Ethics statement

This study utilized de-identified data from the UK Biobank. UK Biobank received ethical approval from the NHS National Research Ethics Service North West (11/NW/0382).

UK Biobank sample

Our data came from the UK Biobank, a population-based sample from the United Kingdom. In total, 502,682 individuals were recruited from 2006–2010 from 22 centers across the UK. Participants were given a touchscreen interview that included questions about demographic characteristics, health history, and lifestyle information (e.g. diet, alcohol intake, sleep habits), and some anthropometric and physical measures were collected. DNA was extracted from whole blood and genotyped using either the Affymetrix UK Biobank Axiom array or the Affymetrix UK BiLEVE Axiom array. Detailed genotyping and sample QC procedures are described in Bycroft et al. [23] We analyzed data from the second release of up to 400,000 individuals (exact N varied by phenotype) with genotypes available.

Phenotypes

We examined 26 traits related to health, fitness, or sociodemographic characteristics (see S1 Text for full description and field ID of individual measures). These included 17 continuous traits (age at first sexual intercourse, waist to hip ratio, height, body mass index (BMI), basal metabolic rate (BMR), diastolic and systolic blood pressure (BP), hand-grip strength (taking the maximum of left and right grip strength measurements), county-wide socioeconomic status (SES) as measured by the Townsend Deprivation Index (TDI), total household income (an ordinal variable of income brackets recoded to be numeric, ranging from 0–4), years of educational attainment (coded using ISCED classifications as in Okbay et al. [24]), fluid intelligence score (FI), forced expiratory volume in 1 second (FEV1; a measure of lung function), FEV1 over forced vital capacity (FEV1/FVC), birth weight, neuroticism score, and body fat percentage) and 9 binary traits (ever smoked, ever drank alcohol, whether or not they were breastfed as a baby, whether or not they completed college, whether they specified participation in a religious group as a leisure activity, whether or not they had ever been diagnosed with diabetes, probable bipolar and/or major depression status, and whether they live in an urban or rural area). These traits were chosen because they had either been previously studied in ROH analyses (fluid intelligence, grip strength, FEV1, FEV1/FVC ratio, waist-to-hip ratio, height, BMI, diastolic and systolic blood pressure [7–10]), were diagnoses or binary traits of psychiatric and biomedical interest with large enough Ns to reasonably include in our study (diabetes, probable bipolar or major depressive disorder diagnoses, ever smoked, ever drank), or were our only available proxy measure of reproductive success (age at first sexual intercourse). Our “sociodemographic” traits were chosen because of their hypothesized (regional poverty via Townsend Deprivation Index [25], breastfed as an infant as a proxy for mother’s socioeconomic status [26], income, urbanicity) or previously demonstrated (education, religious assortment [15,27]) influence on ROH–trait associations. We excluded individuals who weighed less than 36.28 kg (~80 lbs), weighed more than 6.8 kg (~15 lbs) at birth, had systolic BP readings >200 mmHg or diastolic BP readings >120 mmHg, were shorter than 120 cm (~3.93 ft), had a hip circumference <50 cm or >175 cm, had a waist circumference <40 cm or >160 cm, had grip strength >70 kg, or reported having had sex before 12 years of age. These exclusion criteria were chosen based on thresholds typically defined as being boundaries of normal physiological, anthropometric, or behavioral ranges and by checking for obvious outliers that may have been incorrect data entries. More information on specific phenotype derivations and calculations are included in the supplemental material. We standardized all quantitative phenotypes (within sex) before calculating their relationship with F_ROH for ease of comparison with Joshi et al.’s and others’ results [7].

Quality control (QC)

Because the sample was predominately European ancestry, we restricted analyses to individuals of European ancestry (N = 436,065) as identified by visual inspection of plots of genomic principal components. We followed sample and genotypic quality control that has become typical in ROH analyses. In particular, we excluded SNPs if they a) deviated from Hardy-Weinberg equilibrium at p<1×10⁻⁶, b) missing call rate > 0.02, or c) had a minor allele frequency (MAF) < 0.05. We also excluded individuals with discordant self-reported gender and genetic sex, a missing genotype call rate > 0.02, and we removed the minimum number of individuals so that all remaining subjects were unrelated at > 0.2 (using GCTA’s—grm-cutoff option [28]) (n = 31,541 removed in total). We also repeated these analyses after using a stricter relatedness cutoff, removing all individuals related at > 0.05 (n = 103,389 removed in total; see S1 Text), to ensure the robustness of our results.

ROH calling procedures

After QC, we pruned out SNPs that were in strong linkage disequilibrium with other SNPs by removing those that had a variance inflation factor > 10 (equivalent to an r² of 0.90) between target SNPs and 50 surrounding SNPs (plink command:—indep 50 5 10). After these procedures, 263,609 SNPs and 404,524 individuals remained. For our main analysis, we called ROHs as being ≥65 homozygous SNPs in a row, set the minimum KB length very low (essentially ignoring the length requirements), with no heterozygote calls and three missing variant calls allowed (5% of the SNP threshold), per recommendations from Howrigan et al. (2011) for genotype data of similar SNP density. We required ROHs to have a density greater than at least 1 SNP per 200 kb (the average density across the genome in the SNPs used in the analysis was 1 per 10 kb) and split an ROH into two if a gap >500 kb existed between consecutive homozygous SNPs. These analyses used the following commands in Plink 1.9 [29]:—homozyg-window-snp 65—homozyg-snp 65—homozyg-kb 10—homozyg-gap 500—homozyg-window-missing 3—homozyg-window-het 0—homozyg-density 200 (see S1 Text for further discussion of parameter choice). After calling ROHs, we summed the total length of all autosomal ROHs for each individual and divided that by the total SNP-mappable distance (2.77x10⁹ bases) to calculate F_ROH, the proportion of the genome likely to be autozygous.

We also tested the relative importance of recent vs. distant inbreeding by calculating F_ROH from longer ROHs (indicative of closer inbreeding) and comparing to the effect of F_ROH from shorter ROHs (a proxy for more distant inbreeding). We defined recent inbreeding as the proportion of the genome contained in autozygous regions longer than 8.5 Mb (F_{ROH_long}) and distant inbreeding as the proportion of the genome in autozygous regions shorter than 8.5 Mb (F_{ROH_short}), as F_{ROH_long} and F_{ROH_short} had approximately equal variances (4.5e-6 and 4.3e-6, respectively) in our sample. An autozygous segment spanning < 8.5 Mb should originate from a common ancestor at least 6 generations ago on average [30].

In addition to calling ROHs, we also calculated a measure of SNP-by-SNP homozygosity (F_SNP) for each individual, using the—het flag in Plink 1.9 [29]:

F_SNP = [observed homozygous count—expected count] / [total observations—expected count]

Because it is calculated with genotyped SNPs, F_SNP is a measure of excess homozygosity at common SNPs (see S3 Table for the correlation matrix between F_SNP, F_{ROH_long} and F_{ROH_short}).

ROH burden analysis

F_ROH was used as the primary predictor of the traits of interest in analyses described below. The distributions of ROH lengths and F_ROH are shown in S1 Fig. We regressed each trait (Y) on F_ROH using the model in the equation below, where is the intercept, C is a matrix of covariates (including e.g. the first 20 principal components) and ε represents the residual error term.

As noted above, all quantitative phenotypes were standardized to intra-sex z-scores for ease of comparison with previous findings in the literature. In addition, for ease of interpretation, we reverse-coded some of the phenotypes such that lower values represented what we thought were likely to be lower fitness and/or less desirable outcomes (e.g. disease diagnosis was coded as ‘0’ while no diagnosis was coded as ‘1’, and TDI was reverse-coded such that lower values represented greater material poverty). We were primarily interested in the estimate of , which represents the association of F_ROH with the trait, controlling for covariates (although in one set of models, described below, we were also interested in the effect of F_SNP on the trait). For binary traits, we ran logistic regression models with the same covariates as in the linear regression models for quantitative traits.

We ran a total of three sets of models for each trait. The first set of models was designed to test for a simple relationship between F_ROH and the traits listed above. Because confounding factors such as population stratification, SNP missingness, call quality, and plate effects can influence F_ROH, we included the batch number, percentage of missing SNP calls per sample in the non-imputed genotype data, and the first 20 ancestry principal components (calculated within individuals of European ancestry), as well as age, age², and sex, in all of the regression models unless explicitly stated.

In our second set of models, we tested whether background sociodemographic characteristics mediated F_ROH-trait relationships. In addition to the above covariates, in these models we also included income, years of educational attainment, Townsend Deprivation Index (a measure of the amount of material deprivation in a given region [25]), and whether subjects attended college, lived in an urban area, participated in a religious group as a leisure activity, and reported being breastfed as an infant. Although the covariates of true interest are those measured on the parents (whose sociodemographic traits may influence mate choice), parental information was unavailable (other than breastfeeding, which is associated with mother’s socioeconomic status [26]), and so we used the subjects’ own values on these traits as the best available proxies for characteristics of their parents.

In our third set of models, we tested the degree to which observed F_ROH-trait relationships were due to homozygosity at common versus rare alleles. To do this, we included F_SNP as a covariate in addition to the covariates from the second set of models above. Because common SNPs can often predict (are in linkage disequilibrium with) other common SNPs but typically poorly predict rare SNPs, F_SNP captures effects of homozygosity at common SNPs only whereas F_ROH captures the effects of homozygosity at both common and rare SNPs [5]. In the S1 Text (and S4 Table), we demonstrate via simulation that entering both F_SNP and F_ROH as predictors simultaneously in the regression equation allows insight into the degree to which observed inbreeding effects are due to homozygosity at common versus rare alleles.

Overview of findings

We replicated several previous associations between F_ROH and fitness-related traits, identified a novel association between F_ROH and a reproductive phenotype (age at first sexual intercourse), and found weak evidence that background sociodemographic characteristics may be partially mediating a few of the observed relationships between F_ROH and complex traits (Fig 2). In particular, we found robust evidence that fluid intelligence (FI), forced expiratory volume in one second (FEV1), and age at first sexual intercourse (AFS) are associated with F_ROH (Tables 1 and 2), while grip strength and height’s relationships with F_ROH were attenuated enough to become non-significant after controlling for background sociodemographic variables. The associations of F_ROH with FI and FEV1 were especially robust, with the inclusion of sociodemographic covariates having little to no consequence on these effect sizes. The strength of F_ROH associations for more recent inbreeding was similar or stronger than those for more distant inbreeding, except, interestingly, for AFS, FI, and participation in a religious group. When we accounted for SNP-by-SNP homozygosity in the model, AFS and FI were still significantly associated with F_ROH, consistent with their relationships with F_ROH being more strongly driven by homozygosity at rare rather than common variants. Certain other associations were likely due to social rather than genetic causes; for example, it is much more plausible that non-religious individuals tend to outbreed at higher rates and have less religious offspring than that autozygosity causes individuals to be more religious.

Comparison with previous results

Our results largely agree with recent reports [7–10] on the relationships between estimated autozygosity and complex traits in population-based samples. Replicating Howrigan et al. [8], Joshi et al. [7], and Yengo et al.[10], as well as previous pedigree studies [31], we found a significant, negative relationship between F_ROH and fluid intelligence. In addition, we replicated Joshi et al.’s [7] finding of a significant relationship between increased F_ROH and decreased FEV1. We initially observed a significant association between increased F_ROH and decreased height, as did Joshi et al. [7] and Verweij et al. [9], but this association was attenuated in our sample after controlling for background sociodemographic variables and did not meet statistical significance after Bonferroni corrections (Table 1 and Fig 2). Our initial results (Table 1) were consistent with previous findings for an effect of inbreeding depression on grip strength [9], though this association appears to be more likely due to homozygosity at common rather than rare variants (Table 2).

Despite the general consistency across reports on F_ROH-complex trait associations, there were two differences between our results and those from earlier studies. First, educational attainment (in years of education) was not significantly associated with F_ROH in any of our models, contrary to several previous reports [7,9], but consistent with Yengo et al. [10]. We found a significant (p = 2.18e-4) relationship between F_ROH and income (which itself was correlated with years of education at r = 0.37), but we found no evidence for an association between F_ROH and either years of education or the binary variable measuring whether or not an individual attended college. The reason for the discrepancy in findings for education is unlikely to be due to sampling variability because the two confidence intervals do not overlap (Fig 2). One possibility is that educational attainment is less correlated with geographic mobility (and the tendency to outbreed) in the UK compared to other countries previously investigated, and Joshi et al. [7] did report significant heterogeneity of the F_ROH-education association across sites. Moreover, of the 5 cohorts from the UK investigated by Joshi et al. [7], two (GRAPHIC and LBC1936) showed associations in the opposite direction of the overall association (see their Extended Data Fig 2). Thus, it is possible that the F_ROH -educational attainment relationship might be different in the UK than is typical in other societies. Furthermore, the association we found between height and autozygosity was attenuated (by ~37%) when we accounted for sociodemographic covariates, and was somewhat smaller than that found by previous studies even when we did not control for sociodemographic variables (e.g. a 1% increase in F_ROH predicted a decrease of ~.03 s.d. of height in previous studies [7,32] versus a decrease of ~.02 s.d. in the current study). Nevertheless, the confidence intervals for Joshi et al.’s [7] and our observed association between height and F_ROH overlapped (Fig 2), suggesting that sampling variability could be a reason for the discrepant height findings.

In comparing results across recent publications and the current one, it is important to note the differences in populations, samples, and measurements across studies. Both Howrigan et al. [8] and Joshi et al. [7] took a meta-analytic approach, conducting F_ROH analyses in each contributing sample separately, and then combining across samples, controlling for relevant covariates (e.g. dataset, country of data collection). Joshi et al. in particular analyzed a much more diverse overall sample than the present study, including multiple cohorts from European, African, and Asian populations. Another difference is in the measurement of intelligence across studies: our measurement for general cognitive ability was the unweighted sum of the number of 13 fluid intelligence questions answered correctly, given as part of the UK Biobank’s cognitive function assessment, while Howrigan et al. [8] converted the scores from each contributing sample’s measure of general cognitive ability (e.g. WAIS-R, Cattell Culture Fair Test) into z-scores (to avoid bias from different measurement schemes across samples), and Joshi et al. used g as their measure of general cognitive ability, “calculated as the first unrotated principal component of test scores across diverse domains of cognition”. Furthermore, our regression models controlled for the first 20 ancestry principal components, while Howrigan et al. controlled for the first 10 and Joshi et al. the first 3.

Possible evolutionary interpretations

There are two major evolutionary theories for why inbreeding depression occurs [4]: the overdominance hypothesis posits that an overall loss of heterozygosity at loci governed by heterozygote advantage leads to inbreeding depression, while the partial dominance theory postulates that inbreeding depression occurs as selection acts most efficiently on the most additive and dominant deleterious mutations, purging those from the population while leaving behind the more rare, partially recessive deleterious alleles. This second hypothesis, partial dominance, is widely accepted as the more likely mechanism of inbreeding depression [3,33]. The robust associations we observed between F_ROH and AFS, FI, and FEV1, even after controlling for homozygosity at common variants with F_SNP, suggest that the variants contributing to lower trait values are biased toward being rare and recessive, consistent with predictions from a partial dominance model of inbreeding depression [5] and consistent with the hypothesis that these traits, or traits genetically correlated with them, have been under directional selection over evolutionary time. Cognitive ability, including intelligence test scores, is a predictor of multiple Darwinian fitness-related outcomes, including overall health and lifespan [8,34]. FEV1 is correlated with mortality and lifespan [35–38], traits that are components of fitness and thus more likely to have been under directional selection over evolutionary history [39]. Thus, our replication of the associations between autozygosity and FEV1 and FI adds to a body of evidence that these traits, or traits genetically correlated with them, have been under directional selection over evolutionary history, consistent with the expectation that variants that influence them are biased toward being rarer and more recessive than expected under a neutral drift model.

The positive relationship we observed between AFS and F_ROH is a novel finding, to the best of our knowledge, though associations between F_ROH and reproductive phenotypes have been observed previously, for different proxy measures (e.g. number of children fathered, in Yengo et al. [10]). The F_ROH-AFS association was attenuated but remained statistically significant after controlling for sociodemographic variables and homozygosity at common variants (F_SNP). Our finding is consistent with a body of research suggesting that reproductive traits, like AFS, in non-human populations are under more intense selection pressures than non-fitness traits [5,40]. If autozygosity causally influences AFS (see “Limitations” below), there are two possible evolutionary interpretations. First, it is possible that early sex itself was advantageous in ancient human history due to a prolonged reproductive period. A second possibility is that the observed association between autozygosity and AFS is due to selection on a genetically correlated trait, such as sexual attractiveness [41,42]. However, it is important to note that the original effect size for the association between F_ROH and AFS decreased by ~20% after accounting for sociodemographic variables. Furthermore, sociodemographic variables were not measured on parents (the more relevant control; see below), and the single dichotomous variable for religious group participation that we used as a proxy for religiosity is unlikely to capture the full confounding effects of religious observance on sexual behaviors. Thus, the association between inbreeding and age at first sexual intercourse that we observed in this study should be interpreted with caution.

Limitations

There were three central limitations in the current study. The most important one, which applies equally to all other F_ROH studies that we are aware of, is that ROH associations might be due to third-variable explanations. Unlike GWAS analyses, where parental or offspring sociodemographic traits are unlikely to be associated with allele frequencies and therefore are unlikely to bias GWAS results, it takes only a single generation of parental inbreeding to strongly influence F_ROH levels in offspring. For example, higher income might be associated with greater opportunities to meet mates of diverse origins and to higher outbreeding; offspring of higher income parents might thereby have not only lower levels of autozygosity, on average, but might also differ on any traits influenced genetically or environmentally by parental income. While sociodemographic confounding is particularly problematic in ascertained samples where cases and controls are drawn from different populations (e.g. cases drawn from a psychiatric hospital, controls from a nearby university), the possibility of confounding cannot be eliminated, even in population-based samples, unless relevant sociodemographic variables among parents are measured and controlled for or other (e.g., within-family) designs are used. For example, in a study of approximately 2,000 individuals of Dutch ancestry, Abdellaoui et al. [27] found only a weak association between F_ROH and the subjects’ own educational attainment (p = 0.045), but found highly significant negative associations between the subject’s F_ROH and their parents’ educational attainment (p_father < 10⁻⁵, p_mother = 9e^-5). These relationships were entirely mediated by the geographic distance between parents’ birthplaces, such that parents with higher educational attainment tended to be more geographically mobile, increasing their chances of mating with someone genetically dissimilar from themselves and thus creating systematic differences in levels of inbreeding across levels of educational attainment in their offspring.

Having information on parents’ birth location, education, income, mobility, level of religious involvement, and so forth is important in order to control for the possibility that these sociodemographic variables are associated with both higher levels of (distant) inbreeding and lower offspring trait values. Unfortunately, the UK Biobank has limited parental information other than indirect measures such as whether one was breastfed. In the current study, we used sociodemographic responses of the offspring as imperfect proxies for parental responses, which is effective only to the degree that offspring values on these sociodemographic variables are positively correlated with their parents’ values. For example, parental educational (r = 0.25–0.40; [43,44]), income (r = .60; [44]), and religiosity [45] are imperfectly correlated between parents and offspring in Great Britain. These imperfect correlations imply that the true mediating influences of the sociodemographic variables on observed F_ROH -trait relationships were likely to be underestimated in the present report, and thus causal interpretation of our results may not be warranted.

Still, while F_{ROH_short} is an imperfect measure of truly distant inbreeding (as recent inbreeding can also produce short ROHs), it may be that F_{ROH_short} is less susceptible to confounding from recent assortative mating than F_{ROH_long} (a measure of more recent inbreeding). The three traits (AFS, FI, and FEV1) that were significantly associated with F_ROH even after the inclusion of sociodemographic covariates were also significantly associated with F_{ROH_short} (S6 and S8 Tables). This bolsters our hypothesis that autozygosity is causally influencing these three traits through inbreeding depression, while the evidence for height and grip strength (which were associated with recent but not distant inbreeding, S6 Table) is weaker.

A second limitation is the potential for a collider bias in the models in which we controlled for sociodemographic variables. Collider bias occurs when a covariate in a linear model is causally influenced by both the outcome and the predictor of interest, which creates a bias in the estimated association between the predictor and the outcome [46,47]. In the context of the current study, it is possible that income is negatively influenced by autozygosity and positively influenced by intelligence. If so, then the regression coefficient for F_ROH predicting fluid intelligence would be upwardly biased when income is included as a covariate. Unfortunately, it was not possible to know the degree to which our results were influenced by a collider bias given that the directions of causation between variables are unknown. Nevertheless, it has been argued that the potential bias from failing to adjust for a covariate is likely to be greater than the bias introduced when controlling for a collider [48,49]. Thus, our solution was to present results from models that both controlled and did not control for the sociodemographic covariates, and to highlight the potential for biases in models that controlled for sociodemographic variables.

The third limitation to the current study is that we did not have access to all of the phenotypes studied in recent articles such as Yengo et al. [10], Verweij et al. [9], Joshi et al. [7] (e.g. the cholesterol measures in Joshi et al.), so we could not attempt to fully replicate these previous investigations. In addition, as mentioned earlier, the covariate that we had available as a proxy for religiosity was unlikely to account for all possible confounding of religiosity on F_ROH-trait associations.

Summary

We found several significant associations between estimated autozygosity and several sociodemographic, anthropometric, health, and otherwise fitness-related traits. All effects were in the direction that would be predicted by evolutionary hypotheses (i.e. higher inbreeding associated with lower fitness). When controlling for measures of background sociodemographic characteristics (educational attainment, college education, income, urbanicity, TDI, religious participation, and whether an individual was breastfed)–which should at least partially reflect parental characteristics–we found that two (height and grip strength) of the five significant F_ROH-trait associations were attenuated and became non-significant, while AFS, FI, and FEV1 remained significantly associated with F_ROH. The fact that the associations between estimated autozygosity and both grip strength and height were reduced after controlling for the additional covariates suggests that these relationships might not hold up if relevant confounder variables in parents had been controlled for, and we cannot eliminate the possibility that the other F_ROH-trait associations we report here would not also be attenuated or eliminated in this situation.

Nevertheless, our results generally replicate previous findings in humans [7–9] and are consistent with similar ones from non-human populations [40,50,51]. This cumulative evidence may well reflect the detrimental effects of autozygosity on complex traits, revealing ancient selection pressures on these or correlated traits. However, the fact remains that even in very large, well-powered, unascertained samples such as this one, it is exceedingly difficult to make definitive statements about the underlying causal mechanism of observed relationships between F_ROH and complex traits.