Scientist Stories: Ryan Collins, Dosage Sensitivity Map of the Human Genome

Good morning. My name is Ryan Collins, and I'm a PhD student in Mike Talkowsky's lab at Mass General Hospital and the Broad Institute in Boston. And I'm excited to tell you this morning about some of our recent efforts to understand dosage sensitivity in the human genome. So let's start with some definitions. What is dosage sensitivity? Well, it's a genomic phenomenon that describes the intolerance of individual genes or loci to changes in copy number through through deletion or duplication. Dosage sensitivity comes in two different flavors. The first, which has been extensively studied in humans and model organisms, is haploinsufficiency, where the loss or deletion of a single copy of DNA has a damaging effect. The second and less understood type of dosage sensitivity is called triplosensitivity, where the duplication or gain of additional copies of DNA results in a deleterious outcome. And we also now know that not all loci tolerate copy number variation equally well. So of course, many genes have been identified where deletions or duplications lead to a disease phenotype, but the reciprocal copy number change is observed frequently in the general population and has no obvious effect on overall fitness. However, there are also a few dozen loci described to date that exhibit what are known as mirror phenotypes, where deletions lead to one set of phenotypes and duplications of that exact same gene or locus lead to the opposite or mirrored phenotype. And while I think these patterns are overall very fascinating, we don't really have a solid understanding of why they're observed. So for the last several years, we've been trying to understand these patterns using large human genetic datasets. And today I'm going to focus on one of our recent studies where we compiled rare copy number variants from existing microarray datasets like the UK Biobank. More specifically, we've assembled a dataset of over 3/4 of a million individuals across 13 sources and 30 primarily neurodevelopmental disease phenotypes and have mined this dataset for new insights into dosage sensitivity. And we're, of course, actively pursuing similar efforts in large sequencing-derived datasets, and those studies are ongoing. But I'm hoping to convince you today that although microarray data does have plenty of limitations, there's still a lot we can learn through large-scale data aggregation in efforts such as this. A critical first step in any large-scale data aggregation, of course, is stringent quality control and harmonization. And not surprisingly, there was extensive technical heterogeneity across the 13 CNV call sets included in our study shown here. Including the distributions of raw CNV calls from the 13 different sources, where on the x- and y-axes I'm showing the average number of CNVs per sample or the median CNV size. And you can see that there's close to 2 orders of magnitude variability between studies in those metrics. Now, to combat this, we developed a stringent filtering pipeline shown on the right to harmonize CNVs across these sources. This pipeline aimed to retain large subchromosomal events between 100 kilobases and 20 megabases in size with sufficient array probe density, and also that all appeared at below below 1% frequency in every genetic ancestry or population documented in existing reference datasets like NOMAD or the 1000 Genomes Project. Now, after applying this harmonization pipeline, you can see that this effectively reduced the interstudy variability by up to 2 orders of magnitude across different cohorts. Similarly, we needed to standardize phenotype data across these cohorts, which we accomplished by first mapping all the available phenotype data we had onto the structured Human Phenotype Ontology and then clustering these HPO terms to define a set of 30 distinct terms for downstream association testing. In total, this allowed us to define a set of 300,000 samples as cases, which matched one or more of these 30 HPO terms, and a set of 461,000 controls that effectively matched none of these terms. From this curated dataset, we applied a battery of analytical approaches to identify and characterize dosage-sensitive segments and genes throughout the genome. Now, in the interest of time, I really only have the time to briefly touch on two of our analyses today, but I do want to mention that we have a preprint up on medRxiv with the full details of the study, including many analyses that I'm not going to be able to even mention today. So I do encourage you to check out our preprint if you're interested in this work. In the first analysis I'll be presenting today, we conducted genome-wide association tests to identify loci where large C and V segments were associated with disease. More formally, we performed fixed effects meta-analyses separately for deletions and duplications in 200-kilobase windows sliding across the genome in 10-kilobase steps. And after collapsing our results across all 30 HPO terms, we arrived at a list of 49 distinct CNV segments that were associated with at least one disease at a strict genome-wide significance threshold. Now, among these 49 segments were many interesting examples and some interesting biological anecdotes, but unfortunately due to time, I can't really get into the details of individual loci today. But I do want to mention that we highlight many of them in our, in our preprint if you're interested in more information. But just to convince you that our methods are recapitulating known pathogenic CNVs, I'm going to show two quick analyses, both of which use a permutation-based strategy where we shuffled these segments randomly throughout the genome 100,000 times. And in so doing, we find that our empirical CNV-associated loci are massively enriched for previously reported genomic disorder CNVs. And even at the level of individual genes, we found a weaker but highly significant enrichment for phenotype-matched disease genes when querying OMIM for the exact same HPO term used in our association tests. We next wanted to understand what features of these segments were responsible for their disease association. And so using the same permutation-based strategy I just described, we assessed this permuted null distribution against a variety of genomic features. Not surprisingly, protein-coding genes emerged as a particularly prominent feature. So for example, not only did these 49 CNV segments include twice as many genes as expected by chance throughout the genome, but even after conditioning on this greater gene density, we still found that mutationally constrained genes were modestly enriched beyond expectations in these regions. And we also saw some interesting patterns when stratifying our dataset by the number of phenotypes associated per locus, where we're defining loci with 2 or more associations as pleiotropic. So for example, we didn't find any obvious differences in the density of all genes between single phenotype and pleiotropic CNVs, but we did observe a significantly greater density of constrained genes among pleiotropic CNVs. And while I don't have the time to show the results here today, we do also find that these same driver genes are significantly enriched in de novo damaging point mutations in individuals with related or similar phenotypes, And in fact, we can actually nominate a single monogenic driver for approximately one-third of all of these disease-associated CNVs segments. So collectively, I think that these results point pretty clearly to the presence and the number of dosage-sensitive constrained genes per CNV as an important factor in their overall pathogenicity. Given this prominence of individual dosage-sensitive genes in the analyses I just showed, we next wanted to see if we could predict dosage sensitivity for each gene in the genome. And so to do this, we took advantage of the relatively large size of our aggregated CNV dataset and trained an ensemble of machine learning models. So like most supervised machine learning, this process first started by defining training sets of haploinsufficient and triplo-sensitive genes based on existing literature evidence alone. And we next asked a variety of machine learning architectures if they could predict the empirical CNV effect sizes we observe in severely affected cases versus controls in our dataset from these training genes based strictly on the 129 gene-level features which we added as annotations per gene and no other information. And then remarkably, the answer was a resounding yes, actually, not only for haploinsufficiency but also for triplo-sensitivity. And while you can see that this was true for a variety of different architectures, the overall ensemble averaging model across all architecture was the most performant, and so we used that moving forward as our final probabilities per gene. So when you compare the probability of triple sensitivity per gene to the probability of haploid sufficiency, the first thing you'll see is that our prediction suggests that most genes are either haploid sufficient and triple sensitive or neither. And while there are a few examples, this suggests that genes that are haploid sufficient are also very likely to be triple sensitive. We also wanted to make sure that these scores were recapitulating what we believe to be biological ground truth. And so we did this a variety of different ways. First, by looking at gene set enrichment analyses where we found against a variety of completely independent gene sets enrichments in our predicted haploinsufficient and triplo sensitive genes in gene sets that we expect and finding depletions where we don't. We also looked at an independent dataset of de novo copy number variants identified through exome sequencing in individuals with autism and their unaffected matched siblings. And we find that just by prioritizing these CNVs based strictly on the triple sensitivity or haploinsufficiency scores at the level of individual genes, we can stratify risk conferred by de novo copy number variants by up to an order of magnitude, both for deletions and duplications. And finally, we find that our scores are inversely correlated with the rate of protein-truncating deletions and whole gene duplications in NOMAD, all of which suggests that these scores are predicting true haploinsufficient dosage-sensitive genes. And finally, we wanted to use these scores to define sets of dosage-sensitive genes. So how we did this is we took a gold standard set of genes known to be constrained against protein truncating loss or loss of function point mutations and asked at what score threshold of our dosage sensitivity scores does the effect size of rare deletions in our dataset match that of the— for the gold standard constrained genes. And so using this approach, we can calibrate cutoff for about, you know, let's say 3,000 haploinsufficient genes based on our scores. But more interestingly, when you do this for duplications, the first thing you'll notice is that the effect size of deletions for most genes on average is stronger than duplications. But if you zoom in on the far tail of this duplication score or triplo-sensitivity score distribution, we can identify a set of 295 genes where rare duplications in our dataset have the same effect size as loss of function of gold-standard mutationally constrained genes. And so, unfortunately, I have to wrap up there in the interest of time. There was a lot that I wasn't able to present today. It's all contained in our preprint. I'm happy to go back and forth on email afterwards or take questions now. And so I, of course, would also like to thank everyone who contributed to this study. It's really been a wonderful collaborative experience. And I also thank you for your attention, and I'm happy to take any questions. Thanks.