Scientist Stories: Jolanda Van Leeuwen, Analysis of genetic interactions from yeast to human

Yes, there we go. Okay, so I want to thank the organizers, even though Alex is not here, for inviting me as a yeast geneticist to a human genetics meeting. So my talk will be around 80% or so yeast, but there will be 20% or so human genetics. And so what I hope is that what I can I want to tell you is that with yeast genetics, we can— thank you— we can actually also learn something, some things that are relevant for human genetics as well. And so, what we do in my lab is we study genetic interactions where we have a specific focus on genetic suppression. And this is driven by this observation that normally when we think about Mendelian disease, we think about cases where a person that carries a disease allele is sick, and a person that carries the wild-type allele is healthy. However, now that more and more healthy individuals are being sequenced as well, we can start to identify people that are perfectly healthy despite carrying mutations that have been directly associated with severe early-onset Mendelian diseases, such as cystic fibrosis. And so the idea is that one possible explanation for this can be that these people carry additional mutations elsewhere in the genome that can compensate for the deleterious effects of the disease mutation. And this is what we call genetic suppression. And such suppressive loci have actually been identified for, for example, a recessive deafness disorder. Now, why we're interested in this is that if we would know in which genes these suppressive mutations occur, then that would give us an understanding of the disease, and maybe in the long term would help in drug development. However, each individual carries millions of different variants, and at this moment we just don't have the knowledge to identify the suppressive mutations amongst these variants. Millions of variants. And so this is what we're trying to study in my lab. Now, just to give you a toy example of how suppression, for example, can occur is if you have two genes, A and B, that have an opposite effect on a gene or pathway, C, that is required for normal fitness. Now, in the absence of A, due to continued inhibition by B, the levels of C will be reduced, which leads to a fitness defect in this case.. And one possible way in which this could be overcome is now by additional loss of B as well, because now the inhibition is gone, levels of C can be restored, and so the fitness can be restored as well. Now, the opposite can also happen, where two mutant alleles on their own do not lead to a phenotype, but if you combine them both, this can lead to disease. And this is what we call synthetic lethality. And this, for example, can happen between two parallel pathways. Where the inactivation of one doesn't lead to a phenotype because the other one is still there. So in this case, in the absence of A, B is still there to activate C. But in the absence of both, this leads to lethality because now there's no longer the activation of C. And so with GWAS studies, we've been able to map thousands of variants to common diseases. But as you all know, this often explains only a very small fraction of the disease phenotype. And so there are several different causes or potential causes for this missing heritability. One possible explanation is that not all variants have been identified because they're rare or they have a very small effect. But another possibility is that the effect of some of the variants might not be additive and that genetic interactions may occur. And there has been some evidence in literature that this is actually the case. And so we know that genetic interactions can can be important for complex diseases, for Mendelian disease inheritance, but we at this moment do not have the power to detect them in GWAS or in other types of human genetics data. And so what we try to do in the lab is to use model organisms to really systematically study genetic interactions so that we can learn something that will help us decipher genotype-to-phenotype relationships. So we do this on quite a large scale so that we can really draw connections between all the genes in the genome and make what we call genetic wiring diagrams of a cell, and I'll show you some examples of that. Then, by doing this large-scale, we define general properties of suppression interactions, and then we can want to use these properties to find or maybe even predict genetic interactions in genome sequences of yeast, but also of other species, and I'll talk about this a little bit as well. Then, another thing that we want to do is to use these genetic interactions that we identify to understand which mutational trajectories are available during evolution. And I will not really mention this today, but I'd be happy to talk about this offline. And then, a third thing we do in the lab is that we directly identify genes that can overcome fitness defects of specific disease alleles. And I will just briefly mention this a little bit at the end of the talk. So, we use different model organisms. I already mentioned that one of our favorite model organisms is budding yeast. And budding yeast is a eukaryote. It's stable both as a haploid and a diploid, so we can do crosses. We have several mutant collections available, so there's 6,000 genes, 5,000 nonessential genes for which we have collections like the one you see here on agar plates, where each colony represents a different mutant in which a different nonessential gene has been deleted. And there are about 1,000 essential genes for which we have temperature-sensitive mutants available. Now, one of these plates has 1,500 different colonies, 1,536 to be exact, and we can process these using robotics. And so in my lab alone, we have a throughput of around 300,000 mutants a day. And so this is the main reason also why we use yeast, because even with CRISPR-Cas9 studies in human cells, we don't even come close to the throughput that we can have in yeast. However, yeast is a little bit simple compared to a human cell, and so we also have some cell culture work So, HEPA1 is a human cell line. It's a haploid cell line. Again, this facilitates the genetics that we do, and this is how it looks like. Okay, so today I want to talk about 3 different projects. First, I'll talk about some large-scale genetic interaction networks that we mapped in yeast. This is all published work. And then I want to talk briefly about some unpublished work in which we look at genetic context-dependent gene essentiality. And then finally, I'll talk a little bit about how we're trying to move all our yeast work into human cells. So, during my postdoc in Charlie Boone and Brenda Andrews' lab at University of Toronto, one of the main goals of the lab was to test all 6,000 genes for genetic interactions. And we wanted to do this in pairwise, so we wanted to test 6,000 mutant alleles against 6,000 mutant alleles. So this meant measuring 36 million pairs for genetic interactions. And to be able to do this efficiently, they developed a method that's called synthetic genetic array, or SGA analysis. Now, what we do when we do SGA is we take a mutant of interest, and we cross it into a collection of mutants, and this can, for example, be the collection of 5,000 nonessential gene deletion mutants. These are both haploid, we cross them to create a diploid, and then we drive them to sporulation, which is meiosis basically, and then we isolate haploid cells where each cell carries the mutation of interest that comes here from what we call our query strain, and then one other mutation that comes from this collection of, for example, nonessential gene deletion mutants. So then we have here this collection of double mutants. As I mentioned, we can do this on agar plates, so we take pictures like this. And then we measure the fitness of our double mutants by just measuring the colony size. So we take a picture and then we basically count how many pixels each colony is. So what we then do is we set the fitness of a wild-type colony to 1, and if we have two single mutants, in this example with a fitness of 0.8 and 0.5, we use a multiplicative model to have an expected value of the fitness of the double mutant. So in this example, if this mutant X has a 20% fitness defect in wild type, we would expect it to also have a 20% fitness defect in this mutant Y background. So we would expect a fitness of 0.8 times 0.5 is 0.4 in case there's no genetic interaction. If the actually observed double mutant fitness is lower than this, we call this a negative interaction, such as synthetic lethality.. And if the fitness is higher than this, we call this a positive genetic interaction. Now, I told you in the beginning in my introduction that we're really interested in suppression interactions, and so these are a very extreme type of positive interaction in which you go from a single mutant that has a fitness defect to a double mutant that really grows a lot better. Okay, so we used this method to map our 36 million pairs, or actually we did 18 million because you don't have to do everything in two directions.— and this is what the final results look like. So, what you're looking at here is this bit of a hairball where each node represents a different gene. So yeast has 6,000 genes, and all of them are represented here as 6,000 dots on this network. Now, they're connected if they share similar sets of genetic interactions. And what this means is if you think about two genes that encode two proteins that, for example, encode subunits of the same protein complex, Because these two proteins are in the same protein complex and they basically have the same function, you would expect them to have a similar set of genetic interactions, to have interactions with the same genes. And so if genes share similar genetic interaction profiles, as we call them, this is a very strong measure of a very similar function of these two genes. And so you can see these dense clusters appear on this network. And if we do unbiased GO term analysis on this for GO term enrichments, we can see that each of these clusters corresponds to a different biological function. Now, what you can see here is also that similar biological functions or related function map close together on this network. So, for example, here, this green cluster is ribosome biogenesis. It's right next to ribosomal RNA processing. Which is next to mRNA and tRNA processing. So, now what can we do with a map like this? One of the things is that we have— if we have an uncharacterized gene and we measure its genetic interaction profile, we can see where it maps onto this map, and this gives us a very strong prediction of what the function of this gene actually is. Another thing we can do is if we haven't measured the genetic interaction profile of a gene, but we know that it's in the same complex as a few other genes that we actually did test, we can predict its genetic interactions quite accurately. So this work is now published a few years ago, and it mapped hundreds of thousands of positive and negative genetic interactions. But those suppression interactions that are especially interesting when we think about human disease were only very poorly represented in this dataset, and this has several technical reasons. And so, we wanted to have a second look at suppression interaction specifically. And so, for that, we took a different approach. So, one of the things you can do with yeast is you can take a mutant that has a fitness defect and then grow it for several generations to isolate faster growing cells that may have acquired a spontaneous suppressor mutation somewhere in the genome. So, on agar plates, this looks something like this, where you have the small colonies that represent the starting strain, And then you have the bigger ones that are those that may have acquired a suppressor mutation. And then we sequence this whole genome to identify the genes in which those mutations have occurred. So we've done this for 200 different mutants, and we also collected data from the literature in which other people had done this for an additional 1,800 interactions. So with this set of 2,000 suppression interactions, we mapped another network. So I'm showing you this only very small in the corner here, because it's a little sparse still for a network to really learn a lot from the— from the global picture, but if we zoom in, I don't know if this is— okay, it's better visible for you than it is for me. What I did here is I grouped the genes by protein complex or pathway membership. And so these genes here all have a role in DNA damage repair. Now, I'm showing the suppression interactions as arrows that go from the suppressor gene to the gene with the fitness defect that's being suppressed. And what I hope you can see here is the vast majority of the arrows all point in the same direction, which is summarized by these bigger arrows here on top. And so what this means is that by mapping these interactions systematically, we can build a hierarchy between complexes and pathways, in this case involved in DNA damage repair. Now, what does this mean? So we can see there are some complexes that only function as suppressors, there are some complexes that can only be suppressed, And we looked into what these relationships actually mean. Now, in this example, it all represents the rerouting of DNA damage repair to alternative repair pathways in case one pathway is inactivated. And we can do something similar for genes involved in chromatin and transcription, where we can build similar hierarchies, and in that case, all the suppression relationships represent the activation or repression of transcription. Now, another thing we can do is, now that we have a set of 2,000 interactions, is try to define those general rules I talked about. So we try to divide these 2,000 interactions into distinct classes, or mechanistic classes. So what I'm showing you here is for all suppression pairs, so all 2,000 interactions, whether they had a functional connection to each other. So all the different parts in blue here represent, for example, that the two genes were members of the same complex, of the same pathway, or had some other type of functional connection. Now, this is way higher than you would see for any other type of genetic interaction, or for example, for physical protein-protein interactions. And because we see such a strong functional connection between the suppressor gene and the gene with the defect it is suppressing, we can use this to assign function to previously uncharacterized genes. And so we've done this, for example, here for this gene that we named any1, that we found as a suppressor for a neo1 mutant. Now, neo1 is essential, so we're using a temperature-sensitive mutant here that grows perfectly fine at low temperature, but has a severe fitness defect at high temperature. Now, in the absence of this any1 suppressor gene, the cells grow perfectly fine also at this high temperature. And so, the fitness defect of this NEO1 mutant is suppressed at this high temperature. Now, ANY1 encodes a membrane protein that localizes to delayed endosomes in the Golgi, and NEO1 is a phospholipid flipase that specifically flips phospholipids to one side of bilayer membranes, such that in a NEO1 mutant, this membrane asymmetry is gone,, and we can measure this in the lab using several chemical probes or GFP-tagged proteins. Now what we can see if we do these experiments is that any one overexpression allele behaves like a NEO1 loss-of-function allele, and any one NEO1 double loss-of-function mutant behaves like wild type. And so this made us suggest that any one had an opposite function to NEO1 and might function as a scramblase that randomly distributes phospholipids over both leaflets of a bilayer membrane. Such that in a NEO1 mutant, because of ANE1 activity, this membrane asymmetry is lost. But then if you inactivate ANE1 as well, there are some minor phospholipid flippases that can rebuild membrane asymmetry, but that in the presence of ANE1 would not be able to compete with the ANE1 scramblase activity. So, this was an example of a functional relation between the suppressor and the gene that was being suppressed. But we also find some more general ways in which suppression can occur. And so if we look specifically at hypomorphic alleles, so these are partial loss-of-function alleles that have a point mutation somewhere in the gene that makes the protein or the mRNA less stable. Now, what we can see is that these alleles are suppressed by mutations in the protein or the mRNA degradation pathways. You can see that almost half of the alleles are actually suppressed by a mutation in one of these. Pathways. Now what this means is that you have an allele that has a mutation that makes it less stable, that makes the protein less stable. Then one way to overcome this is by inactivating part of the protein degradation pathway, so you have more of your mutant protein. So this has nothing to do with the function of the gene that has this mutation, but just by the fact that the mutation makes the protein less stable. Now, I'm not talking here about mutations that completely knock out the proteasome. This would be lethal to any type of cell. But these are very subtle mutations that just sort of— that are chaperones or ligases that just alter a little bit the efficiency of protein or mRNA degradation. And so we thought, because most human disease alleles are hypomorphic alleles as well, maybe some subtle variation in these pathways might be associated with a protective effect against very specific disease mutations. So to be able to test this, and this is basically what it's showing here, so that reduced protein mRNA degradation can be beneficial in some cases. So to test this, and this is data from a collaborator, Chad Myers, what he did is he looked at GWAS data where he looked for the co-occurrence of SNPs in people both in the control and in the disease cohorts. Then he drew these SNP-SNP interaction networks. So basically, you have a line here between two SNPs if they co-occur within the same person. Then he overlaid pathway data, and then he looked for enrichments of pathways that had interaction either in the control cohort that he would say are associated with a protective effect, or in the disease cohort, which we would call risk-associated. Now, I won't actually show you the data here. I'm sorry, but it's Chad's data. So it's not really up to me to show this to you. But what we can see is, so we tested over 800 pathways in more than 10 disease cohorts, that we can actually find both protein and mRNA degradation machineries across the top protective-associated pathways across very many different diseases. Okay, so just to summarize this first part, I showed you how we can map very large-scale genetic interaction networks in yeast that we can use to draw these maps of a functioning cell. If we zoom in to especially suppression interaction networks, we can build hierarchies of interactions between complexes and pathways. And if we map our genetic interactions into different mechanistic classes, we can learn about general mechanisms of suppression that might be conserved across species and may be relevant to human disease as well. So, for the second part, I want to switch gear a little bit and talk about genetic context-dependent gene essentiality. So, we all know that some disease mutations have a different outcome in different people, and this is true for yeast as well. If you knock out a gene in one strain, it might be perfectly fine, but in another strain, it might be lethal. So, for example, I told you about this interaction of Neo1 and Any1, where I told you that loss of function of Any1 could suppress a Neo1 temperature-sensitive mutant. However, this suppression is so strong that Neo1— so this is an essential gene, normally you knock it out, the cells are dead— If you don't have AN1, the cells are perfectly fine. So this means that with one genetic mutation, one variant somewhere in the genome, you completely switch the requirement from this gene from being absolutely required for cell viability to not really being necessary at all. This, this NEO1 mutant in the absence of AN1 basically grows like a wild-type cell. And so what we wanted to know is how frequently is gene essentiality actually dependent on just a single genetic variant somewhere in the genome? Now, we wanted to test this systematically, and even in yeast, this is not very easy because how do you suppress something that's dead? And so for this, I'll spare you all the yeast genetics, but we used a little trick in which we use haploid cells in which an essential gene is deleted in the genome, but the cells are viable because they carry a temperature-sensitive allele of the same essential gene on plasmid. Now, these cells are now temperature sensitive, which means we can grow them at high temperature to isolate suppressors, which is like the bigger and the smaller colonies like I showed you before. And what we can then do is test if these colonies can survive loss of the plasmid. So we've done this for over 700 genes, so around 75% of all essential genes in yeast, and we found 124 genes. So this is more or less 15%, a little more, of the essential genes for which we can isolate strains in which the essential gene is lost, but the strains are still viable. So what we did is we sequenced these strains' whole genome to identify the suppressors. But first I want to talk to you a little bit about what these 124 essential genes are that are no longer essential in specific genetic backgrounds. So if we look at the function of these dispensable essential genes, as we call them, we can see that they are enriched for different processes that have a role, for example, in transport or in signaling, and they're depleted for things like translation or ribosomal RNA processing. And so you can imagine that a cell simply cannot live without a ribosome, but there might be alternative ways in which something can be transported or by which signaling can occur. Now, because we tested over 700 different genes, we could also look at this dispensability on a complex level. Because if a complex is essential, often a lot of the subunits are essential, and so we could see if the different subunits behaved consistently in our assay. So what I'm showing you here is every row will be a different complex. I'm showing the number of essential subunits, where in blue are the indispensable subunits. So we tested these genes in our assay. The cells could never grow without this gene. And the gray ones are the ones that we didn't test in our assay,, and some yellow ones will occur later, which are the dispensable essential genes. So, for the vast majority of the complexes that we looked at, we could not identify any suppressors of a deletion allele of any of these essential genes. So, these are truly required, and these, for example, include the general transcription factors, the spliceosome, et cetera. Now, there was this subset of 11 complexes for which each of the subunits that we tested, we could actually get viable mutants in the absence of this essential subunit.. And these include, for example, TOR complex II, which is involved in signaling, signal recognition particle, etc. Now, what was interesting is that we also found this middle class where these are complexes where some of the subunits we could isolate these bypass suppressors for, so they were dispensable, and some other subunits were not. Now, in some cases, this represented different genetic properties. So for example, if we look at the ribosome, the only subunits that were dispensable were those that had a paralog in the genome. And so the suppression in this case occurred by duplication of the paralog so that the actually deleted gene was no longer required because there were two copies of the paralog present now. In other cases, these dispensable subunits represented a different submodule of a complex. So for example, if you look at the nuclear pore complex, so I'm using the same coloring as here, the yellow ones are those that we found were the subunits that we found were dispensable. The blue ones are the ones for which we could not find any bypass suppressors. So these are the really core essential genes. What we can see is that none of the structural core components of the nuclear pore complex could be lost. But these ones that line the inside of the pore, these subunits have a role in pore specificity. So basically, these subunits determine what gets through the pore and what doesn't., and this can be lost. So, you cannot live without a nuclear pore, but you can live with a pore that is slightly less specific. Similarly, if we look at RISC, this is a chromatin remodeler, none of the core subunits could be lost, but cells could survive loss of these regulatory proteins. And the way they did this is by gain-of-function mutation in Sth1, which is the catalytic subunit of this complex. And so, normally Sth1 gets activated by these regulatory proteins, but this gain-of-function mutation made sure that the complex it was active and no longer needed these proteins for its activation. So, we see that there are certain genes for which the essentiality is dependent on genetic context in yeast. And so, what we wondered is if we know that the essentiality of these genes is dependent on genetic context in yeast, then maybe this means that these genes are more frequently absent or at least nonessential in other species as well. And so what I'm showing you here, this is for the indispensable essential genes. This is the vast majority of yeast genes. If we look in different species, so in the other budding yeast, Candida albicans, fission yeast, Aspergillus pompejani, worm— hi, Alex— or in human cell culture, what we can see— sorry, let me go back— is that for about half of these essential genes, they are also essential in these other species. But for about the other half of these essential genes, they're either absent, nonessential, or duplicated in these species. Now, if we do the same for our dispensable essential genes, you can see that these blue bars, where they were here about 50% of the genes, they're here about 25%. And so this means that indeed for the genes whose essentiality is context-dependent in yeast are also more frequently nonessential or absent. In these other species. So, so far I only talked about these dispensable essential genes, but what about the suppressors in this case? Sorry, I'm going too fast. First, I'm going back to can we actually predict gene dispensability? So one of the things that we wanted to use is the data that I showed you on the previous slide, so the functional categories, the evolutionary properties, but we had a lot of other different properties like essentiality, in other yeast strains, et cetera. And we wondered if we could use all the properties that we found to actually predict gene dispensability amongst the essential genes that were not tested in our assay. And so, we did this here where we took the 200 or so genes that we did not test, and we used all the properties that we defined to calculate a dispensability probability score for each of these 200 genes. Then we took the top 20 or so and the bottom 20, and we tested these in our experiment to see if they were actually dispensable or not. And as you can see, about 7 of our top-scoring genes indeed were, and none of our lowest-scoring genes. And so this is a very significant difference that shows that we can actually use what we learn from doing these experiments to predict essential gene dispensability. Okay, so now the suppressors. So when we look at them, one of the things that we can see is like before, we see a very strong functional connection between the suppressor gene and the gene with the fitness defect that's being suppressed. And so we can look at it as different things like colocalization of the proteins, shared GO terms, co-expression of the genes, or if they encode members of the same pathway or complex. And all of these are very highly enriched. Also, again, if we look at these different mechanistic categories like I showed you before, we can see that about two-thirds indeed are members of the same complex, same pathway, or have another type of functional relation. Now, one other thing that we could do with this new data that we couldn't do before is because we now have multiple independent suppressor genes for each essential gene deletion mutant, we could ask, are they always suppressed by mutations within the same gene, or do we see multiple different genes that can act as suppressors? And so if we do this, we can see that about a third of the essential genes are always suppressed by mutations in the same suppressor gene. So every time we do the experiment, we might get a different mutation, but it always hits the same gene. Then we have about 25% or so where we see different suppressor genes, but they have a functional relation.. And I'm showing you an example here for a TORC2 mutant. So TORC2, one of its many roles is that it activates sphingolipid biosynthesis in yeast. So if you have a TORC2 mutant, you don't have sphingolipid biosynthesis, and this is lethal for a yeast cell. And this can be suppressed by mutations at different genes in this pathway. So for example, YPK1 here, LCB1 or 2 here, or ORM1 and 2 here.. And so, if you do isolate new suppressors for this complex, you can see suppressor mutations in different genes, but they basically all do the same. And so, this is this light blue category here. And then there is this last category in which we can find functionally diverse suppressors for essential genes. And so, one of the things we can see is that the essential genes that have multiple suppressors for which we don't find a functional relation among them, often have multiple functions themselves. So, if an essential gene already does multiple things in the cell, it can often also be suppressed by multiple different suppressors. Okay, so just to quickly summarize the second part, I showed you how we isolated 380 suppressors of 124 essential yeast genes, and these are suppressors of the complete null allele of the essential gene so that it's no longer required for viability. We see that these dispensable essential genes, as we call them, often have a role in transport or signaling. They have distinct gene and protein level properties, and they're often nonessential in other species. And so we can use these properties that we saw to predict gene dispensability amongst untested genes in our assay. Then about the suppressors, they often have a functional connection to the dispensable essential gene. I didn't show you the data, but they're often essential themselves. I showed you how there's only a few routes to suppression in general, but sometimes there are multiple genes involved. I didn't show you this data, but there are a few cases in which we find two suppressor mutations that have to occur simultaneously, and we also often see aneuploidy. So if you're interested in that, I'd be happy to talk about this offline. So this was mainly work that we've done in yeast, but of course, now that we have CRISPR-Cas9, what we really like to do is also do screens for suppression or other types of genetic interactions in cultured human cells. Because although we can learn so much from yeast, there's a lot of more gene regulatory things that are not present in a yeast cell. And so, when I was still in Toronto during my postdoc, what we started up was we developed a genome-wide lentiviral knockout library. For the experts here, this is TKOv3. Which is a genome-wide library of guide RNAs that targets basically each gene in the genome. And so what we can do with this library is that if we start with a wild-type cell and we inactivate genes of interest, so we knock out gene A, we knock out gene B, and we take a wild-type as a control along as well, we can infect these with our genome-wide knockout library so that we get a collection of double mutants. So in each of these cells, our gene A will be inactivated and one other gene in the genome. And then we can passage these for several generations, and then by sequencing, identify what the mutants are that are present here and what the mutants are that were present in our double mutant pools. And then we can look for gene knockouts that specifically drop out of one mutant pool but not of our wild type or others. So in this example, so if we look at our wild type, so for wild type, of course, we get a single mutant population. We here have 3 genes, the yellow, pink, and blue one, and all 3 of them are still present at our final time point. But if we look at our mutant B background, the blue mutant is not present here, while it was present here, and it's also present in another screen. And so what this means is that this gene, this blue gene, gene Z, probably has a synthetic lethal interaction with mutant gene B that was the background in which we did the screen. And similarly here for the pink gene, gene Y, and mutant A. And so, we've done about 100 different of these screens now for 100 different mutant genes, and I'm just showing you example data for one of these genes. So, this is for ARID1A, which is a component of the SMY/SNF chromatin remodeling complex. And so, what I'm showing you here each dot again is a gene, is how quickly does this gene, when it's knocked out, drop out of an ARID1A screen compared to a wild-type screen. And so the vast majority of genes just maps right onto the diagonal, so it doesn't have— it has an equal effect when you knock it out in wild-type than when you knock it out in an ARID1 mutant cell line. But I highlighted the genes here in pink that drop out more quickly in wild-type than in an ARID1A mutant. So these are our positive interactions, including our suppressors. And these are the blue genes here are those that drop out more quickly in an ARID1A mutant background than in a wild type. So these are our negative interactions, such as synthetic lethal partners. So I highlighted 4 of the most significant positive and negative hits. And what you can see is that many of those also have a role in transcriptional regulation. So similar to what we've seen in our yeast genetic interaction network, also in human cells, when we look at the interaction partners, they often have a functional relation to the query gene itself. And so, just to highlight one example, we find here ARID1B, which is the paralog of ARID1A, so you would expect this would be a synthetic lethal interaction. Now, although we find some positive interactions, it's similar to the early high-throughput screens that we did in yeast. It's very hard to find really true good suppressors in these screens. And so, one of the things that my lab is currently working on is to optimize the screening technology to really specifically look for suppressor mutations. And just to show you one example here is we did a screen for ADSL, which is a gene involved in adenine biosynthesis, and it converts a toxic metabolite to a non-toxic compound. This means that if you don't have ADSL, you build up this toxic metabolite, which leads to a fitness defect in cells or to ADSL deficiency in human. And we know from human genetics that modifiers likely exist. Exist in the population. Now, this gene is also highly conserved, and we know that the gene in yeast is very suppressible. And so we thought this would be a good test gene in our human suppression screens. And so we did this screen, and amongst the top 10 hits, we find many genes with a role in apoptosis, as you would expect. If you inactivate these apoptosis genes, your cells live longer, and so they show up as hits in our assay. But we also find this gene that has a role in purine metabolism, and especially in the detoxification of toxic purine metabolites. And so we still have to confirm all of this, but it seems that based on these data, that our suppression screens are working in human cell culture as well, and that we can use it to find functionally relevant suppressor genes for human disease mutants. And so just to very briefly summarize this last part, I showed you how we can do synthetic lethality screens in human cells, and that we're currently optimizing the technology to do suppression screens. So just to acknowledge everyone in my lab, Carlos Ponce in Patrick Aloy's lab at IRB in Barcelona has done a lot of the statistics. Charlie Boone and Brenda Andrews is right in my postdoc, so the first part of the large-scale genetic interaction networks was done there. Fritz Roth at University of Toronto sequenced a lot of the strains that I talked about. Chad Myers at the University of Minnesota is involved in looking at GWAS data and trying to find some of the suppressor mechanisms that we found in yeast in human GWAS as well. And Jason Moffett at the University of Toronto was involved in the development of the TKOv3 genome-wide knockout library. And so, the University of Lausanne and the Swiss National Science Foundation for funding. Now, if you really like this work, there's positions available in my lab. And if you really like high-throughput screening, we're organizing a meeting in Cambridge this September, and I hope you will all come. Thank you.