The Influence of Demographic History and Genetic Architecture on Complex Traits via Runs of Homozygosity

Abstract

Runs of homozygosity (ROH) are contiguous genomic regions where all sites are homozygous, inherited from identical haplotypes due to shared ancestry. The number and length of ROH in individuals varies based on population history and sociocultural behaviors. Although often discussed in the context of inbreeding, ROH are ubiquitous in putatively outbred human populations, and their prevalence are associated with multiple complex traits, including height and measures of lung function. Importantly, ROH have been shown to be enriched for deleterious alleles, suggesting a mechanism by which ROH prevalence can influence traits. Here we employ realistic forward-in-time population genetic simulations and a flexible quantitative model of a generic complex phenotype to explore how population history and genetic architecture influence ROH associations with a generic quantitative phenotype. We show that ROH are important for all simulated demographic histories and genetic architectures but especially when phenotypes have a recessive component. This is even more prominent when the rare-allele contribution to the phenotype is upweighted and in high-diversity populations (e.g. African). For a fully recessive phenotype, ROH can account for 25–45% of an individual’s total phenotype score, depending on demographic history and rare-allele weight. Our results emphasize the utility of ROH in helping to explain phenotype variation across different population histories and genetic architectures.

Introduction

Runs of homozygosity (ROH) are contiguous genomic regions in an individual where all sites are homozygous, resulting from the inheritance of identical haplotypes from both parents due to common ancestry at some point in the past. In 1999, Broman and Weber¹ first identified numerous long homozygous chromosomal segments using genomic data from the Centre d’Étude du Polymorphisme Humain (CEPH). At that time, researchers hypothesized that these long homozygous segments are prevalent even in outbred populations, potentially providing more insights into human health and broader population genetics than previously expected. Several years later, in 2006, Gibson² analyzed the number, length and distribution of long homozygous chromosomal segments from outbred HapMap populations and identified three outlier individuals with extremely long and abundant homozygous segments. In the same year, Li et al³ reported similar observations in a Taiwanese population and suggested these long homozygous chromosomal segments likely represent autozygosity instead of the effects of gene deletion, recombination, or other genomic events. These findings prompted the formal definition of Runs of Homozygosity (ROH)⁴ and the associated metric F_ROH,⁵ a novel estimator of the inbreeding coefficient.

Building upon this early work, researchers have gained a greater insight into the distribution of ROH and the factors that create them. Studies have shown that ROH are ubiquitously distributed across putatively outbred global human populations.^6,7 For different populations, their unique demographic history and sociocultural preferences, for example, consanguineous marriage, can shape the length and number of ROH among the individuals of their contemporary populations.^8,9 Indeed, ROH prevalence has been used as a powerful tool in population genetics to provide insights about demographic events such as bottlenecks, founder events, and inbreeding in humans and other species.^10–15

ROH analyses have deepened our understanding of the influence of inbreeding depression on complex traits in various species.^16,17 For example, a study in wild Soay sheep¹⁸ found that long ROH reduce the likelihood of surviving the first winter more than short ROH. Across both wild and managed populations, ROH have also been used to study the impact of inbreeding and inbreeding depression on milk production traits^19–21 in livestock. ROH also informs selection analyses in horse breeds²² and supports the design of management programs for endangered populations such as wild wolves,²³ North American Thoroughbred horses,²⁴ and Indian tigers.²⁵

ROH are also informative in medical genomics studies to elucidate the genetic basis of human diseases.²⁶ Previous studies have demonstrated the associations between ROH and increased risk of schizophrenia,^4,27 autism,²⁸ human height,^29,30 cancer,^31,32 blood pressure,^33,34 depression,³⁵ and cardiovascular diseases,³⁶ which highlights the value of ROH in assessing genetic health risks within populations and across populations.³⁷ Total ROH levels have even been associated with direct fitness effects in humans.³⁸

Indeed, previous studies have suggested the mechanism by which ROH may be associated with traits. Szpiech et al.³⁹ demonstrated that long runs of homozygosity are enriched for deleterious mutations worldwide. These results demonstrated the potential of long runs of homozygosity to harbor significant number of rare deleterious mutations, which are usually are low frequency in the population but get paired into homozygotes within ROH. In a study of ROH in admixed populations, Szpiech et al.⁴⁰ demonstrated that ROH overlapping African ancestry regions accumulated deleterious homozygotes at a higher rate compared to those overlapping European or Native American ancestry segments, demonstrating that demographic history influences the strength of enrichment. These results suggest that demographic history may play an important role in understanding the relationship between ROH and the genetic architecture of complex traits.

Unlike Mendelian traits driven primarily by variants at single loci, complex traits possess a polygenic architecture, arising from the cumulative effects of multiple genetic variants and their interplay with environmental factors. Genome-wide association studies (GWAS) and polygenic risk scores (PRS) have provided robust quantitative frameworks for studying complex traits, providing statistical tools to understand the genetic basis of complex traits. However, their effectiveness is still limited by their inability to detect very rare causal alleles⁴¹ and by population biases. Although many rare variant association tests (RVATs) have been developed,^42,43 these methods often fail to account for the evolutionary forces that help to shape genetic architecture. Existing studies have shown that when realistic evolutionary factors such as population growth and natural selection are included, the statistical performance of these methods diminishes.⁴¹ As a result, it remains difficult to detect associations between rare variants and phenotypes or to draw general conclusions. In contrast, ROH, by capturing homozygous regions enriched for rare, recessive deleterious alleles³⁹ and reflecting unique population demographic histories, offer a complementary perspective for studying genetic contributions to complex traits.

However, previous research has rarely considered a comprehensive analysis of how distinct population histories systematically influence ROH patterns and their relationship to the distribution of phenotypes. To address this gap, we use detailed forward-in-time population-genetic simulations to characterize how demographic history influences the relationship between ROH and the genetic architecture of a simulated complex phenotype. Instead of simulating a specific phenotype, we adopted an omnigenic model. The omnigenic model proposes that complex traits are determined by a vast number of genes distributed across gene regulatory networks and the whole genome, rather than by a small number of specific genes.⁴⁴ The existence of phenotype correlations with total ROH levels implies that causal alleles for these phenotypes are widely distributed across the genome.⁴⁵ Thus, by simulating these genetic effects across the entire genome, our work establishes a generic quantitative framework for assessing the relationship between ROH and “omnigenic” phenotypes. Additionally, we investigate the influence of dominance coefficients and assess the significance of rare allele contributions to phenotype scores, offering potential guidance for future methodological developments in population and medical genetics.

Methods

Simulating Genetic Data

We simulated genetic data using SLiM 4.1⁴⁶ and msprime⁴⁷ under a three-population Out-of-Africa demographic model from Gravel et al.⁴⁸ (for demographic parameters see SLiM recipe 5.4). For a realistic genome structure, we simulated 100 Mbps genome segment with an exon structure based on the CCDS⁴⁹ coordinates (GRCh37) from the first 100 Mbps of human chromosome 1. We used a variable recombination rate map based on the HapMap3⁵⁰ recombination map. Deleterious mutations were restricted to exon regions; thus, the deleterious mutation rate was set to 2.36×10⁻⁹, which is 10% of the mutation rate of 2.36×10⁻⁸ for the entire region. Deleterious selection coefficients were sampled from a gamma distribution with shape 0.2 and mean −0.03. We considered four different scenarios for dominance: fully recessive (h = 0), fully additive (h = 0.5), fully dominant (h = 1), and a mixed model including all three.

To determine the relative proportions of recessive, additive, and dominant mutations in the mixed dominance model, we proceeded as follows. First, we downloaded the GRCh37 gene annotation file^51,52 and restricted it to only autosomal genes. Next, we downloaded two gene lists^53,54 generated by the MacArthur group (https://github.com/macarthur-lab/gene_lists) corresponding to genes that are deemed to follow autosomal dominant inheritance or recessive inheritance based on Online Mendelian Inheritance in Man (OMIM) database. We downloaded a third list (“universal”⁵⁵) and assumed that genes present in the universal list but absent from both the dominant list and the recessive list can be classified as additive. For each gene, we calculated the maximum transcript length across the different transcript versions. Then, we summed the total transcript length among all genes classified as dominant, recessive, or additive, and then divided by the total transcript length of all the genes considered. For the simulations, the relative proportion of dominance coefficients for new mutations is given by these proportions, calculated as dominant:additive:recessive = 0.04549:0.8839:0.07059.

To speed up our simulations, we did not use SLiM to simulate neutral mutations. Instead, we generated tree sequences and used msprime to recapitate and overlay neutral mutations at a mutation rate of 2.124 × 10⁻⁸. For each population, we sampled 500 diploid individuals, and we simulated 500 replicates per parameter combination.

Simulating Phenotypes

We simulated phenotypes under two separate scenarios, one where causal mutations are deleterious and one where causal mutations are neutral. For phenotypes generated from deleterious mutations, we assumed all deleterious mutations contribute to the phenotype. We used a quantitative model derived from Eyre-Walker, Adam⁵⁶ and Uricchio, Lawrence H et al,⁴¹ which explicitly relates the selection coefficient of the mutation to an effect size. For a given selection coefficient, $s$, the phenotype score ${𝒵}_s$ is given by

$${𝒵}_s=\left\{\begin{array}{ll}\delta |s{|}^{\tau }& \hfill \mathit{with}\mathit{probability}\rho \\ \delta {\left|s_r\right|}^{\tau }& \hfill \mathit{with}\mathit{probability}1-\rho \end{array}\right.$$ (1)

In this model, $\delta$ and $\tau$ allow the marginal distribution of effects to differ from the marginal distribution of selection coefficients. As we are interested in directional effects, we set $\delta =1$. When $\tau$ increases, it up-weights the significance of rare-allele contributions to phenotypes, and, when $\tau$ decreases, it up-weights the significance of common-allele contributions to phenotypes. The parameter $\rho$ allows for the introduction of pleiotropy without altering the overall marginal distribution of effects, as $s_r$ is generated from a new draw from the marginal distribution of selection coefficients. For phenotypes generated from deleterious alleles we set $\rho =1$. For each sampled individual, the total phenotype score, $Z$, was computed as follows.

$$Z=\left\{\begin{array}{ll}\sum _{j}I\left(X_j==2\right)Z_{s,j}& \hfill \mathit{if}\mathit{recessive}\\ \sum _j\frac{1}{2}{X}_j{Z}_{s,j}& \hfill \mathit{if}\mathit{additive}\\ \sum _{j}I\left(X_j\ge 1\right)Z_{s,j}& \hfill \mathit{if}\mathit{dominant}\end{array}\right.$$ (2)

where $X_j$ is the number of effect alleles at locus $\mathrm{j}$, and $I\left(A\right)$ is an indicator function that takes 150 the value 1 when proposition A is true and 0 when false. The sum is across all effect loci. The inheritance pattern of the phenotype effect (i.e., recessive, additive, or dominant), is determined by the dominance coefficient of the selection coefficient. We considered $\tau \in \{0.7,0.8,0.9,1.0,1.1,1.2,1.3\}$ to examine the influence of changing the significance of rare alleles.

For phenotypes generated from only neutral alleles, we used the above phenotype model but set $\rho =0$, so that we randomly drew effects from the marginal distribution of selection coefficients and transform accordingly. However, as there are many more neutral alleles than deleterious ones in the genome, we down sampled the number of neutral effect alleles that contribute to the phenotype as follows.

For a given parameter combination and dominance regime, we calculated the distribution of the total number of deleterious mutations for all populations combined across simulation replicates. We fit a normal distribution and employed the Shapiro-Wilk test for testing goodness of fit. The results of this statistical test confirmed that the data fits well. (See Supplemental material. Table S7) Then, we sampled from this distribution to determine the number of neutral effect alleles for each replicate with a matching parameter and dominance regime. Using this number of effect alleles, we randomly choose neutral mutations that fall within exons and assign an effect size using the above model. Although most loci are biallelic, occasionally a locus with >1 derived allele is encountered. In this case, we randomly choose a derived allele at that locus to be the effect allele. Importantly, the dominance of phenotype for neutral alleles follows the dominance pattern of the deleterious alleles that were simulated. In the case of mixed dominance simulations, we randomly assign a neutral allele’s phenotype dominance based on the mixed proportions described above.

Finally, for each parameter set and replicate, we standardize each individual’s phenotype score by dividing by the standard deviation of the distribution of phenotypes across all sampled individuals in all three populations. By placing the phenotype scores from all sampled individuals on a common, global variation-based scale, standardization facilitates meaningful comparisons between simulations under different parameter sets.

Runs of Homozygosity

To identify runs of homozygosity in the genotype data, we first generated tped files from VCF file produced by msprime using PLINK⁵⁷ (v1.90b6.21) (www.cog-genomics.org/plink/1.9/). Then, we called runs of homozygosity by using GARLIC.^7,58,59

Based on the output of GARLIC, we obtained the size class and location of all the ROH regions in all the sampled individuals of all the replicates. We also calculated the total length of ROH for each size class in all the sampled individuals of all the replicates. These results are used for analyzing the relationship between ROH and phenotype score. The parameters used in GARLIC are set as follows: --map plink.chr1.GRCh37.map, --size-bounds 0.25 1, --centromere dummy.centromeres.txt, --winsize 10, --auto-winsize, --auto-overlap-frac. By using this software, we classified the ROH in genotype data generated from each simulation into short ROH (shorter than 0.25cM), medium ROH (longer than 0.25cM but shorter than 1cM), and long ROH (longer than 1cM).

Statistical Analysis

For each replicate, we place individuals into three categories: individuals with no ROH, individuals with at least one ROH that contributes to the phenotype, and individuals with ROH but none contribute to the phenotype. Each of these categories are further stratified by ROH size class, and then we calculate the mean across replicates.

For all population models and genetic architectures, we calculated the phenotype score per-cM contributed by all ROH and non-ROH regions which contribute to the phenotype. The per-cM contribution $\left(C\right)$ for a specific genomic region type ( $k$, where $k\in \{longROH,mediumROH,shortROH,non-ROH\}$) in an individual $\left(i\right)$ was calculated as the ratio of the total phenotype score contributed by that region $\left(P_{i,k}\right)$ to the total length of region in $\mathrm{c}\mathrm{M}\left(L_{i,k}\right)$:

$$C_{i,k}=\frac{P_{i,k}}{L_{i,k}}$$

To obtain a single summary value for each replicate ($r$), we then calculated the mean of the per-cM contributions from all 500 sampled individuals of each population $(N=500)$. When calculating the mean, we only considered individuals that have ROH of the given size class and only ROH that contributed to the phenotype score. This mean value, denoted as $\overline{C_{r,k}}$, represents the result for that replicate:

$$\overline{C_{r,k}}=\frac{1}{N}\sum _{i=1}^N{C}_{i,k}=\frac{1}{N}\sum _{i=1}^N(\frac{P_{i,k}}{L_{i,k}}),P_{i,k},L_{i,k}\ne 0.$$

Where $\mathrm{N}$ is the total number of individuals satisfying the conditions above and is not necessarily 500. Moreover, we calculated the difference in per-cM contribution between different categories of ROH and non-ROH regions under various scenarios as a function of tau values. The difference in per-cM contribution $\left(D_r\right)$ between different categories of ROH and non-ROH regions is calculated as follows:

$$D_{r,\mathit{short}ROH}=\overline{C_{r,\mathit{short}ROH}}-\overline{C_{r,non-ROH}}$$

$$D_{r,\mathit{medium}ROH}=\overline{C_{r,\mathit{medium}ROH}}-\overline{C_{r,non-ROH}}$$

$$D_{r,\mathit{long}ROH}=\overline{C_{r,\mathit{long}ROH}}-\overline{C_{r,non-ROH}}$$

Finally, we examined the proportion $\left({Prop}_{i,k}\right)$ of the total phenotype score attributed to different categories of ROH and non-ROH regions across these simulations. We first calculated the total phenotype score for each individual ($P_{i,\mathit{total}}$) by following:

$$P_{i,\mathit{total}}=P_{i,\mathit{short}\mathit{ROH}}+P_{i,\mathit{medium}\mathit{ROH}}+P_{i,\mathit{long}\mathit{ROH}}+P_{i,\mathit{non}-ROH}$$

Then, the proportion of phenotype scores attributed to different ROH and non-ROH regions are calculated as follows:

$${\mathit{Prop}}_{i,k}=\frac{P_{i,k}}{P_{i,\mathit{total}}}$$

The summary value for each replicate is calculated by following:

$$\overline{{\mathit{}\mathit{Prop}}_{r,k}}=\frac{1}{N}\sum _{i=1}^N{\mathit{Prop}}_{i,k}$$

To statistically evaluate the impact of different genomic regions on phenotype scores, for each combination of population model, dominance scenario, and tau value, we applied a series of pairwise comparisons on the distributions of per-cM contributions generated from the 500 replicates. We used the Wilcoxon signed-rank test to account for the paired nature of the data within each replicate. This test was applied to compare: (i) the contribution of each ROH category (Short, Medium, and Long) against that of the non-ROH regions, and (ii) the contributions between different ROH categories themselves (e.g., Long ROH vs. Medium ROH). To correct for the multiple comparisons performed across all pairwise tests and experimental conditions, the resulting p-values were adjusted using the Holm-Bonferroni method. (See Supplemental material. Table S1 and S2.)

Results

Proportion of individuals with ROH contributing to the phenotype

Runs of Homozygosity were called in each individual and categorized into three size classes, Short ROH (<0.25 cM), Medium ROH (0.25–1.0 cM), and Long ROH (>1.0 cM). Figure 1 shows the mean proportion of individuals across simulation replicates for each ROH size category, where blue represents the proportion of individuals that have no ROH of a particular size, red represents the proportion of individuals with at least 1 ROH of that size that also contributes to the phenotype, and yellow represents the proportion of individuals that have ROH of that size but none contribute to the phenotype.

Figure 1.

Proportion of individuals with ROH contributing to phenotype across various population histories and genetic architectures. Phenotype scores are only contributed by deleterious mutations. Blue bars represent the individuals lacking ROH (thus no phenotype score contribution); red bars represent the individuals having at least one ROH with a phenotype score contribution; yellow bars represent the individuals having ROH with no phenotype score contribution. Region definitions: Short ROH (< 0.25cM); Medium ROH (0.25–1cM); Long ROH (> 1cM).

We find that phenotype architecture doesn’t have strong influence on these proportions, but demographic history does influence the proportion of individuals with ROH > 0.25 cM that contribute to the phenotype. In the African population, about 50% of individuals have at least one medium ROH with non-zero phenotype score contribution. The proportion of individuals in the European population having medium ROH with non-zero phenotype score contribution is higher, reaching 75%, and the proportion in the East Asian population is higher still, reaching 87.5%.

Interestingly, for long ROH, the differences between populations are small than those of medium ROH. As these populations are simulated with no explicit inbreeding, the proportion of individuals with long ROH are low, and approximately 10% of individuals from each of these three populations have long ROH with non-zero phenotype score contribution. Interestingly, the proportion of individuals in the African population with at least one long ROH contributing to the phenotype is slightly larger than that in the European and East Asian populations (approximately 12% vs. 10% and 10%, respectively), even though the total proportion of individuals with long ROH is lower (approximately 25% vs. 27% and 28%, respectively).

Phenotype score contribution per centimorgan

Next, we consider the relative contributions of ROH to the phenotype across a range of architectures conditional on an ROH having non-zero contribution. In order to fairly compare ROH of different sizes, we normalize by ROH length, thus we analyze the phenotype contribution per centimorgan of total ROH of different size classes.

Additive Model

Under the additive model the functional allele has a phenotype effect regardless of zygosity, with heterozygotes having half the effect of homozygotes. Figures 2 and S1 shows that short ROH contribute significantly more per cM to phenotype scores than non-ROH, regardless of demographic history and the value of tau. Interestingly, short ROH also contribute significantly more per cM to phenotype score than long ROH across these parameters, and long ROH contribute more than medium ROH, except for larger tau in the European and East Asian populations. In contrast, long ROH and medium ROH were more sensitive to the value of tau and demographic history. When emphasis is shifted to common alleles (small tau), long and medium ROH contributed significantly more to phenotype score than non-ROH, however, when emphasis is shifted to rare alleles (large tau), this pattern flips.

Figure 2.

Per-cM phenotype score contribution (log-transformed; deleterious mutations) for various regions, population histories, and genetic architectures. Regions are: Short ROH (< 0.25cM); Medium ROH (0.25–1cM); Long ROH (> 1cM); Non-ROH. Phenotypes are generated from deleterious mutations. The parameter tau is varied for different weighting of rare alleles in phenotype score calculation. (A) $\tau =0.7$, (B) $\tau =1.0$, (C) $\tau =1.3$. Significance levels for adjusted P-values: $\mathrm{***}{P}_{\mathit{adj}}<0.001,\mathrm{*}\mathrm{*}{P}_{\mathit{adj}}<0.01,\mathrm{*}{P}_{adj}<0.05$, ns $P_{adj}\ge 0.05$.

Dominant Model

Under the dominant model the functional allele has a phenotype effect regardless of zygosity, however in this case both heterozygotes and homozygotes have full effect. Figures 2 and S1 shows that ROH contributions per cM to phenotype score follow similar patterns as in the additive model. Short ROH contribute significantly more per cM to phenotype scores than non-ROH and long ROH. For medium ROH, the influence of demographic history is clear. For the African population, medium ROH have a significantly higher contribution to phenotype score per cM compared to non-ROH except for the largest tau, however for European and East Asian populations, medium ROH generally contribute lower compared to non-ROH. Long ROH have significantly more contribution per cM compared to non-ROH only when emphasis is shifted to common alleles (small tau), but this pattern flips when emphasis is put on rare alleles (large tau).

Recessive Model

Under the recessive model, the functional allele has a phenotype effect only in homozygous form. Figures 2 and S1 shows that under the recessive model, all three ROH categories across the three populations exhibited significantly higher contributions to polygenic phenotype scores than non-ROH regions, with these differences being consistently statistically significant. Among the three ROH categories, long ROH regions exhibited an outstanding signal, having significantly higher phenotype score per cM than short, medium, or non ROH regions. This pattern was robust to changes in emphasis on common vs rare alleles (changing tau).

Mixed Dominance Model

Under the mixed dominance model, the relationship between functional allele phenotype effect and zygosity is randomly determined, with approximately 88.4% additive, 7.1% recessive, and 4.5% dominant (see Methods). Figures 2 and S1 shows trends that generally reflect the patterns of the single inheritance models mixed at the above proportions. Interestingly, as the phenotype score per cM for recessive inheritance models dramatically favors ROH over non-ROH regions, even a relatively small recessive component drives the importance of ROH regions across several parameters. For models with an emphasis on common alleles (small tau), short, medium, and large ROH show significantly higher phenotype score per cM compared to non-ROH. As emphasis shifts to rarer alleles (large tau), the influence of demographic history begins to differentiate patterns somewhat. Under an African demographic history, all ROH regions show significantly higher phenotype score per cM compared to non-ROH, but this pattern changes for European and East Asian demographic histories. For European histories only short and medium ROH have a significantly higher phenotype score per cM, and for East Asian histories only short ROH have a significantly higher phenotype score per cM.

Changing emphasis from common to rare alleles

Equation 1 transforms a mutation’s selection coefficient into a phenotype effect, and the parameter tau indirectly modulates whether a phenotype score is relatively more influenced by common variation (small tau) versus rare variation (large tau). This occurs since large tau accentuates large absolute selection coefficients more than small ones, and large absolute selection coefficients (i.e., highly deleterious alleles) are more likely to be rare. However, by changing the exponent tau, both the mean and variance of the phenotype effect distribution is changed, which may obscure important comparative patterns when looking across tau values. To address the changing variances, the total phenotype scores per individual are normalized by the standard deviation across all populations, however mean centering would disallow a log transform. To explore how changing tau influences the relative per-cM phenotype distribution among ROH and non-ROH regions, figure 5 plots the difference between ROH and non-ROH log mean per-cM phenotype contributions as a function of tau. In this figure, positive values indicate that the mean per-cM phenotype contribution of the ROH is higer than the non-ROH, and negative values indicate that non-ROH per-cM phenotype contribution is higher.

Figure 5.

Difference (in log space) in per-cM phenotype score contribution between ROH and non-ROH regions with varied tau, various population histories and genetic architectures. Phenotype scores are only contributed by deleterious mutations. Panels show the difference in contribution between the specific ROH regions and non-ROH regions: (A) Short ROH regions (< 0.25 cM); (B) Medium ROH regions (0.25–1cM); (C) Long ROH regions (> 1 cM).

We find that genetic dominance has the largest effect on these patterns, with the largest positive values occurring with the recessive phenotype model across all ROH sizes. Furthermore, as tau increases weight on rare alleles, we find that ROH regions contain substantially more per-cM phenotype effect compared to non-ROH regions, with long ROH showing the highest differences and biggest changes. Demographic history also plays a role, with African demographies having the highest difference in ROH vs non-ROH per-cM phenotype contribution.

In contrast to the recessive phenotype model, the additive and dominant phenotype models show more modest differences between ROH and non-ROH per-cM phenotype scores, with the dominant phenotype model generally showing the smallest positive differences. In some cases, especially for large tau among medium and long ROH, we find that non-ROH regions have an increased and higher per-cM phenotype score compared to ROH. Interestingly, changing tau has minimal effect on the difference between short ROH and non-ROH per-cM phenotype effects for these phenotype models, and, for non-African demographies under the dominant model, there is even a small increase in the per-cM phenotype effect difference for larger tau.

The proportion of phenotype explained by ROH

Next, we examine how much of the total phenotype score is explained by ROH versus non-ROH regions. Different from the per-centimorgan phenotype score calculation, we considered all the individuals here, with no regard to ROH content of the genomes.

Figure 7 illustrates these patterns divided by short, medium, long, and non-ROH regions. Generally, for dominant and additive phenotypes, non-ROH regions explain the vast majority of a phenotype’s total score. This is particularly notable for dominant phenotypes under the African demographic history, where ~90% of the phenotype is explained by mutations in non-ROH regions. European and East Asian demographic histories, which have experienced a historical bottleneck, see somewhat more total phenotype explained by ROH for dominant and additive phenotypes. Notably, under a purely recessive phenotype model, total ROH explains a substantially larger proportion of the phenotype, ranging from approximately 25% to 55%, depending on the demographic history and emphasis on common versus rare alleles (low versus high tau).

Figure 7.

Proportion of total phenotype score per individual explained by different types of regions with varied tau, various population histories and genetic architectures. Phenotype scores are only contributed by deleterious mutations. Panels show the proportion of phenotype score explained by (A) Short ROH regions (< 0.25 cM); (B) Medium ROH regions (0.25–1cM); (C) Long ROH regions (> 1 cM); (D) Non-ROH.

Indeed, while changing the emphasis of common versus rare variation doesn’t have an appreciable effect on the proportion of phenotype explained for other inheritance regimes, for recessive phenotypes there is a clear trend. When the phenotype effect is shifted to more heavily favor rare alleles (large tau), ROH regions increasingly explain more of the phenotype score. In particular, the contribution proportion of short ROH regions to the phenotype scores ranged from 22 to 45%, medium ROH ranged from 4 to 10%, and Long ROH ranged from 0.5 to 2%.

Phenotypes from neutral alleles

The previous results presented here considered a phenotype that is determined purely from deleterious alleles and whose total score is a function of the underlying selection coefficients of those alleles. We also considered the case where a phenotype was determined purely from neutral alleles (of similar number) simulated in the presence of deleterious alleles (i.e., background selection). In this case the phenotype score was determined by sampling from the same marginal distribution of selection coefficients and computing the phenotype effect using the same functional transformation, but with the actual selective effect of the alleles remaining neutral.

Similar to Figure 1, Figure 3 shows the mean proportion of individuals across simulation replicates for each ROH size category, where blue represents the proportion of individuals that have no ROH of a particular size, red represents the proportion of individuals with at least 1 ROH of that size that also contributes to the phenotype, and yellow represents the proportion of individuals that have ROH of that size but none contribute to the phenotype. As before, the differences within the results are primarily observed across the ROH size classes and population histories, while there are minimal differences across genetic architectures.

Figure 3.

Proportion of individuals with ROH contributing to phenotype across various population histories and genetic architectures. Phenotype scores are only contributed by neutral mutations. Blue bars represent the individuals lacking ROH (thus no phenotype score contribution); red bars represent the individuals having at least one ROH with a phenotype score contribution; yellow bars represent the individuals having ROH with no phenotype score contribution. Region definitions: Short ROH (< 0.25cM); Medium ROH (0.25–1cM); Long ROH (> 1cM).

Notably, although the proportion of individuals with ROH are approximately the same, the proportion of individuals with at least 1 ROH contributing to the phenotype is higher for medium and long ROH when phenotypes are determined from neutral alleles. In the African population, 87.5% of individuals have medium ROH with non-zero phenotype score contribution. About 10% of individuals have medium ROH with zero phenotype score contribution, while a very small proportion do not have medium ROH. All individuals in the European and East Asian populations have medium ROH with non-zero phenotype score contribution. Slightly below 25% of individuals in each of these three populations have long ROH with non-zero phenotype score contribution. In all these populations, a small number of individuals having long ROH without phenotype score contribution. In particular, the proportion in the East Asian population is slightly higher than that in the European population, while the proportion in the European population is also slightly higher than that in the African population. Still, regardless of genetic architecture, all the individuals from these three populations have short ROH with non-zero phenotype score contribution.

Figures 4 and S2 show the phenotype score per cM for short, medium, long, and non ROH under this model for varying genetic architectures, demographic histories, and inheritance regimes. Generally, none of these variables have as conspicuous an effect compared to the case where phenotypes are determined by deleterious alleles.

Figure 4.

Per-cM phenotype score contribution (log-transformed; neutral mutations) for various regions, population histories, and genetic architectures. Regions are: Short ROH (< 0.25cM); Medium ROH (0.25–1cM); Long ROH (> 1cM); Non-ROH. Phenotypes are generated from deleterious mutations. The parameter tau is varied for different weighting of rare alleles in phenotype score calculation. (A) $\tau =0.7$, (B) $\tau =1.0$, (C) $\tau =1.3$. Significance levels for adjusted P-values: $\mathrm{***}{P}_{adj}<0.001,\mathrm{*}\mathrm{*}{P}_{adj}<0.01,\mathrm{*}{P}_{adj}<0.05$, ns $P_{adj}\ge 0.05$.

Under the recessive model, all ROH regions in the African demographic history had a significantly higher phenotype score per cM compared to non ROH regions, regardless of the emphasis on common or rare variant (value of tau). Under this model, in the European and East Asian demographic histories, only short ROH and medium ROH regions have a significantly higher phenotype score per cM compared to non ROH regions, regardless of tau. These patterns contrast with the patterns in Figures 2 and S1, where long ROH consistently had the highest phenotype score per cM.

Similarly, under the additive, dominant, and mixed models, short ROH were the only ROH region that consistently had a significantly higher phenotype score per cM compared to non ROH regions across all populations and tau values. Under these models, the phenotype contribution per cM of long ROH regions were consistently either equal to or less than that of the non-ROH region. However, medium ROH was only significantly higher compared to non-ROH regions in the African population under the additive and mixed models, and this pattern was consistent across tau values.

Notably, the European and East Asian demographic histories showed similar patterns to each other across all models, but they differed from the African demographic history. Specifically, under all models and tau values, the per-cM phenotype score contribution of long ROH in the African population was slightly higher than that in European and East Asian populations (See Supplemental material. Table S1 and S2.). Although the per-cM phenotype score contribution of long ROH was equal to or significantly less than non-ROH regions in all models other than the recessive model.

Figure 8 shows the phenotype score contribution proportion by ROH and non-ROH for phenotypes with neutral effect alleles. The range of contribution proportion for ROH and non-ROH regions was similar to that in the deleterious scenario, with the exception of the recessive case. Not only do ROH contribute proportionally less to the phenotype under the recessive neutral model compared to the recessive deleterious model (Figure 7), but the proportions do also not change with increasing tau, suggesting that shifting phenotype contribution from common to rare variants doesn’t influence ROH importance.

Figure 8.

Proportion of total phenotype score per individual explained by different types of regions with varied tau, various population histories and genetic architectures. Phenotype scores are only contributed by neutral mutations. Panels show the proportion of phenotype score explained by (A) Short ROH regions (< 0.25 cM); (B) Medium ROH regions (0.25–1cM); (C) Long ROH regions (> 1 cM); (D) Non-ROH.

Under the neutral model, figure 6 shows that varying tau does not have much effect on the difference between ROH and non-ROH per-cM phenotype effect. This is likely due to the fact that under this model the phenotype effect and the selective effect of the allele are decoupled, therefore decoupling the relationship between phenotype effect and allele frequency. However, we still see that under the recessive model ROH generally have larger per-cM phenotype effect than non-ROH, although this difference is largest for short ROH and smallest for long ROH. In fact, for long ROH, it is only under the African demography that long ROH have higher per-cM phenotype effect compared to non-ROH.

Figure 6.

Difference (in log space) in per-cM phenotype score contribution between ROH and non-ROH regions with varied tau, various population histories and genetic architectures. Phenotype scores are only contributed by neutral mutations. Panels show the difference in contribution between the specific ROH regions and non-ROH regions: (A) Short ROH regions (< 0.25 cM); (B) Medium ROH regions (0.25–1cM); (C) Long ROH regions (> 1 cM).

Discussion

Runs of homozygosity (ROH), as widely observed genomic features, are gradually emerging as a key factor in understanding the genetic basis of complex traits.^{6,26,37,39,40,45} However, the specific mechanisms by which ROH influence complex trait phenotypes, particularly how these mechanisms are systematically modulated by the combined effects of population history and genetic architecture, remain largely unknown. To address this gap, our study systematically investigated the interactions between these factors using large-scale realistic simulations. Our results reveal that the contribution of ROH to complex traits phenotype score is important but variable. Under the recessive model, ROH, especially long ROH, play a significant and central role, but under other genetic architectures, the contribution patterns differ based on the underlying model.

Proportion of individuals with ROH

As observed in Figures 1 and 3, many individuals lack ROH regions that contribute to phenotype score. While overall levels of ROH are generally determined by the demographic history, the contrast between Figures 1 and 3 indicate that when the phenotype is determined by deleterious alleles, ROH that have a phenotype effect are subject to purging. Importantly, functional alleles were simulated only within exon regions, which induce a clustering of functional alleles along the genome. If intronic and intergenic alleles can also contribute to phenotypes than we may expect ROH to contain further enrichments of functional variation beyond coding variation. As the three populations we simulated followed random mating with only two populations experiencing a historical bottleneck, long ROH were relatively rare across populations. However, we note that modern human populations are more complex than these demographic models, and founder effects,⁹ recent bottlenecks,⁶⁰ and localized fine structure⁶¹ can influence long ROH levels in the genome.

Previous studies have focused on the relationship between the total length of ROH (sROH) and total number of ROH (nROH) of sampled individuals in relation to phenotypes.⁴⁵ Our study shows that ROH, especially the long ROH with non-zero phenotype score contribution, exist in only a small proportion of individuals. If we combine the phenotype score contributed by these ROH across all sampled individuals, it will inevitably lead to substantial downward bias in the relationship between ROH and phenotype. We propose that future work on the relationship between ROH and phenotype attempt to model these two types of ROH, those with no phenotype contribution and those with phenotype contribution.

The recessive model

Among all the genetic architectures tested, ROH contribute the most to the complex trait phenotype score under the recessive model. Under this model, our results demonstrate significant cross-population applicability: regardless of population history, in all the populations, the per-cM phenotype score contribution of all ROH regions, and especially long ROH, is systematically higher than that of non-ROH regions (Figure 2, S1). Furthermore, this pattern exacerbates when the weight of rare-allele contribution to phenotype score is increased (Figures 2, 5, S1).

The core mechanism behind this lies in the difference in the probability of recessive alleles forming homozygotes in different genomic regions. In non-ROH regions, for an effect allele with frequency $p$, homozygotes are expected to form at a rate proportional to $p^2$. However, within ROH regions homozygotes are expected to form at a rate proportional to $p$, and, for a rare allele, $p\gg p^2$. Furthermore, changing the weight of rare-allele contributions to the phenotype underscore this pattern. As rare alleles become relatively more important contributors to phenotype (increasing tau), the difference between the per-cM phenotype score contribution of ROH regions compared to non-ROH regions increases, with long ROH increasing the most (Figure 5). Under the recessive model with deleterious effect alleles, only homozygotes are exposed to negative selection, and it is these factors that drive the significant differences between the per-cM phenotype score contribution between ROH and non-ROH regions.

These mechanisms also explain the pattern that long ROH have a significantly higher per-cM phenotype score contribution than medium and short ROH. This is closely related to the age of the underlying haplotype and the effect sizes of deleterious mutations they carry. Long ROH are comprised of young haplotypes and are the direct result of recent inbreeding or bottlenecks and other recent evolutionary events. Due to their age, natural selection has not had time to eliminate any deleterious mutations they may carry. This leads to long ROH having a higher chance of being enriched for the deleterious mutations with high effect. Thus, they have the highest per-cM phenotype score contribution. In contrast, the older short and medium ROH are comprised of haplotypes that have segregated for a longer time, with more opportunity to be exposed to negative selection. As a result, deleterious mutations with strong effects would be mostly purged by natural selection, resulting in their relatively low per-cM phenotype score contribution.

However, when we shift focus from per-cM phenotype score contribution to phenotype score contribution proportion (Figure 7), the order of contribution of ROH regions is changed, with short ROH > medium ROH > long ROH. This is mainly because we have simulated outbred populations and long ROH are relatively infrequent and generated due to small effective population size only. Indeed, long ancestral haplotypes are constantly broken up by recombination events, and thus ROH are biased toward being comprised of shorter haplotypes that have persisted to the present day. Therefore, though the per-cM phenotype score contribution of short ROH is the lowest among all the ROH categories, their large number in the genome allows them to have the highest phenotype score contribution proportion. Although long ROH have the highest per-cM phenotype score contribution, they are rarer and younger and thus have the lowest phenotype score contribution proportion (Figure 3). However, these patterns also highlight that, in the presence of a harsh bottleneck or increased consanguinity, increased prevalence of long ROH could account for a substantial proportion of a phenotype. Indeed, these patterns and expectations likely explain the importance of ROH in highly endogamous populations.^38,62–64

The additive and dominant models

When the genetic architecture is changed from the recessive model to the additive and dominant models, the patterns of ROH contribution to phenotype score is significantly altered, with differences observed in per-cM phenotype score contribution, phenotype score contribution proportion, and its response to changing rare-allele weight (tau value).

Whereas, under the recessive model, ROH had higher per-cM phenotype contribution compared to non-ROH regions, under the additive and dominant models, the pattern of per-cM phenotype score contribution of ROH regions is reversed, with short ROH regions typically showing the most prominent effect (Figures 2, S1). More importantly, under these two models, as the tau value increases, only short ROH regions consistently show a higher effect per-cM over non-ROH regions (Figure 5).

We propose these patterns are due to the fundamental differences in the efficiency of natural selection under different genetic architectures. Under the additive and dominant models, regardless of zygosity, the deleterious effect of mutations is exposed to the pressure of natural selection, resulting in the elimination of mutations with strong effect. In contrast to the recessive model, long ROH, which are comprised of “young” haplotypes, cannot effectively accumulate a large number of deleterious mutations with high effect size because they are efficiently eliminated by selection under the additive and dominant models, leaving only weak effect alleles to persist. Hence, the per-cM phenotype score contribution of long ROH regions exhibits significant decline compared to those under the recessive model. As the ancient haplotypes that comprise short ROH survived in the evolutionary process, the persistent existence of short ROH may itself imply certain characteristics of the regions in which they are located, such as a slightly lower recombination rate. In population genetics, the lower recombination rate decreases the efficiency of natural selection in eliminating weak deleterious mutations. Therefore, these short ROH regions have accumulated more weak deleterious mutations over a long period of time, which may explain the higher per-cM phenotype score contribution compared to long ROH and non-ROH regions.

Notably, the relative per-cM phenotype score contribution of long and medium ROH is dependent on the tau value. When tau is small, emphasizing common variants, the long and medium ROH have higher per-cM phenotype score contribution than non-ROH regions across all populations under additive and dominant models, but as tau increases, emphasizing rare variants, this advantage reverses. Thus, when tau is low, the phenotype score of long and medium ROH is highly dependent on the common alleles they carry. When tau increases, because they lack the long-term accumulation of rare alleles, their phenotype scores decrease sharply. Although non-ROH also experience a decline in phenotype scores due to the increasing tau, their genetic diversity allows rare variants to buffer the decline in phenotype scores. It is precisely this difference in vulnerability to changes in tau that leads to the reversal of contribution differences between long/medium ROH and non-ROH. However, we also noticed an exception: under the dominant model, the medium ROH in European and East Asian populations didn’t show the advantage in per-cM phenotype score contribution over non-ROH even when tau is small. This significant population-specific pattern is discussed in a subsequent section.

Additionally, we observed that the value of tau for which longer ROH explain less phenotype than non-ROH under the dominant model is lower than that under the additive model. This is likely because of the different efficiency of natural selection under these models. Under the dominant model, the full effect of a deleterious allele is fully exposed in heterozygotes, which leads to the higher purging efficiency. Thus, the phenotype score per-cM under the dominant model is more sensitive to tau values.

The additive and dominant models show consistent patterns with increasing tau leading to a decrease in per-cM phenotype score contribution for both ROH and non-ROH regions, which is in contrast to the recessive model. On the other hand, the phenotype score contribution proportion of ROH regions under the additive and dominant models is not only lower than that under the recessive model, but also remains generally unchanged with the increasing tau. The former is because the deleterious heterozygotes in non-ROH regions can contribute to phenotype score too, thereby increasing the total phenotype score and decreasing the proportion of ROH regions. The lader is because, under these two models, as shown in Figure 5, the relative per-cM phenotype score contribution between ROH and non-ROH regions does not change much. Therefore, their phenotype score contribution proportion remains stable even when tau increases. These results clearly demonstrate that the contribution of ROH regions to polygenic phenotype scores is highly dependent on the underlying genetic architecture of the trait.

Population history: a key factor influencing ROH phenotype contribution patterns

Population history is another factor that influences the ROH phenotype contribution patterns. Based on the classic out-of-Africa model, we systematically investigated its effect on “African”, “European”, and “East Asian” populations. These effects are mainly reflected in two aspects: per-cM phenotype score contribution and the phenotype score contribution proportion of each ROH and non-ROH region.

First, regarding the per-cM phenotype score contribution, the results under the recessive model stand out the most. The per-cM phenotype score contribution of ROH regions in the African population is significantly higher than those in the European and East Asian populations (Figure 2), while the European and East Asian populations have similar patterns. These results directly reflect the distinct population histories. Due to its long-term stable effective population size, the African population has the highest genetic diversity. This includes more rare deleterious mutations that are masked as heterozygotes but can potentially be enriched as homozygotes when paired identical-by-descent within ROH. In contrast, European and East Asian populations experienced a historical bottleneck event, which greatly reduced their genetic diversity and likely induced purging of strong deleterious alleles.

These patterns are exacerbated when changing tau to modify the relative influence of rare and common alleles. We observed that under the recessive model, with an increase in tau values, the relative contribution of ROH to the phenotype compared to non-ROH regions increases significantly more for African populations compared to European and East Asian populations (Figure 5). This underscores our expectation that the African population has a more abundant pool of rare variants. This means when the weight of rare alleles in phenotype score calculation increases, compared to the other two populations, the ROH regions in the African population are enriched for rarer/more deleterious variation.

Notably, an interesting pattern arises under the dominant model when examining the influence of population history. In the African population, when tau is low, medium ROH have significantly higher per-cM phenotype contribution compared to non-ROH regions (Figure 2). However, across all tested tau, in the European and East Asian populations these ROH have significantly less per-cM phenotype contribution compared to non-ROH regions. This pattern likely reflects the population bottlenecks experienced by these two populations, along with the accompanying founder effect, may have created a preponderance medium ROH that carry a lower-than-average load of dominant deleterious mutations.

Population history also shows a notable influence under the mixed dominance model. The mixed model is special because it includes three types of deleterious mutations with different genetic structures in proportion, approximately 88.4% additive, 7.1% recessive, and 4.5% dominant (see Methods). However, even though additive mutations have such a high proportion, the per-cM phenotype score contribution of ROH and non-ROH regions is only similar to those under the additive model in European and East Asian populations, but in the African population they similar to the recessive model (Figure 2). Once again, this is likely the result of high genetic diversity of the African population with many rare recessive alleles masked as heterozygotes, that get paired into homozygotes within ROH. In contrast with the European and East Asian populations for which many of these alleles were purged during their bottlenecks. Therefore, although the proportion of recessive alleles is low, in the African population, the impact they produce on phenotype scores when enriched in ROH is so great that their overall effect is disproportionately amplified, even enough to override the effects of additive mutations that dominate in terms of absolute numbers.

Finally, population history also has a consistent influence on the proportion of phenotype explained by different ROH regions. Specifically, we observed that the proportion of phenotype explained by short and medium ROH under all the models follows the order of East Asian > European > African (Figure 7A-B). However, the phenotype score contribution proportion of long ROH shows the reverse pattern, African > European > East Asian (Figure 7C). The pattern exacerbates when tau increases. These results highlight the importance of long ROH for explaining phenotypes in populations of high genetic diversity.

Phenotypes from neutral mutations as a comparison

To contextualize the results from above where the phenotype is derived from the fitness effects of individual mutations, we also considered a scenario where neutral mutations were the causal alleles as a reference. In this scenario, when we calculate the phenotype score of neutral mutations, the effect size of mutations is sampled from the marginal distribution of selection coefficients while the actual selection coefficient of the mutations remains 0. To control for any effects of background selection, we continue to simulate deleterious alleles in these cases, but they do not have any phenotype effect. Therefore, in Figure 3, the observation that all the individuals from the European and East Asian populations have at least one medium ROH with non-zero phenotype score is due to the bottleneck event and the random distribution of neutral mutations in exons. The random distribution also explains the increased proportion of individuals with non-zero contribution long ROH across all three populations, as well as the increased proportion of individuals with non-zero medium ROH from the African population.

Furthermore, under the neutral scenario, compared to non-ROH regions, the higher per-cM phenotype contribution observed in ROH regions (especially long ROH) under the recessive model disappeared. Similarly, all the key patterns observed in the deleterious model, such as the advantage of long ROH, the distinct patterns of African populations in mixed models, or the systematic effects of tau values, are either completely absent or greatly weakened in the neutral mutation scenario (Fig 4, 6, 8). These contrasts strongly demonstrate that the observations we obtained from deleterious scenarios are the results of the combined effects of natural selection, genetic architecture, and population history.

Limitations and Future Directions

Although these simulations are designed to be realistic, this study still has several limitations. Although this study systematically considered multiple different genetic architectures, our model did not include epistasis or gene-environment interactions (GxE^65,66), which are wide-spread in biology. The out-of-Africa model we used is widely used in forward-time simulations, but it is also an idealized and simplified model that does not fully capture the more complicated migration and admixture events in the actual human evolution history. In this case, the complicated evolutionary history⁶⁷ underlying the real human genomic data may further influence the distribution of ROH and their phenotype contribution patterns. The genomic structure we simulated in our model is based on the first 100 Mbps of human chromosome 1, which cannot fully represent the entire human genome. Furthermore, we only simulate causal alleles within exon regions and assume all alleles contribute to the phenotype. However, despite these limitations, we believe that the core conclusions of our study, that ROH regions contribute robustly and significantly to polygenetic phenotype under different genetic architectures and population histories, particularly in recessive models and populations with high genetic diversity, remain valid.

Our study provides new mechanistic insights into existing work in the field. Previous studies have found that long ROH can enrich deleterious mutations.^39,40 In this project, we have extended their work to broader systems with more complicated genetic architectures and population histories. In previous studies, ROH has often been used to investigate its association with diseases such as schizophrenia^4,26,27,35 and Alzheimer’s disease.^26,68–70 However, these associations have shown inconsistent conclusions across different populations,^6,71–74 leading to doubts about the utility of ROH. Our work shows that such inconsistencies are actually expected, as the effect of ROH is closely related to the population history of the sampled population and the genetic architectures of the trait.

In the future, the conclusions and observations from this study can be further verified in real human genome data. At the same time, the sensitivity of ROH to evolutionary history suggests its potential to be incorporated into inference methods. We also believe that the contribution patterns of ROH, especially the long ROH under the recessive model and short ROH under other models, can be helpful for method development in finding rare variants in genome data and exploring the missing heredity. In addition, future simulations can build upon our work to integrate more biological realities to explore the role of ROH in more realistic genetic landscapes.

This study systematically shows that ROH in the human genome are an underestimated but powerful source for understanding the genetic basis of complex traits. The significance of ROH goes far beyond that of an inbreeding indicator. ROH are dynamic features shaped by population history, whose function is defined by the genetic structure of traits. Our study demonstrates that the polygenetic phenotype score contribution of ROH is the result of both natural selection and the accumulation of deleterious mutations, with the relative importance of these two mechanisms determined by both genetic architecture and population history.

Code and data availability

All simulation scripts and analytical pipelines described in this study are available at the GitHub repository (https://github.com/Mingzuyu-Pan/ROH_phenotype_simulation).