Discovery of runs-of-homozygosity diplotype clusters and their associations with diseases in UK Biobank

  1. Ardalan Naseri
  2. Degui Zhi
  3. Shaojie Zhang

  • Curated by eLife

    eLife logo

    eLife assessment

    This study is of potential interest to readers in human genetics and quantitative genetics, as it presents a new method for homozygosity mapping in population-scale datasets, based on an innovative computational algorithm that efficiently identifies runs-of-homozygosity (ROH) segments shared by many individuals. Although the method is innovative and has the potential to be broadly useful, its power and limitations have not yet been adequately evaluated. The application of this new method to the UK Biobank dataset identifies several interesting associations, but it remains currently unclear under what conditions the new approach can provide additional power over existing genome-wide association study methods.

Reviewed by

Read the full article

Listed in

Log in to save this article

Abstract

Runs-of-homozygosity (ROH) segments, contiguous homozygous regions in a genome were traditionally linked to families and inbred populations. However, a growing literature suggests that ROHs are ubiquitous in outbred populations. Still, most existing genetic studies of ROH in populations are limited to aggregated ROH content across the genome, which does not offer the resolution for mapping causal loci. This limitation is mainly due to a lack of methods for the efficient identification of shared ROH diplotypes. Here, we present a new method, ROH-DICE (runs-of-homozygous diplotype cluster enumerator), to find large ROH diplotype clusters, sufficiently long ROHs shared by a sufficient number of individuals, in large cohorts. ROH-DICE identified over 1 million ROH diplotypes that span over 100 single nucleotide polymorphisms (SNPs) and are shared by more than 100 UK Biobank participants. Moreover, we found significant associations of clustered ROH diplotypes across the genome with various self-reported diseases, with the strongest associations found between the extended human leukocyte antigen (HLA) region and autoimmune disorders. We found an association between a diplotype covering the homeostatic iron regulator (HFE) gene and hemochromatosis, even though the well-known causal SNP was not directly genotyped or imputed. Using a genome-wide scan, we identified a putative association between carriers of an ROH diplotype in chromosome 4 and an increase in mortality among COVID-19 patients (p-value = 1.82 × 10 −11 ). In summary, our ROH-DICE method, by calling out large ROH diplotypes in a large outbred population, enables further population genetics into the demographic history of large populations. More importantly, our method enables a new genome-wide mapping approach for finding disease-causing loci with multi-marker recessive effects at a population scale.

Article activity feed

  1. Version published to 10.7554/elife.81698 on eLife
  2. eLife

    Author response:

    Reviewer #1 (Public Review):

    In this manuscript, Naseri et al. present a new strategy for identifying human genetic variants with recessive effects on disease risk by the genome-wide association of phenotype with long runs-of-homozygosity (ROH). The key step of this approach is the identification of long ROH segments shared by many individuals (termed "shared ROH diplotype clusters" by the authors), which is computationally intensive for large-scale genomic data. The authors circumvented this challenge by converting the original diploid genotype data to (pseudo-)haplotype data and modifying the existing positional Burrow-Wheeler transformation (PBWT) algorithms to enable an efficient search for haplotype blocks shared by many individuals. With this method, the authors identified over 1.8 million ROH diplotype clusters (each shared by at least 100 individuals) and 61 significant associations with various non-cancer diseases in the UK Biobank dataset.

    Overall, the study is well-motivated, highly innovative, and potentially impactful. Previous biobank-based studies of recessive genetic effects primarily focused on genome-wide aggregated

    ROH content, but this metric is a poor proxy for homozygosity of the recessive alleles at causal loci. Therefore, searching for the association between phenotype and specific variants in the homozygous state is a key next step towards discovering and understanding disease genes/alleles with recessive effects. That said, I have some concerns regarding the power and error rate of the methods, for both identification of ROH diplotype clusters and subsequent association mapping. In addition, some of the newly identified associations need further validation and careful consideration of potential artifacts (such as cryptic relatedness and environment sharing).

    1. Identification of ROH diplotype clusters.

    The practice of randomly assigning heterozygous sites to a homozygous state is expected to introduce errors, leading to both false positives and false negatives. An advantage that the authors claim for this practice is to reduce false negatives due to occasional mismatch (possibly due to genotyping error, or mutation), but it's unclear how much the false positive rate is reduced compared to traditional ROH detection algorithm. The authors also justified the "random allele drawing" practice by arguing that "the rate of false positives should be low" for long ROH segments, which is likely true but is not backed up with quantitative analysis. As a result, it is unclear whether the trade-off between reducing FNs and introducing FPs makes the practice worthwhile (compared to calling ROHs in each individual with a standard approach first followed by scanning for shared diplotypes across individuals using BWT). I would like to see a combination of back-of-envelope calculation, simulation (with genotyping errors), and analysis of empirical data that characterize the performance of the proposed method.

    In particular, I find the high number of ROH clusters in MHC alarming, and I am not convinced that this can be fully explained by a high density of SNPs and low recombination rate in this region. The authors may provide further support for their hypothesis by examining the genome-wide relationship between ROH cluster abundance and local recombination rate (or mutation rate).

    Thanks for this insightful comment. Through additional experiments, we confirmed that the excessive number of ROH clusters in the MHC region is due to the higher density of markers per centimorgan. As discussed above at Essential Revision 2, we took this opportunity to modify our code to search for clusters with the minimum length in terms of cM instead of sites. We have also provided the genetic distance for reported clusters in the MHC region with significant association (genetic length (cM) column in Tables 1 and 2). We include the following in the main text:

    “We searched for ROH clusters using a minimum target length of 0.1 cM (Figure 3–figure supplement 1). As shown in the figure, there is no excessive number of ROH clusters in chromosome 6 as was spotted using a minimum number of variant sites.”

    Methods section, ROH algorithm subsection:

    “We implemented ROH-DICE to allow direct use of genetic distances in addition to variant sites for L. The program can take minimum target length L directly in cM and detect all ROH clusters greater than or equal to the target length in cM. The program holds a genetic mapping table for all the available sites, and cPBWT was modified to work directly with the genetic length instead of the number of sites.”

    1. Power of ROH association. Given that the authors focused on long segments only (which is a limitation of the current method), I am concerned about the power of the association mapping strategy, because only a small fraction of causal alleles are expected to be present in long, homozygous haplotypes shared by many individuals. It would be useful to perform a power analysis to estimate what fraction of true causal variants with a given effect size can be detected with the current method. To demonstrate the general utility of this method, the authors also need to characterize the condition(s) under which this method could pick up association signals missed by standard GWAS with recessive effects considered. I suspect some variants with truly additive effects can also be picked up by the ROH association, which should be discussed in the manuscript to guide the interpretation of results.

    We added a new experiment in the Results section “Evaluation of ROH clusters in simulated data” under Power of ROH-DICE in association studies. We compared the power of the ROH cluster with additive, recessive, and dominant models. Our simulation shows that using ROH clusters outperforms standard GWAS when a phenotype is associated with a set of consecutive homozygous sites. We added the following text:

    “...We calculated the p-values for both ROH clusters and all variant sites. We used a p-value cut-off of 0.05 divided by the number of tests for each phenotype to determine whether the calculated p-value was smaller than the threshold, indicating an association. For GWAS, only one variant site within the ROH cluster, contributing to the phenotype, was required. We tested for all additive, dominant, and recessive effects (Figure 1–figure supplement 3). The figure demonstrates that ROH-DICE outperforms GWAS when a phenotype is associated with a set of consecutive homozygous sites. The maximum effect size of 0.3 resulted in ROH clusters achieving a power of 100%, whereas the additive model only achieved 11%, and the dominant and recessive models achieved 52% and 70%, respectively. The GWAS with recessive effect yields the best results among other GWAS tests, however, its power is still lower than using ROH clusters.”

    1. False positives of ROH association. GWAS is notoriously prone to confounding by population and environmental stratification. Including leading principal components in association testing alleviates this issue but is not sufficient to remove the effects of recent demographic structure and local environment (Zaidi and Mathieson 2020 eLife). Similar confounding likely applies to homozygosity mapping and should be carefully considered. For example, it is possible that individuals who share a lot of ROH diplotypes tend to be remotely related and live near each other, thus sharing similar environments. Such scenarios need to be excluded to further support the association signals.

    We acknowledge that there could be confounding factors that may affect the association's results. To address this, we utilized principal component (PC) values and additional covariates while using PHESANT after our initial Chi-square tests. We also included your comments in our Discussion section:

    "We used age, gender, and genetic principal components as confounding variables in the association analysis. Genetic principal components can reduce the confounding effect brought on by population structure but it may be insufficient to completely eliminate the effects of recent demographic structure and the local environment45. For example, individuals sharing excessive ROH diplotypes may share similar environments since they are closely related and reside close to one another. Since we did not rule out related individuals, some of the reported GWAS signals may not be attributable to ROH.”

    1. Validation of significant associations. It is reassuring that some of the top associations are indirectly corroborated by significant GWAS associations between the same disease and individual SNPs present in the ROH region (Tables 1 and 2). However, more sanity checks should be done to confirm consistency in direction of effect size (e.g., risk alleles at individual SNPs should be commonly present in risk-increasing ROH segment, and vice versa) and the presence of dominance effect.

    The beta values for effect size are now included in all reported tables. All beta values for ROH-DICE are positive indicating carriers of these ROH diplotypes may increase the risk of certain non-cancerous diseases. Moreover, we conducted the suggested sanity check to confirm the consistency of the direction of risk-inducing ROH diplotypes and risk alleles.

    We also computed D’ as a measure of linkage between the reported GWAS results and ROH clusters. We found that most of the GWAS results and ROH clusters are strongly correlated. However, in a few cases, D' is small or close to zero. In such cases, the reported p-value from GWAS was also insignificant, while the ROH cluster indicated a significant association. We included these points in the Results section.

    Reviewer #3 (Public Review):

    A classic method to detect recessive disease variants is homozygosity mapping, where affected individuals in a pedigree are scanned for the presence of runs of homozygosity (ROH) intersecting in a given region. The method could in theory be extended to biobanks with large samples of unrelated individuals; however, no efficient method was available (to the best of my knowledge) for detecting overlapping clusters of ROH in such large samples. In this paper, the authors developed such a method based on the PBWT data structure. They applied the method to the UK biobank, finding a number of associations, some of them not discovered in single SNP associations.

    Major strengths:

    • The method is innovative and algorithmically elegant and interesting. It achieves its purpose of efficiently and accurately detecting ROH clusters overlapping in a given region. It is therefore a major methodological advance.

    • The method could be very useful for many other researchers interested in detecting recessive variants associated with any phenotype.

    • The statistical analysis of the UK biobank data is solid and the results that were highlighted are interesting and supported by the data.

    Major weaknesses:

    • The positions and IDs of the ROH clusters in the UK biobank are not available for other researchers. This means that other researchers will not be able to follow up on the results of the present paper.

    We included the SNP IDs, positions, and consensus alleles for all reported loci in the main tables. Moreover, additional information including beta and D’ values were added. The current information should allow researchers to follow up on the results. Supplementary File 2 contains beta, D’ values for all reported clusters.

    Supplementary File 3 contains the SNP IDs and consensus alleles for all reported clusters in Tables 1 and 2. The consensus allele denotes the allele with the highest occurrence in the reported clusters.

    • The vast majority of the discoveries were in regions already known to be associated with their respective phenotypes based on standard GWAS.

    We agree that a majority of the ROH regions are indeed consistent with GWAS. However, some regions were missed by standard GWAS (e.g. chr6:25969631-26108168, hemochromatosis). Our message is that our method is a complementary approach to standard GWAS and will not replace standard GWAS analysis. See our response to Reviewer #2 Point Six.

    • The running time seems rather long (at least for the UK biobank), and therefore it will be difficult for other researchers to extensively experiment with the method in very large datasets. That being said, the method has a linear running time, so it is already faster than a naïve algorithm.

    Thank you for your input. The algorithm used to locate matching blocks is efficient and the total CPU hours it consumed was the reported run time. Since it consumes very little memory and resources, it can be executed simultaneously for all chromosomes. We also noticed that a significant time was being spent parsing the input file and slightly modified our script to improve the parsing. We also re-ran it for all chromosomes in parallel and reported the elapsed time which was only 18 hours and 54 minutes.

    “This was achieved by running the ROH-DICE program, with a wall clock time of 18 hours and 54 minutes where the program was executed for all chromosomes in parallel (total CPU hours of ~ 242.5 hours). The maximum residence size for each chromosome was approximately 180 MB.”

  3. eLife

    eLife assessment

    This study is of potential interest to readers in human genetics and quantitative genetics, as it presents a new method for homozygosity mapping in population-scale datasets, based on an innovative computational algorithm that efficiently identifies runs-of-homozygosity (ROH) segments shared by many individuals. Although the method is innovative and has the potential to be broadly useful, its power and limitations have not yet been adequately evaluated. The application of this new method to the UK Biobank dataset identifies several interesting associations, but it remains currently unclear under what conditions the new approach can provide additional power over existing genome-wide association study methods.

  4. eLife

    Reviewer #1 (Public Review):

    In this manuscript, Naseri et al. present a new strategy for identifying human genetic variants with recessive effects on disease risk by the genome-wide association of phenotype with long runs-of-homozygosity (ROH). The key step of this approach is the identification of long ROH segments shared by many individuals (termed "shared ROH diplotype clusters" by the authors), which is computationally intensive for large-scale genomic data. The authors circumvented this challenge by converting the original diploid genotype data to (pseudo-)haplotype data and modifying the existing positional Burrow-Wheeler transformation (PBWT) algorithms to enable an efficient search for haplotype blocks shared by many individuals. With this method, the authors identified over 1.8 million ROH diplotype clusters (each shared by at least 100 individuals) and 61 significant associations with various non-cancer diseases in the UK Biobank dataset.

    Overall, the study is well-motivated, highly innovative, and potentially impactful. Previous biobank-based studies of recessive genetic effects primarily focused on genome-wide aggregated ROH content, but this metric is a poor proxy for homozygosity of the recessive alleles at causal loci. Therefore, searching for the association between phenotype and specific variants in the homozygous state is a key next step towards discovering and understanding disease genes/alleles with recessive effects. That said, I have some concerns regarding the power and error rate of the methods, for both identification of ROH diplotype clusters and subsequent association mapping. In addition, some of the newly identified associations need further validation and careful consideration of potential artifacts (such as cryptic relatedness and environment sharing).

    (1) Identification of ROH diplotype clusters.
    The practice of randomly assigning heterozygous sites to a homozygous state is expected to introduce errors, leading to both false positives and false negatives. An advantage that the authors claim for this practice is to reduce false negatives due to occasional mismatch (possibly due to genotyping error, or mutation), but it's unclear how much the false positive rate is reduced compared to traditional ROH detection algorithm. The authors also justified the "random allele drawing" practice by arguing that "the rate of false positives should be low" for long ROH segments, which is likely true but is not backed up with quantitative analysis. As a result, it is unclear whether the trade-off between reducing FNs and introducing FPs makes the practice worthwhile (compared to calling ROHs in each individual with a standard approach first followed by scanning for shared diplotypes across individuals using BWT). I would like to see a combination of back-of-envelope calculation, simulation (with genotyping errors), and analysis of empirical data that characterize the performance of the proposed method.

    In particular, I find the high number of ROH clusters in MHC alarming, and I am not convinced that this can be fully explained by a high density of SNPs and low recombination rate in this region. The authors may provide further support for their hypothesis by examining the genome-wide relationship between ROH cluster abundance and local recombination rate (or mutation rate).

    (2) Power of ROH association. Given that the authors focused on long segments only (which is a limitation of the current method), I am concerned about the power of the association mapping strategy, because only a small fraction of causal alleles are expected to be present in long, homozygous haplotypes shared by many individuals. It would be useful to perform a power analysis to estimate what fraction of true causal variants with a given effect size can be detected with the current method. To demonstrate the general utility of this method, the authors also need to characterize the condition(s) under which this method could pick up association signals missed by standard GWAS with recessive effects considered. I suspect some variants with truly additive effects can also be picked up by the ROH association, which should be discussed in the manuscript to guide the interpretation of results.

    (3) False positives of ROH association. GWAS is notoriously prone to confounding by population and environmental stratification. Including leading principal components in association testing alleviates this issue but is not sufficient to remove the effects of recent demographic structure and local environment (Zaidi and Mathieson 2020 eLife). Similar confounding likely applies to homozygosity mapping and should be carefully considered. For example, it is possible that individuals who share a lot of ROH diplotypes tend to be remotely related and live near each other, thus sharing similar environments. Such scenarios need to be excluded to further support the association signals.

    (4) Validation of significant associations. It is reassuring that some of the top associations are indirectly corroborated by significant GWAS associations between the same disease and individual SNPs present in the ROH region (Tables 1 and 2). However, more sanity checks should be done to confirm consistency in direction of effect size (e.g., risk alleles at individual SNPs should be commonly present in risk-increasing ROH segment, and vice versa) and the presence of dominance effect.

  5. eLife

    Reviewer #2 (Public Review):

    The authors have proposed a computational algorithm to identify runs of homozygosity (ROH) segments in a generally outbred population and then study the association of ROH with self-reported disorders in the UK biobank. The algorithm certainly identifies such segments. However, more work is needed to justify the importance of ROH.

  6. eLife

    Reviewer #3 (Public Review):

    A classic method to detect recessive disease variants is homozygosity mapping, where affected individuals in a pedigree are scanned for the presence of runs of homozygosity (ROH) intersecting in a given region. The method could in theory be extended to biobanks with large samples of unrelated individuals; however, no efficient method was available (to the best of my knowledge) for detecting overlapping clusters of ROH in such large samples. In this paper, the authors developed such a method based on the PBWT data structure. They applied the method to the UK biobank, finding a number of associations, some of them not discovered in single SNP associations.

    Major strengths:
    • The method is innovative and algorithmically elegant and interesting. It achieves its purpose of efficiently and accurately detecting ROH clusters overlapping in a given region. It is therefore a major methodological advance.
    • The method could be very useful for many other researchers interested in detecting recessive variants associated with any phenotype.
    • The statistical analysis of the UK biobank data is solid and the results that were highlighted are interesting and supported by the data.

    Major weaknesses:
    • The positions and IDs of the ROH clusters in the UK biobank are not available for other researchers. This means that other researchers will not be able to follow up on the results of the present paper.
    • The vast majority of the discoveries were in regions already known to be associated with their respective phenotypes based on standard GWAS.
    • The running time seems rather long (at least for the UK biobank), and therefore it will be difficult for other researchers to extensively experiment with the method in very large datasets. That being said, the method has a linear running time, so it is already faster than a naïve algorithm.

  7. Version published to 10.1101/2020.10.26.20220004v1 on medRxiv