Scientist Stories: Andrew Sharp, Epivariations in the Human Genome

Hello, my name is Andrew Sharp. I'm an associate professor in the Department of Genetics and Genomic Sciences at Icahn School of Medicine, Mount Sinai in New York, and today I'm going to talk about epi-variations in the human genome. I have no financial disclosures to make. Over the last two decades or so, efforts such as the International HapMap Project The 1000 Genomes Project and efforts such as the Exome Aggregation Consortium have curated deep catalogs of both common and rare genetic variation that have really acted as the catalysts to give us numerous insights into the role of genetic variation in human disease. In contrast, our insight into the role of epigenetic variation is much more limited. Arguably, the seminal study on this came about 2 decades ago, where studies showed that in the Agouti mouse, both the variation in coat color and perhaps more significantly variation in susceptibility to metabolic disease and various cancers was caused by epigenetic variation at the Agouti locus. Since that time, in humans, it's now been identified that about 2% of cases of hereditary colon cancer are actually due to promoter methylation of the MLH1 gene. There have been a few families identified with hereditary breast and ovarian cancer caused by promoter methylation of BRCA1, and it's also been found that a number of cases of metabolic defects are similarly due to promoter methylation events. However, despite these individual studies into specific diseases, Up until a few years ago, our knowledge of the prevalence of rare epigenetic variation and its contribution to human disease was really very limited. And what I was going to cover today is a number of studies that my lab has put out in the last few years that have really focused on trying to answer these questions. And in these three papers that I'm showing you here, we describe characterization of epigenetic variation in about 25,000 individuals, including patients with intellectual disability and congenital anomalies, autism, and schizophrenia, as well as a large number of individuals from the general population. These studies have identified thousands of loci where we observe epivariations in humans, including hundreds of what appear to be inactivating promoter methylation events. Including numerous different known disease genes. Also, these have given us quite deep insights into the causes and consequences of the variations. Now what I'm going to talk about here are a study of around 23,000 individuals that we collected together. And here we took an approach very similar to that used by the Exome Aggregation Consortium, which studied coding mutations, where we simply gathered basically any available cohort we could find from a variety of publicly available databases. And in total, we accumulated together data from 22 different cohorts, each of which have been profiled using the Illumina 450k array, and at least 300 independent individuals. Most of these individuals, the DNA was derived from peripheral blood. And what was significant, as you'll, you'll see in the next few minutes, is that these studies included a number of cohorts composed of either monozygous or dizygous twins. Several of them had gene expression data, and several also had genetic variation data, either through SNP arrays or whole genome sequencing. In total, these comprise nearly 25,000 individuals, including 700 pairs of monozygous twins and thousands of individuals with expression or genetic variation data. Now, one thing that's very important to appreciate if you're someone who's considering performing these kinds of studies is quality control of your data. And over the last few years, we've sort of optimized this approach. We now remove any sample that has more than 1% of probes that fail detection p-values if they have low signal intensity. We also perform principal component analysis to remove outlier samples. Another important factor is that epigenetic measurements can be heavily dependent on the underlying composition of cells within your sample, and in particular, in peripheral blood, this is an important consideration to account for. So we remove any samples that have outlier cell counts that we estimate directly from the methylation data. And we also remove samples that appear to have sex mismatches or sex chromosome aneuploidies. And in this particular study, out of the 25,000 individuals we started with, we removed about 7.5%, leaving us with just over 23,000 individuals left. After performing sample-level filtering, we then normalize the data using standard approaches and also perform probe-level filtering to remove probes that overlap common variation, have non-unique map positions, or show a high level of failure across the cohorts. In order to identify epi-variations, we settled on using a sliding window approach as a particular particularly robust method to identify epi-variations in a population. And here we can basically identify either gains or losses by considering regions where we see 3 or more probes that each lie within the extreme tail of the population and show a minimum difference from population mean. So here is one example of an epi-variation that we detect. Just to orientate you, on the x-axis here we have physical position. So here we're looking at a region of around 2 Kb. The y-axis is the methylation value, where 0 corresponds to unmethylated, 1 is fully methylated, and a value of approximately 0.5 is consistent, at least with one of the two alleles being methylated. In this particular plot, we're seeing all 23,000 individuals that we tested. The shades of gray represent 1, 2, and 3 standard deviations from the mean. Individuals who lie outside this distribution are shown by the dotted black lines, and our epivariation carrier is shown by the red line. Here is a second locus where we see both gains and losses at the same locus. In fact, this is an imprinted gene where we see what appear to be imprinting defects recurrently in the population. Now, one of the major observations we saw is that most of the epimutations we observe are rare events. In total, we found nearly 14,000 epimutations, as around one-third of the samples tested carry one or more epi-variants. And these were observed at around 4,500 loci. We see about a 2:1 excess of gains to losses. That is, gains of methylation are much more common than losses, although this could be may be an artifact of the placement of probes on the 450K array. We also observe a small number of loci where we see both gain and loss at the same location. And on the right here, this plot is showing the distribution of the frequency of epimutations in the cohort, and you can see that the vast majority of events are rare, that is observed in only a single individual in the entire cohort. So one of the first things that we wanted to do was study the effect of epi-variations on local gene expression, using about 3,500 individuals who had RNA-seq data available. And focusing specifically on individuals who had hypomethylation, that is losses of methylation at gene promoters, we observed a clear trend where many of these individuals showed unusually high expression levels. In contrast, individuals who had gains of promoter methylation tended to show unusually low expression levels. Indeed, many of these carriers have some of the lowest expression in the entire cohort. And this clear effect of promoter methylation is consistent with the known repressive effects of DNA methylation at gene promoters. Taking this one step further, we then used a cohort of individuals where we had both DNA methylation, RNA sequencing, and whole genome sequencing data that enabled us to look at allele-specific expression events. What I'm showing you here is an individual with a promoter hypermethylation event at the UBE2T gene. This is associated with an autosomal recessive form of Fanconi anemia. And here we found this same individual showed the lowest expression of UBE2T in the whole cohort.. And using transcribed polymorphisms within the UBE2T gene, while most of the other individuals we studied showed clear biallelic expression, the epivariation carrier showed exclusive expression for only one of the two alleles. And therefore this promoter methylation event is clearly associated with monoallelic expression, and therefore at the functional level can be considered to be equivalent to a loss of function mutation. Now, when we looked more broadly, we in fact observed hundreds of known disease genes where we observed presumably inactivating promoter hypermethylation events. This included several ones that have been reported previously in the literature, such as hypermethylation of the MMACHC gene, which is associated with hereditary vitamin B12 metabolic defects. We observed these to occur at about a frequency of 1 in 1,000 in the general population. And similarly, hypermethylation events of the XYL-T1 gene associated with Barrett-Ellis-Scott syndrome were observed at about 1 in 2,300 in the population. However, we also observed several, as I mentioned, several hundred other genes where we believe ethnic mutations likely to contribute to the mutational spectra underlying these disorders. Now, perhaps some of the most interesting of these are listed in this table here. In previous years, the American College of Medical Genetics has published a list of several dozen genes that they consider to be clearly pathogenic but also medically actionable and recommended to be returned to any subjects where these are found within any study. For example, we observed 7 individuals with gains of promoter methylation of the BRCA1 gene, which is associated with dominant forms of inherited breast and ovarian cancer, and 4 individuals here with gains of promoter methylation at the LDLR gene, which is associated with familial hypercholesterolemia. Now, promoter methylation of BRCA1 has been reported before in a number of pedigrees and is known to be associated with hereditary breast and ovarian cancer. These are thought to be generally very rare events. Now, in our cohort of 23,000 individuals, we detected 7 unrelated carriers of this epi-variation. It gives a prevalence of around 1 in 3,300. If we compare this with the prevalence of coding mutations, damaging coding variants of BRCA1, these are estimated around 1 in 200 in large biobanks from the US. And putting these two numbers together, this suggests overall that perhaps around 5% of pathogenic BRCA1 alleles could be due to promoter hypermethylation events. Therefore, we believe may be significantly more common than is currently appreciated. Now we then wanted to investigate the underlying causes of epi-variations, and the first study that we did to investigate this was by association analysis between common or relatively common epi-variations and local SNPs. What I'm showing you here is essentially a Manhattan plot. So the x-axis is showing a region of around 2 megabases each side of an individual epivariation we found in that was in 9 different carriers. The y-axis is the minus log 10 p-value. And by associating presence of the epivariation with local SNP markers, we identified a cluster of SNVs within which the epivariation was contained.. And the peak SNP here, while it has a minor allele frequency of only around 1% in the general population, 8 of the 9 epivariation carriers we identified were heterozygotes for this SNP, giving a p-value of around 10^-40. What this data indicates is that this particular epivariation is clearly a secondary event segregating on a local haplotype. And overall, around 70% of the epivariations that we were able to study in this way were associated with local SNPs at a 1% FDR level. Now we then looked at this more closely by performing studies of rare sequence variation using cohorts of individuals where we had whole genome sequencing data. And doing this, what we observed is a strong co-occurrence of rare sequence variation to occur with many of the epi-variations we were studying. So we found that about 10% of epi-variants were associated with rare underlying SNVs, and this was shown in the plot on the left here, about a 10-fold enrichment versus the random null expectation. And here you can see that these rare SNPs are strongly clustered around the midpoint of these epi-variations.. When we then restrict this analysis only to SNVs that overlap transcription factor binding sites, this relationship gets even stronger. And again, you can see very clearly that there's an association of these SNPs to occur right at around the midpoint of these variations. Here's one particular example. This is an individual with a gain of methylation, and we found this person carried a rare SNP that was located right within the consensus sequence for CTCF, which is a known transcription factor that has roles in chromatin organization, in particular organizing chromatin into loops and therefore dictating epigenetic patterns in chromatin. In addition, we are also able to study the contribution of rare CNVs, that is copy number variants,. And even though the absolute number of these that we detected was relatively small and appeared to account for around 3% of IRP variations, this was strongly enriched versus random. Here again, I'm showing you one particular example of, in this case, a 36-kb duplication that we identified specifically in a carrier for an epi variation.. And suggest that indeed a fraction of epivariants are caused directly by underlying sequence mutations, whether that be rare single nucleotide variants or rare copy number variants. Now, one of the other things that we hypothesized we might find when we set about doing these studies was repeat expansion. Now, as many of you will know, perhaps the best known example of a repeat expansion is the Fragile X locus. As its name suggests, this is located on the X chromosome, and Fragile X is actually caused by expansion of a hypermethylated CGG repeat located within the promoter of the FMR1 gene. In rare individuals, this repeat expands to greater than 200 copies, and in that state it becomes highly methylated and is also associated with the production of a fragile site, which can be directly visualized if cells from carriers of this CGG expansion are cultured in a folate-reduced media. Hence the name fragile X. So what I'm showing you here is a table of the top 20 most frequent hypermethylated epi-variations we found in this study.. And what we observed is that in fact, like, several of these were associated with underlying CGG repeats at the epivariation locus. And in fact, several of these repeats are known to undergo rare expansions and associated with hypermethylation. So there's FRAP10AC1, C11ORF80, which is associated with the fragile site FRAR11A, and FRAR12A fragile site, which is associated with CGG expansions of the DIP2B gene. However, we also noticed that there were several other recurrent epi-variations that we found that also had underlying CGG repeats at the same locus, and so we hypothesized from this that they— these may be sites of novel CGG expansion. So what I'm showing you here is one of these loci. This is the LINGO3 gene, and here we had an individual with a clear gain of methylation directly overlapping an annotated CCG repeat. Using Illumina whole genome sequencing data from this individual, what we observed is a very striking pattern where when we visualize the mapped reads to this locus in the IGV browser, We see many reads that have this pattern of soft clipping where you can see the XX bases are composed of what appears to be additional copies of the CCG repeat. And this pattern was absent in other control individuals who did not carry this gain of methylation. So we then obtained a DNA sample from this individual, performed long read sequencing using Pacific Biosciences,, and we're able to confirm that indeed this individual carries an expansion of this CCG repeat with around 220 copies that clearly is undergoing hypermethylation. Here is an additional example of another novel CCG expansion that we found based on our epivariation data. This was located within intron 1 of the FZD6 gene. We identified 14 individuals, so this, this CCG expansion has a frequency of about 1 in 1,600, and also FZD6 lies within the FRA8A fragile site locus. So we obtain DNA from one of these individuals with a gain of methylation, and here we use Oxford Nanopore whole genome sequencing, and identified an insertion or expansion of around 1,800 CCG repeats in one of these epi-variation carriers. And then in addition, one of the advantages of Oxford Nanopore data is you're able to profile DNA methylation directly from the reads themselves. And by phasing the reads into those that either contained or did not contain the CCG expansion, we're able to show that the the expanded allele clearly shows a high gain of methylation, proving that this epivariation event is specifically caused by an underlying CCG expansion of this locus. Now, if we then go back and take all of the gains of methylation that we found and overlap them with what we believe is a set of unstable CCG CCG repeats. We identify in total 31 candidate loci where we believe these are likely to be caused by underlying CCG expansions. 6 of these 31 loci are in fact already known sites of repeat expansion. We've now been able to validate 13 novel CCG expansions using either long read sequencing or Southern blots. These are shown in green here, and we're able to predict at least another 12 loci in the genome that we've not yet been able to validate, but we believe are highly likely to be caused by underlying CCG expansions. Now, if we then go back and overlap these novel CCG expansions that we're detecting with known fragile sites, folate-sensitive fragile sites in the genome, 10 of these folate-sensitive fragile sites have already been found to be caused by CCG expansions. 9 of the novel repeat expansions that we are identifying here in fact coincide with unmapped fragile sites, strongly suggesting that these novel CCG expansions are what is the molecular defect underlying these fragile sites. In addition to these studies of the sequence dependency of epivariations, we also use the presence of several hundred pairs of monozygous twins in our cohorts to look at potentially other effects on epivariations. So here in particular, we studied the concordance rate of epivariation in these MZ twin pairs. What we observed is that in 63% of the cases, As we might expect, monozygous twins show the same pattern of epivariation. That is, where one twin carries an epivariant, the other twin also carries it. We were quite surprised to see that in about 30% of cases of an epivariation, this was completely absent in their co-MZ twin. And then in a small fraction, we observed cases that we classified as partially concordant. That is, one twin carried a clear outlier. Methylation event, and their cotwin was an outlier to a lesser extent. So overall, this suggests that around 1/3 of epi-variations likely have a somatic origin, is postzygotic, and this would also suggest that many epi-variations therefore are potentially mosaic events. To corroborate this further, we then looked at the prevalence of epivariations in comparison to an individual's age. And here we saw a very strong correlation that as individuals carried more and more epivariations within their genome, they tended to be significantly older than individuals that had zero or very few epivariants. This therefore indicates that epi-variations do indeed likely accumulate with age, likely due to somatic, either epi-mutation or genetic mutation which has secondary epigenetic effects. We then followed this up further by looking at the familial inheritance of epi-variation within families. So really the question here was if Epivariants are potentially somatic events. Are these heritable between, from parents to offspring? So in total, we had 117 parent-offspring families, and in total we had 89 informative meioses for, that we could look at. Here, what I'm showing you is one particular example of a clearly inherited event. Here, the mother shown in the red profile has clearly passed this on to one of her three offspring in green. And this was observed in around 31% of the meiosis we could study. However, the more common state we found was examples like this, where here on the right, the father in blue carries an outlier epigenetic event which has not been passed on to any of his four offspring. And overall, this is significantly different from the Mendelian expectation of 50% and suggests indeed that some epi-variations are somatically mosaic. Therefore, while we may detect them in the blood, they could, for example, be absent from the germline. Or alternatively, this observation is also consistent with the fact that some epi-variations may be wiped and reset during the epigenetic reprogram— reprogramming that occurs during gametogenesis. Okay, so now I want to talk very briefly just about two additional studies that we've performed. The first was looking in a cohort of around 500 individuals with intellectual disability and congenital anomalies, the majority of which have previously been tested by array CGH and exome sequencing. No mutations have been found. And secondly, this other study where we looked at around 1,000 individuals with either autism or schizophrenia. Here the idea was we wanted to investigate the potential contribution of epivariants to congenital anomalies and neurological disease. Now, one of the major hallmarks of causality when when studying genetic disease is looking for the de novo status of the mutation. In particular, de novo mutations are often a hallmark that can indicate causality when studying a trio. And here again, using these familial cases of population controls, we found that around 13% of the epivariations we studied appear to be de novo events, that is, absent in both parents and appearing as a new mutation in offspring. In contrast, when we studied this cohort of 500 trios where the child had intellectual disability or congenital disease, we observed that 42% of these variants we found in these children were de novo, which is about a threefold excess and is statistically significant. This suggests strongly that at least a subset of these epivariants are likely contributing to the disease in these individuals. Then moving on to our studies of autism and schizophrenia. Here we profiled around 1,000 individuals and specifically looked for epivariations that were absent in unaffected controls. What we observed is a clear enrichment for epivariations at known genes associated with autism and schizophrenia. One example here is the SHANK3 locus. This is a well-known autism gene. Here we observed the red line is a proband with autism. Here we found two losses of methylation overlapping CpG islands within SHANK3. And a second clear example was the IGGYF1 gene, again another known autism gene, where again we observed here one gain and also a large loss of methylation within GIGYF1. What's particularly striking is in both cases the unaffected sibling shown in green did not carry this epivariation. Looking on a population level, we observed also an increased burden of epivariations in affected versus unaffected sibs with autism. In our schizophrenia cohort, we observed epivariations in schizophrenia cases were associated with genes with significantly higher expression within the brain. So just to conclude, our studies show that epivariants are relatively frequent in the human genome, and many of these disrupt gene expression patterns. Around two-thirds of epivariations result from underlying sequence variants. And around 1/3 appear to be sporadic or somatically acquired with age, suggesting many of these are associated with mosaicism and therefore reduced heritability. We believe that epivariations likely contribute to the mutational spectra underlying a number of both dominant and recessive human diseases. And it's important to emphasize that many of these will be missed by purely sequence-based approaches, unless you're looking at the epigenetic profile, you're not going to see these events. And therefore, overall, we believe that the application of epigenome profiling has the potential to increase the diagnostic yields in unsolved families, and also, as we've shown, can be a powerful approach to uncover functional rare sequence variation, including CCG expansions. I'd just like to thank the members of my lab and collaborators and funders of this work. Thank you very much.