Scientist Stories: Manolis Kellis, Integrative Analysis of 10,000 Epigenomic Maps

Hello everyone. I'm Manolis Kellis, and today I'll be presenting our work on integrative analysis of 10,000 genomic maps across 800 samples for regulatory genomics and disease dissection. This is primarily the work of Carlos Boix, a graduate student in my lab, in collaboration with Ben James, Youngjin Park, and Wouter Meijerman. So he's been combining ENCODE, Roadmap Epigenomics, genomics of gene regulation, systematic chromatin imputation, genome-wide association studies to understand gene regulation and disease. And the paper is on bioRxiv. You can just Google that and you will find it. So what is our goal? Our goal is to enable personalized medicine, enable precision medicine, and enable us to recognize the molecular basis of human disease. And the way to do that is to start with human genetics and then try to understand the mechanism through which genetic variants are acting to develop new target genes, new therapeutics, and enable precision medicine and personalized medicine. The challenge, of course, is that 93% of disease association studies are in fact lying— their hits are lying in non-coding regions where the target gene is not known out of many genetic variants in association. In this particular case for FTO, we have shown that the true target gene is actually 1.2 million nucleotides away and 600,000 nucleotides away. So the target gene is not known. The causal variant is not known. In this example, 89 common variants are co-inherited, but only one of them appears to be the causal variant. The cell type of action is not known. In this particular case, these regions could be active in many different locations, and it's not known from non-coding regions how do you systematically find the true target genes. And of course, the relevant pathways are not known and the mechanism is not known. So our team and many others in the field have basically taken this approach for dissecting the mechanism underlying these genetic associations. So we start with disease genetics across both common and rare variants. We then systematically profile RNA and the epigenome in healthy and disease samples And we integrate, of course, the data systematically to predict driver genes, regions, and cell types. And of course, we validate our predictions in human cells and mouse models. For EpiMap, the work that I'm presenting today, what Carlos did is he systematically sought to integrate ENCODE, Roadmap Epigenomics, and GGR as a way to build a resource that allows us to systematically enable these studies with large-scale datasets. And in particular, he sought to uniformly process 3,000 different chromatin and RNA tracks, systematically impute an additional 14,000 tracks, and then carry out a lot of QC and validation, followed by a high-resolution annotation of enhancer regions based on their presence in enhancer chromatin states based on H3K27 acetylation signal and of course DNA hypersensitivity sites to get at the high resolution. He then studied the activity patterns of all those elements and then inferred both the relationships of the different samples to each other and modules of activity—of coordinated activity—across all these samples. He then built links from transcription factors to their downstream target enhancers using their motif enrichment, and from the enhancers to their target genes using correlated activity between RNA and epigenomics. He then integrated all of these data with genome-wide association studies to infer the tissue, the target, and the variants underlying these traits, and then also studied the interaction and integration between pairs of traits acting together and pairs of tissues acting together. So the contributions are threefold. Number one, a data resource which you can find at compire.mit.edu/epimap. And this is browsable, interactive. You can download specific combinations of data as well as browse all of the regulatory genomics and disease dissection results. The second contribution is a lot of insights on regulatory genomics, namely what are the modules of enhancers acting together, what are their target genes, what are their upstream regulators and motifs and their combinations across many different human tissues. And then systematically dissecting the circuitry of complex trait loci. So in this particular case, you see some examples of coronary arteries, for example, looking at both liver and heart loci, as well as a combination of liver and heart acting together across multiple variants or a single variant in some cases. So here's the combined resource. We basically have 14,000 epigenomic tracks across 833 samples and 18 different assays. You can see here the life stage, sex, type, and the project, as well as the group of tissues that we're talking about. These are the 33 different groups that we have defined across 4 different types of samples, 5 different life stages, the 2 sexes, and of course the 5 projects here, as well as the imputation status of these datasets, where most of them, as you can see, are imputed, but also some of them are both imputed and observed, and others are only observed. We've also grouped these into different tiers based on the density of coverage. So Tier 1 is quite densely covered from observed data, whereas Tier 2 and Tier 3 are at intermediate coverage, and Tier 4 is at much, much lighter coverage from observed data. The basic methodology is ChromImpute. This is a tool that Jason Ernst developed in my lab a few years ago, looking at the combination of within sample and within-MARC inferences that can together allow us to infer a target predicted dataset before actually observing it in very high resolution. Here are some examples of what these predictions look like, followed by the observation here in gray, which you can see matches quite nicely thanks to the similarity of of different marks within the same cell type, as well as the similarity of different cell types for the same mark. So doing this systematically, this is what the epimap data looks like across many different epigenomic marks and many different loci. So here's 3 different loci selected at random at 25 kb, 200 kb, and 1.5 megabases. You can see that very strong agreement between a large diversity of marks and observed and imputed data. And here's across 2,000 different locations. This is similarly this very strong agreement in the genome-wide occurrences of these maps across ubiquitous enhancers and promoters as well as these tissue-specific ones. What's really cool is that the agreement between observed and imputed data is usually found, but when it's not found, it's not that the imputed data is wrong, it's usually that the observed data is wrong. So we have used discrepancies between observed and imputed data as a way to flag low-quality datasets. And you can see here that the flagged ones are in fact showing very low strength correlation quality control scores compared to the ones that show strong agreement where these scores are pretty high. We've also used independent experimental data generated after our imputations to basically show that the predicted datasets are in fact of very high quality and then they surpass these very stringent benchmark of the nearest track. The nearest track imputation basically means that I'm going to choose the observed dataset which is closest after observing that data. Of course, this is unattainable in practice because if you can't find the nearest dataset by observing the data, well, you've observed the data, so we have the data. So here we're basically saying even if we knew the entire track and we chose the closest one, how well would we do? And the answer is that imputation, in fact, without knowing the observed track, does better in 77% of cases for broadmarks and in 96% of cases for punctate marks, which is quite remarkable even compared to this unrealistic comparison. You can see that the imputed data is surpassing in quality in nearly all cases. We've also used these very high-quality imputations to now study the cross-sample relationships using both activating marks and repressive marks and then understanding what are the primary drivers between tissues and embryonic versus adult cell lines. And we're finding in general that the active marks are primarily driven by lineage, whereas the repressed marks are primarily driven by stage of differentiation. This is the tree that we get based on the sharing of these high-resolution enhancers across the 833 samples. You can see that the biological functions of the tissues that we have now classified are in fact making a lot of sense. We've also annotated chromatin states using ChromHMM, also developed by Jason Ernst, previously applied across 127 tissues of the Roadmap Epigenomics project, now applied across the 833 tissues of APIMAP. And you can see here the imputed data is quite substantial in its coverage and quite high quality in its resolution. And you can see here a browser that Wouter Mellemans has built called Epilogos, initially developed in my lab, now in his own lab, where you can see the tracking of these samples and the coverage of different chromatin states in a very rapid and high-quality interface across the genome. This is what the modules look like. You see basically these 300 active enhancer clusters that we call modules. 97% of them are tissue-specific. 3% of them are in fact highly shared across broad tissues. You can see here both across organs, across primary cells, across brain, embryonic stem cells, the clustering and activity patterns of these modules. And you can see here their true size on the top and then the 300 modules uniformly spaced right below. These modules are very enriched in gene ontology terms for the genes nearby. So you can see here the specific functions associated with these multiply active modules across multiple different tissues or these tissue-specific modules. For heart, for example, you see a lot of very specific functions associated with the nearby genes. And I've highlighted some of those here. We can also look at the upstream regulation of these enhancer modules that sort of probably drives the way that they act together. And you see 89— sorry, 86 different motif family enrichments. These are group motifs. But then on the website, you can separate them out into the individual members of those motifs. And then you can see that there's a small subset of very broadly active motifs. That are partnering up and collaborating with multiple tissue-specific motifs to define these tissue-specific gene regulatory patterns. These tissue categories are also partitioning these biosamples based on their relevant structures, motifs, and combinatorial motif interactions. So this is a snapshot from our interactive browser where you can click on individual motifs, individual factors, individual tissues, and look at in detail the regulatory inferences that are resulting from this. Ben James has basically worked with CARTAS to predict enhancer gene links systematically across a large number of tissues. And in a subset of those, we've been able to validate them based on gold standard comparisons and to also compare them with other predictions. And what you can see is that the blue predictions shown here from EpiMap are in fact performing quite, quite well in most of the benchmarks compared to all of these other methods. We then turned to use these regulatory circuitry, these enhancer annotations, and these motif annotations to prioritize genetic variants across hundreds of traits. Here I'm showing 250 different traits that are showing enrichment for individual enhancers, for individual tissues, for all of the enhancers across individual tissues.. And you can see the enrichment is quite cell type specific for the vast majority of these traits, which we call unifactorial traits. But then for a subset of polyfactorial traits and multifactorial traits shown in gray and in red here— and in blue— you can see that the enrichments are in fact spanning many, many different tissues, indicating that these traits are in fact acting quite, quite broadly. And you can see some of those traits here. The number of enriched tissues is dramatically increased from previous annotations to now more than 200 with this approach and more than 500 using this tree-based approach that we also developed. This analysis allows us to now dive into individual loci and start looking at what are the genome-wide association signals at the level of individual SNPs. And how those are in fact overlapping specific enhancers and which target genes are these enhancers linked to. In this particular case, you can see this breast cancer association is in fact pinpointing association with specifically breast cancer samples as well as breast cancer— sorry, breast epithelium. Here's another example with a schizophrenia association pinpointing this CAC calcium receptor. In brain-specific enhancers. And you can see these very broad regions of association corresponding to these broad enhancers, in contrast to these very narrow regions of association corresponding to these narrow enhancers. We've systematically looked at the interactions and co-enrichments of many different traits across many different tissues. So every circle here is one trait, where the pie chart represents the tissues that are enriched for that trait. You see a lot of tissue-specific traits which have uniform coloring. For example, here in brain, you see these self-reported math abilities, schizophrenia, and other psychiatric disorders localizing quite specifically in brain only, as opposed to Alzheimer's, which is in fact localizing primarily in immune traits, as we have previously shown— as we and others have previously shown. You can see these filtering functions in kidney, these heart and pulse and blood pressure function in heart and these cholesterol functions in liver. But you also see in the middle these multifactorial traits, which are quite spread across multiple tissues. One striking example of that is coronary artery disease, which is in fact showing the largest number of enrichments across tissues and is an extremely polyfactorial trait. The primary tissues are liver, coronary artery disease, but a whole host of additional tissues appear to be enriched. What's really exciting about this is that if you look at the partitioning of these genetic loci across the different tissues, you can see here that the liver and overlapping SNPs— the SNPs that overlap enhancers active in liver— are in fact enriched in a very different set of biological processes than the SNPs overlapping cardiac enhancers and so on and so forth. So you see this very nice partitioning of the biological functions underlying coronary artery disease based on the tissues, but also this partitioning of the comorbidity patterns with other traits that are also enriched in the same SNPs. And you can see here in greater detail this partitioning. This partitioning is also found for individual loci. So you can see that some of the loci associated with coronary artery disease are in fact enriched only in coronary artery, where others are enriched only in liver. And of course, some loci are enriched in combinations of those. Here's a liver-specific enrichment where you can see this liver enhancer targeting this liver gene in liver. And you can see this cardiac enhancer again linking to cardiac active gene. But in other examples, it's not as simple. You basically have these primarily liver-like looking location but it also has several enhancers active in heart, which are also linked to a heart gene that appears to also play a role in coronary artery disease. So this is very encouraging. Basically provides a way of starting to dissect these complex traits by partitioning them into tissues. And these are the same loci in just higher resolution. So overall, we've basically built this tool that is very helpful for uniform processing of all these datasets, generating these large resources, but also using it for in regulatory studies and for disease studies. And these are some snapshots from this interactive website that you can find here, enabling you to look at custom track hub creation, modules, motifs, networks, GWAS enrichments, and per locus visualization across 30,000 loci. You can basically browse by GWAS region, by GWAS study, and then GWAS region, and then individual SNPs within these regions, and then the specific enhancers, tissues where they're active, and target genes where they're linked. So I really encourage you to explore this. This is teamwork by— led by Carlos Boix, but engaging several members of my lab, both current and former, as well as these resources. So very happy to take any questions. Thank you so much for your attention. Bye-bye.