A phenome-wide association study (PheWAS) in the Population Architecture using Genomics and Epidemiology (PAGE) study reveals potential pleiotropy in African Americans

We performed a hypothesis-generating phenome-wide association study (PheWAS) to identify and characterize cross-phenotype associations, where one SNP is associated with two or more phenotypes, between thousands of genetic variants assayed on the Metabochip and hundreds of phenotypes in 5,897 African Americans as part of the Population Architecture using Genomics and Epidemiology (PAGE) I study. The PAGE I study was a National Human Genome Research Institute-funded collaboration of four study sites accessing diverse epidemiologic studies genotyped on the Metabochip, a custom genotyping chip that has dense coverage of regions in the genome previously associated with cardio-metabolic traits and outcomes in mostly European-descent populations. Here we focus on identifying novel phenome-genome relationships, where SNPs are associated with more than one phenotype. To do this, we performed a PheWAS, testing each SNP on the Metabochip for an association with up to 273 phenotypes in the participating PAGE I study sites. We identified 133 putative pleiotropic variants, defined as SNPs associated at an empirically derived p-value threshold of p<0.01 in two or more PAGE study sites for two or more phenotype classes. We further annotated these PheWAS-identified variants using publicly available functional data and local genetic ancestry. Amongst our novel findings is SPARC rs4958487, associated with increased glucose levels and hypertension. SPARC has been implicated in the pathogenesis of diabetes and is also known to have a potential role in fibrosis, a common consequence of multiple conditions including hypertension. The SPARC example and others highlight the potential that PheWAS approaches have in improving our understanding of complex disease architecture by identifying novel relationships between genetic variants and an array of common human phenotypes.


Introduction
Pleiotropy, however defined, has long been recognized as a feature of genomes with respect to their relationships to individual traits and outcomes that characterize phenomes [1][2][3]. Interest in human pleiotropy has spiked in the last decade owing to the availability of large genotypephenotype datasets generated from genome-wide association studies (GWAS). The analysis and catalog collection of one phenotype versus many genotypes studies revealed that a sizable proportion of common genetic variants are associated with multiple related and independent phenotypes [4,5]. These observations have led to the development of more systematic approaches to identify variant-level pleiotropy [6,7], many of which have been applied to populations of mostly European-descent individuals ascertained in clinical settings (e.g., [8]).
Here, we describe a phenotype wide association study (PheWAS), a systematic approach to identify cross-phenotype associations, in the Population Architecture using Genomics and Epidemiology (PAGE) I study. The PAGE I study was established by the National Human Genome Research Institute (NHGRI) in 2008 with the intent to characterize GWAS-identified variants discovered in European populations using more diverse populations drawn from epidemiologic [9] and clinical [10] studies. The scope of the PAGE I study was subsequently expanded to include discovery and fine-mapping efforts using the Metabochip [11], a fixedcontent array of~200,000 variants designed to interrogate previously-identified GWAS variants as well as select genome regions related to cardio-metabolic traits for fine-mapping [12].
In this PheWAS, we investigated the associations between the 144,740 common genetic variants assayed on the Metabochip and 273 phenotypes collected in 5,897 African Americans participating in three epidemiologic PAGE I studies: the Atherosclerosis Risk in Communities (ARIC) [13]; Multiethnic Cohort (MEC) [14]; and the Women's Health Initiative (WHI) [15]. We identified 133 potentially pleiotropic variants, defined as associated with two or more phenotype classes at p<0.01 in two or more PAGE I study sites. We functionally annotated PheWAS-identified variants and characterized the local genetic ancestry in this admixed population. From these data, we highlight variants likely to be pleiotropic and worthy of further statistical and functional studies. These data also underscore the necessity of diversity in study populations and study designs in PheWAS to ensure that all possible genotype-phenotype human relationships are considered.

Results
For this PheWAS (Fig 1), we comprehensively tested for associations between 114,740 SNPs assayed on the Metabochip with up to 273 phenotypes (S1 Table) available for 5,897 African American participants from three PAGE I studies: Atherosclerosis Risk in Communities (ARIC); Multiethnic Cohort (MEC); and the Women's Health Initiative (WHI) ( Table 1). Due to variations in the data collected across these epidemiologic studies, some phenotypes were available in more than one study, such as C-reactive protein (CRP) and low density lipoprotein cholesterol (LDL-C), while other phenotypes were only available within a single study, such as albumin level measurements. In Methods we describe further the studies included in this PheWAS, details of Metabochip genotyping and quality control, and the PheWAS approach including phenotype classification and filtering by statistical significance.

Replication of previously described genotype-phenotype associations
We first performed comprehensive single SNP tests of associations for each PAGE I study across all SNPs with a minor allele frequency >1% on the Metabochip that passed quality control and all phenotypes available (Fig 2). Of note are the two association peaks on chromosomes 1 and 19. These peaks represent two previously known genotype-phenotype associations, and their identification here attests to the quality of this high-throughput PheWAS approach. The first association peak on chromosome 1 between OLFML2B rs6676438 and natural log-transformed white blood cell count (Table 2) recapitulates a known association in African Americans along this chromosomal region. OLFML2B rs6676438 is located on the short arm of chromosome 1 in a 90MB region known to be in linkage disequilibrium with the Duffy null allele (DARC rs2814778) and associated with hematological traits in African Americans [16]. The second most significant association peak on chromosome 19 (Fig 2) represents the known association between APOE rs7412 and natural log-transformed apolipoprotein B (Table 2) [17 -19]. Apolipoprotein B is the primary apolipoprotein of LDL-C, a phenotype heavily scrutinized by candidate gene, GWAS, and sequencing studies. From these studies, APOE rs7412 is known to be associated with LDL-C in multiple populations [20][21][22][23][24][25][26][27] including European Americans [18,19,[28][29][30] and African Americans [18,19,28,[30][31][32] as well as with related phenotypes such as response to statin therapy [33][34][35][36][37], small dense LDL-C [38], and lipid metabolism phenotypes for LDL-C and free cholesterol [39]. In the present PheWAS, APOE rs7412, along with nearby SNPs, were within 100kb of previously-reported GWAS associations and associated with the following lipid-related traits in a single PAGE I study (at p<1.0x10 -4 ): total cholesterol, LDL-C, response to statin therapy, lipid metabolism phenotypes, and hypertriglyceridemia (Fig 3).

Biology and the Clinical and Translational Science
Collaborative of Cleveland UL1TR000439 from the National Center for Advancing Translational Sciences (NCATS) component of the National Institutes of Health (NIH). Its contents are solely the responsibility of the authors and do not necessarily represent the official view of NCATS or NIH. SAP is employed by and receives salary from Genentech Inc. JMJ is employed by and receives salary from Illumina, Inc. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: SAP is employed by and receives salary from Genentech Inc, and JMJ is employed by and receives salary from Illumina, Inc. This does not alter our adherence with PLOS ONE policies on sharing data and materials. DCC is an academic editor of PLOS ONE. DCC is not involved in the review of this manuscript per journal policy. potential pleiotropy in African Americans, we grouped similar phenotypes measured in PAGE I studies into 30 phenotype classes regardless of genetic associations (Methods and Table 3). A PheWAS-identified variant then represented a variant associated with two or more phenotype classes meeting the significance threshold (Methods and S1 File). After phenotype class binning, we noted 133 SNPs associated with two or more distinct phenotype classes with the same direction of effect within a given phenotype class (S3 Table). As expected, the phenotype class combination 'LDL-C/total cholesterol levels' was associated with dozens (53) of the same SNPs. Also, 37 SNPs were associated with white blood count (WBC) coupled with other phenotype classes on chromosome 1, results likely driven by the Duffy polymorphism [16,44].  Table. Some of the phenotypes were measured in more than one study, some phenotypes were related to phenotypes of another study, and some phenotypes were unique measurements for a single study. Not all phenotypic measurements were available for all participants within each study. Maximum sample size and minimum sample size are dependent both on which individuals were genotyped and which individuals also had a specific phenotype measured. See Materials and methods for more information.

Study
Age Range (in years)  Table 2). The y-axis for each chromosome is the −log10(p-value), and the x-axis is chromosomal base pair location.
The remaining 43 PheWAS-identified associations (Table 4; Fig 6) represent 38 independent associations at r 2 �0.80 based on African population data from the 1000 Genomes Project [45]. Of these, seven (18.4%) PheWAS-identified variants were associated in the opposite direction between phenotype classes. Approximately half (20) of the phenotype-class combinations were associated with a single variant; the remainder were associated with more than one variant (Table 4). These multiple-associated phenotype classes were associated with two (insulin/height, body mass index/C-reactive protein, smoking/myocardial infarction, hypertension/ smoking) and three (smoking/LDL-C, hemoglobin/hematocrit, smoking/alcohol consumption) variants each. One PheWAS-identified variant (rs9349379) was associated with three phenotype classes (smoking/diabetes/hypertension; Table 4 and Fig 6).
Apart from the expected pleiotropic associations represented by the LDL-total cholesterol and white blood cell phenotype classes, this PheWAS in African Americans from PAGE revealed potentially novel pleiotropic relationships, notably with phenotype classes that represent common exposure, lifestyle, or environmental variables. For example, rs568938 was associated with both LDL-C and smoking phenotype classes (Table 4 and Fig 6). The LDL-C/ rs568938 association has been previously described in diverse populations [20,46]; however, the association with the smoking phenotype class is novel regardless of population. The direction of effect for these associations suggests that the coded allele of rs568938 is associated with both increasing LDL-C and duration of smoking reported (Table 4), results that are consistent with epidemiological studies that describe a relationship between smoking and increased LDL-C [47]. Likewise, DOCK7 rs10889334, previously associated with total cholesterol [48] and cardiovascular disease [49] via linkage disequilibrium, was also associated with LDL-C and smoking phenotype classes in the same direction (Table 4). Among the non-LDL-C associations, PheWAS-identified PHACTR1 rs9349379 was associated with the three phenotype classes of smoking, diabetes, and hypertension in opposing directions. The PHACTR1 association with hypertension in this PheWAS is supported by the recent GWAS literature for blood pressure [49][50][51]. In contrast, the opposite-direction-of-effect association observed for smoking and diabetes is not yet supported by genetic data but instead supported by some of the epidemiologic literature where those who report current smoking have lower blood pressure and less hypertension compared with non-smokers (e.g., [52]). Other exposure, lifestyle, and environmental phenotype classes implicated in this PheWAS include alcohol consumption and hormone use (Table 4).

Functional and ancestral annotation of potentially pleiotropic SNPs
To better understand the functional impact of the 38 potentially pleiotropic SNPs (Table 4), we implemented several in silico annotation approaches for these as well as proxy SNPs (in linkage disequilibrium at r 2 �0.8) using various public resources, including HaploReg v4.1 [53], RegulomeDB v1.1 [54], and the SNP and CNV Annotation Database (SCAN) [55]. Almost all (94.7%) of these 38 PheWAS-identified variants are intronic (20) or intergenic (16), with the remaining two classified as synonymous (rs114374279) and missense (rs76394293; Table 5). We note that most (24) PheWAS-identified variants were annotated as associated with gene expression or as expression quantitative trait loci (eQTL) in at least one resource used here (Table 5).
We also estimated local genetic ancestry at these 38 PheWAS-identified loci given that African Americans are admixed, with varying proportions of African, European, and other ancestral alleles throughout the genome (Table 5). Consistent with reported global estimates of African and European ancestry proportions [56][57][58][59], PAGE African Americans have on average 78.8% African ancestry and 21.1% European ancestry for Metabochip variants (S1 Fig).
For ancestry proportions at specific PheWAS-identified loci, we found that the majority of the annotated SNPs, such as SNPs within CELSR2 for example, were truly admixed and are consis-

Discussion
We conducted here a large-scale PheWAS for >5,000 African Americans using dense array data and carefully collected and curated epidemiologic data. With these data, we replicate previous GWAS findings from mostly European-descent populations as well as identify novel pleiotropic associations. Because the PAGE study and other efforts have focused or are focusing on multi-population discovery efforts [20,50,[60][61][62][63][64][65] as well as replication, generalization, and fine-mapping of GWAS-identified signals [43,[66][67][68][69][70][71][72][73][74][75][76][77], we focus the remainder of our Discussion on the potential novel pleiotropic associations identified in this African American PheWAS. Potential pleiotropic common variants were identified via single SNP tests of association by the PAGE I study followed by statistical significance filtering and comparison across phenotype classes. For tests of association with consistent statistical evidence across PAGE I studies, we further characterized the PheWAS-identified variants using functional and local genetic ancestry annotations to better understand possible mechanisms or explanations underlying the evidence for pleiotropy in this population. Of the 133 PheWAS-identified findings, we bring to attention those with the most statistical and in silico functional evidence.
Three PheWAS-identified variants were consistently associated with two phenotype classes in two or more PAGE study sites at p<0.01, and they or their proxies were identified as possible eQTLs and were previously associated with one of the phenotype classes in GWAS: DOCK7 rs10889334, APOB rs568938, and PHACTR1 rs9349379. All three were associated with the smoking phenotype class, and none of the three have been implicated in GWAS for Plotted are single SNP tests of association in the Atherosclerosis Risk in Communities (ARIC) for LDLR rs6511720 and nearby SNPs within 100kb of previously-reported genome-wide association study (GWAS) associations also associated here at p<1.0x10 -4 with lipid-related traits. Data shown are sample size, coded allele frequency (CAF), genetic effect size (beta), -log10(pvalue), and ARIC phenotypes on the y-axes. Each SNP is plotted on the x-axis at the top of the figure from 5´to 3 0 , and genomic positions along chromosome 19 along with annotated genes are given above. Data are color-coded by phenotype and displayed as a square (for SNPs), a triangle (p-values), or closed circles (betas, CAFs, and sample size). Direction of the triangle represents the direction of the effect size.
https://doi.org/10.1371/journal.pone.0226771.g004 Plotted are single SNP tests of association in the Atherosclerosis Risk in Communities (ARIC) for CETP rs3764261 and nearby SNPs within 100kb of previously-reported genome-wide association study (GWAS) associations also associated here at p<1.0x10 -4 with lipid-related traits. Data shown are sample size, coded allele frequency (CAF), genetic effect size (beta), -log10(p-value), and ARIC phenotypes on the y-axes. Each SNP is plotted on the x-axis at the top of the figure from 5´to 3 0 , and genomic positions along chromosome 16 along with annotated genes are given above. Data are color-coded by phenotype and displayed as a square (for SNPs), a triangle (p-values), or closed circles (betas, CAFs, and sample size). Direction of the triangle represents the direction of the effect size. https://doi.org/10.1371/journal.pone.0226771.g005 Pleiotropy in African Americans from PAGE any of the smoking categories curated by the NHGRI-EBI GWAS Catalog nor have they been implicated in recent gene-environment studies for lipid traits [78]. The other phenotype classes represented in these associations (LDL-C, hypertension, and diabetes) all have complex relationships with smoking, and these PheWAS data do not provide a clear causal pathway that defines the potentially pleiotropic variants' relationships with the phenotype classes or between the phenotype classes themselves.
Among those variants without evidence of previous GWAS relationships, one example of a novel and potentially pleiotropic variant is the intronic SPARC rs4958487-A associated with increased glucose levels and hypertension. The secreted protein acidic and rich in cysteine (SPARC) gene product modulates the interaction between the extracellular matrix and surrounding cells and is highly expressed in fibrotic tissues [79]. Fibrosis is a clinical feature of hypertension, and both human and animal models support a relationship between SPARC and type 2 diabetes pathogenesis [80,81]. While intronic, annotation of SPARC rs4958487 suggests that it is a significant eQTL in tibial artery (GTEx p = 7.0x10 -11 ) and coronary artery (GTEx p = 5.0x10 -7 ) tissues among others, with the A allele associated with higher SPARC expression compared with the G ancestral allele. Local genetic ancestry estimates for this locus suggest no deviations from expected proportions of European and African ancestry at this locus. Although SPARC rs4958487 has not yet been associated with any phenotype (including glucose or hypertension) at p<10 −8 in the NHGRI-EBI GWAS Catalog, it was included on the Metabochip genotyping array for replication based on early meta-analyses of mean platelet volume in European-descent populations at p<1.0x10 -3 [82][83][84]. To our knowledge, the present PheWAS in African Americans provides the first statistical and in silico evidence for pleiotropy for this locus, which has already been noted as likely pleiotropic based on its possible roles in type 2 diabetes, obesity, cardiovascular disease, bone strength, tendinopathies, and cancers [81,85].
Among the annotations examined for these PheWAS-identified associations, local genetic ancestry was among the least informative. Genetic ancestry and admixture are widely recognized as useful markers of human migration [56,58] and disease associations [86], including potential genetic interactions [87]. Here we note several PheWAS-identified variants with fewer (DARC "Duffy" locus, JAZF1 rs216922, MTMR11 rs2205303, and TLL2 rs94208) or more (RBKS locus) African-derived alleles than expected. While some have interpreted deviations such as those likely to be due to natural selection since admixture [88], recent large-scale studies have suggested that most local ancestry deviations are due to chance [89].     The present study has several limitations as well as strengths. A major limitation of this and other PheWAS is sample size and power for any individual test of association, a limitation compounded by the multiple testing penalty. An ideal PheWAS would be one conducted in a large sample size of hundreds of thousands of uniformly genotyped (or sequenced) and phenotyped participants. The PAGE I study PheWAS represents a collaboration across several, independent epidemiologic cohorts each genotyped on the Metabochip, necessitating a strategy that emphasized within-study tests of association and across-study patterns of consistent results. The phenotype class assignments made here, while facilitating the within and acrossstudy comparisons, were based mostly on study data labels interpreted by human curators rather than formal statistical examination of the phenotypic data. As a result, some correlated phenotypes were considered separate phenotype classes rather than a single large class. It is unclear, however, how to best classify the multiply-related phenotypes given that the phentoypic correlations are imperfect and the current GWAS-based evidence of overlapping but not completely identical genetic architectures for many of the phenotypes considered here.
A second major limitation of this and other PheWAS is interpretation of the observed associations. These data only include genetic variants targeted by the Metabochip [11,12], a fixedcontent array of GWAS-identified variants and fine-mapping regions from cardio-metabolic studies of mostly European populations. It is likely that other population-specific and transpopulation variants not assayed here are associated with many of the phenotypes tested. For the significant associations identified in the present study, a PheWAS-identified association can be interpreted as evidence of true pleiotropy, true comorbidity, or confounding, among others [90]. The PAGE I study PheWAS-identified associations involving the phenotype class smoking illustrate this major limitation: LDL-C is associated with smoking, and the genetic variants are associated with both phenotype classes. It may be that these PheWAS results are highlighting the correlation between phenotype classes, revealing a novel causal pathway, or representing confounding. These PheWAS results could also be due to chance. Further statistical (e.g., independent statistical replication, mediation analysis, effect modification) and functional data will be required to properly interpret complex PheWAS associations. While we acknowledge that this PheWAS has major limitations, it also has considerable strengths that complement other reported PheWAS. This PheWAS was conducted in African Americans using all Metabochip variants and phenotypes available whereas some previous HaploReg: located within enhancer and promoter regions in multiple cell types; index variant significant eQTL in GTEx transformed fibroblast cells, skin (sun exposed lower leg), tibial artery, and aorta artery tissues; USP1 rs10158897 associated with LDL-C in Chinese population at p = 9.0x10 -6   PheWAS were conducted in European Americans or using specific variants or class of variants and/or a limited set of phenotypes [90][91][92][93][94][95][96]. The few genome-wide, phenome-wide reported PheWAS are based on clinical data extracted from electronic health records (EHRs) [8,97,98]. EHR-based PheWAS rely on structured phenotype data such as International Classification of Diseases codes (ICDs, otherwise known as billing codes) and laboratory values. While EHR data represent real-world clinical phenotypes, these data are not uniformly collected across all patients and are associated with known and unknown biases [99]. Also, EHR PheWAS have yet to consider unstructured exposure, behavioral, or lifestyle variables, which are known to be highly relevant to human health and disease risk but are notoriously difficult to extract from clinical free text [99]. The PAGE study is the first to introduce exposure, behavioral, and lifestyle data to the PheWAS landscape, and results suggest these variables may be relevant in describing the complex genetic architecture of traits and disease risk in humans. These results also provide useful data towards Mendelian randomization studies, which aim to use instrument variables to establish causal relationships. The ideal instrument variable is free of pleiotropy; thus, PheWAS could serve as a test of this important assumption of Mendelian randomization [100].

Conclusions
Our work reinforces the potential of PheWAS in epidemiologically collected, diverse populations. We confirm known genetic associations as well as identify potentially pleiotropic common variants across the genome in African Americans. These data reveal complex genetic relationships between common, complex disorders and, in some cases, exposures as-of-yet undetected in univariate analyses common in GWAS, underscoring the need for phenotypewide studies to better understand the multiple dimensions of genotype-phenotype relationships in humans.

PAGE study sites: Designs and populations
Summary descriptions for each PAGE study site are presented in Table 1. All study protocols were approved by Institutional Review Boards at their respective study sites (S2 File).

Causal Variants Across the Life Course (CALiCo) and the Atherosclerosis Risk in Communities (ARIC) study.
CALiCo is a consortium of six demographically diverse populationbased studies comprising of 58,000 men and women ranging in age from childhood to older adulthood and a central laboratory. The ARIC study is one of the six studies included in CAL-iCo and is a multi-center prospective investigation of atherosclerotic disease in a predominantly bi-racial population. European American and African American men and women aged 45-64 years at baseline were recruited from four communities: Forsyth County, North Carolina; Jackson, Mississippi; suburban areas of Minneapolis, Minnesota; and Washington County, Maryland m. A total of 15,792 individuals participated in the baseline examination in 1987-1989, with follow-up examinations in approximate 3-year intervals, during 1990-1992, 1993-1995, and 1996-1998. After the institutional review board at every participating university approved the ARIC Study protocol, written informed consent was obtained from each participant. A subset of ARIC participants was selected for genotyping and inclusion in these PAGE analyses. Data dictionaries for ARIC are available on their website (https://sites.cscc. unc.edu/aric/) as well as the database of Genotypes and Phenotypes (dbGaP) [101].
Multiethnic cohort (MEC). The MEC is a population-based prospective cohort study consisting of 215,251 men and women, and comprises mainly five self-reported racial/ethnic populations: African Americans, Japanese Americans, Latinos, Native Hawaiians and European Americans [14]. The MEC was designed to provide prospective data on exposures and biomarkers potentially involved in cancer initiation and progression across groups with distinct cultural and dietary patterns. Between 1993 and 1996, adults between 45 and 75 years old were enrolled by completing a 26-page, self-administered questionnaire asking detailed information about dietary habits, demographic factors, level of education, personal behaviors, and history of prior medical conditions (e.g. diabetes). Between 1995 and 2004, blood specimens were collected from~67,000 MEC participants at which time a short questionnaire was administered to update certain exposures and collect current information about medication use. Study protocols and consent forms were approved by the institutional review boards at all participating institutions. A subset of MEC participants were selected for genotyping and inclusion in these PAGE analyses. Data dictionaries for MEC (https://www.uhcancercenter. org/mec) are available in dbGaP.
Women's health initiative (WHI). WHI is a long-term national health study that focuses on strategies for preventing heart disease, breast and colorectal cancer and fracture in postmenopausal women. A total of 161,838 women aged 50-79 years old were recruited from 40 clinical centers in the US between 1993 and 1998 [102]. WHI consists of an observational study, two clinical trials of postmenopausal hormone therapy (estrogen alone or estrogen plus progestin), a calcium and vitamin D supplement trial, and a dietary modification trial. Trial exclusion criteria have been described previously [15]. Study protocols and consent forms were approved by the institutional review boards at all participating institutions. A subset of WHI women were selected for genotyping and inclusion in these PAGE analyses. Data dictionaries for WHI are available on their website (https://www.whi.org/researchers/data/ WHIStudies/StudySites/Pages/home.aspx) as well as dbGaP.

Metabochip content and genotyping
The Metabochip has SNPs selected as GWAS replication targets for cardio-metabolic traits as well as SNPs in fine mapping regions around target SNPs [12]. The remaining SNPs on Metabochip include coverage of the HLA region, SNPs associated at genome wide significance with any human trait from the NHGRI GWAS catalog at the time of chip development, mitochondrial SNPs, SNPs on the X and Y chromosomes (not used in this study), and a series of "wild card" SNPs. Further details of this chip are available at the following URL: http://www.sph. umich.edu/csg/kang/Metabochip/.
Full Metabochip genotyping and quality control details are available in Buyske et al. [11]. Briefly, DNA samples were genotyped at the Human Genetics Center of the University of Texas-Houston (ARIC), the University of Southern California Genomics Core (MEC), and the Translational Genomics Research Institute (TGen) (WHI). Ninety HapMap YRI (Yoruba in Ibadan, Nigeria) samples were genotyped in each of the three sites for cross-site quality control. Genotypes were called separately for each PAGE study site at the PAGE Coordinating Center under a common protocol, using both the Genome Studio GenCall 2.0 algorithm as well the GenoSNP genotyping algorithm [103], which is a sample-based approach for capturing some of the rarer genotypes represented on Metabochip. Discordance between the results of the two algorithms were used as a quality control filter. A total of 0.9% of samples were removed based on sample quality control measures. A total of 14,328 (7.3%) SNPs was considered technical failures because of the GenCall or cluster separation score, call rate, Mendelian error rate, replication error rate, or deviation from Hardy Weinberg Equilibrium. An additional 5,248 (2.7%) SNPs were not used in this study because the probe sequence matched poorly to the reference genome. Identification of related individuals were identified using PLINK [104] and the calculations of identity by descent (IBD) for all pairs, up to 2 nd degree relatives. For pairs identified as related, one from each pair was dropped out of further analysis based on which individual had the higher call rate. Overall, 5,897 samples and 161,097 SNPs on the Metabochip passed the quality control criteria of the PAGE I study. A total of 144,740 of these SNPs passed the present study allele frequency threshold (1%).
To adjust for population stratification across study sites, principal components were determined separately for each PAGE I study using the smartpca package of the Eigensoft software [105]. The first two principal components were used as covariates in all analyses. Full analysis details for the ancestry adjustments are also available in Buyske et al. [11].

Genetic tests of association
All tests of association were performed separately for each PAGE I study site in PLINK [104] and adjusted for the first two principal components and sex (except for the women only WHI). A total of 144,740 SNPs were used in the PheWAS for 273 phenotypes. S1 Table lists the 273 phenotypes used in this study. Linear or logistic regression was performed for continuous or categorical dependent variables, respectively, assuming an additive genetic model (0, 1, or 2 copies of the coded allele). For variables with multiple categories, binning was used to create new variables of the form "A versus not A" for each category, and logistic regression was used to model the new binary variable. Linear regressions were repeated following a y to log (y+1) transformation of the response variable with +1 added to all continuous measurements before transformation to prevent variables recorded as zero from being omitted from analysis. The total numbers of associations calculated for this PheWAS where the coded allele frequency was greater than 1% were ARIC 22x10 6 , MEC 8x10 6 , and WHI 26x10 6 . Data were visualized using PhenoGram [106].

Phenotype Class Matching
All 273 individual PAGE study phenotypes were grouped into categories within sites and then grouped into categories across sites regardless of genetic association. As an example of within study collapsing, WHI had four separate phenotypic measurements related to diabetes, including "Diabetes ever (Y/N)" and "treated diabetes (Y/N)", all binned together in the same phenotype class. Across PAGE, specific phenotypes clearly were collected for more than one study site, such as for the phenotype "Hemoglobin". Other groups of phenotypes that fell within similar phenotypic domains but were not represented in the same form across PAGE study sites (e.g., hormone use, smoking) were also collapsed into phenotype classes. Phenotype classes were developed by one curator, and a second curator reviewed the resultant phenotypes and phenotype classes for consistency and accuracy. Neither curator used genetic association results in the development or review of the phenotype classes. The end result was a total of 30 phenotype classes.

Permutation testing
To determine an empirically derived p-value threshold, we used permutation testing. PheWAS is exploratory and thus incurs a substantial multiple hypothesis testing burden depending on the number of associations being calculated. Dependent on individual PAGE study, a Bonferroni correction would have resulted in an adjusted p-value threshold between~4x10 -9 and 7x10 -9 (S1 File). Bonferroni correction is not suitable for this and other PheWAS as there are correlations between phenotypes as well as correlations between SNPs (i.e., linkage disequilibrium). Therefore, the multiple associations of this PheWAS cannot be considered independent. To determine an empirical p-value threshold, we took a two-step approach. The first step was to permute the data within each study separately (ARIC, MEC, WHI): 1. Randomize the association between the genotype matrix and the phenotype matrix 1000 times, generating 1000 individual datasets.
• This preserved the relationships between genotypes • This preserved the relationships between phenotypes 2. Perform PheWAS-comprehensive tests of association between all the phenotypes and genotypes-for each of the 1000 permuted datasets • Output: Results for 1000 permuted PheWAS datasets The second step was to determine how often SNP-phenotype associations were significant across two or more studies by chance alone (in the permuted null data). These results were then compared to the results from the unpermuted data. Our definition of replication was two or more studies with an association for the same phenotype class at a specific p-value threshold, and in S1 File we present results across 1000 permutations at various p-value thresholds. When requiring replication at any of our p-value thresholds, any single permuted data set did not have a total number of results equal to or greater than the total number of results in the unpermuted data, indicating that requiring replication and using a p-value threshold of 0.01 would allow us to explore the data while still maintaining a stringent-enough threshold to reduce our type-1 error rate. As we wanted to explore pleiotropy, we wanted to explore how many different phenotype classes we would expect for a single SNP by chance alone. Thus, we also compared the results of the permuted data versus the non-permuted data, when requiring replication for any single SNP result and more than one phenotype-class at different p-value thresholds. S1 File also presents results across 1000 permutations, requiring replication for each individual phenotype class, for more than one phenotype class at various p-value thresholds. At a p-value threshold of 0.01, we found only three permuted data sets with more SNPs associated with more than one phenotype class, compared to the 188 results of the non-permuted data. It is important to note that within the non-permuted data the 188 results were further refined, we removed any results that did not have the same direction of effect across studies.

Functional annotation
For each independent PheWAS-identified variant as well as SNPs in linkage disequilibrium (r 2 �0.80), we annotated individual variants using HaploReg v4.1 [53], RegulomeDB v1.1 [54], and the SNP and CNV Annotation Database (SCAN) [55] (http://www.scandb.org/ newinterface/index.html). HaploReg v4.1 [53] (http://www.broadinstitute.org/mammals/ haploreg/haploreg.php) annotates SNPs with ENCODE and GENCODE, GTEx [107], and NHGRI-EBI GWAS Catalog data [108]. We supplemented GWAS annotations using the more recent (2019-June 20) version of the NHGRI-EBI GWAS Catalog. RegulomeDB [54] annotates variants based on evidence for transcription factor binding. SCAN is a database that provides summary information from eQTL experiments, mapping HapMap SNPs to gene expression in European Americans from UT, USA (CEU) and Yoruba people from Ibadan, Nigeria (YRI). The database provides a list of genes showing local and distant associations to the SNP in these two HapMap populations along with p-values calculated using quantitative trait linkage disequilibrium test (QTDT) method. The database also provides functional summary information available from other databases as well as other GWAS summary information for the SNPs used for annotation.

Genetic ancestry
We estimated both global and local genetic ancestry for all PAGE African Americans in this study. Global ancestry was estimated using~196k SNPs on the Metabochip array and the ADMIXTURE software assuming K = 2 populations [109]. Although this test was unsupervised, HapMap YRI samples were included and a 5-fold cross validation was used to ensure accuracy. Local estimates of ancestry were calculated using LAMP-LD [110] for~175,600 SNPs after LD pruning in PLINK [104]. Phased haplotypes for CEU and YRI reference samples from the 1000 Genomes Project were used. We calculated local ancestry using a sliding window of 50 SNPs (200 kb) and 10 states per SNP per recommended by the LAMP-LD manual for maximal accuracy and minimal computational time [110].
Supporting information S1 Table. List of phenotypes included in the Population Architecture using Genomics and Epidemiology (PAGE) I phenome-wide association study (PheWAS), by PAGE study. Given are the abbreviations and brief descriptions of phenotypes available for the PAGE I study PheWAS in African Americans by PAGE study. (XLSX) S2 Table. All PheWAS tests of association for the Population Architecture using Genomics and Epidemiology (PAGE) I study in African Americans. We tested 144,740 SNPs assayed on the Metabochip with up to 273 phenotypes available for 5,897 African American participants from three participating PAGE I studies: Atherosclerosis Risk in Communities (ARIC); Multiethnic Cohort (MEC); and the Women's Health Initiative (WHI). A total of 5,424 tests of association were significant at p<0.01 in two or more PAGE studies and in the same direction for the same phenotype. For each of these significant tests of association, we give the SNP (rs number, chromosome, position), the number of PAGE studies with significant results (at p<0.01), the names of the PAGE studies tested for each phenotype listed regardless of significance, the phenotypes tested (short names, long names, and phenotype classes), the statistics (p-values, betas, standard errors), sample size, coded allele and allele frequency, and reason why the variant was assayed by the Metabochip. (XLSX) S3 Table. Significant PheWAS tests of association for the Population Architecture using Genomics and Epidemiology (PAGE) I study in African Americans. After significance threshold filtering and phenotype class binning, we identified 133 SNPs associated with two or more distinct phenotype classes with the same direction of effect within a given phenotype class. For each of these PheWAS-identified variants, we give the SNP (rs number, chromosome, position), the number and name of PAGE studies with significant results, the phenotypes associated, the statistics (p-values, betas, standard errors), sample size, coded allele and allele frequency, reason why the variant was assayed by the Metabochip, and nearest genes. Global genetic ancestry was estimated using ADMIXTURE (unsupervised, assuming K = 2) and all single nucleotide polymorphisms (SNPs) assayed on the Metabochip that passed quality control measures (~196K). Data from the HapMap YRI reference samples were included, representing West African ancestry, and a 5-fold cross validation was used to ensure accuracy. Plotted are the individual samples' (x-axis) estimated global ancestry (y-axis) for the HapMap YRI reference dataset (left) and the PAGE African American Metabochip dataset (right). European ancestry is color-coded blue while African ancestry is color-coded red. Average African ancestry (European) ancestry for PAGE African Americans in this phenome-wide association study is 78.8% Local estimates of ancestry were calculated using LAMP-LD for~175,600 SNPs after LD pruning in PLINK. Phased haplotypes for European (CEU) and West African (YRI) reference samples from the 1000 Genomes Project were used. Local ancestry was calculated using a sliding window of 50 SNPs (200 kb) and 10 states per SNP per recommended by the LAMP-LD documentation for maximal accuracy and minimal computational time. For each PheWAS-identified single nucleotide polymorphism (SNP), we characterized the estimated genetic ancestry as number of (y-axis) copies of African-or European-derived allele (x-axis). A locus was considered "admixed" if the proportion of African-derived alleles was consistent with global genetic ancestry estimates (78.8% West African, S1 Fig); else, it was classified as either "African-derived" or "European-derived." Examples of PheWAS-identified Europeanderived loci include a) DARC "Duffy" locus b) JAZF1 rs216922 c) TLL2 rs94208 d) MTMR11 rs2205303. (TIF)

S4 Fig. PheWAS-identified African-derived locus RBKS based on local genetic ancestry estimates.
Local estimates of ancestry were calculated using LAMP-LD for~175,600 SNPs after LD pruning in PLINK. Phased haplotypes for European (CEU) and West African (YRI) reference samples from the 1000 Genomes Project were used. Local ancestry was calculated using a sliding window of 50 SNPs (200 kb) and 10 states per SNP per recommended by the LAMP-LD documentation for maximal accuracy and minimal computational time. For each PheWAS-identified single nucleotide polymorphism (SNP), we characterized the estimated genetic ancestry as number of (y-axis) copies of African-or European-derived allele (x-axis). A locus was considered "admixed" if the proportion of African-derived alleles was consistent with global genetic ancestry estimates (78.8% West African, S1 Fig)