Scientist Stories: Mark Gerstein, Using ENCODE Data for Cancer Genomics

Hi, so I'm going to talk about using the— using ENCODE data for cancer genomics. So first of all, I'd just like to point out that lots of the ENCODE data is relevant to cancer. Many of the ENCODE cell lines, as you probably know, are associated— are cancerous cell lines. For instance, HepG2, A549, MCF7, and so forth. And these often can be easily paired with a normal, and they give you a sense of the oncogenic transformation. Furthermore, a number of these cell lines, K562, HepG2, and so forth, are extremely data-rich and have a tremendous number of assays on them and are really unprecedented. In that way for studying oncogenesis. So I'm going to talk about using this data in 4 ways to look at cancer genomics. First, I'm going to talk about background mutation rate correction. Then I'm going to talk about studying network wiring, variant prioritization, and then looking at drivers of differential expression. So first of all, background mutation rate estimation. One of the main ways we find drivers in cancer is through looking at mutational recurrence within a cohort. One of the problems though is that often this mutational recurrence can be confounded by genomic covariates such as replication times. We have a lot more mutations in late replicating regions than early replicating regions, so if you look at recurrence, well, you might be fooled and think, for instance, this region is a driver region because it has a lot of mutations when it just simply has lots of mutations because it's late replicating. So we have to take this into account. Now the mutation rate is correlated with many different types of genomic signals in addition to replication timing. For instance, it's also correlated with open chromatin. And ENCODE has, of course, a wealth of this genomic signal data. So we've shown how you can put a lot of this data together into a model to estimate the background mutation rate. Basically, we do principal components of all the different types of ENCODE signal data. And then we use these principal components to estimate the background mutation rate And one of the main things we see is that we need lots of PCs, often 10, 20 PCs, to accurately estimate background mutation rate. And then, of course, we can look at the signals that really are driving these PCs. There are things, of course, like replication timing, but a lot of the histone mark data really also is very important for this estimation. So now what do we do with an accurate estimation of background mutation rate? Well, we can use that in a model to look at recurrent mutations and to find drivers. Now, a simple model A model that doesn't do that would be, for instance, a binomial model, which assumes a constant rate of mutation across the genome. We're going to contrast things to this. We have two different models, a beta binomial model and a negative binomial model, that allow the mutation rate to vary across the genomic events and this mutation rate to vary with genomic covariates, so replication timing, histone marker data, and so forth. So what do we get when we look at this? Well, here's for instance the empirical distribution of mutations in genomic bins. Here's the simple binomial model. It obviously doesn't fit very well. And here's our correction model. You can see it fits a bit better. And this is for the beta binomial. You can do the same thing for the negative binomial. And then when you come to actually finding recurrently mutated regions of the genome, having this better model deflates your p-values. You don't get these inflated p-values versus if you simply have a simple binomial model, you'll call lots of regions of the genome as recurrently mutated. Whereas if you have our beta binomial model that takes into account the genomic covariates, you'll be much more stingy with finding recurrently mutated regions. You can see this, of course, in other types of graphical formats. Okay, so let's go on to the next topic, which is genomic rewiring in cancer. So let me sort of tell you about this. I'm going to focus on one particular transformation that from K562 to GM12878. GM12878 represents a normal blood cell, K562 represents a white blood cell tumor. And so now for these cell lines where we have tremendous amounts of encode TF-ChIP data, we can actually look at the changes in the regulation and how that regulatory network rewires. So for instance, if we look at the tumor, we have some connections there. If we look at the normal, we'll have different connections. If we compare them, we'll see some of the connections are lost some are gained, and some are retained. Now here's a kind of global picture of that. It's a very complicated hairball when you look at this, of course, across all the different targets. This just shows the 109 regulators, not all their targets, because it's too complicated of a picture. But even here you can see the things that rewire a lot and all the changed connections. It's a useful picture to look at; however, often these pictures are complicated, and particularly if you have all the targets, it'll be very complicated. So we try to develop ways of simplifying simplifying these pictures. And one way we have of doing this is not looking at the change of regulation to all the targets but grouping our targets into pathways or modules or gene communities and then looking at the change in regulation of these communities. Okay, and the idea is there's far fewer communities than there are targets. Now there's a nice mathematical trick for doing this, it's called latent Dirichlet allocation, which is a method often used in text processing where you can look at the underlying topics in documents. Where we think of the documents as like transcription factors, one of the topics of these modules. So we repurpose this calculation methods, and here's kind of what we find when we look at things. Let's just focus on this GM12878 to K562 transition first. We have all the different transcription factors here, and for instance, we can look at— for instance, let's take a look at MYC. We can look at the edges that are gained in cancer. This is a model for CML, chronic myelogenous leukemia. We can see lots are gained, very few are and conversely NBN is a factor where it loses a lot of things. We can just sort all the factors by how much they gain or lose edges, but perhaps more useful is we can also sort them in terms of these gene communities. And this is a kind of useful thing to understand which of the factors are really associated with the change in oncogenesis. Okay, let's go on to the next topic. Just giving you a quick run-through. We're going to look at variant prioritization. Now variant prioritization obviously is a big topic for ENCODE in general. People are always using the ENCODE data to prioritize variants I'm going to focus here on just one particular type of data that's new in ENCODE. This is the RNA binding proteins. Before the last ENCODE, there was not very much of this data. Now there's quite a bit from the eCLIP experiments. Two things that are useful to know about this. First of all, the eCLIP experiments cover a large fraction of genomic real estate, actually more than CDS. But these experiments are actually very precise. The binding sites in them are much more precise actually than TF binding sites. So lots of annotation. Data, and so we tried to build some simple pipeline we call RADAR for putting all this data together to annotate variants. We obviously put all the data together, we put it together with conservation data both cross-organism and within the human population. Since we're looking at RNAs, we have to include secondary structure. We use all these to make a score. We have this entropy weighting scheme for putting all these things together to make a score, and then we also combine with kind of tissue-specific score that's related to the cell line that you're looking at, the tissue context that you're looking at. Just a little bit more on these features, you can see that each of the binding protein sites is conserved really to a different degree, which really highlights the fact of why you want to use conservation. And also a lot of the binding proteins, as I'll show you in a bit, they actually come together to make a network, and there's a lot of co-binding in these proteins. And so we obviously want more highly-weighted network hubs in the RNA binding protein network. Than not, and we take that into account in our procedure. And briefly, we find that our procedure gives higher scores to cosmic genes known as cancer genes than not. And also, it tends to give higher scores to recurrently mutated regions, for instance, in breast cancer than not, which gives us some confidence that the score is useful in a cancer context. And now to the last topic, which is finding regulatory drivers of differential expression. So I've told you a second ago in oncogenesis how you might have the actual regulatory connections between the TF and the target change, you know, rewiring, but you could also have a situation where the targets remain the same but the gene expression of the target genes changes a lot. Okay, and here what we want to do is find the TFs that are associated with the target genes that change most in cancer. So we use the TCGA data, which shows how genes changing cancer, and we use the ENCODE regulatory network to associate regulators with targets. Then we build a simple regression model that tries to explain the target— the change in gene expression from all the regulators. We can find the regulators that are most highly weighted in this expression model, and that's what we plot here. So we have all the different regulators. We have all the different cancers. And this just shows these regression coefficients, which ones get most highly weighted. And you can see that some factors really are associated with upregulation in cancer. MYC is very famous for being associated with that. And so, some are more differential. We can also do the same for RNA binding proteins and their network. And here we find that SUB1 is the RNA binding protein that's most associated with upregulation in cancer. And it turns out you can also look at survival curves, a well-known one for MYC, for instance. We have more active MYC, you tend to have differential survival. You can see the same thing for SUB1 for the RNA binding proteins. And also now to put the RNA binding proteins and the TFs together, we can build up a regulatory hierarchy just like we showed for the TFs. And here we find that the things at the top of the hierarchy, both the RNA and also the TFs, tend to drive gene expression the most. So here's the TFs, and here's the RNA, and here's the average correlation with change in gene expression. There's a lot of crosstalk between the TF and RNA networks. And just to sort of highlight here, here's MYC and sub1, and we just point out that these can really act together on the same genes, in fact, promoting transcription and also post-transcriptional regulation. OK, so with that, I'm going to conclude on my 4 topics. And let me just go through them now in more detail. So what have I told you? I've told you how the ENCODE data is useful for cancer genomics. First way, it's useful for developing accurate models for background mutation rate. We've developed these parametric models for doing this. One of the key things is we have to use lots of data, often 10, 20 different principal components to accurately model background mutation rate. Next thing I've told you about is how lots of the encode cell lines are obviously associated with cancer, and we can pair them with normals, and we can look at the rewiring of the TF regulatory network in some of these cell lines with lots of data, particularly those associated with CML. Here we see lots of complex rewiring. It's very interesting, and we've developed this procedure associated with latent Dirichlet allocation to simplify this. To look at the changes in modules or changes in gene communities. And the third thing we talked about is using the ENCODE data for variant prioritization. We have a pipeline called RADAR which uses the RNA binding protein data along with secondary structure conservation and so forth and prioritizes variants. And then finally I talked about using the ENCODE data to find the regulators of RNA binding proteins or TFs that are associated with driving differential expression by using a simple regression model and then putting these regulators into a hierarchy and seeing that the ones that drive the differential expression tend to be at the top and there's also a bit of crosstalk between the RNA and the TFs. Okay, with that I think I'll conclude. I thank you for your attention. I want to acknowledge the people that worked on this. I think the principal scientist that led most of this work was Jing Song, who was an associate research scientist with me for a number of years. She's now moved on to a faculty position at Irvine. She worked very closely with two graduate students in the lab, Jason Lu and Dongmin Lee. I show all the different tools they developed with the URLs here. And also I'll say we have lots of openings for people who wanted to do ENCODE-related research in the lab, and you can just go to our jobs page if you're interested. Thank you very much for your attention.