Scientist Stories: Paul Flicek, Tissue Specific Enhancer and Promoter Evolution in Mammals

Thank you. And thank you to the organizers, and thank you to all of you this morning. I'm going to spend about the next 40 minutes talking about some of the work from my lab and from our collaborators about trying to understand how regulatory elements work, which regulatory elements are important, by using information contained in evolution. I think that you'll see that some of the things that we're doing can provide— actually follow on nicely from some of the things that Alexis talked about. But our focus and some of the information that we bring in here using evolution is somewhat different. So to start with some background, one of the things that's been known now for nearly a decade is that transcription factor binding locations evolve incredibly rapidly. When you look across species and look to see how often a given transcription factor binding site stays in the same place, it changes very rapidly. And it changes over the course of just a couple of million years. And so when we've looked at, for example, when we look at a whole host of mammals, you can see that by the time you get out to human and mouse distance, you only have about 10% of the tissue-specific transcription factor that might be in the same place. And even at shorter distances, when you go out only a couple of million years, about 50% of the transcription factor binding sites have changed places. Now, we and others have looked at this, and it's been looked at in a number of species. But what has come out of this when all the analysis is done carefully and done together— and this was a study from Anne-Rixanda Carvounis when she was a postdoc in Trey Ideker's group— that the rates are indistinguishable when you look at mammals or insects, when you look at embryos or adult tissues, when you look at different factors. So this is not an effect— this changeover of transcription factor binding sites is apparently not an effect of generation time, it's not an effect of development, it's something inherent in how regulation works in the genome. Now, at the same time we see these exponential changes in transcription factor binding sites, we see that gene expression changes only linearly over evolutionary time. And here again, from the same study that looked at insects and mammals and embryonic tissues and adult tissues, we see this is essentially a linear relationship, and it's based on speciation time. So again, not on generation time, or not on factors of development and so on. So we end up in this situation where gene expression levels are evolutionarily stable, and in fact gene expression levels across species cluster by tissue type. When you do that— this has been known now since Henrik Heisman's group in 2011 did this nice study with mini tissues— and you can see that the gene expression clusters by tissue type and not by species. So the gene expression levels are evolutionarily stable, tissue expression levels are maintained, but at the same time, the tissue-specific transcription factor binding is neither stable nor maintained. And so what we assume to be driving these stable gene expression patterns are less stable than we might like. And so it's these questions that have driven our approach and driven some of the questions that we've asked to try to understand what's going on or to understand more about this system. Here's the basic experimental design that I'm going to talk about. So, and I'm going to talk about sort of two studies. One is a study that looked exclusively at liver tissue in 20 different mammals. And then a current study— that was published a couple of years ago— and then a current study, which is unpublished, which looks at 10 species and 4 tissues. In each of these cases, we do chromatin immunoprecipitation, so we do ChIP-seq to identify active promoters and active enhancers. Now, for those of you who have worked on enhancers and promoters, the definition of what an enhancer and what a promoter is can sometimes be challenging. We are using overlap of histone modifications for our definitions. So there are strengths and weaknesses to that approach. We've also done RNA sequencing in matched tissues. So it's exactly the same individual sample where the ChIP-seq and the RNA-seq is done. So I'm going to start by talking a little bit about this 20-mammal study where we mapped active promoters and active enhancers in liver across 6 mammalian orders. So not only do we have primates and rodents, as often takes place in comparative studies, human and mouse, we've also mapped these in cetaceans, in carnivores, in marsupials. To my knowledge, this is the first time functional genomics was ever done in whales. And so we've put together a combination of how the mammalian regulatory architecture works. And we can see it kind of sort of the things that you might expect. There are some regions of the genome that are absolutely conserved across species, where we see the same patterns of H3K4 trimethylation, H3K27 acetylation, and other regions of the genome that are specific to various clades. And you can— any kind of specificity that you can imagine, from carnivore-specific to cetacean-specific and so on, shows up in the data. Just as a reminder, when we see the H3K4 trimethylation and H3K27 acetylation together, we call that an active promoter. The H3K27 acetylation alone, we call an active enhancer. When I talk about the 10 species, 4 tissue experiment, we're going to modify that definition just slightly, but it largely comes out to the same. So when we focus on what's conserved across all of these species, we find that conserved regulatory regions— again, as I alluded to from how much the transcription factor binding changes— the conserved regulatory regions are rare. And they tend to be promoters. So if you look across, taking the view from human, where we have about 41,000 regulatory regions, promoters and enhancers together, only about 1,800, so roughly 20% of the promoters are conserved across species. But a much smaller number of the enhancers are conserved across species, on the order of about 1/10 of 1%. So to see a conserved enhancer across species is a very rare event, while conserved promoters are much less common. And just as a technical note, in this 20-mammal study, we focused on the 10 mammals with the highest quality genome assemblies. Just because it becomes very difficult to do these cross-species analyses when you have fragmented genome assemblies, something that is fortunately being solved with new PacBio assemblies coming out seemingly every day. So this is what the conserved regulation looks like. We initially focused on the flip side, the parts that occurred, observed in only one species. And from the point of view of liver, typically half of the liver enhancers in any given species are found only in that species. So there's this huge collection of species-specific enhancers. We call these recently evolved. So they are not shared between species. And we set about to try to understand what these recently evolved regulatory regions look like, and then eventually what these recently evolved regulatory regions do. So when we look at recently evolved promoters, recently evolved promoters tend to arise from young DNA. For the— how we define young DNA here is young DNA is not present in the multiple alignment between the species. So when you look into the multiple alignment, if the DNA is new, so maybe it's been inserted from a retrotransposon, maybe it's come about by some sort of duplication, this sort of young DNA sequence is where most of the recently evolved promoters are found. But enhancers are the opposite way. Most of the recently evolved enhancers arise from ancestral DNA, so DNA that is shared between the species, but has been captured in only one evolutionary lineage into this functional relationship. Now, I want to point out a couple of things that are actually very important in this plot. So first of all, it's the difference in the axes. So this is 1,000. This is 10,000. There are many more of these recently evolved enhancers, which basically I alluded to earlier in the amount of conservation. And so even though we make the statement— and it's actually quite clear— that most recently evolved enhancers arrived from ancestral DNA, there are, in fact, thousands of enhancers that— recently evolved enhancers that are arising from young DNA. And, in fact, many more enhancers arise from young DNA than promoters arise from young DNA. And when you look down about the various characteristics of these young enhancers, you see that retrotransposons are a common cause, or at least a common overlap of these young DNA sequences. I'm going to come back a little bit to the role of retrotransposons in regulatory evolution as we get into our multi-tissue studies. Okay, so kind of to sum up, we know that when we look at gene expression between tissues and cell types, the gene expression is the most conserved, the most stable across evolutionary time. Promoters are less stable than genes, and enhancers are less stable still. If I were to plot individual transcription factor binding sites on this plot,, it would drop even further and it would end up below the enhancers. When people have looked at multiple tissues, and this is just to give you a preview of what's coming next, when people have looked at multiple tissues, they tend to do so only across a couple of species and often at very long evolutionary distances. So the mouse ENCODE project did this very comprehensively, looking between human and mouse, but this is just two a two-species comparison. Now, before I go on to the multiple tissue aspect, I want to take a little bit of a diversion and talk about connecting these regulatory regions and this regulatory evolution to gene expression. I mentioned at the start that not only did we collect these active promoters and active enhancers, we collected gene expression in all of the exact same individuals so we can make these connections together. And when trying to understand how gene expression evolves, there's really two characteristics of gene expression that you can effectively compare across species. One is expression level, so normalized expression level across species for orthologous genes, and the other is expression stability, so how variable is the expression of a given gene between species. And some genes are uniformly higher, uniformly lower. Some genes are uniformly more stable. Or uniformly less stable across species. Now these characteristics turn out to be incredibly confounded across species. Highly expressed genes are more evolutionarily stable. And in addition to this confounding factor, assembly quality impacts all of our analyses. And so this annotation quality also impacts our analyses. And this makes for a number of changes. So this is just an example of how these two things are confounded together, so housekeeping genes, ones that are generally expressed in many tissues, you can see here that the correlation, the expression correlation, so this is how stable they are, housekeeping genes are more stable. Core liver genes are also more stable. And all of these are more highly expressed. This is also an example of what the raw data looks like where we have gene expression across tissues. And we have regulatory landscape across various tissues. When you look at all of this together— and I can kind of jump right to the answer— the driving factor of what makes both genes stable and highly expressed is the number of regulatory elements. This is simply the most important factor. The number of regulatory elements is the primary driver of expression level and expression stability. Whether those regulatory elements are conserved plays a role, as I'll show in the next couple of slides, but it is not the primary driver. It's a mass-action, additive sort of force, kind of a strength in numbers approach. And the way we get at this, and I'll just describe this in a couple of ways, is we created— to manage the confounding factors here and the interconnectedness, we created a series of control sets. Where we would match genes with the same expression level or the same number of regulatory regions and then change those in subtle ways. So here, for example, we have a matching of same expression level between genes with different numbers of promoters. And you can see that the expression correlation with 2 or more promoters is higher than when there's 1 promoter. So we compare these sets directly. Similarly, with more enhancers, again matched for same expression level, the expression stability is higher. And we can go over and look here just at a simple— there's an additive effect for gene expression level. Promoters act switch-like. When there is one promoter, essentially the gene turns on. It doesn't matter how many more promoters you have. Where enhancers is much more additive. As you add additional enhancers, the gene expression the gene expression level goes up. So again, it's the number of regulatory elements that drive this. And so once we— when we talk about the number of regulatory elements, we use this concept of regulatory complexity. The more complex a regulatory landscape, the more number of regulatory elements, and then the more highly expressed and the more stable on average these genes are. So this seems to be the mechanism that evolution is using, regulatory complexity. Adding more regulatory regions regardless of their source or their evolutionary age maintains both expression level and expression stability. Now, of course, there are still effects of conservation. Conserved enhancers turn out to be slightly stronger. And so the way we do this is we create various sets where we manage between the number of regulatory elements, and we basically switch between a set where there's the same number of regulatory elements but without them being conserved. And when we see this, we see that gene expression goes up whether or not we do this for promoters or enhancers. And again, we now can double-match for expression level and the same number of regulatory elements and switch from a percent— placental-conserved— so this double matching for expression level allows us to look at expression stability. And when we switch out a placental-conserved for a non-conserved but keep the same number of regulatory elements, you see that the placental-conserved promoters and enhancers drive additional stability. So genes with placental-conserved elements are more highly expressed, slightly. Genes with conserved elements are more stable, slightly. So these are conserved enhancers are slightly stronger than other enhancers. But this doesn't mean that the recently evolved enhancers are functionless. They're just slightly weaker. And when we look at this and add, for example, a recently evolved enhancer, we see additional expression levels and additional expression stability. So this shows us that there is a specific contribution to gene expression when we have these brand new elements that turn up. So, whatever they are, they are not just sort of functional noise, even though they're present in only one species. And we can also look at the relative contribution of these recently evolved enhancers, where we switch a recently evolved enhancer for a non-recently evolved enhancer, and we can see again there's a small dip in the various expression levels. So, adding a recently evolved element increases expression, And exchanging conserved enhancers for recently evolved ones slightly decreases expression. But they still provide both an actual functional contribution. Okay, so we've talked about how recently evolved enhancers— we've talked about how enhancers and promoters drive expression in various ways. We have this very large question on tissue specificity. We see a lot of species specificity. And we were wondering how this connects to tissue specificity. So one of the questions that we initially had was a lot of this species specificity is actually hidden tissue specificity, where if we look down and we could see through many tissues, we wouldn't see these as species-specific enhancers. We would instead see these as tissue-specific ones that are shared So a given enhancer might have taken a liver path in one species, but say a muscle path in another species. And so that was the goal of this project that we're currently working on, where we have, again, mapping of these regulatory elements in liver, muscle, brain, and testes. So actually 4 very different tissues coming from— with both homogeneous cell types and and heterogeneous ones. Again, we do the same sort of thing. Here we've added an additional histone mark, H3K4 monomethylation, and we— but we define active promoters exactly as we did before, just by two marks. We define active enhancers now by having these two marks, K27 and K4 monomethylation. This leaves a small class of K27 marks that are not considered as active enhancers, and then so-called primed enhancers, which just have this H3K4 monomethylation signature. When we look across species and tissues, and you'll notice the numbers are a little bit higher here, we deliberately sequenced much deeper. We sequenced to saturation in all of our marks so that we feel we have a better representation. What you see is these somatic tissues are similar across evolution. All of them have approximately the same number of active promoters, approximately the same number of active and primed enhancers. And in fact, it's really not that different when you look at brain. Brain has slightly more of these regulatory elements due to its cell type heterogeneity, but it's not a dramatically different picture. Testes, though, is a different picture with a an additional group of active promoters and fewer active enhancers overall. So this is what things look like. The tissue-level comparison— or the species-level— the tissue-level comparisons are surprisingly consistent across species. Again, maybe what you would expect knowing that the output of the tissue gene expression is consistent across species. We've also looked through the RNA-seq data, and we know that the number of genes expressed in the somatic tissues are similar across evolution. And in fact, the biggest difference between the genes and the isoforms expressed is simply the quality of the annotation. And so species like mouse, which have basically the highest completeness of the annotation, you see this large gap between the number of genes and the number of isoforms. That's just because we have higher resolution there. And I'll show that in another view in a little bit. As you look at other species where the gap between the number of genes known and the number of isoforms annotated is smaller, these two end up almost on top of each other. Testes, again, is different. It's more transcriptionally active. This is to be expected. Actually, this is well known. But what you can see is that, again, there's a very similar profile across species. So now when we go in and look across tissues instead of looking across species, we see that a large portion of the promoters are active across all tissues. It's by far the largest category. It's one that's shared across all of our tissues. Of course, there is a significant tissue-specific component, especially in the testes. And this pattern is mirrored in the expressed genes. Most of the genes are expressed across all tissues. There's actually a relatively small number of tissue-specific genes in these cases. And just to kind of follow up on something that I said earlier, this is the specific case of mouse. And what you can see is that the matching between the mouse tissue-specific promoters and the mouse tissue-specific genes are much closer when you look at the isoform level in the more completely annotated mouse. And so this sort of drop-off that you see here as compared to over here when we look at all species is really an artifact of the annotation. Enhancers still remain this prominent tissue-specific activity. So there isn't, as we had initially assumed, a large-scale recapture of enhancers into different tissue contexts. Now, there is some, and I'll talk about that a little bit in both cases. But again, the vast majority of these remain tissue-specific even across across all of these various species contexts. So we sought to kind of understand then in a more complete way and dividing now by those that are shared between tissues and those that are tissue-specific, the evolution of regulatory regions. So we have two groups of these regulatory regions, those that are tissue-shared, which have this characteristic, again looking across all of the species, and those that are tissue-specific. When we look at the— and these are built by pairwise comparisons between all of our possible species. When we look at the various active promoters, enhancers, and primed enhancers, you can see again that, as we saw in the single tissue case, the enhancers turn over more rapidly than promoters. But this is looking only at the tissue-shared ones. When we look at the tissue-specific ones, we see that there's a different evolutionary rate. For the active promoters and the active enhancers in that they are evolving faster in both cases. And one of the interesting things here, and it's starting to make us question things, and there'll be even more data that questions this as we come along, is that primed enhancers have no difference in the evolutionary rate between the tissue-shared and the tissue-specific. This is starting to make us wonder if primed enhancers are a thing. If they are just something that exists in the background of the genome. And so— but they do have a different evolutionary characteristic than the active enhancers and of the active promoters. So when we take a look through the lens of recently evolved enhancers, so now trying to bring these two threads together between recently evolved enhancers and tissue specificity— so now we know which ones are tissue-specific and which ones are tissue-shared— we see that the recently evolved enhancers are mostly tissue-specific. And so we set out to try to understand something about this tissue specificity and the difference between tissue-specific and tissue-shared and what other characteristics we find. And the most important characteristic that we find— where here we use a relative enrichment with blue towards tissue-shared and red towards tissue-specific is that different retrotransposons drive different— basically, there's different repeat preferences for tissue-specific and tissue-shared. Specifically, tissue-specific promoters are enriched for L1 elements, strongly so. And tissue-shared promoters are enriched for L2 elements. And this is consistent across all of the species and indeed across every way we look at this. And so there's a very striking difference between how these various retrotransposons are being captured in tissue-specific and tissue-shared contexts. When we look across active enhancers, we see a similar pattern, although not as striking as we do for the active promoters. I think this is something of a consequence of the plots that I showed earlier about how these various aspects of recently evolved and tissue specificity are starting to come together. But again, we see a similar situation where L1 is associated with tissue-specific activity and L2 with tissue-shared activity, as well as some other characteristics in there. When we looked at the primed enhancers, and again, I alluded to this in questioning what primed enhancers are, we just don't see these characteristics. And so we're starting to really wonder what these primed enhancers are and whether or not they are a thing at all that we might be interested in as we're trying to interpret eQTLs or other pieces of the genome and how we might go about doing that. I think these data at least suggest that they should be. Probably, you know, these are the types of things that the machine learning algorithms find. Promoters are very important. Shared enhancers are very important. What to do about some of these other things? The machine learning algorithms are starting to know that they're not very important. We're trying to layer on potentially why they might not be very important. So, the other question that we asked is how often these regulatory regions are repurposed or co-opted across tissues during evolution. So we have a situation where something is tissue-specific, it maintains its tissue specificity across various species, or something more complex. This is what I described earlier. When it does switch between tissues in different evolutionary lineages, we call this concept repurposed or co-opted across tissues during evolution. This is a relatively rare event, but of course it does happen, and so we we've been trying to follow up on what this means. When we look at this, we see that the species-shared elements tend to be mostly tissue-shared. That's the dominant mode of species-shared. So these two things work together. When something is species-shared, it's also likely to be tissue-shared. But there is a component that's repurposed. And in fact, the repurposed component It is, for promoters, smaller than the tissue-specific component. For enhancers, it's about the same size. And for primed enhancers, it's a much more kind of fuzzy picture of these repurposed aspects. So now I can put these two slides up together and note that although the recently evolved enhancers are at times tissue-shared, that it is a much more common situation when there are species shared to be also tissue shared. So these two things are working together in concert. So finally, we sought to try to understand something about whether we can find some evolutionary basis of repurposing. What's going on? How do these things start to switch back and forth? Are there sequence preferences? Are there other things that are giving us clues as to what's happening? And so here we focused on the difference between tissue-specific and repurposed enhancers in these various contexts. So again, tissue-specific in every species exists in the same tissue, where repurposed is switching in various tissues in various species. And one of the ways we sought to do this is we used Kmer analysis, so sequence signatures that are identified by, in this case, an SVM. So it's been known for a little while using a package called Kebabs, which was published in 2015. And in fact, Chen et al. did an analysis— actually, a couple of groups did an analysis on the liver 20 mammals data that I presented earlier. So these are other groups that have reanalyzed our data. And showed that they could predict enhancers across species by using CAMER analysis signatures. So even though you can't see the enhancer through the alignment, by looking at the specific CAMERs in the enhancers in one species, you can find the enhancers in another species. Of course, the reason that this is possible is because what's being conserved is not the regulatory regions themselves, but the transcription factors and the transcription factors sequence preferences, which are being captured in the k-mer analysis. So we wondered initially whether this was a specific characteristic of liver. It's not. You can do this basic analysis exactly the same in all species. And there are specific k-mers that I'll show on the next slide that show up as your most informative ones for determining— for driving this cross-species prediction of regulatory regions. But importantly, there's other groups of K-mers that are associated with the repurposed enhancers. When we look at those K-mers, we can't make any sense of them, so I don't have them on the next slide because they're just letters. But you can see that we are able to predict a repurposed enhancer— we are able to predict a repurposed enhancer, which is kind of right up here with all the others. By looking at the unique CAMER signatures in those enhancers. And these are— as I mentioned, we expected them to be some sort of combination of key transcription factors between one or the other, but that's not what they are. And so we're not quite sure on where we're going with that. But it does seem to be that there's some sequence characteristics that are pushing this evolution between tissues in various species. And so, just to show that, in fact, when we look at a single tissue context, some of this actually makes quite a bit of sense. These are the— basically, the most informative k-mers. They sit right down here for both the active promoters and the active enhancers in these various tissue types. Or these are among the most active— most informative k-mers. And you can see that they map relatively well to key transcription factors in these various tissue contexts. Like I said, I haven't put up the various K-mers that are associated with repurposing because we basically can't make any sense of them at this point. And so to wrap up, there's a number of things that I've tried to present and tried to pull together here. Tissue-specific regulatory regions evolve more rapidly than tissue-shared regulatory regions. We know that enhancers evolve more quickly than promoters. And in fact, only a small number of enhancers are deeply conserved across mammals. The conservation of enhancers, if you want to call it that, is taking place at the level of their sequence specificity of the transcription factors being bound, not because necessarily they're driving they're deriving from an orthologous region in a common ancestor. Of course, most tissue-specific recently evolved enhancers are derived from ancestral sequence, but thousands come from young sequence. And so this allows us to look at both of them in these cases. We see that there's extensive reuse of tissue-specific enhancers. And this occurs in different locations across species. In some cases, we have these repurposing situations.— and that tissue-specific regulatory evolution is associated with specific transposable element families as to whether they are tissue-shared or tissue-specific. And finally, when we translate from these regulatory regions to the gene expression itself, it looks like that this is an additive process, one that we call kind of a strength in numbers process. The more complex the regulatory landscape, the stronger the— again, on average, the stronger the gene expression and the more stable that gene expression is over evolutionary time. And this is reflective both in single species, in single species, as well as across mammals. And finally, for regulatory regions, the conservation status is informative, but the conservation status is not definitive. And so I think as we learn to understand how regulatory elements are conserved and what the conservation means, where it's built into the system, we can make better predictions of which regulatory elements can help us understand the ones that may contribute to various phenotypes. And so with that, I just want to basically thank the people who did this. I talked about published results from two papers that Diego and Camille were the leads on over time. The unpublished results come from Masha and Erica. This has been in collaboration with Duncan Odom's group. And of course, we are enabled to do a lot of this cross-species work because we leverage the Ensembl infrastructure, which makes many complex things much easier. And so thank you very much. I'm happy to take some questions.