Scientist Stories: Eran Segal, Unraveling the functional impact of thousands of p53 mutations

Is that— yeah, okay, good. Okay, good morning, everybody, and thanks for the very kind invitation to speak here today. So my lab has two main areas of focus. One is on microbiome-related research, and the other is on functional genomics and variant interpretation. And given the audience and the context of the conference, I'll focus on the latter today. So a very longstanding mission that we had in the lab was to try and understand how genomic information encoded in various genomic regions. And over the years, we thought that the most powerful approach to do that is what we call a kind of reverse genetics approach, where the idea is that it would be, we thought, very powerful if one can have an approach whereby you can design thousands, even tens of thousands of variants of your choice, integrate them into different regions of interest, into your cells of interest, and then very accurately assay the effect of making these different sequences. And we thought this would be powerful because if you had that ability, you can do many things. For example, if there's a region that you don't know what its function is, you can go and perform mutagenesis of this region and then assay the effect of every mutation that you make. If you want to learn about rules of gene regulation, for example, the effect of changing the location of a binding site for a transcription factor, you can do that in multiple sequences and assay the result. And if you have genetic variants that you want to see what their causal effect is, you can synthesize them within the same cell and then assay your phenotype of interest. And so, with that in mind, several years ago in 2012, we pioneered this approach called massively parallel reporter assays in parallel to the development of this approach by the lab of Jay Shendure and Tarja Mickelson. And in our case, we started with work in yeast, so we designed 6,500 different promoter sequences, integrated them into yeast, and then very accurately en masse in a single experiment measured their expression. So the way this works, and this is the general paradigm of how these mPRAs approaches work, is that we synthesize on programmable microarrays. These 6,500 different variants. We then chop them off of the array, integrate them into, in our case, a plasmid upstream of a fluorescent reporter, amplify them in E. coli, transform them into yeast such that at this point you have one population of cells. Each cell has a different one of these variants, and you don't know which yeast cell has which. But now, because they're all driving a fluorescent reporter, you can actually FACS sort them and then because your variable region is always going to be flanked by universal regions, you can use universal primers to amplify this and then use next-generation sequencing to sequence the entire pool from every bin. And so what you get in the end from this experiment, in a single experiment, you basically get for every one of the promoters that you design, you get a different number of reads coming from each expression bin. And so you can take the weighted mean of these different reads as the average expression of the promoter, but you also get the distribution, so you also actually have cell-to-cell variability or what's called expression noise. And you get all these measurements from a single experiment. And we showed actually that this approach works extremely accurately, so it's actually in near-perfect agreement to what you would get if you were to synthesize each of these sequences one by one, and that's true for the mean expression, and also on the right here for transcriptional noise, cell-to-cell variability. There's also extremely good concordance between the measurements that you get en masse in this single experiment and what you get if you design these variants one by one. So, over the years, we learned a lot from these. So, for example, in the first library in yeast, if we made a plot of the mean expression versus the noise of the entire library, we see that that overall noise decreases with expression, but then you can have many different promoters that express at the same exact mean expression, but they have up to 2 orders of magnitude difference in the cell-to-cell variability of expression that they drive. We learned a lot about the, for example, dose-response curve of expression as a function of number of binding sites. So here we designed promoters for 1, 2, 3, and so on up to 7 different binding binding sites for GCN4 with all possible combinations at multiple positions and on two different promoter backgrounds, and similarly for another transcription factor, GAL4. And you can learn the dose-response curve. So, in this case, you see that expression goes up with the first three or four binding sites and then kind of saturates. And we did this also for multiple different transcription factors, and we saw that the dose-response curve of expression as a function of binding sites varies across different transcription factors. We also saw that as you increase the number of binding sites, expression on average increases, but noise also increases. And even at the point when expression— the mean expression saturates, even at that point, adding more binding sites increases the transcriptional noise. And this was true for multiple different transcription factors. So, you can see how you use this approach to really learn a lot about rules of gene regulation. And over the years, we've had many different projects using this basic approach, adapting it to different settings. So, in addition to putting the libraries in promoters, we also put it in different regions to learn about rules of how sequences in these regions affect the gene expression. In addition to putting these libraries into yeast, we adapted the system to also work in human cells. I'll tell you more about projects along— in this setting. And then also beyond measuring just gene expression, we also adapted the system to measure DNA binding at single-cell, single-molecule resolution, and also to measure effects on DNA binding in vitro. And I invite you to look at a summary of the works shown in these papers here. And then, in addition to measuring expression or measuring very local phenotypes, we also looked at global cellular phenotypes, and I'll talk about two projects using this type of approach in the talk today. So, just also very briefly, one more project in yeast, so just to also demonstrate the power of this approach. So, we looked at 3' UTRs, which prior to our work there was not a lot of information known about how DNA sequence in this region affects expression. And because not a lot was known, we took a mutagenesis approach. So we took multiple different 3' UTRs and then designed mutations, scanning mutations along these different regions. So here's one 3' UTR in which we performed this, and then you get the measurements from the library of expression. Overall, we had 15,000 different variants in this library. So for this particular 3' UTR, you can see that there's one region here where when we make mutations, there's a strong reduction in expression, but then another region which is tolerant to mutations. Here's another 3' UTR again, one single region here where making a mutation has an effect. And overall, looking at the entire library, so 15,000 different constructs summarized in this heat map here, every row is a 3' UTR. Blue indicates a reduction in expression. You can see that this picture was pretty much true for most 3' UTRs. There's a single region in which making a mutation has a detrimental effect on expression. In our library, it has to be the DNA sequence driving these changes. That's the only thing that varies. And so we can analyze the sequence. In this case, something very simple, the AT content of the 3' UTR tracked very well with the effect on expression, we were also able to find a more refined motif which also had a stereotypical distance between the location of this motif and the end of the transcript. So we believe that this uncovered a lot of the principles behind how 3' UTRs in yeast dictate and affect gene expression. So moving to one project that we did in human cells. So this is the first project that we did in human cells to look at rules by which sequences within the promoters affect expression. And we took an approach whereby we integrated the library into a fixed genomic location so that would reduce a lot of the other external variability that you would get and really allow you to tease out the effect of making small sequence changes on expression. So, the first question that we asked is whether there's a difference in how sequences that originate from promoters or enhancers drive expression. So, as you know, the classical textbook view of enhancers is that enhancers work by recruiting RNA polymerase to the promoter and by that facilitating and driving expression. But this classic textbook view has been challenged in recent years by the finding that actually you have local transcription being initiated from enhancers, so the enhancers have an ability to directly recruit RNA polymerase and not just through looping and recruiting it to promoters. And it was shown that you can also have bidirectional transcription from these enhancers. And so, we wanted to study this in our library, so we selected multiple different enhancer sequences and promoter sequences from the human genome and subjected them to the assay in our library. And the first thing that we saw is that sequences that in vivo had more binding of the pre-initiation complex also drove higher expression in our assay, suggesting that there's an internal sequence determination of the ability to recruit the PIC complex. And by that, that's at least one of the ways that determines how you drive higher expression.. But then comparing promoters and enhancers, we took promoters and enhancers that actually have the same level of PIC binding in vivo, and nevertheless, in that case, promoters were able to drive higher expression in our assay as compared to enhancers. And so, this result suggests that in addition to some inherent sequences that are able to recruit the pre-initiation complex, there's additional sequences located within the promoter sequences that are able to drive higher expression compared to the expression that you get from enhancers. With respect to the question of bidirectional transcription, one question was whether that's actually being driven by the same promoter that drives expression in both directions or by two different promoters. And so, what we did in our library was to take these different promoters and actually put them in both orientations. Both for enhancer sequences and promoter sequences. And we also— we actually saw no correlation between the expression that you get from the same sequence if you put it in both orientations. And so, that suggests that the bidirectional transcription that's seen both in promoters and enhancers is actually driven by two distinct promoters that each drive expression in opposite directions. Another question that we looked at with this library is the effect of different elements in the promoter, different core promoter elements that have been described in the literature, most notably by the lab of Jim Kadonaga, who identified multiple elements including— so the TATA box is a known one, but then the initiator element, DPE, MTE, and several other BRE elements, but no systematic study has really tried to dissect the direct effect of each of these different sequences. And so what we did with the power of this library was to take these 6 different elements at their predefined positions and then generate all 2 to the 6 or 64 configurations of presence and absence of these elements along the promoter and also plant them on 5 different promoter backgrounds and then assay the expression within the system. So, here's a summary of these results. So here, every row corresponds to a promoter, and the blue and red indicate the presence or absence of each of these elements. So, for example, this row here corresponds to a particular promoter that has these particular three elements, and the promoters are sorted according to the expression that they drive in our assay from high to low. And so, when we look at this library, we can see now if there's anything special, any particular elements that are enriched in those promoters that drive high expression versus those promoters that drive low expression. And so, in this case, we found that if you look at the TATA element, TATA elements tend to occur in those promoters that drive high expression and they tend to be depleted from promoters that drive low expression. And indeed, when we compare these promoters with or without a TATA box, we see higher expression, so TATA box contributes as would be expected positively to expression. And this was true also not just on synthetic promoter backgrounds that we designed, but also if we took native promoters from the human genome and we either added or deleted TATA boxes from them, we could see in our assay that inclusion of a TATA box increases expression. And we didn't see this— and you don't see this effect if you just compare, like, CAGE-seq measurements in vivo of promoters, presumably because in vivo you have many differences between promoters that have or don't have a TATA box, which would mask out the signal. So again, this shows you the power of the approach. If you want to really dissect the functional effect of a variant, there's nothing like just adding or removing it from the same variant and asking the effect. And overall, when we looked at all of these different 6 core promoter elements, we only found evidence that TATA box and initiator have an effect on expression, and for all the other 4 elements, whether they're present or absent, that, in our assay at least, did not have any effect on expression. The other thing we realized is that we could use this system to actually generate these types of sensors for the activity of different transcription factors. So, within that same experiment and library, we designed for over 100 different transcription factors, we designed multiple different binding sites for them, and we designed them in both orientations and then put them in the library and then assayed the expression driven by each of these different promoters. And so, we can take these promoters and we can rank them according to the expression, and we found that actually two-thirds of the transcription factors had drove expression above our background levels, indicating that in the K562 cells to which we integrated the library, over two-thirds of the transcription factors actually show activity. Interestingly, there was near-perfect correlation between taking the same construct and putting it in both orientations, indicating that the directionality of the transcription factor has no effect on expression, at least for for these transcription factors that we tested. One other interesting question is also looking at the dose-response curve, so the effect of number of binding sites on expression. A previous work from the lab of Nadav Ahituv argued that if you have heterotypic clustering of binding sites, so sites for different transcription factors, that would drive higher expression than sites, multiple sites for the same transcription factor or homotypic clustering. So, to look at that, we first looked at the organization, the native organization of binding sites within the human genome, and we saw that there's really a lot of differences between some sites, some transcription factors tending to occur in clusters within the human genome, like sites for SP1, which tend to occur in clusters. In contrast, sites for CREB tend to appear more, much more in singletons. So, we reasoned that perhaps sites for these different transcription factors would actually show different dose-response curves. And so, what we did was to take a bunch of these different transcription factors, select 7 predefined locations within promoters, and then generate within the library all 2 to the 8, or 2 to the 7, 128 different combinations of promoters at different multiplicities and presence-absence along promoters. And so when we look at the dose-response curve for one transcription factor, YY1, which actually has a fairly moderate clustering, native clustering of sites in the human genome, we see that expression does increase as you add more binding sites, but it's much more pronounced if you do the same type of experiment for SP1,, which shows a much higher degree of clustering of sites in the genome. You see that expression goes up much more dramatically with adding more sites. And interestingly, for CREB, that doesn't show really clustering of its sites in promoters in the human genome. You see no effect of adding multiple sites. And we saw that also on the background of two different promoters. So, this is very nice. It shows that kind of evolutionarily where you see clustering of binding sites for the same transcription factor, there's probably a reason because in our functional assay, that actually shows that it drives higher expression, whereas when you look at that for sites that don't show this clustering, you don't see that. The final question that we asked with this library here was the effect of sequences that disfavor nucleosome formation, because previously we have shown and others have shown in yeast that these poly(A)D(T) elements that are very rigid and block nucleosome formation, they can exclude nucleosomes from promoters and promote expression, but that was not tested in humans. So what we did was to add these nucleosome-disfavoring sequences to promoters that contained two binding sites for various different transcription factors, and we compared adding that to adding additional binding sites for these transcription factors. And we saw that in most cases, Adding these nucleosome-disfavoring sequences also in human promoters also increased expression and actually sometimes increased expression even more than when you add additional transcription factor binding sites. So this, again, shows you the power of this approach for trying to understand the function driven by different promoter elements. Different genetic variants within promoters. But all the work that I showed you so far is looking at the relationship between changes in the sequence and local changes like changes in expression. But perhaps an even broader question is whether these changes in expression actually have any functional meaning. So, in another project, the first— I want to show you two such projects. The first one we did in yeast. Where really the question was, what happens if we perturb expression? Does that have an effect on some more global cellular phenotype? So, we reasoned that perhaps the best approach to do that would be to somehow have a way of modifying the expression levels and then assaying the phenotype. In our case, this is going to be cellular fitness in yeast. So, what we did was to select 100 promoters that from previous work that we did we knew spanned the entire range of possible attainable expression in yeast from very low to very high in relatively constant intervals, which we measured previously. And then we selected 100 interesting genes and we swapped for each of those 100 genes, we swapped its endogenous promoter with each of these 100 different variants. So essentially we're driving the expression of each of these 100 genes at all possible levels. And then we have this pool of about 10,000 strains, and we can sequence it before and after letting it grow for 11 generations. And then by next-generation sequencing of the library, we can really look at the growth rate of each of these 10,000 strains within a single experiment. And from this experiment, you would then get these dose-response curves. So now on the x-axis, you have expression measured at 100 different levels, and on the y-axis, we can have the growth rate, so a measure of fitness of the yeast cells. So, we can cluster these fitness curves that we measure for all of the different genes here. So, here the x-axis is the expression level. Every row corresponds to a different gene. And when we look at that, we see that the expression-to-fitness curve is very different depending on the gene you're looking at. So, for example, SPT4 here has a high sensitivity to low expression, so at low expression levels you get a very dramatic hit on fitness. PHD1, on the other hand, actually has high sensitivity to high expression. Then you have other genes that are just very sensitive to any change that you make to their expression. You can see that even very small changes from the wild-type expression, which is shown here by the dashed red line. Any change that you make from that actually has an effect on fitness. Then other genes that either show a gradual response of fitness with expression or a more step-like function whereby moving from the wild-type expression has a really dramatic effect on fitness. And then other genes like this one here that's completely insensitive. You can change its expression levels and there's no effect on fitness. And even some intriguing bimodal type of behavior. But even though these genes exhibit very different fitness to expression to fitness curve, when you cluster them by the function of the genes, then you actually see much more concordance, like the genes belonging to the Leloir pathway or proteasome genes or topoisomerase genes have very similar types of behaviors when they come from a very similar pathway. We can also ask whether low expression or high expression is more detrimental to fitness, and we see that when we express these genes at a very low level, so this kind of mimics more or less a deletion type of situation, we see that almost all genes have a very profound effect on fitness, and this is much less pronounced when you overexpress genes, so most genes when you overexpress them, actually there's very little effect on fitness. There are some genes that are affected. These are mainly transcription factors that when you express them highly have an effect on fitness, presumably because they regulate many different targets, so you're going to have massive effects on the genome. Now, we also saw that if you really study carefully these functions, these expression-to-fitness curves, you can learn a lot about the function of the genes. So here's, for example, the galactose metabolism pathway, and we can see that 3 genes here that exist in a single copy, they really have these step-like functions whereby you reduce their expression a little bit, you get a huge penalty on fitness, whereas for GAL2, another transporter in this pathway, you see a much more mild effect, presumably because GAL2 actually has this HXT transporter genes which can back it up. Similarly, when we look at regulators, we see that for a transcriptional activator, if you lower its expression, you get a reduction in fitness, but for a repressor, if you increase its expression, you get a reduction in fitness. So, if you didn't know whether these regulators are activators or repressors, these expression-to-fitness curves would tell you that. Here's another example where we also studied the central carbon metabolism pathway. So this, as you know, can run glycolysis top-down, fermenting glucose into ethanol, or it can run gluconeogenesis backwards, bottom-up. So in our library, we measured this in galactose, so this is going to be glycolysis going downwards. And very nicely, we saw that genes that actually increase the glycolytic glycolytic flux, if you increase their expression, as you may expect, the fitness goes up, and genes that actually reduce the glycolytic flux, when you increase their expression, the fitness goes down. So this you can see by the curves of the different— the fitness expression to fitness curves of these different genes, and also if you look more carefully at each of these genes, you can see that, for example, this gene PFK2, which increases the glycolytic flux if we increase its expression, so increasing the flux, we get an increase in fitness. In contrast, FBP1, which synthesizes a reaction in the opposite direction, so reducing the flux, if you increase its expression, you get a hit in fitness, and a similar picture for these other Eno1 and Eno2 genes. Tubulin is also a nice example. So here for TUB2, which has a single copy, we see a very high sensitivity of changes to the expression and the effect on fitness. But for TUB1 and TUB3, which can back each other up, we see a much more mild correspondence. If we actually add them up, so now we're looking at the total either TUB1 or TUB3 expression, then we get a very similar behavior to that which we see on TUB2. Which also tells you that these two proteins should be in a 1-to-1 stoichiometry, which actually is known in the literature. So, again, if you didn't know this before, these assays would tell you that. The other thing we did was to measure this entire library on different conditions to see whether the effect of changes in expression of genes varies by condition. So, we made a histogram of the distance in fitness for every gene between measuring its— doing this assay for it in two different conditions, either in glucose or galactose. And then we saw some genes that had very similar behavior, very similar effects when you measure them in the two conditions. These were a lot of the housekeeping genes, which presumably are needed in both conditions. But then we saw genes that, like these glucose catabolism genes, that actually were entirely insensitive in galactose where presumably they are not needed, and in blue here you can see very high sensitivity to changes in expression of these genes when you grow them on glucose where these genes are needed. And conversely, genes that belong to the galactose metabolism pathway showed the opposite behavior. They're not needed under glucose, so changing their expression in glucose has no effect on fitness, but changing their expression in galactose has a huge effect on cellular fitness. A related question is to ask whether the wild-type expression is actually expressing at the place where you actually achieve optimal growth. So basically, if you look at the expression-to-fitness curve, it's asking the question of whether wild-type expression is here, which would indicate that it's at the optimal expression level for growth, or whether it's at some non-optimal place. And interestingly, what we found is that for all genes that we tested, with one exception, we saw that their expression, their wild-type expression, is actually optimal for growth in glucose, which is the most prevalent condition that yeast encounter in the wild. And in contrast, if we grow the cells in galactose, then we find that about 20% of the genes are not expressed expressed in their wild-type level at their— at an expression which is optimal for growth. And when you change their expression to optimal growth, it makes a lot of sense. So, again, if we look at the galactose metabolism pathway, so galactose metabolism feeding into central carbon metabolism, we see that genes that increase galactose flux, when you increase their expression from their wild-type expression to a different location, when you increase their expression, therefore increase the flux into central carbon metabolism, you also get an increase in fitness. And conversely, if you decrease the level of a repressor, you also see that in similar behavior for genes that increase glycolysis. So, finally, with this library, we also looked at cell-to-cell variability or transcriptional noise, which in our library varies about 100-fold between different genes. And we hypothesized that perhaps if you have a gene that's very sensitive to changes in expression, so it has this curvature, if you look at the expression to fitness curve, meaning that changes in its expression will have a very big effect on fitness, then presumably noise would be much less tolerated in these genes because if you had variability in expression, you would have a huge impact on fitness, whereas if you have genes that are insensitive in changes of their expression to growth, then in those genes perhaps you would be able to tolerate noise much more. And so, indeed, if we look at genes that have low transcriptional noise, these genes tend to actually have this sensitivity to have this curvature. When you change their expression, there's an effect on fitness.. And in contrast, these genes that have very high noise, they tend to be insensitive to changes in their expression and how that affects fitness. So of the 4 possible states that genes can be in, either low or high noise or low or high curvature, we believe that the configuration of having high noise and high sensitivity to expression is very strongly disfavored by evolution, and we really don't see see these types of behaviors. So, finally, in the last 10 minutes or so that I have in my talk, I want to switch gears and tell you about a different project that we did with this library, also assaying a global cellular phenotype, and this is a project that we did in the area of cancer. So, we decided to focus on the p53 gene, which, as you very well know, is the most heavily mutated gene in human cancer, mutated in about 50% of all tumors. And there's been— but even though there's been thousands of different mutations that have been clinically isolated, much of the research has focused on only just 6 hotspot mutations, even though these hotspot mutations account for only 30% of all the clinical variant mutations that you find, and these thousands of other mutations that account for 70% of all tumors have been much less heavily studied. So we wanted to ask whether we can actually use our library to assay the function of all of these different P53 variants, because— and maybe move away from just the dogma of whether the tumor is mutated for P53 or not, and by looking at the actual mutation and seeing whether that actually has different functions. Because the reasoning behind it was that if we were to see some different behavior depending on the mutation, then that may lead to a different course of action or different ways in which we would like to treat and understand each of these different mutations when we see them in patients. And so what we did was to use the power, again, of this library to synthesize 10,000, over 10,000 different variants, so all the clinically isolated variants in the DNA binding domain of p53, and then multiple additional variants. And we took all of those variants and we introduced them into p53-null H1299 cells, and we then took all that collection of 10,000 strains and we let them grow for a period of 2 weeks, and then we used next-generation sequencing to sequence library before and after this growth, so similar to what we did in yeast, and that gives you a quantification of the— basically the curve of the abundance of each of those 10,000 strains. Now, because p53 has an anti-proliferative activity, for all the variants that retain p53 activity, what we expect to see is a behavior like shown here in blue, where we would see a lowering of the abundance over time. And if we had a mutation that actually abolished P53 activity, so losing the anti-proliferative activity, then we would see that abundance of that particular strain would remain. That's what's shown here in red. So here's a summary of all of those results. So 10,000 different growth curves for all the 10,000 different P53 variants. So let me walk you through this. So the x-axis here is the position along the DNA binding domain of p53. The y-axis is all 20 amino acids. So we made, for every position, we made a change to all other possible amino acids. And at the bottom, there's two frameshift mutations done at each position and also introduction of a stop codon. And then the coloring here is according to what I showed before. Blue indicates wild-type P53 activity, so if you zoom in on each colored entry like this colored entry here in blue, then the growth curve behind it would look something like this, whereby we would see P53 activity, so reduction in the abundance of this variant, whereas if we look at a red-colored entry here, this would be loss of P53 activity, so we would see retention and high abundance of the strain over time. So this is the entire library measured, and then we overlaid different functional data on top of that to make sense of it. So the first thing that we did was to look at the structure, the known structure of p53, and quite remarkably, if we look at the structure and now we color the structure according to the average effect of making a mutation in that residue in our assay,, then we see that residues in the DNA binding domain of p53 that touch the DNA are colored red, indicating that these positions are highly sensitive to mutations, and any change that you make there really abolishes p53 activity, and that's much less so the case if you make mutations in residues of the DNA binding domain that are not directly in contact with the DNA. Even more remarkably, when we overlay evolutionary sequence conservation, which obviously was not information available to us or in the assay, we see a very high correspondence between the effect that we see for these mutations in our assay and evolutionary conservation of signals. And all the 6 hotspots sit in regions that both are evolutionarily conserved and also have a strong effect in our assay. We can also cluster the position, so now this is the same data, but now we're clustering the positions according to the behavior in our assay, and when we do that, we see three different types of behaviors, either positions here that are very highly sensitive, so essentially any mutation that you make at these positions abolishes P53 activity, then another region here of positions that are tolerant in general, so you can make almost any mutation and there will be no effect on P53 activity. And then there's this intermediate region where some mutations that you make to some amino acids will have an effect on P53 activity and others will not. As you may expect, we find the hotspots to be enriched in these sensitive regions, but quite surprisingly, even for one of those hotspots, we find it to be in the intermediate region indicating that even for a very well-studied hotspot mutation, at least some of the mutations that would occur in that hotspot would still render the p53, in our assay at least, fully active. Another question that we asked is whether there is a correlation, correspondence between the effect that we see in our assay and the prevalence of these mutations in tumors, and indeed, we see a correlation, so in other words, the more prevalent the mutation is in tumors isolated from patients, the higher we see it having an effect in our assay. So all the hotspot mutations, which are the most prevalent, also exhibit, as you can see here, an effect in our assay. But interestingly, there's this other region here of mutations that are very detrimental to p53 activity, but they're not very prevalent, presumably because to attain these, you actually have— need more than one nucleotide change. So it's two nucleotide changes that are required for these amino acid changes. So that's why probably they're more rare, but if you see them in the clinic, they will likely abolish p53 activity. If we remove those and look at just those mutations that are easily attainable with a single base pair mutation, we see a an even better correlation between the prevalence of these mutations in tumors and the effect in our assay. And here I want to highlight this interesting region here. So here you have mutations, some of which are actually quite prevalent. You find them in the clinic, but actually they, at least in our assay, retain full P53 activity. So perhaps we reason that if you find these types of mutations in the clinic, these are passenger mutations not really having an effect on p53 activity, and if you find them, you probably need to search for other mutations that are the ones driving the cancer. So the final result I want to show, it's well known that there's actually natural polymorphism in p53, so naturally occurring variants that occur across the human population in P53, which in our assay, when you introduce these SNPs, these naturally occurring SNPs, there's no effect on P53 activity, but we wanted to test whether adding now the mutations on the context of those different genetic backgrounds, those polymorphisms, whether that has some kind of interaction and whether that somehow enhances or retains or subtracts from the effect of making the mutations.. And indeed, we found that— we found several such cases. So one is demonstrated here for a naturally occurring SNP. When we introduce the mutations that we made in our library on top of this naturally occurring SNP, we see that actually now the effect of these mutations is going to be enhanced, so you're going to see greater loss of p53 activity compared to when you introduce these exact same mutations but just on a different, the more consensus genetic background. So this really also shows you the power of this type of approach for dissecting the function of these variants also in the context of different genetic backgrounds that naturally occur in the population. So with that, I will end. I'll just put up the slide of my great and very talented group of students and postdocs that are responsible for driving this work. The work on human promoters that I showed you was done by Shira Gabay from the lab who's now doing a postdoc at Harvard. The work on the phenotypes of expression to fitness curve in yeast was done by Liat Keren, who's now a postdoc at Stanford, and Eilon Sharon was the PhD student in in my lab who 7 years ago pioneered the mPRa approach in my lab, also now a postdoc at Stanford. And with that, I'd like to thank you for your attention. [APPLAUSE]