Rare and low-frequency variants and their association with plasma levels of fibrinogen, FVII, FVIII, and vWF

Fibrinogen, coagulation factor VII (FVII) and factor VIII (FVIII) and its carrier protein von Willebrand factor (vWF) play key roles in hemostasis. Plasma levels of these hemostatic factors are associated with risk of arterial and venous thrombosis, and fibrinogen is also a marker of inflammation.^1-6  Previous genome-wide association studies (GWASs) mainly interrogated common genetic variation and identified variants of modest effect across these phenotypes,^4,7-14  with the largest studies identifying 23 loci for fibrinogen,⁹  5 each for FVII¹³  and FVIII,¹³  and 8 for vWF.¹³  Nonetheless, the associated variants still explain little about the trait heritabilities.^9,12,15  An additional proportion of the missing heritability may be attributed to association with rare variants, which are not captured by the conventional genome-wide marker arrays or imputation panels that have been used for GWASs.¹⁵  In addition, investigating rare genetic variation is important to understanding individual variation in the biology underlying hemostasis pathways.

The aim of this study was to identify low-frequency and rare variants, analyzed individually or at the level of the gene, that influence plasma concentrations of fibrinogen, FVII, FVIII, and vWF. To this end, we meta-analyzed phenotype-genotype associations of low-frequency (minor allele frequency [MAF], 0.01-0.05) and rare (MAF < 0.01) exonic variants in 76 000 individuals of European (EUR), African (AFR), Hispanic (HIS), or East Asian (ASI) ancestry from 16 studies within the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) Consortium.¹⁶  We restricted our analyses to variants that were predicted to alter the coding sequence of the gene product to enhance the likelihood of identifying causal variants and to reduce the burden of multiple testing.

This study was organized within the CHARGE Consortium Hemostasis Working Group and included 16 cohorts of EUR, AFR, ASI, or HIS ancestry. Descriptions and ancestry composition of participating cohorts are found in the supplemental Data available on the Blood Web site).

Hemostasis phenotypes included plasma measures of fibrinogen, FVII, FVIII, and vWF. Fibrinogen (g/L) was available in all 16 studies; FVII activity (% or IU/mL × 100) and FVII antigen (% or IU/mL × 100) were available in 7 studies; FVIII activity (% or IU/mL × 100) was available in 5 studies; and vWF antigen (% or IU/mL × 100) was available in 8 studies. Methods used by each study are noted in Table 1.

Fourteen studies were genotyped by using the HumanExome BeadChip v1.0 (Illumina, Inc., San Diego, CA) whereas one was genotyped by using BeadChip v1.1 and another by using BeadChip v1.2. Variant calling and quality control procedures are described in the supplemental Data and in previously published articles.^17,18  Prior to analysis, individual studies recoded variants to additive coding by using the minor allele derived from the CHARGE joint calling.

In each study fibrinogen measures were natural-log (ln) transformed. For untransformed FVII, FVIII, or vWF, participants with values 3 standard deviations above or below the population mean were removed prior to cohort-level analysis. Study-specific regression analyses were adjusted for sex, age, study design variables, and population substructure by using principal components. MAF thresholds were defined by using the ancestry-specific allele frequencies derived from the CHARGE joint calling.¹⁷  Variant annotation was performed centrally within CHARGE by using dbNSFP v2.0.^19,20  All association analyses were performed by using the R package seqMeta (http://cran.r-project.org/web/packages/seqMeta/index.html). Details of the genotyping chip and version of statistical software used by each study are provided in supplemental Table 1.

We investigated low-frequency and rare variants individually by using standard single-variant association analyses. From among the functional variants on the array (defined as missense, stop-gain, stop-loss, or splice-site changes), we selected variants with an MAF <5% and an expected minor allele count of ≥5 in the total meta-analysis sample for single-variant association of autosomal chromosomes. Because commonly occurring variations on the X chromosome have not previously been investigated for some of the phenotypes, no upper MAF threshold was used when testing for associated variants on this chromosome. The Y and mitochondrial chromosomes were not interrogated. Bonferroni-corrected P value thresholds of statistical significance were based on the number of single-variant tests performed, and they varied by ancestry: 2.5 × 10⁻⁷ (ALL cohorts), 2.6 × 10⁻⁷ (EUR + AFR cohorts), 2.9 × 10⁻⁷ (EUR only), 3.3 × 10⁻⁷ (AFR only), 1.7 × 10⁻⁶ (ASI only), and 4.7 × 10⁻⁷ (HIS only) (see supplemental Data).

Analytical methods that aggregate the effect of multiple rare variants across a gene were used to test for association. This resulted in a P value for a gene rather than for a single variant. Both unidirectional and random effects tests were used; unidirectional texts are more powerful when rare variant effects within a region are in the same direction, and random effects tests are more powerful when rare variants affect a phenotype in opposite directions or when many variants have null effects.

All gene-based tests were again restricted to include only functional single nucleotide variants. Random effects (sequence kernel association test [SKAT]²¹ ) and unidirectional²²  (T5) gene tests were performed using only variants with an MAF <5%. The T5 burden was defined as the total number of rare alleles among variants in the gene with an MAF <5%.²³  All genes were required to contain more than 1 variant to be included in the analysis and to have a cumulative MAF greater than the frequency such that the meta-analysis sample size would have an expected minor allele count of 5. A Bonferroni-corrected, gene-based P value threshold of 1.9 × 10⁻⁶ was used for gene-based tests (0.05/26 965 genes).

Meta-analyses of single variants and gene-based analyses were performed by using seqMeta v1.3. The primary analysis was to meta-analyze all ancestries together, with a secondary set of ancestry-specific analyses performed to complement and inform the results of the primary analysis. All significant nonsynonymous variants were re-annotated by using an updated version of dbNSFP (v.3.0).^19,20,24,25 

To test for independence of the new discoveries from variants previously demonstrated to be associated with the phenotype at that locus, conditional analyses were performed and meta-analyzed. These analyses were undertaken for EUR and AFR ancestry cohorts only, and in some cases, the single nucleotide polymorphisms (SNPs) that were conditioned on differed between ancestry groups, generally because of the conditional SNP being monomorphic in 1 population. A description of conditional analyses undertaken is included in supplemental Table 3.

Single-variant and gene-based tests for all 4 hemostatic factors identified significantly associated loci for all phenotypes. The Q-Q plots for all association analyses are found in supplemental Figures 1-3. Functional annotations for all significant nonsynonymous single variants can be found in supplemental Table 2.

Exome array genotyping and fibrinogen measures were available for 76 316 participants across 16 cohorts and 4 ancestry groups.

Associations for 6 rare or low-frequency variants that exceeded array-wide significance were observed within 4 genes: 2 fibrinogen structural genes (FGB and FGG) and 2 other genes (PDLIM4 and HNF4A) (Table 2 and supplemental Figure 4).

Two rare variants within FGB, rs6054 (Pro235Leu; MAF, 0.0042; P = 1.8 × 10⁻⁴³) and rs201909029 (Lys148Asn; MAF, 0.00077; P = 3.2 × 10⁻¹³) were associated with lower fibrinogen levels. Both variants had similar effect sizes (–0.111 and −0.139 ln[g/L]) and the magnitude and direction of the association was similar for both variants in all ancestry groups (Table 2). Fibrinogen levels were lower by 10.5% and 13.0%, respectively, per copy of the minor allele when other model factors were fixed (see supplemental Data). The rs6054 association has been reported previously,¹⁰  but the rs201909029 variant association is new. Two rare variants within FGG were also associated with fibrinogen levels: rs148685782 (Ala82Gly; MAF, 0.0033; P = 9.2 × 10⁻¹⁵²) and rs145051028 (Ser219Phe; MAF, 0.00016; P = 4.8 × 10⁻⁰⁹). In this study, rs148685782 had an effect size of −0.238 ln(g/L), which translates to a 21.1% lower fibrinogen level per copy of the minor allele. The direction and magnitude of the effect was similar across all ancestry groups in which it was polymorphic (Table 2). The FGG Ala82Gly variant has previously been associated with low plasma fibrinogen levels.^26-28  The rs145051028 variant has an effect size of −0.239 ln(g/L) or a 21.3% lower level of fibrinogen per copy of the minor allele and was polymorphic only in AFR ancestry cohorts. This association has not been previously reported.

To determine whether the newly and previously identified associations within the fibrinogen gene cluster were independent of one another, 3 separate conditional analyses were undertaken: (1) adjustment for previously associated common variants in FGB (rs4220 and rs6056),¹⁰  (2) adjustment for the significant rare variants in FGG (rs148685782 and rs145051028; AFR only), and (3) adjustment for the most significant rare variant in FGB (rs6054) (supplemental Table 3). Results demonstrated independence of all variants from one another (Table 3). In total, the rare variants within the fibrinogen gene cluster explained ∼1.3% and ∼0.12% of the trait variance in the EUR and AFR populations, respectively. The majority of the variance in the EUR population (∼0.9%) was attributed to FGG rs148685782.

The association of low-frequency variants within the PDLIM4 and HNF4A genes supports prior reported associations. The PDLIM4 SNP was in high linkage disequilibrium with previously reported IRF1 SNP rs11242111 (r², 0.85; D′, 1 within 1000 Genomes Map Pilot 1 v.3, CEU) on chromosome 5,⁹  and the HNF4A SNP rs1800961 has been previously reported, although it was just below the genome-wide significance threshold in that study.¹⁰  The effect size for each was 10-fold smaller than those for FGB and FGG.

Single variants in KCNT1 and in HID1, located in regions not previously reported to be associated with fibrinogen levels, reached array-wide significance in the exploratory AFR only analysis of fibrinogen (Table 2 and supplemental Figure 4). KCNT1 rs151272083 (MAF, 0.00072; P = 2.7 × 10⁻⁰⁷) codes for an Arg891Gln change (also reported as the same amino acid change at position 865, 877, or 910 because of transcriptional variation) and was predicted to decrease fibrinogen by 0.330 ln(g/L) or approximately 28.1% per copy of the minor allele in the AFR population. This SNP was also polymorphic in EUR populations but did not reach statistical significance, and the estimated effect was 20-fold smaller (β, 0.017; P = .13). HID1 rs141869748 (Ile421Thr/Ille193Thr; MAF, 0.0013; P = 4.0 × 10⁻⁰⁸) was associated with 0.252 ln(g/L) lower fibrinogen (22.3% decrease per copy of the minor allele) in the AFR population. This SNP was monomorphic in the EUR and ASI populations, and its estimated effect in the HIS population, although small, was not in the same direction despite a similar MAF (MAF, 0.0011; β, 0.008; P = .94).

When we further explored these characteristics of the novel associations in the AFR population, we found no evidence for heterogeneity across studies (P_het, 0.07 [rs151272083] and 0.91 [rs141869748]; supplemental Figure 5), and we confirmed that carriers of the variant allele in AFR cohorts had lower mean plasma fibrinogen levels than noncarriers (supplemental Table 5). The variants explained approximately 0.7% (rs151272083) and 0.4% (rs141869748) of the trait variance.

SKAT and T5 tests yielded gene-level associations with all 4 genes described earlier: FGB, FGG, PDLIM4, and HNF4A (Table 4). Gene-based testing did not identify other genes that contributed to plasma-level variation in fibrinogen.

Exome array genotyping and coagulation FVII measures were available for 25 372 participants across 7 studies of EUR and AFR participants.

Five exome-wide significant coding rare-variant associations were observed in F7 as well as nearby genes MCF2L and PROZ. When conditioning on the common, previously reported coding variant rs6046 in F7,¹³  3 previously unreported rare variants within F7 remained exome-wide significant, whereas the variants in MCF2L and PROZ were no longer significant (Table 3). The minor alleles of F7 variants rs150525536 (Arg117Gln; MAF, 0.0010; P_cond = 1.0 × 10⁻²²), rs121964926 (Arg342Gln; MAF, 0.0015; P_cond = 1.5 × 10⁻¹⁴), and rs3093248 (Glu423Lys; MAF, 0.00085; P_cond = 1.4 × 10⁻⁰⁷) were all associated with significantly lower plasma FVII levels (Table 2 and supplemental Figure 4). The three variants explained ∼0.06% of the trait variance in EUR participants and 4.5% of the trait variance in AFR participants. For all identified variants, the MAF was lower in EUR than in AFR populations but the direction of effect was the same even if the magnitude varied (Table 2). Sensitivity analyses that removed the 2 studies with FVII antigen rather than activity measured did not have an impact on the findings.

SKAT and T5 tests yielded gene-level associations with F7 (Table 4). No other gene was associated with plasma levels of FVII.

As reported by our prior GWASs, association results for plasma levels of FVIII and vWF were similar, so they will be presented together.¹³  FVIII measures were available from 28 291 participants from 5 cohorts across all ancestry groups, whereas vWF was available in 23 272 EUR and AFR participants from 8 cohorts.

Genome-wide significant rare and low-frequency variants are presented in Table 2, and cluster plots for the associated SNPs are found in supplemental Figure 4. Five novel low-frequency and rare variant associations were found for FVIII and vWF levels, most within loci with previous FVIII/vWF associations.¹³ 

Low-frequency variant rs7962217 (Gly2705Arg; MAF, 0.046; P = 2.5 × 10⁻¹³) and rare variant rs41276738 (Arg854Gln; MAF, 0.0040; P = 2.2 × 10⁻¹³) in VWF were significantly associated with lower plasma levels of FVIII but not vWF (P = .96 and P = .03, respectively). Only the association of rs7962217 has been previously reported,²⁹  and conditioning on the most significant common VWF variants associated with FVIII levels (rs1063856 and rs62643635¹³ ) did not materially alter these results (Table 3). Ancestry-specific analyses yielded effects with the same direction and similar magnitudes, although the MAFs varied by up to 2 orders of magnitude (Table 2).

A single rare variant in STAB2 rs141041254 (Glu2377Lys; MAF, 0.00087) was significantly associated with FVIII (P = 2.1 × 10⁻⁰⁸), and vWF levels (P = 2.4 × 10⁻⁰⁷) and the new signal remained unchanged when adjusting for rs2271637, the most highly associated STAB2 common variant on the array. In the 2 ancestries in which the variant was polymorphic (AFR and EUR), the direction and the magnitude of the effects diverged (Table 2). This association has not been previously reported.

For FVIII and vWF levels, 11 significant single-variant associations were observed with rare or low-frequency variants within ABO and surrounding genes on chromosome 9. However, after conditioning on common variants tagging the major ABO blood types (A1, A2, B, and O), none of the 11 associations identified in this region remained. A description of these conditional analyses is presented in the supplemental Data and supplemental Table 4.

In exploratory analyses for the FVIII phenotype only, there was a significant association with a common variant on the X chromosome in F8, the gene encoding FVIII. This coding variant, rs1800291 (Asp1260Glu; MAF, 0.27; P = 8.2 × 10⁻⁰⁸), had an MAF and effect direction that varied across ancestry groups (Table 2).

For the FVIII phenotype only, a rare variant in KATNB1, a gene not previously associated with FVIII levels, achieved array-wide significance in the AFR population. This variant, rs142508811, was rare in both EUR and AFR populations and was monomorphic in ASI and HIS populations; the estimated effect size was 80-fold larger in AFR than in EUR populations. Across the studies with AFR populations, there was no evidence of heterogeneity (P_het, 0.74); a forest plot for these associations is presented in supplemental Figure 5. Levels of FVIII in carriers of the variant allele had a higher mean FVIII than noncarriers (supplemental Table 5).

For the FVIII phenotype, the 5 variants explained approximately 0.9% of the phenotype variation in both EUR and AFR populations. For the vWF phenotype, the STAB2 variant explained 0.2% and 0% in EUR and AFR populations, respectively.

For FVIII levels, ABO, VWF, and STAB2 yielded gene-wide significant associations with SKAT testing, whereas ABO and VWF were significant with T5 testing (Table 4). For vWF levels, ABO and STAB2 yielded gene-wide significant associations with SKAT testing, whereas ABO was significant with T5 testing; the VWF gene did not achieve significance for vWF. No new associations were identified through gene-based testing.

We identified 12 novel associations of low-frequency (n = 2) and rare (n = 10) variants across the fibrinogen, FVII, FVIII, and vWF traits that were independent of previously identified associations. Nine of the variants were within genes previously established as associated with the trait; findings for associations in 3 new candidate loci were detected in people of AFR ancestry, possibly because of monomorphic or much lower frequency characteristics of these variants in all other ancestries. These newly identified associations accounted for modest amounts of the variance explained and suggest that, at most, a small proportion of the missing heritability can be attributable to them. The gene-based tests did not reveal new loci.

Associations of rare variants with fibrinogen levels were found in gene regions previously associated with fibrinogen by common variant GWASs. The association of FGB rare variant rs6054 with lower fibrinogen has been previously reported.¹⁰  Although the association of FGB rs201909029 is a novel finding in this context, it has been reported in mild hypofibrinogenemia cases²⁶  in clinical databases (MERIVALE II),³⁰  although it has not been reported to cause hemorrhage or thrombosis.³⁰  The rare FGG variant rs148685782 was associated with hypofibrinogenemia and hemorrhage^26-28  in multiple affected individuals. FGG rs145051028, which was associated with fibrinogen levels in AFR cohorts only, has not been reported in clinical databases or population studies. This may be a result of the low MAF and also a lack of studies that included AFR participants. Conditional analyses showed that the common and rare variant associations across the fibrinogen gene cluster were independent, an observation supported by their low r² for the pairwise linkage disequilibrium. Within the fibrinogen gene cluster, the 4 significant FGB and FGG rare variants explained two- to fourfold more trait variance than the common FGB rs4220 variant,^7,9,10,14,31  which had an effect size of 0.029 ln(g/L) or a 2.9% higher level of fibrinogen per copy of the minor allele in this study.

In exploratory ancestry-stratified analyses, the associations of KCNT1 and HID1 with fibrinogen in AFR participants were the only findings that identified new candidate loci tha influence fibrinogen regulation. These findings can only be considered hypothesis generating and require replication.

We identified 3 rare coding variants in the FVII protein structural gene F7 associated with plasma levels of FVII, none of which were previously reported in the epidemiologic literature. rs150525536 was rare in the AFR population and had a 10-fold lower frequency in the EUR population. A previous case report of this variant was found in a male with an EUR ancestry homozygote who had severe FVII deficiency and who also carried another F7 mutation (Arg212Gln).³²  Both mutations were thought to contribute to the phenotype. The mutation reported here is found in the first epidermal growth factor-like domain and is required for binding to tissue factor, its cofactor. It causes reduced binding to tissue factor and reduced clotting ability in a concentration-dependent manner as well as slower activation.³²  Variant rs121964926 was also more common among the AFR population than in the EUR population. It has been observed clinically in both asymptomatic and symptomatic individuals with FVII activity <5% from Germany and France as well as patients with reduced FVII activity from Costa Rica, Venezuela, and the United States.³³  Nothing has been reported regarding clinical consequences of the rs3093248 variant.

The findings for the vWF trait consisted of a subset of the FVIII results. None of the associations between variants within the ABO gene region and FVIII/vWF were independent of established ABO blood group alleles. Two rare variants in VWF, rs7962217 and rs41276738, were associated with plasma FVIII levels. rs7962217 was associated with higher FVIII levels whereas rs41276738 was associated with lower levels and had an effect size similar to that of the strongest genetic predictor of FVIII levels, the O-deletion tagging SNP (rs657152). rs41276738 has been reported in patients with von Willebrand disease type 1^34,35  and type 2N,^36-43  but the association with vWF levels did not reach exome-wide significance, although its direction was consistent with the direction of effects on FVIII. The STAB2 variant rs141041254 was associated with higher plasma levels of both FVIII and vWF. The effect size was more than 10-fold larger than that reported for the more common STAB2 variant rs2271637 (β_FVIII, 1.95%; β_vWF, 2.47%). The common F8 coding variant rs1800291 was associated with a much smaller effect on FVIII compared with the ABO O-deletion variant. It has been previously reported,^29,44,45  and in the European Association for Haemophilia and Allied Disorders (EAHAD) Coagulation Factor Variants Database, it is annotated as unlikely to be pathogenic. The KATNB1 rs142508811 variant and FVIII association was restricted to the AFR population, although MAF and direction of effect were similar across the 2 polymorphic populations.

Inferring causality of uncommon and rare variants with a phenotypic expression is challenging and requires strong statistical evidence combined with experimental data.⁴⁶  Inferring clinical implications from the causal variants requires additional evidence⁴⁷  not available in our approach. In this article, we identified rare variants associated with higher or lower phenotype levels in 4 hemostasis measures. Some of the variants have been found in patients with diseases related to blood clotting, which suggests that these genes and their uncommon and rare genetic variation may play a role in a clinical phenotype.^26-28,32-43  The distribution of the phenotypes within our research populations were within the extremes of a clinically important range (FVII: 0.80-11.40 g/L [fibrinogen], 26% to 441% activity, and 2% to 297% antigen; FVIII: 14% to 500% activity; and vWF: 2% to 374% antigen). Furthermore, the magnitude of difference in the phenotype associated with the variant was mostly modest, although some were larger and were associated with a change equivalent to half the size of the estimated population mean for the phenotype of interest. Therefore, the magnitude of any clinically relevant effects of these variants would be expected to be small to modest. The findings from our study suggest that the contribution of the uncommon and rare variants to complex clinical phenotypes, such as arterial or venous thrombosis or hemorrhagic stroke, should be evaluated in large populations. This article identifies several variants which may be good potential candidates.

We decided a priori to use all the phenotype-genotype association data for discovery to reduce false-negative findings,⁴⁸  but this approach provided us with no replication setting. Although these candidate variants are now well characterized, the rare allele frequencies will create challenges for replication in the absence of additional large phenotyped populations. However, our findings provide strong rationale for further functional genomic follow-up, and some of our observations confirm associations for several rare variants that have been reported in patients with the corresponding congenital clotting factor deficiencies. This investigation of low-frequency and rare variants on the 4 phenotypes was limited to the variants included on the BeadChip. Differing sample sizes for the meta-analysis between phenotypes likely affected our power to detect associations, but this may also be influenced by biological differences. In addition, we did not have the statistical power to test for differences in associations across the 4 ancestries. Although it was not an aim of our study, a subsequent effort with this objective would be worthwhile to better understand the genetic architecture of the phenotypes. Finally, although we enriched our variant population with those predicted to be causal, we cannot attribute causality to the variants with novel associations.

The quality of rare variant genotype calling was maximized by the joint clustering performed within CHARGE on thousands of samples.¹⁷  By incorporating individuals of non-European ancestry in the primary analysis, we increased our power to detect association in which variants may be more frequent or genetic diversity greater in one ancestry group than another. It also allowed us to broadly look at ancestry-specific gene and rare-variant associations but was vastly underpowered to draw any strong conclusions.

In conclusion, in meta-analyses of 4 hemostatic factors and functionally enriched exonic variants, novel associations of low-frequency and rare variants were identified in 16 studies that included 4 ancestries. Novel variant associations were found within previously reported genes, and they had effect sizes that were often independent of and much larger than previously reported common variants. In addition, rare variant associations at KCNT1, HID1, and KATNB1 identify new candidate genes related to hemostasis for follow-up replication and functional genomic analysis.

A complete list of acknowledgments for each cohort is found in the supplemental Data.

Contribution: D.V., D.M.B., R.A.M., O.P., A.F.W., C.H., U.V., D.I.C., P.M.R., J.M.S., I.J.D., B.M., B.M.P., N.L.S., W.K., A.H., O.H.F., C.J.O., A.P.R., E.B., and J.I.R. designed the research; J.E.H., O.P., C.H., T.K., U.V., D.I.C., L.M.R., S.E.H., J.M.S., I.J.D., M.S.-L., A.R.F., J.A.B., K.L.W., K.D.T., K.S., A.H., O.H.F., D.L., C.J.O., P.L.A., M.F., N.P., X.G., and J.Y. performed the research; A.T., D.I.C., F.G., M.L.G., A.G., R.J.S., B.S., A.S., M.M.-N., M.W., W.K., M.P.M.d.M., F.R., A.G.U., P.L.A., L.Q., and C.K. contributed vital new reagents or analytical tools; L.R.Y., D.V., D.M.B., R.A.M., O.P., C.H., A.G., D.I.C., F.G., S.E.H., J.M.S., H.W., A.H., A.R.F., B.M.P., M.W., W.K., M.P.M.d.M., F.R., A.G.U., A.H., O.H.F., D.L., G.H.T., C.J.O., P.L.A., C.K., A.P.R., E.B., M.F., M.C., C.-C.H., and N.Z. collected data; L.R.Y., D.V., D.M.B., R.A.M., J.E.H., C.H., M.L.G., T.K., D.I.C., L.M.R., F.G., R.E.M., S.E.H., M.S.-L., A.C.M., J.A.B., K.D.T., B.M., B.M.P., N.L.S., H.R., P.S.d.V., A.D., M.-H.C., G.H.T., C.J.O., C.K., A.P.R., M.F., N.P., X.G., J.Y., W.T., and J.I.R. analyzed and interpreted data; L.R.Y., J.E.H., T.K., D.I.C., L.M.R., R.E.M., M.S.-L., A.C.M., J.A.B., M.M.-N., H.R., P.S.d.V., A.D., M.-H.C., P.L.A., L.-A.L., X.G., and J.Y. performed statistical analysis; and J.E.H., R.E.M., S.E.H., I.J.D., P.S.d.V., G.H.T., A.C.M., A.P.R., C.J.O., and N.L.S. wrote the manuscript. All co-authors were given the opportunity to revise and comment on the text and content of manuscript. This manuscript was approved by all relevant cohort Publication and Presentation committees prior to submission.

Conflict-of-interest disclosure: D.V. is a consultant for MBC, Inc. D.I.C. and P.M.R. have received funding for exome chip genotyping in the Women’s Genome Health Study and collaborative scientific support from Amgen. B.M.P. serves on the data and safety management board for a clinical trial of a device funded by the manufacturer (ZOLL Lifecor) and on the Steering Committee of the Yale Open Data Access Project funded by Johnson & Johnson. O.H.F. is employed by ErasmusAGE, a center for aging research across the life course funded by Nestlé Nutrition (Nestec Ltd.), Metagenics Inc., and AXA. Nestlé Nutrition, Metagenics Inc., and AXA had no role in the design and conduct of the study, the collection, management, analysis, and interpretation of the data, or in the preparation, review, or approval of the manuscript. The remaining authors declare no competing financial interests

Correspondence: Nicholas L. Smith, 1730 Minor Ave, Suite 1360, Seattle, WA 98101; e-mail: nlsmith@u.washington.edu.