DNA methylation: from array to sequencing

Editor's Notes

• #2: Good afternoon & thanks to the Winter School for inviting me to tell you all about how we measure DNA methylation.
• #3: My talk today will start off with a general introduction about what epigenentics is and then focus specifically on DNA methylation and how we measure it. I will cover the two most popular platforms for measuring methylation: arrays and sequencing and how they work.
• #4: This is me. I am a post doc at the Murdoch Childrens Resarch Institute at the Royal Children’s Hospital in Melbourne. Somewhat unsurprisingly, I work mostly with human data focusing on disease and development, but sometimes I work on data from mice and other model organisms. So, a lot of what I will talk to today will be about humans, but many of the principles relate to other species as well. In my day to day work, I mostly analyse epigenetic data from a variety of sequencing and array platforms. I also analyse a lot of gene expression data from RNAseq and microarrays. And, when I’m not analysing data, I occasionally develop methods and write software. Along with my colleague Belinda Phipson, I have developed a software package for analysing methylation array data called missMethyl, which is freely available online from the Bioconductor website, which I’ll tell you a little more about later.
• #5: So, what is epigenetics? Epigenetics typically refers to stable heritable traits that can’t be explained by changed to the DNA sequence. This Greek prefix “epi” literally means “on top of” genetics, so it refers to another layer of control on top of the DNA sequence. Typically when we refer to epigenetics, we are talking about chromosomal modifications that affect the expression of genes; usually histones and DNA methylation. But, epigenetics can pretty much refer to anything that isn’t DNA. Epigenetics is important because it’s essential for development, can be modified by the environment and can be disrupted in disease.
• #6: I like to think that epigenetics is what brings DNA to life. DNA itself is just a static code but it’s epigenetics that executes the code to produce many different cellular phenotypes from the same DNA. If we look at human development, for example, we all start of as a single cell – the zygote – with a bunch of instructions written into the DNA. But then this single cell goes on to become the hundreds of different cell types that make up an entire human being – like the one I’ve got wriggling around in here! And all of this is made possible through epigenetics!!! Although this is a human example, epigenetics is absolutely vital for all species.
• #7: Because epigenetics lets DNA turn a single cell into an entire multicellular, living organism it is actually CRAZY complicated! And many of the new sequenceing and microarray technologies that have appeared in the last decade have enabled us to learn a lot more about all sort of epigenetics phenomena. This has required a lot of different types of data and different types of analysis but, Lucky for me, today I’m only going to be focusing on one thing – DNA methylation.
• #8: At this point some of you may be wondering, what is DNA methylation? Well, if we zoom into one of our many cells, and unpackage the DNA and then look at the individual bases, the As, C, Ts, and Gs, DNA methylation is the chemical modification of the cytosine C base through the addition of this CH3 methyl group. This primarily occurs on Cs in the context of CG dinucleotides, which we call “CpGs”.
• #9: The human genome contains about 30 million CpGs, which is 1% of the total number of bases in the genome. But, this varies quite a lot between different species. A thing to note about CpGs is that they are not evenly spaced across the genome, they tend to be present in clusters that we call CpG islands. You can see that in this diagram here where the lollipops represent CpGs. You can see that they are sort of randomly far apart and then appear in a tight cluster here. This represents a CpG island in a promoter, which is quite common as these islands are often colocalised within regulatory regions. Also of note is that CpGs that are closer together tend to be correlated in their methylation status. This plot at the bottom shows that methylation within about 500bp is between 60 and 90% correlated but that the correlation is reduced with greater distance.
• #10: Methylation is important because it can regulate the expression of genes. So, sometimes we see a relationship between methylation and gene expression that looks like what you see here in this plot on the right. Each point in this plot is from one sample and DNA methylation is shown on the x axis and gene expression is on the y And you can see that as DNA methylation increases gene expression decreases and vice versa. Now the reason that this can happen is shown here on the left. You can see that in a normal cell, the CpG lollipops in the body of the gene are methylated, while the ones in the promoter are unmethylated and so this gene is expressed in a normal cell. However, in cancer for example, we can lose the methylation in the gene body and gain methylation in the promoter, which can lead to this gene being supressed. This is when you see the sort of relationship show in this plot.
• #11: A neat example of how methylation can control gene expression and change phenotype is the coat colour of agouti mice. The agouti mouse has this PS1A gene that controls its coat colour. The promoter of this gene contains a cluster of CpGs that can modify the expression of the gene depending on how methylated they are. There mice in the photo are all genetically identical but have different coat colour because the ones on this side a less methylated at this gene than the ones on this side. What’s even more interesting is that this difference was produced solely by differences in the diet their mothers were fed!
• #12: Another very cool example of how methylation can produce wildly different phenotypes from the same DNA is in bees. In bees, the female larvae are genetically identical and whether they become a queen or a worker depends on their diet, which changes their methylation. The larvae fed on royal jelly, are very hypomethylated and go on to become queens. However, the larvae that are fed worker jelly, gain a lot of methylation and become workers.
• #13: Ok, so I’ve give you a few examples of how methylation is pretty cool, So, what you may be wondering now if what we usually want to know about it?
• #14: When we study methylation, we are usually interested in finding methylation differences. Because methylation is critical in determining cell types, we often want to know what the differences between cell types are. Methylation can also be disrupted in disease, for example cancer. So, we may want to compare cancer and normal tissue. Methylation has also been shown to be affected by environmental factor such as smoking. So, we may want to compare smokers and non smokers. In a typical methylation study, we firstly collect the appropriate samples, for example normal and cancer cells. Then we extract their DNA and measure the methylation using an appropriate platform. Once we have the data then we can do some statistical analysis to answer the biological question of interest.
• #15: A common type of study that people of then do is called an epigenome wide association study or EWAS. Now, if this sounds familiar it’s because it’s similar to a genome wide association study or GWAS. These kind of studies usually involve comparing lots of cases to lots of controls, to find some difference in methylation. EWAS are often looking for small effects such as those that might be present in a complex disease or due to environment. For these studies to work, they require lots and lots of samples; Usually you would hope to have 100s or 1000s of cases and controls.
• #16: Ok, so now we know that methylation is cool and interesting and that we often want to know about differences between things, But how do we actually measure it? Well, there are two main types of approaches: The first is using bisulfite conversion, which lets us create “SNPs” which can tell us if a CpG was methylated or not. This technique provides us with single nucleotide resolution and we can use it with both arrays and sequencing. The other approach is to do some sort of enrichment of methylated DNA. This typically uses a restriction enzyme or some sort of affinity pulldown to enrich for the methylted DNA. In contract to bisulfite conversion, this really only gives us a regional resolution, so we can only tell if chunks of DNA were methylated or not. We can also combine this with both arrays and sequencing. For the remainder of this talk, I’m only going to focus on bisulfite based approaches.
• #17: At this point you may be wondering what is this bisulfite you speak of, what is bisulfite coversion? Bisulfite conversion is a chemical process that was originally published by Sue Clark, who will be speaking after me. So, you are actually very lucky to have a DNA methylation pioneer speaking to you today. Ok, so bisulfite conversion works like this: You start with some DNA that has some methylated Cs and some unmethylated Cs and you mix it with sodium bisulfite. Bisulfite conversion then happens: Any unmethylated Cs are converted to uracils, while the methylated Cs are protected and remain Cs. Then after PCR, the Us become Ts. So effectively we’ve created C=>T SNPs that will let us call methylation.
• #18: Now that we know about bisulfite conversion, I’m going to talk about how its used in conjunction with methylation arrays and how they work
• #19: These give us the ability to measure the methylation at a subset of the CpG sites in the human genome, for a several samples on a single chip. An early version of the array, released in 2009, enabled us to measure about 27000 or the 30million sites in the genome for a single sample. A newer version, released in 2011, increased the number of sites measure to 450 thousand out of 30 million. The newest array, released in 2015, gives us the ability to measure >850 thousand out of 30 million CpGs. The Illumina arrays are for humans only and the CpGs on them have been carefully selected to be in and around genes and regulatory regions As these are likely to be most relevant to human development and disease.
• #20: Because bisulfite conversion lets us create methylation SNPs, methylation arrays are actually based on SNP genotyping arrays. SNP arrays work via a whole bunch of probes that bind to the sequence near SNPs of interest,. Then when you apply your DNA sample to the array, bits of DNA will stick to the probes and then we want to know what the SNP in the sample is. To find this out, fluorescently labelled nucleotides are then incorporated. By measureing the fluorescence we can tell what SNP base was in the sample. Methylation SNPs C=> are created by bisulfite conversion and we measure them in a very similar way. Then we can compare the intensity of the C to T SNPs to get the proportion of methyhlation at a single CpG.
• #21: Since were feading fluorescence intensities from the array, you may be wondering what values we can actually get for methylation? Let’s just think about these 3 cell here that have a pair of chromosomes – one for mum and one from dad. And we’re interested in methylation at a single CpG. In this first cell, there is no methylation on either chromosome. This cell in the middle here, is methylated at this CpG on one chromosome but not the other. And finally, this third cell is methylated at the CpG on both chromosomes. Can anyone tell me what methylation values they think we will get for each of these cells? Yes, these cell will have methylation values of 0, 0.5 and 1. However, on an array are not measuring single cells, we are measuring methylation in a population of cells – a whole smoosh of cells. So, even though and individual cell can only be either 0, 0.5 or 1, across a population we can get any continuous meaure between 0 and 1.
• #22: Due to bisulfite conversion we measure both methylated C and unmethylated Ts to calculate the proportion of methylation at a CpG This is called the beta value and is the proportion of the methylated signal over the total signal. Looking at the density distribution of the CpGs measured on the array we can see that most are either unmethylated or methylated with a smattering of signal in between. This is a very intuitive measure that is easy to interpret and is great for visualisation, but it’s not great for statistics. For statistical analysis we tend to use M-values, which also have a bimodal distribution but aren’t constricted between 0 and 1, the distribution extends from –inf to +inf M-values have been shown to have better statistical properties and are recommended for statistical testing. These two measures are also related through a logit transformation, which is quite useful.
• #27: Take out variability
• #28: Another type of analysis that you can do on methylation data is looking for differentially methylated regions. If you recall, at the beginning of the talk I said that CpGs that are close together are correlated. What we’re doing in this type of analysis is looking for CpGs that are close together that consistently display group-average level differences like seen in this plot here. Some of the populat functions for doing this are bumphunter in minfi and dmrcate from the DMRcate package.
• #29: Once you’ve done some statistical analysis it may give you some clues ideas about follow up analyses could be interesting. In our immyne cell type differences analysis we took all the DMRs that we found between two of the cell types and then extracted all of the underlying DNA sequences and looked for enriched sequence motifs. We found that the unmethylated regions from regulatory Tcells were enriched for a consensus sequence motif that matched the FoxP3 binding motif. Foxp3 is known to be a master regulator of Foxp3 and this showed that methylation is involved in the binding meshanism.
• #30: To sum up the section of my talk on methylation arrays, I’d like to highlight that methylation array analysis is actually very mature and that there are lots of published methods freely available. Most of them are written in R and available from the Biodoncutor website. One of these is our package missMethyl which we published in Bioinformatics last year. We also recently published an example workflow for methyhlation array analysis in F1000 research. It that takes you through and entire methylation array analysis using real data, including all of the required software packages and R code. So, it’s an excellent resource if you are starting out in this field.
• #31: Moving on now to methylation sequencing, by which I actually mean bisulfite sequencing and all of the associated trials and tribulations.
• #33: Before I move onto the nitty gritty of how to analyse and map bisulfite sequencing data, I’d like to refresh your memory about how bisulfite conversion works. We start of with a DNA fragment that has some methylated Cs and some unmethylated Cs. Then we subject it to bisulfite conversion, which converts the unmethylated Cs to T and the methylated Cs stay Cs. We then end up with two sequences that are no longer complementary that we can PCR. After PCR, we end up with four possible bits of sequence that we can get sequence reads from. There are some protocols which let you sequence only the original top and bottom strands, However, it is generally equally likely that you will get reads from all 4 of these sequences.
• #34: So, given this complexity, what are the challenges that we face? F
• #35: Given that I’ve just outlined that mapping bisulfite sequencing reads is pretty challenging, I’m going to take you through how it actually works. Specifically, the approach that the Bismark alignment tool takes. So, if we start of with this particular DNA fragment and then bisulfite convert and PCR it, we can get reads from these 4 sequences. Let’s just focus on how we would map this one from the original top strand.
• #36: Given that I’ve just outlined that mapping bisulfite sequencing reads is pretty challenging, I’m going to take you through how it actually works. Specifically, the approach that the Bismark alignment tool takes. So, if we start of with this particular DNA fragment and then bisulfite convert and PCR it, we can get reads from these 4 sequences. Let’s just focus on how we would map this one from the original top strand.
• #37: Given that I’ve just outlined that mapping bisulfite sequencing reads is pretty challenging, I’m going to take you through how it actually works. Specifically, the approach that the Bismark alignment tool takes. So, if we start of with this particular DNA fragment and then bisulfite convert and PCR it, we can get reads from these 4 sequences. Let’s just focus on how we would map this one from the original top strand.
• #38: Given that I’ve just outlined that mapping bisulfite sequencing reads is pretty challenging, I’m going to take you through how it actually works. Specifically, the approach that the Bismark alignment tool takes. So, if we start of with this particular DNA fragment and then bisulfite convert and PCR it, we can get reads from these 4 sequences. Let’s just focus on how we would map this one from the original top strand.
• #39: Given that I’ve just outlined that mapping bisulfite sequencing reads is pretty challenging, I’m going to take you through how it actually works. Specifically, the approach that the Bismark alignment tool takes. So, if we start of with this particular DNA fragment and then bisulfite convert and PCR it, we can get reads from these 4 sequences. Let’s just focus on how we would map this one from the original top strand.
• #40: Given that I’ve just outlined that mapping bisulfite sequencing reads is pretty challenging, I’m going to take you through how it actually works. Specifically, the approach that the Bismark alignment tool takes. So, if we start of with this particular DNA fragment and then bisulfite convert and PCR it, we can get reads from these 4 sequences. Let’s just focus on how we would map this one from the original top strand.
• #41: (B) The methylation state of positions involving cytosines is determined by comparing the read sequence with the corresponding genomic sequence. Depending on the strand a read mapped against this can involve looking for C-to-T (as shown here) or G-to-A substitutions.
• #42: (B) The methylation state of positions involving cytosines is determined by comparing the read sequence with the corresponding genomic sequence. Depending on the strand a read mapped against this can involve looking for C-to-T (as shown here) or G-to-A substitutions.