and maybe, you know, as Daniel MacArthur would say, we're getting toward a million genomes. They've been aggregated and they're shared publicly. And these, the minor allele frequencies from these population screening methods or population genomes gives us evidence that, you know, maybe some of these variants are occurring at too high of a frequency to be causative for disease, right? And so that gives us an idea of which one of these might be benign, but it still really doesn't give us much information about these super rare variants, and most of these variants of uncertain significance are actually really rare. Okay, and a third way we could go after this is by functional assays, right? And this has been used to great effect for a long time. You can make, you know, your protein of interest and do an in vitro functional assay. You can even make a mouse, right, and get at, you know, the entire organismal level of what's happening with these variants. And, you know, we can use them to say that, you know, these are functional and maybe this one's not. But how are we going to do this at scale, right? So again, going back to this plot, this is just 59 genes, 59 fairly important genes, but it represents 23,000 variants of uncertain significance, right? And so that's a lot of postdoc hours. And, you know, these genes are also— they represent, you know, a big swath. There's a lot of different protein folds in here, a lot of different functions. We've got ion channels, tumor suppressors, pieces of the sarcomere, as Karen Musinuru showed us yesterday. And so how are we going to really assess these phenotypes? So in order for functional assays to be a way forward, we've got to be able to assess a multitude of phenotypes at an unprecedented scale. So today I'm going to tell you about one of our solutions to this, and this is multiplex functional assays. I'm going to show you two examples, one for BRCA1 and one for PTEN. And I'm going to talk a little bit about the benefits and limitations of using these types— this type of data for variant interpretation, and I want to leave you guys with some future challenges. Okay. So multiplex functional assays, like Erin showed us, test thousands of variants simultaneously. And although they can be different for different types of elements, like Erin was showing, promoters, enhancers, or protein coding genes, they all consist of a few key steps, right? So we have to make a library of all possible variants. And again, we try to make every possible single nucleotide variant or every possible amino acid change. These are all done in a pool. They're never separated. And then we have to get these variants into cells. And today I'm going to show you two different examples, one of putting transgenes, so cDNAs, into cells, and another one where we edit these variants right into the genome. And next we need a functional assay, and I've left this purposely as a black box because the functional assay really only depends on your imagination and possibly your budget. So as Aaron showed you, and I'm going to show you again, you can use cell survival, right? So this would be— but the key to these assays is that you can separate the functional from nonfunctional variants. So survival is one of these types of assays. Reporter assays where you can use fluorescence-activated cell sorting in order to separate functional from nonfunctional variants using a reporter. And what we're going to try to do next, which hopefully I'll have data for in the next couple of years, is kind of hooking these up to single-cell assays. To get more information from these variant effects. And because the genotype is connected to the phenotype, we can use— we can harness the power of next-generation sequencing to identify and count each of these variants before and after selection. So for example, in a survival assay, we would amplify the variable region from the before and after population and identify and sequence and count these variants using sequencing. And then the math is actually pretty easy. We really are only just looking at the ratio of the abundance of that variant in the selected population over the initial population in order to look for variants that are becoming depleted or enriching in these populations, depending on their phenotype. And then we try to look at all this information in sequence function maps, and I'm going to show you one that looks like this today, and I'm going to show you other ones that look just like Aaron's. So I thank him for introducing those to you. Okay. So this is an example of a multiplex functional assay for BRCA1. So as nobody will be surprised, there are a lot of variants of uncertain significance in BRCA1, right? This is one of the most tested for genes, therefore it's got some of the most variants of uncertain significance. But what you might not know is that there's actually more variants of uncertain significance significance in this gene than variants for which there is no effect, right? And there's about 2,000 of them. So we— and it's really important to know for these, right? Because there's been, you know, there's real-world problems with this, right? There's been surveys of surgeons in the United States— again, crappy healthcare system, I know— but some of them will actually offer surgery to women with variants of uncertain significance when the recommendation is that they don't do anything at all. Okay, and I want to start by with the acknowledgments here. So this project was led by an extraordinarily talented graduate student named Greg Finley, who's actually looking for a job, and two amazing technicians, Riza Daza and Beth Martin. And it wasn't actually done by, you know, the rest of these 30 people in Jay Shendure's lab, So, you know, these assays are so powerful that basically 3 people can do them in a year. And this is Malta right here in the middle. Okay. So you need to know just 2 things about genome editing for this. And as you know, Cas9 is an RNA-guided nuclease. So you load a little guide RNA onto the Cas9, and you can basically cut anywhere in the genome. And taking advantage of non-homologous end joining. So if you cut the genome and don't give the cells a repair template, you wind up with an insertion or deletion because of non-homologous end joining. And this has been used to great effect to make knockouts, right, to make indels in exons. And the other, and if you want to make a precise edit, what we do is we give the cells a repair template, which is basically a homologous sequence with a specific change in order to make, you know, just a single base pair change. And I'm going to show you an example of both of those in the next couple of slides. Okay, so this whole thing started when Greg noticed that in haploid— in screens for essential genes in haploids, BRCA1 and other members of the homology-directed repair pathway were coming up as essential. And you can imagine that your double-strand break repair pathway might be fairly essential in a haploid genome. And so that was all done by gene traps, and so we wanted to know if targeting BRCA1 using Cas9, if we could actually kill the cells as well. And here you can see that when we express Cas9 and a guide to exon 17, that we just kind of wipe out all the cells. Cells, whereas if we express a guide to a nonessential gene, in this case HPRT1, really nothing happens. So we figured we had a functional assay, right? Maybe we could use the survival of these cells to figure out which variants in BRCA1 were functional and which ones aren't. Okay, so again, we have to get all the variants into the genome, and we're going to do this by a method called saturation genome editing that Greg actually developed. In his rotation, which is kind of insane. Okay, so here we express Cas9 and a guide RNA to an exon of interest. In this case, I'm showing you exon 18, and that Cas9 causes a double-strand break. We also give these cells a library of repair templates, right? So these are plasmids with single nucleotide variants on them and big homology arms. As these repair templates get incorporated into the genome, we wind up with a pool of cells that all have a different variant of BRCA1 edited into their genome. The nice thing about editing is that we can look at the effect of the coding sequence, but we can also look at the effect of variants that are kind of a little bit into the introns, so we can get an idea of what's going on with splicing as well. And this has not been the case for most of these types of assays in the past. OK, so now we just basically— we've edited these cells, and now we just wait to see which variants have become depleted from the population. And then we have to sequence— we sequence the library to see which variants we could have made. We sequence the cells at day 5— or we sequence the variable region at day 5. To see which variants got edited into the genome successfully. And we sequenced them again on day 11 to see which variants had become depleted. And again, because we used editing, we can look at the effects of these single nucleotide variants on RNA expression. And so we sequenced the RNA as well as the DNA on day 5. And to get an expression score or to figure out the effect of these variants on RNA, So all we do is just measure the abundance of the variant in the RNA over its abundance in the DNA on day 5. And to get an overall functional score, so this is protein function and splicing, we just look at the ratio of the variant on day 11 to day 5. It's a little bit more complicated than that, but that's the gist. Okay, and we have to do this, we had to do this 13 times. So we made repair template libraries and guide RNA constructs, for each of the 13 exons that we targeted. So BRCA1, even though it's this really massive gene, really only has 2 functional domains, and so we wanted to target those, and it's also where the controls are, right? And so we aimed to make 4,000 possible single nucleotide variants, and we wound up with scores in the end after doing this for 3,958, so we almost got all the way there. And you'll notice that these scores are fairly bimodal with wild-type-like or most synonymous variants here scoring around 0, and then the nonfunctional variants scoring around -2. And for the most part, categories of variants scored as we would have expected, right? So nonsense codons are bad for function, and, you know, this, when you're doing these types of things, is a nice sanity check, right? And next, synonymous variants. These are variants that don't change the coding sequence but not the amino acid sequence. And most of the time they're functional, and about 1% or a little bit over 1% cause enough damage to splicing where we actually saw them drop out of the population. And this is interesting because most testing companies don't even report synonymous variants. Canonical splice variants, these are the two essential nucleotides on either side of the exons, and again, most of the time those are damaging, except when they aren't. And then going into the introns, so this is more than 8 bases into the introns, you can see that, you know, that those variants are usually not, are are usually functional, again, unless they're not, right? So we're kind of getting to why it's really hard to predict which of these variants are going to cause defects or damage without actually doing these experiments. If we get closer to these canonical splice sites again, the scores become bimodal. And then the missense variants are kind of completely bimodal with about 20% in these really essential regions of BRCA1 causing enough damage to become depleted from the population. Okay, but what you really want to know probably is how well we do versus the known pathogenic and benign variants that were already in ClinVar. And so what these little lines are here are how we classified these variants are functional or not, and we actually didn't use the ClinVar variants, we used the synonymous whether or not the variants overlap the synonymous population, and we called those functional, and the variants that overlapped the nonsense population were the ones that we called nonfunctional. And here we're just mapping these variants from ClinVar on top of that. So for the likely pathogenic and pathogenic, we get 206 out of 209 of them score correctly, and for the likely benign and benign, I think we get 163 out of 166. And this is an AUC here of about 0.98, and we— it's about 98% specific and sensitive. And that's important because we have to know the sensitivity and specificity of these kinds of assays in order to be able to use them as evidence in the clinic, okay? And it turns out that we can use this as strong evidence if we calculate the odds of pathogenicity. And applying these scores to the variants that were— variants of uncertain significance that were in ClinVar at the time we published this, we're actually able to separate 90% of them into either the functional or nonfunctional categories. So these guys, these 91 in the nonfunctional categories, given the sensitivity of this assay, very well may have enough damage in order to be pathogenic. Okay, so these are all the variants that are in ClinVar here on the bottom, and I think this represents the most powerful point or part of these multiplex assays, right? So we have another 3,000 scores that are for variants that have not yet been seen in the clinic, but could basically show up at any time, right? And, you know, as we were writing this paper, when we first downloaded ClinVar, ClinVar, another— there was an extra 100 variants of uncertain significance from the time we started this and the time we finished it, right? So this is important because, you know, if we did these one at a time, you're kind of chasing these variants of uncertain significance as they pop up, but here we kind of have comprehensive scores for all of them. And again, this is a kind of a different kind of sequence function map. So I kind of just showed you all the scores in aggregate. But this is them in a more granular form, okay? And if I blow this up, you can see the way we did this was every column is a nucleotide position, and every row is either the wild-type base or the 3 single nucleotide variants. And the boxes are colored by the category of variant, right? So green is missense, and orange is these canonical splice sites, so you can use them to tell where the ends of the exons are. And if they're filled in, they're nonfunctional. And if they had red stripes in them, it means that they're nonfunctional and the RNA was depleted as well. So here we have not only— we have a granular map of the function of BRCA1 with the mechanism by which it was depleted. So this is, I think, pretty powerful tool. And again, we have this for all 13 exons, and you can find all this information on our server, sge.gs.washington.edu. Okay. So I want to tell you about another gene we did this for, and that's another tumor suppressor called PTEN, right? So PTEN, when It's a dominant mutation found— when it's a dominant variant, a pathogenic variant found in the human genome in the germline, it can cause Cowden syndrome, which is, you know, a cancer syndrome, and it's also a frequent driver of sporadic cancers as well. And so we started this course of experiments by thinking about whether all proteins were special, right? So one of the most expensive and time-consuming parts of developing multiplex functional assays is actually developing a new assay for every protein, right? So we don't want to do that. So we're kind of, we were thinking of, well, what are certain protein properties that all proteins have to have in order to be functional? And first I'm going to start with acknowledgements. I did this in collaboration with Doug Fowler's lab, also at UW, Kenny Matrejek, drove a lot of the PTEN work, and of course Beth Martin and Martin Kersher helped as well. Okay. So, we posited that all proteins need to be correctly folded and stable and, you know, at a certain abundance in order to function well in the cell. When variants cause missense mutations that cause the protein to become unstable, The protein usually decreases in abundance because the cellular machinery kind of chops up that protein, and the proteasome will destroy it. And so in order to look at this, we did the dumbest thing in the world. We just fused them to GFP. And so when we fuse a wild-type protein to GFP, there's a lot of it around, the cells are bright green, and we can And you can see by flow cytometry that there is a lot of this protein and the cells are bright green. However, when we fuse GFP to an unstable protein, the protein itself and the GFP get degraded and the cells become dim or dark, and you— and by flow cytometry we're able to separate unstable from a wild-type-like protein. Here, stability is a known factor in whether or not PTEN is functional, and there are two known unstabilized or destabilized PTEN variants, and by flow cytometry, we can see that we're able to separate those two variants from the wild type. And so we thought, again, well, maybe this is a good functional assay. We'll be able to use fluorescence-activated cell sorting to separate stable from unstable proteins. Okay, so we made a library of all possible amino acid swaps of PTEN, and it was fused to GFP with a cherry, a red fluorescence protein, as a reporter of expression. And we had to get these into cells, and so we recombined these. We have to get one and only one of these copies into the cell, and so we use a recombinase landing pad system to get these transgenes into the cells. And so we made about 4,000 variants of PTEN, and you can see by this flow cytometry plot that the wild-type protein is— this is the same one I showed you before. The wild-type protein is bright green. The unstable proteins are dark, but the library of 4,000 missense variants kind of spans that whole range. And using fluorescence-activated cell sorting, we separate the brightest variants into bin 4, the dimmest variants into bin 1, and the intermediate variants into bins 2 and 3. And then by high-throughput sequencing, we can count and identify each of these variants, and using the same type of math that Aaron showed you before, we just take a weighted average of where the variant is in these fax plots. And we call this variant abundance by massively parallel sequencing, or BAMSeq, and these are the scores for these variants, and it looks very similar to the fax plot you saw before. Most of the nonsense variants— or we set the nonsense variants to 0. We sent the synonymous variants to 1, and you can see that the missense variants kind of range across both of those distributions. So this is the sequence function map for PTEN. Every residue in PTEN is along the x-axis, and every amino acid change is on the y. We made about half of all possible variants in PTEN, and there were bottlenecks in some of this library construction, and we can talk about those sorts of things later. But you can see in some cases, there really was no effect of mutating P10, and in these two regions, they're actually not even in the crystal structure, they're too floppy, and you can see that that doesn't— these amino acid positions are very tolerant to mutation, whereas in these phosphatase and folded domains, they were very fairly sensitive to mutation, which makes a lot of sense. So what we did was we made functional classes to describe these abundance phenotypes from wild-type-like all the way down to low abundance, given their distributions across these categories. And then we wanted to— and then we also, again, because in human genetics we really only look at single nucleotide variants, we took out the 1,300 of these that were possible by a single nucleotide change. And this is the distribution of what they all looked like, right? So we've got low abundance variants, high abundance variants, and those in between. And we wanted to look at the sensitivity of the assay, but as it turns out, there are no benign variants of PTEN in ClinVar. And this is because one of the reasons that this there's just not a lot of variation at all in PTEN. It's a very important gene apparently. But these three variants are in NOMAD at a higher frequency than they could be causative for Cowden syndrome. So we kind of took these as benign-like variants, and those are all stable, which was nice. But looking at the variants that were pathogenic in ClinVar, 60% of them were in this low abundance phenotype. Another 20% were impossibly low. So, we were able to capture 80% of the pathogenic variants just by looking at protein stability alone, suggesting that if we looked at another gene, we might be able to capture, you know, a good percentage of the pathogenic variants just by looking at the stability of the protein. And again, because of this, so we were able to look at the pathogenic variants, But there's, you know, obviously 20% of the variants we didn't find, and these were actually in the active site of the protein, and we didn't look at the protein activity, right? So this type of assay has a very good positive predictive value, or a good predictive value for pathogenicity, but basically almost none at all for benign, for looking for benign variants. Okay, so for variants of uncertain significance then, we would really only look at the variants that were unstable, And these— this class of variants right here, these variants might actually have sustained enough damage to be unstable enough to actually be pathogenic possibly if found in a human, or they were found in a human if we reinterpreted these variants. Okay, so to show that this assay was actually generalizable, we applied it to another gene. So TPMT is an important pharmacogene. It metabolizes thiopurine drugs, and these drugs are actually given to children with leukemia, and they have a high risk of overdose, which causes hepatotoxicity and actually bone marrow failure. And so we wanted to look at this gene. You can see that more of these TPMT variants were stable, and And that's, we think, because this protein is just more well-folded than PTEN to begin with. And you can see that these— again, this is a sequence function map, so this is all 250 positions in TPMT and all the possible amino acid changes. I think we covered about 80% of this protein because it's shorter. And these VAMP-seq scores correlated really well with the average dose intensity that these kids were able to tolerate. So it looks like, you know, there's not pathogenic and benign in pharmacogenes, but we're able to correlate this with the activity of, or the ability of these kids to tolerate the drug. Okay, and we call this VAMP-seq. Again, I am like really proud of this logo, and all of these scores are available at abundance.gs.washington.edu. At stanfordmedicine.edu. Okay. All right. So I've shown you two of these assays. We think they're pretty powerful, but there are benefits and limitations for using these for variant interpretation. So first, the first benefit, and, you know, I think this is fairly obvious, but I think it can't be overstated, and that's all of the data is in the same place, right? There's no hunting through the literature for specific variants. Which also means that there's no lab-specific batch effect, which is really nice. And it's also comprehensive coverage, right? We actually might have scores for these variants before they're seen in the clinic, right? These can be like a prospective lookup table for understanding or getting a first idea of what might be going on with these variants. The next is that these assays are internally controlled, right? These stop codons and these synonymous variants, can tell you whether or not an assay like this is actually working, and this is something that somebody could look at when they were trying to use these scores to reassess a variant. And another thing is that they're unbiased, right? So we didn't pick a set of pathogenic variants or a set of benign variants to look at the function. We did all of them, right? So we didn't, you know, we don't have some selection bias for which variants we chose. They're also continuous and quantitative scores, right? So right now I think people think about pathogenicity and benignity, which isn't a word, as binary, right? But there's all sorts of different penetrance and things like that. And right now I don't think we actually even for genes like BRCA1 and 2, there's some anecdotal evidence of low penetrance missense variants, but there really just aren't enough to compare to this kind of data. And I think these functional assays could really start getting at or calibrating some of these low penetrance variants eventually when we get there. And if you have enough gold standard controls, as we did for BRCA1, you can really well calculate specificity and sensitivity or the negative and positive predictive value, which is what you would want in order to reinterpret variants. And this brings us to limitations because beyond BRCA1 and 2, and as I showed you for PTEN, in most cases these big gold standard sets in order to compare these data just doesn't exist, right? And we actually started with BRCA1 because because that's where the controls were, which is, you know, kind of bad, right? So when we get beyond BRCA1 and 2, I think we're going to have really difficult time actually telling whether or not our assays are terribly predictive just because we're lacking the controls. Another is that the data can be noisy, right? So again, these experiments can be a little bit expensive, but as the price of sequencing Multiplexing has gone down. We've been able— a lot of these labs are doing a lot more replicates. We can report error estimates for all of these scores, and I think that's going to become even more important going forward. And again, the inference distance from the function queried to molecular function underlying disease, right? We're using really oversimplified assays that are amenable to multiplexing and high throughput, And, you know, green or not green or dead or alive, you know, crazy haploid cell line are not terribly informative phenotypes, right? But they're good enough and they're predictive enough. But that leads us again to future challenges where the scale and phenotypic complexity of these types of assays are kind of the central challenge, right? So there are many existing technologies that are low scale and high complexity. Complexity, right? If you make a humanized variant in a mouse, you can tell its effect on all organ systems over differentiation and through time. However, you're really not going to do a lot of those at once, and I can't see actually doing this in a mouse. Other assays like the massively parallel reporter assays and deep mutational scanning are very high scale, right? We can look at thousands or even hundreds of thousands of variants, But the phenotype is really, really simple, right? Is the gene on or off? Is the cell alive or dead? Is the protein green or not? And what we're trying to do is push these assays into this space here, right? And we think we did that a little bit with saturation genome editing, because at least we could look at splicing and protein function using the same assay. But I think it— we'll have to see what the future holds for these sorts of assays, right? Can we get these into multicellular organisms? Can we maybe start using iPSC cells to get at more complex phenotypes? And then again, how will these be incorporated into clinical variant interpretation workflows? And I think that still actually remains an open question. So thank you, um, for inviting me, and I'm happy to take questions.