Genome-wide association analysis of pulse wave velocity traits provide new insights into the causal relationship between arterial stiffness and blood pressure

Background The pathophysiology of arterial stiffness is not completely understood. Pulse wave velocity (PWV) is an established marker for arterial stiffness. We compare genetics of three PWV modes, namely carotid-femoral PWV (cfPWV), brachial-ankle (baPWV) and brachial-femoral (bfPWV), reflecting different vascular segments to analyse association with genetic variants, heritability and genetic correlation with other biological traits. Furthermore we searched for shared genetic architecture concerning PWV, blood pressure (BP) and coronary artery disease (CAD) and examined the causal relationship between PWV and BP. Methods and results We performed a genome-wide association study (GWAS) for cfPWV, baPWV and bfPWV in LIFE-Adult (N = 3,643–6,734). We analysed the overlap of detected genetic loci with those of BP and CAD and performed genetic correlation analyses. By bidirectional Mendelian Randomization, we assessed the causal relationships between PWV and BP. For cfPWV we identified a new locus with genome-wide significance near SLC4A7 on cytoband 3p24.1 (lead SNP rs939834: p = 2.05x10-8). We replicated a known PWV locus on cytoband 14q32.2 near RP11-61O1.1 (lead SNPs: rs17773233, p = 1.38x10-4; rs1381289, p = 1.91x10-4) For baPWV we estimated a heritability of 28% and significant genetic correlation with hypertension (rg = 0.27, p = 6.65x10-8). We showed a positive causal effect of systolic blood pressure on PWV modes (cfPWV: p = 1.51x10-4; bfPWV: p = 1.45x10-3; baPWV: p = 6.82x10-15). Conclusions We identified a new locus for arterial stiffness and successfully replicated an earlier proposed locus. PWV shares common genetic architecture with BP and CAD. BP causally affects PWV. Larger studies are required to further unravel the genetic determinants and effects of PWV.

Beside genetic pre-disposition, biological aging causes arterial stiffening due to structural changes in the medial layer of the arterial walls [13]. While for many years hypertension was considered as causal for arterial stiffness, during the last decade several studies showed an inconclusive picture suggesting a reverse causality [14].
As so far no GWAS measured all three PWV modes in parallel on individuals of a large population-based study, we gain new insights on how the underlying genetic factors affect different segments of the arterial tree. Furthermore we analyse the causal relationship between arterial stiffness and blood pressure as well as the common genetic effects of PWV, blood pressure and CAD.

Cohort description
LIFE-Adult is a randomly sampled population based study of 10,000 participants from Leipzig, a city in the east of Germany [15]. Most of the participants are aged between 40 and 79, with a small sub group of 400 participants being between 18 and 39. The study population is of central European descent and the main study goal is to investigate prevalence, genetic predisposition and the role of lifestyle-related factors (such as smoking habits, alcohol consumption, dietary patterns and physical activity) on major civilization diseases including subclinical signs. Initial data collection was performed between 2011 and 2014. LIFE-Adult meets the ethical standards of the Declaration of Helsinki. It has been approved by the Ethics Committee of the Medical Faculty of the University Leipzig, Germany (LIFE-Adult: Reg. No 263-2009-14122009). Written informed consent including agreement with genetic analyses was obtained from all participants.

Pulse wave assessment
Pulse wave analysis was performed using the oscillometry-based Vicorder device (Vicorder, Skidmore medical, Bristol, UK) as previously described [8]. Validity and intra-and interobserver reliability of the Vicorder device were shown to be good [9]. As recommended in Baier et al. [10], we used the first measurement available for baPWV and cfPWV while for bfPWV, we averaged values for first and second measurement, since variance of bfPWV was significantly higher than it was for baPWV and cfPWV. Measurements were discarded if the difference between first and second measurement was >5m/s (baPWV), >10m/s (cfPWV) and >25m/s (bfPWV), respectively. PWV modes were only moderately correlated.

Genotyping
For the genotyping of 7,838 randomly selected samples, we used the genome-wide SNP array Affymetrix Axiom CEU1, and for calling, Affymetrix Power Tools (APT) Version 1.20.6. We executed the calling according to the Affymetrix Axiom Genotyping Solution Data Analysis Guide and filtered 116 samples failing the dish QC criteria (dish QC � 0.82, technology specific-signal-to-noise-ratio) and sample call rate criteria (SCR � 97%) in the initial calling round [16]. Sex-mismatches were filtered and cryptic relatedness was assessed according to Wang [17]. Ten highly related samples (coefficient of relatedness > 0.99) were filtered. For Xchromosomal analysis, we filtered irregularities of X-Y intensity plots. Genetic principal components ("Swap-One-In"-PCA) were calculated using PLINK 1.9 and we removed ethnical outliers (>6 SD in any of first 10 principal components). For SNP quality control, the following filters were applied: SNP call rate � 97%; p-value of exact test for Hardy-Weinberg equilibrium �10 −6 ; p-value of plate association �10 −7 (i.e. test for dependency of allele frequency on plate); monomorphic SNPs and criteria of cluster plot quality as suggested by Affymetrix. For X-chromosomal SNPs, call rate threshold was increased to 98% and HWE test performed in women was required to achieve p�10 −4 . After QC of the initial 568,519 autosomal and 16,384 SNPs on chromosome X, a set of 7,669 LIFE-Adult subjects and 532,676 SNPs was available for autosomal analyses. For gonosomal analyses, 7,660 samples with 13,476 SNPs were used.

Statistical analysis
Genome-wide association analysis. Pulse wave parameters and blood pressures were logtransformed for all analyses to approximately conform to normality [S1 Table in S2 File]. We performed genome-wide association analyses for each phenotype by linear regression models assuming an additive mode of inheritance with adjustments for sex, age and SBP [S2 Table in S2 File]. Adjustment for SBP was included to avoid identifying blood pressure related SNPs based on the observed correlation between SBP and PWV (all Pearson's R with p-values < 2.2x10 -16 : cfPWV: R = 0.24, bfPWV: R = 0.19; baPWV: R = 0.49) [20]. SBP was log transformed to achieve normal distribution. We tested the log transformed PWV modes for correlation with the first ten genetic principal components (PC). While there was no significant correlation with cfPWV, there was significant correlation with baPWV (PC8: p = 7.3x10 -4 , PC9: p = 7.0x10 -3 and PC10: p = 8.3x10 -4 ) and bfPWV (PC7: p = 2.6x10 -2 , PC9: p = 2.5x10 -4 and PC10: 8.9x10 -2 ). As in the multivariate model neither PC9 nor PC10 were significant for baPWV (PC9: p = 4.4x10 -1 and PC10: 7.3x10 -1 ) or bfPWV (PC9: p = 4.0x10 -1 and PC10: 9.1x10 -1 ), we decided to disregard the PC for all three PWV mode to apply a uniform model to all three modes. We tested the association of antihypertensive and lipid lowering medication on PWVs. They are associated with cfPWV and bfPWV, but do not explain much of the traits variance (r 2 <1%). Therefore, we did not adjust our genetic model for medication. We also considered log-transformed BMI as an additional covariate. Although significant, it explained only a small proportion of phenotype variance. Therefore we did not include it as additional adjustment. Analyses were executed with SNPTEST [21] using expected genotype counts. The threshold for genome-wide significance was set to p<5x10 -8 . Associations with p<1x10 -6 were considered as suggestive and presented as list of top SNPs. In order to determine independent hits, we calculated linkage disequilibrium (LD) between markers in the LIFE-Adult data. Variants, which are in LD of r 2 �0.5 with a SNP of higher significance, were considered tagged by that SNP (priority pruning). As locus we defined SNPs in a region of ±500 kb around the target SNP. Our approach had more than 80% power to detect an association of a SNP with more than 0.62% variance of our phenotype for cfPWV, which is quite well in the range of previously reported associations of comparable phenotypes in a similarly sized study [22].
Replication of known PWV SNPs. We looked up known and suggestive genetic loci for PWV and compared the reported effect sizes with our data [22,[35][36][37]. While the look-up of associated loci is feasible, the comparison of beta estimates is difficult since the other groups did not only use different PWV modes but also different transformations of their data. As our study is the first that performed a GWAS for three PWV modes in parallel, we provide new insights into the comparability of PWV assessments.
As of March 2019 four GWAS studies on PWV had been published [22,[35][36][37]. From these studies we selected SNPs for replication that had passed at least suggestive significance level (p<1.00x10 -6 ) in the screening cohort [ Table 1]. Mitchell et al. [35] and Tarasov et al. [22] identified ten (rs17773233, rs1381289, rs7152623, rs2225442, rs1461587, rs987514, 20p12.3 rs7271920 [36] n/a n/a -0.17 7.20x10 -9 baPWV cfPWV yes T 0.12 -0.012 6.55x10 -2 SNPs highlighted with grey back ground were not successfully replicated in our GWAS as they did not pass the Bonferroni corrected nominal significance level (p<1.67x10 -2 ). SNPs in 14q32.2 are in LD with two independent lead SNPs (rs1381289 and rs17773233). Effect directions are opposite to Mitchell as he considered1000/ cfPWV as phenotype. Summary statistics of all SNPs and phenotypes can be found in S8 Table. https://doi.org/10.1371/journal.pone.0237237.t001 rs10782490, rs1381273, rs8015529, rs10764094) respectively two SNPs (rs1194820, rs3742207) exceeding suggestive significance level, both using cfPWV on European populations. Park et al. [36], using baPWV on an Asian population, identified two SNPs (rs7271920, rs10125157) exceeding suggestive significance level. While they rejected the SNPs as they were not able to show nominal significance in their replication cohort, we included them for replication. Levy et al. (using cfPWV on a North American population) did not identify SNPs that passed at least suggestive significance level [37]. Therefore we identified in total 14 SNPs for replication analysis in our cohort. These SNPs are distributed over six loci. All of these SNPs were available in our data set. SNPs were considered as replicated in our study if at least Bonferroni corrected nominal significance (p<0.05/3) for any of the PWVs modes is achieved. For Mitchell et al. [35] the effect direction was expected to be opposite to ours as he considered 1000/ cfPWV as primary phenotype. Authors used different regression models for their analysis. While Mitchel used a sex specific regression model with 1000/cfPWV as residual adjusted for age, age2, height and weight, Tarasov used inverse normal transformed cfPWV residuals adjusted for age, age 2 and sex. Park used log transformed baPWV data, which adjusted for age, gender, systolic blood pressure and diastolic blood pressure. All models used different transformation for the residuals and did not deliver relevant higher r 2 for our data [S2 Table in S2 File]. Heritability and genetic correlation. We estimated the heritability of our three PWV modes in our study by LD Score regression analysis using LD Hub [38]. To avoid blurring the relationship between PWV and SBP, we used PWV associations only adjusted for age and sex but not SBP. A total of 1,192,954 of our SNPs could be matched to the SNPs available in LD Hub. According to Zheng et al. [38] the minimum criteria to test for genetic correlation are: Heritabilty (H 2 ) Z score > 1.5, mean Chi square of the test statistics > 1.02 and intercept estimated from the SNP heritability analysis between 0.9 and 1.1. This held for baPWV but not for bfPWV and cfPWV. We tested for genetic correlation between baPWV and 517 UK Biobank traits and four blood lipids traits (triglycerides, total cholesterol, LDL-cholesterol, HDL-cholesterol) as provided and categorized by LD Hub.
Common genetic associations with BP and CAD. To analyse potential patho-mechanistic overlaps between arterial stiffness and CAD, we analysed 169 genome-wide significant CAD-SNPs [1][2][3][4][5] for association with PWV modes in our cohort. We determined the empirical distribution of the number of SNPs associated with at least one of the PWV modes by 10,000 permutations of phenotypes and genotypes. To test for an enrichment, we compared this distribution with the number of actually observed associations with Bonferroni corrected nominal significance (p<0.05/3).
We also wanted to evaluate the genetic relationship between blood pressure and arterial stiffness to gain a better understanding of the underlying common pathomechanisms. For this purpose, we considered 885 SNPs with genome-wide significance for SBP, diastolic blood pressure (DBP) or pulse pressure (PP) described by Evangelou et al. [39] [Supplementary Table 24] and tested them with the same methodology as described above for the CAD SNPs.
Finally we searched for overlaps between BP, CAD and PWV at locus level. For this purpose, we determined the loci for the combined set of CAD and BP SNPs from literature as explained above. A locus containing both, CAD and BP SNPs, is considered overlapping. In addition, a BP / CAD locus was considered overlapping with PWV, when the lead SNP was Bonferroni corrected nominal significant (p<0.05/3) for at least one of the PWV mode.
Mendelian randomization. To estimate the causal relationships between blood pressure and arterial stiffness we performed bidirectional Mendelian randomization (MR) analysis considering the genetic associations reported by Evangelou et al. [39] as instruments for blood pressure and the SNPs identified in our GWAS as instruments for PWV. Evangelou et al. used untransformed, medication-adjusted blood pressure traits for their non-genetic model, therefore we decided to recalculate blood-pressure betas in our study for our MR analysis.
To analyse the causal effects of systolic blood pressure on PWV, we excluded the 149 blood pressure SNPs that showed nominal significance for PWV in our data as they represent potential direct effects on PWV which would violate one of the MR pre-conditions. Of the remaining 711 SNPs associated with one or more blood pressure traits, 49 SNPs showed at least nominal significance for SBP in our study cohort and were not in LD with any of our PWV SNPs (all pairwise R 2 < 0.1). We combined the evidence of these SNPs regarding a causal effect of SBP on our three PWV modes by the inverse-variance weighted method [40].
To estimate the reverse causal effect of PWV on blood pressure, we used the independent SNPs from our GWAS, which had passed the suggestive significance level for at least one PWV phenotype. None of these SNPs showed a significant effect on blood pressure for our study participants, but one SNP (rs4016640) was in weak LD with a hit reported in Evangelou (rs13082711, R 2 = 0.11). Therefore we removed this SNP. We then performed MR, applying the inverse-variance weighted method for each PWV mode separately, using SNPs that showed at least suggestive significance for one PWV mode and at least nominal significance for the PWV mode considered. For a secondary MR, we removed the lead SNP, since there is a weak link between this locus and blood pressure, which could violate the bivariate MR assumption.
To the best of our knowledge the selected PWV and blood pressure SNPS do not effect confounders for blood pressure and/or PWV [S3c Table in S2 File]. Therefore the preconditions for a bidirectional Mendelian Randomization were fulfilled (Fig 1).
Using Cochrans Q-test to check heterogeneity, we performed a sensitivity analysis of the results to test for pleiotropy effects. There is an association between the SNP and the risk factor, (ii.) the SNP has no association with the outcome and (iii.) the SNP is independent of factors that confound the SNP-outcome relationship. To analyse the causal relationship of BP on PWV, we use as instruments BP SNPs reported by Evangelou [39], which were also nominal significant in our cohort. https://doi.org/10.1371/journal.pone.0237237.g001

Results
Basic characteristics of LIFE-Adult and their univariate effects on the three PWV modes are presented in Table 2.

GWAS results
None of three genome-wide analyses showed signs of inflation (λ cf = 1.00, λ bf = 1.01, λ ba = 1.02). We identified thirteen independent SNPs with at least suggestive significance level at 10 different loci. Thereof one locus achieved genome-wide significance [ Table 3]. Nine of the eleven suggestive hits reached nominal significance with at least one of the other PWV phenotypes. In all nine cases the beta of the primary and secondary PWV phenotype had the same direction. A Manhattan Plot of all traits is given in Fig 2. Further summary statistics regarding all phenotypes can be found in S4 Table in S2 File. Summary of eQTL look-up and GWAS catalogue look-up can be found in S5 and S6 Tables in S2 File.
To confirm the robustness of our non-genetic model, we tested our best-associated SNP rs939834 under further adjustment for antihypertensive and anti-lipid drugs and found no significant change in effect size. We additionally adjusted our top SNPs for log transformed BMI [S7

Replication of known PWV SNPs
Of the 14 candidate SNPs from literature [S8 Table in S2 File], seven showed Bonferroni corrected nominal significance in LIFE-Adult with any of the three PWV modes. Among these seven SNPs, only two were independent in our data. These two independent SNPs (rs17773233 and rs1381289) identified by Mitchell et al. [35] are located at Cytoband 14q32.2 within a distance of 10kb, i.e. they belong to the same genetic locus. SNPs were replicated for baPWV and not for the phenotype of the original study (cfPWV). A comparison of the betas and standard errors for the SNPs between the original GWAS and ours are shown in Table 1.
As sensitivity analysis we calculated the effect of each of the SNPs for each PWV mode and with all three published models as well as our own non-genetic model [S9

Common genetic associations with BP and CAD
Of the 169 SNPs associated with CAD [S11 Table in S2 File], one was not in our data. Of the remaining 168 SNPs, ten SNPs showed Bonferroni corrected nominal significance (p<0.05/3) for at least one of the PWV modes. This represents no significant enrichment compared to an empirical expectation of eight SNPs which showed Bonferroni corrected nominal significance (p<0.05/3) as a result of our permutation analysis (p = 0.18; results were Poisson distributed).
Of the 885 SNPs associated with blood pressure traits [S12 Table in S2 File], 25 were not available in our data. Of the remaining 860 SNPs, 60 SNPs showed Bonferroni corrected nominal significance for at least one PWV phenotype. This represents a significant enrichment compared to an expectation of 41 SNPs (p = 0.002; results were Poisson distributed). The ten CAD SNPs described above, represented nine overlapping loci between CAD and PWV; the 60 BP SNPs represented the respective number of overlapping BP and PWV loci. Furthermore we identified 110 overlapping loci for CAD and BP, including 9 SNPS that were reported genome wide significant for BP as well as for CAD.
Cochrans Q showed no pleiotropy effects for any of the Mendelian randomization analyses performed.

Discussion
The non-invasively measurable PWV is an established operationalization of arterial stiffness. Performing the first GWAS for three PWV modes in parallel on the same cohort provides unique insights about PWV and the genetic effects on different vascular segments. As the three PWV traits represent different branches of the arterial tree, the significance of the association with the three PWV modes changed for different SNPs. We assume that the association of different SNPs reflect arterial stiffness in different arteries, e.g. SNPs associated with cfPWV act  The hypothesis that Systolic blood Pressure causally effects PWV was tested. Calculation is based on 49 of the blood pressure SNPs described by Evangelou [39], which showed at least nominal significance for blood pressure in our study population, but no association with PWV.
https://doi.org/10.1371/journal.pone.0237237.t004 over candidate genes effecting primarily large arteries, whereas baPWV SNPs effect more the peripheral arteries. Accordingly, the three PWV traits are only moderately correlated so that one cannot expect the same genetic background. In addition the variance in the measurements of the PWV modes is different [10], sample sizes for PWV-GWAS are still limited and assessments are note easily comparable [9]. All these effects introduce heterogeneity in the results. While determining arterial stiffness by measuring cfPWV was recommended by AHA scientific statement from 2015 [44], in our study we found higher compliance and higher explained multivariate variance with age, sex and blood pressure for baPWV (adjusted r 2 = 0.54, p<0.001) compared to cfPWV (adjusted r 2 = 0.20, p<0.001). Interestingly enough baPWV was also the only PWV mode for which we were able to show heritability. The SNPs identified in the new genome wide significant locus we discovered at 3p24.1 are all in LD with cis-eQTL SNPs for SLC4A7 (also known as NBCn1). SLC4A7 encodes a transmembrane protein transporting sodium and bicarbonate ions. According to Boedtkjer et al. [45] SLC4A7/NBCn1 promotes arterial remodelling in mice. In 2017, Ng et al. [46] published a functional study of variants at the SLC4A7 locus in cultures of vascular smooth muscle cells (the cellular components of the blood vessel wall) that showed an association between variation at the SLC4A7 locus and blood pressure. The association of this locus with arterial stiffness is therefore biologically plausible.
Biologically interesting are also two loci with SNPs at suggestive significance level. The locus on 2q23.3 around rs55895454 is close to RND3, a gene that is assumed to regulate cell actin cytoskeleton dynamics as a negative regulator of cytoskeletal organization, leading to loss of adhesion. A recent study has shown that RND3 can affect signalling in endothelial cells by shifting the balance of RhoA and Rac1 signals [41]. Another study [42] suggests that the integrin-Gα13-RhoA-YAP pathway is a possible target for drug testing against atherosclerosis. Endothelial-cell signalling can affect plaque formation and RND3 affects, among other things, the RhoA signals in this pathway. The two SNPs in the locus on 9p24.3 are in LD with a cis-eQTL for KCNV2, which may affect pulmonary artery smooth muscle cells. Downregulated potassium voltage-gated channels are involved in pulmonary vascular medial hypertrophy associated with pulmonary hypertension [43]. KCNV2 is an inhibitor of functional K + channels and therefore influences the resting membrane potential in vascular smooth muscle cells. Further association studies with larger sample size as well as additional experimental evidence are required to corroborate these associations.
From the six literature loci we analysed, we were able to replicate only the locus "in the 3'-BCL11B gene desert" [35] on chromosome 14 (cytoband 14q32.2). A study published in 2014 [47] showed that BCL11B (alias CTIP2) might be associated with cardiac hypertrophy (HCM). In 2015 another study showed the correlation between left ventricular hypertrophy and elevated PWV [48], indicating an indirect association between the locus and PWV. None of the other proposed loci was replicated, indicating the heterogeneity of the underlying genetic causes and confirming that the comparability of effects based on different PWV assessments is limited because it also depends on the PWV modes used representing different segments of the arterial tree.
The heritability described for PWV so far is only moderate. While no heritability estimates for baPWV have been published yet, our heritability estimate for baPWV is within the range reported for cfPWV in other literature.
At the current state of research, the relationship between blood pressure and aortic wall stiffening is still not completely understood. While based on the Framingham cohort, higher aortic stiffness was associated with higher risk of incident hypertension, initial blood pressure was not independently associated with progressive aortic stiffening [49]. However, based on the original Framingham Heart Study it was determined that untreated hypertension may accelerate the rate of development of larger artery stiffness [50]. There are other authors that see arterial stiffening as cause for increased blood pressure [51,52] or expect the functional relationship to be bidirectional [53]. Our permutation analysis confirms a non-random overlap of genetic associations between PWV and blood pressure. Also we showed a positive genetic correlation between high blood pressure and baPWV. Based on the results of our MR we show that blood pressure causally affects PWV. Conversely, we were not able to show a causal effect of PWV on blood pressure.
Limitations: Comparability of our results with those of other studies is still hampered by the different technologies used for PWV assessment. The power of this GWAS differs between our phenotypes due to different sample sizes (between n = 3,462 for bfPWV and n = 6,734 for baPWV). Thus, further GWAS and meta-GWAS are required to identify further robust associations and to corroborate our candidates.
In conclusion, we performed the first GWAS of three PWV modes in parallel and the first for bfPWV at all. We discovered one novel and biologically plausible locus with genome-wide significance and successfully replicated one locus published in literature for another PWV mode. By Mendelian Randomization we established evidence of a causal relationship of BP on PWV but not vice versa. This suggests that arterial stiffness is a consequence but not a cause of increased blood pressure. Analyses of Glucose and Insulin-related traits Consortium), MESA (Multi-Ethnic Study of Atherosclerosis), PGC (Psychiatric Genomics Consortium), Project MinE consortium, Repro-Gen (Reproductive Genetics Consortium), SSGAC (Social Science Genetics Association Consortium) and TAG (Tobacco and Genetics Consortium), TRICL (Transdisciplinary Research in Cancer of the Lung consortium), UK Biobank.

Supporting information
We gratefully acknowledge the contributions of Alkes Price (the systemic lupus erythematosus GWAS and primary biliary cirrhosis GWAS) and Johannes Kettunen (lipids metabolites GWAS).

Author Contributions
Data curation: Katrin Horn.