If you’ve ever struggled to make sense of your ChIP Seq data, then check out this webinar presented by our colleagues at Active Motif, to get some ideas on how to bring some order to the chaos of ChIP-seq.
Presented by Brian Egan, Ph.D, Epigenetic Services Manager at Active Motif
Normalization of ChIP-seq Data
Today, I’m going to talk about a persistent problem we’ve seen for the last couple of years, and what we’ve done to try and solve that problem. And that problem is, that we do not see the expected changes in H3K27 tri-methylation in our ChIP-Seq data, after the addition of EZH2 inhibitors. Before I present that work, I want to get everyone up to speed, by giving a brief introduction to epigenetics and ChIP-Seq, and then I’ll talk a little bit about our ChIP-Seq service. And the majority of the slides will cover the spike-in strategy that we developed to normalize ChIP-Seq data.
Epigenetics refers to the functionally relevant, inheritable modifications that occur on the genome, that are not related to changes in nucleotide sequence. These modifications can concur directly on the genome in the form of methylation of cytosines , but can also occur on histone tails. There have been a number of histone modifications identified, but the most widely studied are methylation and acetylation. These modifications affect the openness, or compaction state, of the chromatin, and eventually lead to changes in gene regulation.
There are three main groups of proteins that regulate these modifications. They are loosely referred to the writers, erasers, and readers. The writers are enzymes that deposit these modifications. The erasers are enzymes that remove the modifications. And readers are a group of proteins that do not have enzymatic activity, but function through binding certain modifications, and recruiting other proteins to chromatin.
I think it’s important to note, that like inducible transcription factor binding, changes in histone modifications are inducible, and can be the endpoints of intracellular signaling cascades. To illustrate this point, I’m showing some ChIP-Seq data that we generated, and I’ve just zoomed in to a single gene. And we show that after interferon gamma treatment, we can recruit stat one to the promoter of this gene. But at the same time, we also induce H3K4 tri-methylation modifications at that promoter. So there’s a very intricate interplay between both transcription factor binding, and histone modifications, that regulate gene expression.
Obviously the field of epigenetics has grown tremendously over the past 10 years, and we’ve reached a point now, where there’s a lot of interest in exploring epigenetic marks, and modifying proteins as biomarkers or drug targets. Epigenetic modifying enzymes, and now even the reader proteins, are hot drug targets, and pharmaceutical companies are making significant investment developing drugs that target these types of proteins. And right now, probably the most well known, or the most popular targets in pharma, are against the methyltransferase EZH2 and the reader protein BRD4.
As of today, there are currently four drugs on the market that have epigenetic modes of action. Two of those are DNMT inhibitors, and two of those are HDAC inhibitors. There are a number of other inhibitors at various stages of clinical development. I think it’s safe to say, that most, if not all, large pharmaceutical companies now have programs dedicated to developing inhibitors against epigenetic modifying proteins. And most of that effort is occurring in the field of oncology But there’s also a number of smaller pharmaceutical companies that are solely dedicated to developing inhibitors against these types of proteins.
ChIP Basics
I’ve highlighted two companies in this list, Genentech and Constellation Pharmaceuticals, and the reason for that, is both of these groups collaborated with us on our project to normalize ChIP-Seq data. One of the main techniques used to detect histone modifications, as well as transcription factor binding at specific genes or across the genome, is chromatin immunoprecipitation, or ChIP. In this technique, living cells, or animal tissues, are fixed with formaldehyde.
Formaldehyde is a small molecule. It readily permeates the cell membrane, resulting in protein-protein, and protein DNA cross linking. Once the cells are fixed, the cells are sonicated to lyse the cells, and to fragment DNA. The DNA is typically about 500 base pairs long with the native proteins covalently attached. That chromatin can be used in immunoprecipitations, using an antibody against the target of interest. So you enriched for that target, but you also enriched for the DNA to which that target is bound.
Once you reverse the cross links, and get rid of the protein, you have a population of DNA that is representative of all the binding sites across the genome. And that ChIP DNA can be analyzed in different ways. The first, on a gene by gene basis using qPCR, but also the whole population can be sequenced and analyzed bioinformatically to identify where those binding sites are.
As I mentioned at the beginning, Active Motif has an epigenetic services group that performs end-to-end ChIP-Seq, and Active Motif’s ChIP technology follows the same basic steps that I outlined in the previous slide. However, over the last 10 years, we’ve developed more protocols to give optimal signal, but probably, more importantly, minimal background.
Additionally, we’ve added a number of quality control steps. So our goal is, by the time we get to sequencing, we want to have complete confidence that we’re going to get a good ChIP-Seq data set. So customers, typically, send us fixed cells, or frozen tissues. We prepare the chromatin. And as our first quality control, we quantitate the chromatin that comes from each reaction. If we’re going to do multi-sample experiments, we want to be sure to add the same amount of chromatin into every reaction.
We make recommendations for the antibodies and perform the ChIP followed by qPCR at positive and negative control sites. if a qPCR looks good, we proceed to library generation. We QC the libraries in three different ways. We then do the sequencing on an Illumina platform and analyze the data.
Epigenetic ChIP-Seq Sevices
Active Motif’s Epigenetics Services Group actually started as a different company back in 2002 called Genpathway. We recognized that chromatin immunoprecipitation was going to be a valuable technique, and started developing our protocols. In 2004, we did our first paid ChIP-qPCR assay. We now generated more than 3,000 ChIP-Seq data sets, which is certainly more than any other commercial group. We’ve successfully performed ChIP-using antibodies to more than 220 different targets. We’ve processed over 12,000 samples for ChIP in that time frame, and we prepped chromatin and done successful ChIP on over 15 different species and 25 different cell types.
Because of the nature of the services business, we have had the privilege of working with a lot of really great scientists over the years, and much of that work has ended up being published. And as a scientist who’s worked in the field of epigenetics for quite some time now, it’s exciting to see epigenetics migrating into other fields. Epigenetics is becoming a core discipline within gene regulation, and is starting to move into other areas. You’ll see there’s a few papers here related to oncology. That’s no surprise. A lot of the work in epigenetics is focused on cancer. We also see things like circadian rhythms, inherited behaviors with epigenetic roots, as well as immunology.
ChIP-Seq Reveals Small, Sample-to-Sample Changes
There’s one more topic that I want to cover today, before I present the spike in normalization strategy that we’ve been working on. And that topic is, how good is ChIP-Seq as a technique at detecting differences across multiple samples. I’ve talked to many people about ChIP-Seq, and I’ve heard a lot of people say, that ChIP-Seq is good for detecting binding versus no binding, but it’s not necessarily good for quantitating or detecting degrees of binding.
I think it’s widely accepted that you can detect on-off situations. I think a good model for this, is in stem cells. As the cells differentiate, you can clearly see both activating in marks, and repressive marks shifting during that differentiation. The story’s not so clear when it comes to detecting degrees of changes at commonly bound sites across multiple samples. And since many people have seemed to have problems detecting these changes, there’s been a number of publications over the last few years trying to address this problem bioinformatically.
And personally, I’m not so sure that these bioinformatic approaches makes huge differences to the data sets. Maybe for low quality ChIP data sets you can improve them by using these types of techniques, but I’m of the opinion that if you’re doing ChIP well, you should be able to detect degrees of binding across multiple samples. Of course, this may be influenced by cell type and antibody, but I think in general it can be done.
Here are a few examples of those types of studies that we have performed. So in this first example, we were looking at a specific histone modification in the presence of an inhibitor that should lead to a decrease in that modification. Without the inhibitor, we see the mark at a few genes here, and as we increase the concentration, we see that mark declining. The entire data sets for this project look very similar. I’ve just zoomed in a few areas to illustrate the point that you can see degrees of binding.
The same is true when looking at transcription factors. We’ve had many assays where we’ve induced transcription factors over time, and have seen those changes. Here’s an example of an assay where we looked at estrogen receptor binding on after estrogen treatment.Without the treatment, we do see a little bit of background binding for estrogen receptor, but at 30 minutes that signal goes up, continues to go up at one hours and six hours. I think these are good illustrations of how you can detect degrees of binding, and not necessarily just on-off situations with ChIP-Seq.
Challenges of ChIP-Seq Data Normalization
In the project I’m going to talk about today, we did not see these differences. And that’s what led us down the road of developing a spike-in system that allows us to see these differences. In that project, we did a ChIP-Seq in the presence of EZH2 inhibitors, and we were not detecting the expected decrease in H3K27 tri-methylation and occupancy. ECH2 is a methyltransferase. It methylates H3K27. ECH2 is also a cancer target. It’s frequently over expressed in cancer, and there are some cancers have activating mutations. And these include diffuse large B-cell lymphomas, follicular lymphomas, and melanomas.
As such, there’s a lot of interest from pharmaceutical companies to develop inhibitors to EZH2, with the idea that you could bring K27 tri-methylation levels back down to something that is closer to normal, and this would have some therapeutic benefit. So there are a lot of groups developing EZH2 inhibitors and characterizing them. And some the questions they’re asking are how do these inhibitors influence global K27 tri-methylation patterns. How do these epigenetic changes influence gene regulation? And how do these inhibitors influence clinical outcome?
The question we can contribute to with ChIP-Seq, is how the EZH2 inhibitors change global H3K27 tri-methylation patterns. We’ve done a number of K27 trimethyl ChIP-Seq projects over the years, using a number of EZH2 inhibitors, and in every single case we have not been able to see the expected decrease in K27 tri-methylation. So here’s a figure that encompasses four different experiments, using four different EZH2 inhibitors. And you can see, in each case in the presence of the inhibitor, we’re not seeing a decrease in K27 tri-methylation.
This is a big problem for the people developing the inhibitors, because they’ve shown in biochemical assays, that the inhibitors actually do inhibit. And they’ve also seen global decreases in K27 tri-methylation as measured by mass spec in western blot. So once we identified this problem, we were lucky enough to team up with two different groups. The one at Genentech, led by Marie Classon, and the one at Constellation, led by Patrick Trojer.
We have some data from those groups. This Western blot, showing that the inhibitor used by Genentech, which is GSK126, results in a global reduction in K27 tri-methylation levels. And at the same time, a slight increase in K27 acetylation levels.
In the constellation samples, one of their compounds that they’re using in this particular experiment, is CPI360. So by Western blot, you can detect global K27 tri-methylation levels. They’re decreased at two days after treatment, and significantly decreased at four days and eight days.
Now the question is, why did we not see these expected decreases in the ChIP-Seq data? Is it a function of the ChIP reaction, that is, are those differences present in the ChIP DNA, or is it a function of the downstream processing and preparation for sequencing?
The answer to the first question is, Yes. We actually can detect those differences the ChIP reaction. So here’s three separate experiments using three different EZH2 inhibitors. We perform ChIP followed by qPCR at selected loci. And in each case, you can see somewhere between a two and four-fold decrease in K27 occupancy in the presence of the inhibitor.
As I mentioned, we’ve done many other inhibitors. Here’s an example of an experiment where we did BRD4 inhibitor. We increased the concentration of the drug, we see a decrease in BRD4 binding at its target chain. So we can clearly see these differences in the ChIP DNA. So it must be a problem in processing the samples for library generation sequencing.
So at what steps might we be losing this difference? Well, the first would be during library generation. Library generation requires PCR amplification. And for us, we do 15-18 cycles of PCR to get enough library to sequence. So you can imagine, that during an amplification, you may be losing those differences between the two samples.
The second opportunity is actually during cluster generation before sequencing. So if you’re doing multiple samples with bar coding, it’s common practice that when you prepare your flow cell, you want to have equal cluster density for each sample. So if there are quantitative differences in the amount of library, you will actually normalize them out during that step of cluster generation.
And the last point is, it could occur bioinformatically. It’s also standard practice when doing multiple samples, after the sequencing is complete, to normalize those data sets to a common number of sequence tags. So for us, if for example, if we’re doing five-sample experiment, we might normalize all those samples so that there’s 20 million sequence tags in each data set.
ChIP-Seq Spike in Controls
So we started exploring ways that we might correct for this problem, and the approach we settled on, was to use a spike-in control that is added equally to all the samples, and we could follow that spike-in through the procedure, and normalize back to that spike-in at the end. There is precedence for using a spike-in strategy in a whole genome setting, in particular, a Rick Young’s group in 2012. They were looking at gene expression in normal cells and in cells over-expressing myc. The cells that over-express myc, have a dramatic up-regulation of the entire transcriptional profile of the cell.
The expectation is that everything in those cells are being up-regulated. And in fact, those cells that over-express myc, have two to three-fold more RNA than a normal cell. When they did the RNA Seq experiment, they actually saw that there was only a subset of the genes that were being up-regulated. Many were not changing, and some where being down-regulated. So they decided to use a spike-in strategy, and in their strategy, they use 92 synthetic polyadenylated transcripts that were added to the RNA prep. When they repeated the RNA Seq experiment with the spike-in, and normalized back to the spike-in, they now see the expected result, which is everything is being transcriptionally up-regulated.
The problem we were experiencing, is not dissimilar to the problem I just described from the paper. And at its most basic level that is, that we’re expecting global changes, and we’re not detecting those in our whole genome read out. So we also decided to use a spike-in strategy. We decided to use a small amount of drosophila chromatin, and spike that chromatin into our human ChIP-Seq reactions. Not only did we just spike-in the drosophila chromatin, we also added a second antibody. And this antibody is against the drosophila specific H2A.V variant.
The idea was, that if we added this antibody, we could consistently pull out a small portion of the drosophila genome for normalization. If we use this antibody in ChIP-Seq using only drosophila chromatin, we get many thousands of peaks across the genome, showing that the antibody does work in ChIP-Seq. If we use that antibody with only human chromatin, we don’t get any peaks, indicating that the antibody doesn’t cross react with human chromatin.
Just to summarize what those reactions look like; in a single tube, we have human chromatin, the H3K27 trimethyl antibody, a small amount of drosophila chromatin, which is roughly equivalent to 1/40th of a human chromatin, and the H2A.V antibody. After we do the sequencing, we take that data, and we map it both to the drosophila genome, and to the human genome. And if there are any differences in the number of tags across samples that are mapped leading to drosophila genome, we normalize those to make them even, and then we’ll normalize the human data set by the same ratio.
Performance of ChIP-seq Spike-in Normalization
When we incorporated that strategy, the first thing we saw was that we consistently saw more tags mapping to drosophila in our H3K27 trimethyl IPs in the presence of the EZH2 inhibitors. So if we look at the Genentech samples, we have tags mapping to drosophila from the DMSO sample, but significantly more tags mapping in the inhibitor treated sample. And we saw the same thing in the Constellation experiments, so, that we have tags mapping in the DMSO sample, but many more tags mapping to drosophila both at four days and eight days of inhibitor treatment.
The idea is, that we’re going to normalize these tag numbers, so that they’re equal. So we’ll normalize down the tag count in the inhibitor treated samples, and we’ll normalize down the tag count in a human sample by the same ratio. We also used control antibodies in these reactions. For Genentech, one of the controls was H3K27 acetylation. I’d mentioned earlier, that we actually would expect to see an increase in signal from those samples. If we map to drosophila, we’re actually getting fewer tags mapping in the inhibitor treated sample. So we’re going to normalize these up, so that they’re even, and we’ll normalize up the human tag count as well. That’s exactly what we would expect for this particular antibody.
The control antibody in the Constellation samples, was H3K9 trimethyl. This mark should not be changing, and we don’t see significant changes in the tag counts across those samples. If we apply that normalization strategy to the samples, this is what the data looks like. So this is with the Genentech inhibitor. If we use our normal ChIP-Seq protocol, we don’t detect differences in K27 trimethyl between DMSO and inhibitor treated cells. We also do not detect differences in the control antibody. However, after spike-in normalization, we now see a dramatic reduction in the signal from K27 trimethyl ChIP-Seqs in the presence of the inhibitor, with no significant effect on the H3K4 trimethyl control antibody.
The first time we ran this experiment, we saw five-fold reduction in signal in the inhibitor treated cell. The technical replicate showed a 4.54-fold reduction. And we just recently got back the biological replicate that also showed a five-fold reduction. And this was right in line with what we were expecting, based on Western blot data.
To show how dramatic this normalization is now, I’ve just zoomed in on a single location in the genome. Again, no major changes in using the normal ChIP-Seq protocol. But after we incorporate globally, we’ve taken all the data and plotted it with box plots and scatter plots.
For the scatter plot, each point represents a peak that was identified in the data set in the EZH2 inhibitor treated sample, versus the DMSO sample. And if these two samples are the same, that data wouldn’t line up nicely along this diagonal. So you can see, using the normal ChIP-Seq strategy, those samples are quite similar.
After we incorporate the normalization, you see a dramatic reduction in the EZH2 inhibitor treated cells. The H3K4 tri-methylation antibody, we don’t see a big difference between EZH2 and DMSO treated cells, and the spike-in doesn’t really affect that data. And for K27 acetyl which were expected to go up, the spike-in actually does normalize that data up.
So to summarize by looking at the box plots, we see no difference using our standard protocol in K27 trimethyl, but a dramatic reduction with the spike-in strategy. The control antibody is not changing, and that also doesn’t change in the presence of the spike-in. And we see no significant difference in K27 acetyl using our normal protocol, but a slight increase if we do the spike-in normalization.
Applying Advanced ChIP-Seq Normalization Strategies
We applied the same strategy using the Genentech samples. Here we had DMSO treated samples, four days of inhibitor, and eight days of inhibitor. Using our normal strategy, we see no differences in binding across those samples. After the spike-in, we see that we now have a dramatic reduction both at four days and eight days.
Again, if we look at that data globally with the scatter plots, we see the same effect, that without normalization those EZH2 and DMS inhibitor, and DMSO treated samples are very similar. And the spike-in dramatically reduces the signal from the EZH2 inhibitor samples. And that’s true both for the four day and an eight day of inhibitor. Our control antibody, which shouldn’t be changing, again, is very similar between treatment and no treatment, and the spike-in doesn’t change that data. And you can see that clearly, again, looking at the box plots.
So what do we think is going on? I think the first thing, is to note that there are differences, depending on what antibody you use. When we do ChIP-Seq in the presence of a BRD4 inhibitor, we can clearly see the reduced signal in the presence of the inhibitor. So we don’t know how globally this is going to be applied to different transcription factors, or antibodies, but we do know that it’s necessary when there is global changes for K27 trimethyl. And we see the shift our data as expected with the K27 acetyl antibody as well.
What we think is going on, is when you use an antibody against H3K27 trimethyl, that antibody is specifically pulling out the majority of the DNA in the reaction. And only a minority of the DNA is coming from background, or from in our case, the spike-in. So in the presence of the inhibitor, you’re pulling out much less DNA, but the background and spike-in is remaining consistent. So in the end, you have more ChIP DNA in one sample, versus the other sample.
During the amplification, the one with less DNA becomes over-amplified, evening out the signal from K27 trimethyl. But during the over amplification, you over amplify, and you can see that over amplification in the drosophila spike-in. So if you normalize this out, you can recover those initial differences.
For BRD4, we actually think that the majority of the ChIP reaction is background, and a minority of BRD4 specifically pulled down DNA, makes up a minority of that reaction. So in this case, the amount of ChIP DNA coming out between an DMSO and inhibitor treated samples is equal. When you do the amplification, they now amplify equally, and you maintain that difference in the BRD4 samples, and the inhibitor treated sample.
Using Spike-ins to Measure ChIP Consistency
The last point I want to make before we wrap things up, is that the drosophila spike-in is not only limited to normalization of global differences. We’ve also used it to monitor the quality and consistency of our ChIP reactions. So there were a lot of steps in ChIP, and a lot of opportunities for errors when processing the samples. And you can monitor the consistency using qPCR of drosophila sites.
So an example is in the data I’m presenting here. We did three separate experiments with K27 trimethylation antibody using the same chromatin. And so we had samples that were DMSO, and again, EZH2 inhibitor treated cells.
In the first two samples, we see the expected decrease in the presence of the inhibitor by qPCR. If we look at the drosophila signal, using drosophila specific primers, we see that we have a very even signal. So that shows us that there is consistency and no loss of sample in our chip reactions.
We actually don’t see the expected differences when we do K27 tri-methylation ChIP followed by qPCR. And if we do the drosophila qPCR, we don’t see that these reactions are not even. So we had sample loss or some technical error in that first sample. So if we normalize these to become even, and normalize the human data by the same ratio, we now see data that is very similar to the experiments that had good technical consistency.
So I think if you’re doing multiple sample reactions, and you’re expecting differences, this is a good mechanism to show that those differences are real, and not due to errors in processing. Additionally, if you’re going to do ChIP-Seq, and you’re going to make the investment in generating libraries and paying for sequencing, it’s worthwhile to do these quality control checks upfront, before you move on to sequencing.
In conclusion, what I’ve shown is that over the years we’ve not been able to detect global differences in K27 tri-methylation by ChIP-Seq. So if we use standard protocols, we don’t see these differences. And we’ve come up with a spike-in strategy that now allows us to consistently detect global changes in K27 tri-methylation. And in the last slide, I showed you that a similar spike-in strategy can be used to actually monitor the quality and consistency of your ChIP reactions. And that can be done both for ChIP-Seq, and in a ChIP qPCR setting.