If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Genome-Wide Association Studies of Retinal Vessel Tortuosity Identify Numerous Novel Loci Revealing Genes and Pathways Associated with Ocular and Cardiometabolic Diseases
Dept. of Computational Biology, University of Lausanne, Lausanne, Switzerland;Swiss Institute of Bioinformatics, Lausanne, Switzerland;Jules-Gonin Eye Hospital, Lausanne, Switzerland
Swiss Institute of Bioinformatics, Lausanne, Switzerland;Center for Primary Care and Public Health (Unisanté), University of Lausanne, Lausanne, Switzerland;
Swiss Institute of Bioinformatics, Lausanne, Switzerland;Center for Primary Care and Public Health (Unisanté), University of Lausanne, Lausanne, Switzerland;Center for Integrative Genomics, University of Lausanne, Lausanne, Switzerland;
Dept. of Computational Biology, University of Lausanne, Lausanne, Switzerland;Swiss Institute of Bioinformatics, Lausanne, Switzerland;Center for Primary Care and Public Health (Unisanté), University of Lausanne, Lausanne, Switzerland;
Dept. of Computational Biology, University of Lausanne, Lausanne, Switzerland;Swiss Institute of Bioinformatics, Lausanne, Switzerland;Dept. of Integrative Biomedical Sciences, University of Cape Town, Cape Town, South Africa
To identify novel susceptibility loci for retinal vascular tortuosity, to better understand the molecular mechanisms modulating this trait and reveal causal relationships with diseases and their risk factors.
Design
Genome-Wide Association Studies (GWAS) of vascular tortuosity of retinal arteries and veins, followed by replication meta-analysis and Mendelian randomization.
Participants
We analyzed 116 639 fundus images of suitable quality from 63 662 participants from three cohorts, namely the UK Biobank (n = 62 751), SKIPOGH (n = 397), and OphtalmoLaus (n = 512).
Methods
Using a fully automated retina image processing pipeline to annotate vessels and a deep learning algorithm to determine the vessel type, we computed the median arterial, venous and combined vessel tortuosity measured by the distance factor (the length of a vessel segment over its chord length), as well as by six alternative measures that integrate over vessel curvature. We then performed the largest GWAS of these traits to date and assessed gene set enrichment using the novel high-precision statistical method PascalX.
Main Outcome Measure
We evaluated the genetic association of retinal tortuosity, measured by the distance factor.
Results
Higher retinal tortuosity was significantly associated with higher incidence of angina, myocardial infarction, stroke, deep vein thrombosis, and hypertension. We identified 175 significantly associated genetic loci in the UK Biobank; 173 of these were novel and 4 replicated in our second, much smaller, meta-cohort. We estimated heritability at ∼25% using linkage disequilibrium score regression. Vessel type specific GWAS revealed 114 loci for arteries and 63 for veins. Genes with significant association signals included COL4A2, ACTN4, LGALS4, LGALS7, LGALS7B, TNS1, MAP4K1, EIF3K, CAPN12, ECH1, and SYNPO2. These tortuosity genes were overexpressed in arteries and heart muscle and linked to pathways related to the structural properties of the vasculature. We demonstrated that retinal tortuosity loci served pleiotropic functions as cardiometabolic disease variants and risk factors. Concordantly, Mendelian randomization revealed causal effects between tortuosity, BMI and LDL.
Conclusions
Several alleles associated with retinal vessel tortuosity suggest a common genetic architecture of this trait with ocular diseases (glaucoma, myopia), cardiovascular diseases and metabolic syndrome. Our results shed new light on the genetics of vascular diseases and their pathomechanisms and highlight how GWASs and heritability can be used to improve phenotype extraction from high-dimensional data, such as images.
] and a major societal health burden. Though several risk factors for CVD development, such as age, smoking, and hypertension, have been firmly established, the degree of importance of vascular properties as risk factors is unclear. Retinal fundus photos allow non-invasive in-vivo assessment of the vascular system of the superficial inner retina, i.e. the central and branch veins and arteries plus the venules and arterioles. These vessels are composed of tightly sealed endothelial cells (ECs), forming the inner blood-retina barrier (BRB), encased by smooth muscle cells (SMCs) form to the vessel wall [
]. Automatic segmentation of retinal vessels in fundus images is well established, and computer-aided image analysis started entering clinical care to screen and diagnose ocular and systemic diseases [
Pathological changes in the retinal vessels often coincide with those in the microvasculature of other organs and may precede the progression of systemic vascular diseases. The retinal vasculature can provide insights into neuro-degenerative diseases, such as Alzheimer’s, Parkinson’s, and vascular dementia [
Retinal vascular image analysis as a potential screening tool for cerebrovascular disease: a rationale based on homology between cerebral and retinal microvasculatures.
]. In addition, abnormalities in retinal parameters, such as vascular calibers and tortuosity, are of diagnostic value for systemic diseases, including increased risk of diabetes [
Sabanayagam C, Xu D, Ting DSW, Nusinovici S, Banu R, Hamzah H, et al. A deep learning algorithm to detect chronic kidney disease from retinal photographs in community-based populations. The Lancet Digital Health. 2020;2: e295–e302.
In recent years, genome-wide association studies (GWAS) have been used to link genes with phenotypes extracted from fundus images, such as vessel size [
] on vascular density and fractal dimensions reported 7 and 13 single nucleotide polymorphism (SNPs) associated with these traits respectively, including OCA2, MEF2C and GNB3. Retinal vessel tortuosity has been associated with SNPs that map to the genes ACTN4 and COL4A2 [
]. These results demonstrated that GWAS on retinal traits extracted at a single time point can reveal genes with a potential role in modulating vascular properties and related pathomechanisms.
Here, we report the results of the largest GWAS on vessel tortuosity to date using images and genotypes from 62 751 subjects in the UK Biobank (UKBB) and from 397 and 512 subjects of the much smaller, yet independent, population-based cohorts SKIPOGH [
The CoLaus study: a population-based study to investigate the epidemiology and genetic determinants of cardiovascular risk factors and metabolic syndrome.
Retinal vascular image analysis as a potential screening tool for cerebrovascular disease: a rationale based on homology between cerebral and retinal microvasculatures.
Tapp RJ, Owen CG, Barman SA, Welikala RA, Foster PJ, Whincup PH, et al. Associations of Retinal Microvascular Diameters and Tortuosity With Blood Pressure and Arterial Stiffness. Hypertension. 2019. pp. 1383–1390. doi:10.1161/hypertensionaha.119.13752
], making further discoveries likely. We constructed an automated image analysis pipeline to extract retinal tortuosity from these data as a biomarker. We report the correlation with patient records, SNPs, genes, pathways (set of genes), tissue expression, pathomechanisms, and causal effects associated with this biomarker. Our findings advance the understanding of the molecular players and mechanisms contributing to retinal vessel morphology, which may be important also for other vasculatures and associated diseases.
METHODS
Data: genotypes, phenotypes and fundus images
The UKBB is a population-based cohort of approximately 488 000 subjects with rich, longitudinal phenotypic data and a median 10-year follow-up [
]. We analyzed 173 837 standard retinal 45° color fundus images from 84 825 individuals, captured using a Topcon Triton 3D OCT 1000. Genotyping was performed on Axiom arrays for a total of 805 426 markers, from which approximately 96 million genotypes were imputed. We used the subset of 15 599 830 SNPs that had been assigned an rsID. We performed an additional quality control (QC) step by filtering out SNPs with MAF < 5×10-4. Our choice of low MAF cut-off was motivated by the large power of the UKBB. With our sample size of 62 751 of subjects after QC (see below) we still expect about 30 subjects to have at least one minor allele, so the effect size estimate is still reasonably robust. Finally, we applied a filtering procedure [
] to remove SNPs with imputation quality < 0.3. In addition to genomic information, the UKBB also provided us with phenotypic information from the patient records, particularly with diagnosis dates for: type-2 diabetes, angina, myocardial infarction, deep vein thrombosis (DVT), stroke, hypertension and smoking status. Age, sex, and principal components of genotypes were used to correct for biases in the genetic associations.
We performed replication via a meta-analysis of two independent, population-based cohorts: the Swiss Kidney Project on Genes in Hypertension (SKIPOGH) [
The CoLaus study: a population-based study to investigate the epidemiology and genetic determinants of cardiovascular risk factors and metabolic syndrome.
]. SKIPOGH is a family-based, cross-sectional study exploring the role of genes and kidney hemodynamics in blood pressure regulation and kidney function in the general population, comprising 1 054 genotyped individuals. 1 352 retinal fundus images were available from 518 participants. The genotyping was performed with the Illumina Omni 2.5 chip. OphtalmoLaus is a sub-study of Cohorte Lausannoise (CoLaus), a population-based cohort comprising 6 188 genotyped individuals. 7 252 fundus images were available from 1 015 subjects. CoLaus has as its objective to investigate the epidemiology and genetic determinants of CVD risk factors and metabolic syndrome: participants were phenotyped accordingly. The genotyping was performed using the 500K Affymetrix chip technology. Like in the UKBB, in both Swiss cohorts retinal fundus images were captured using Topcon Triton devices. Genotype imputation for SKIPOGH and CoLaus was performed using Minimac 3 as algorithm and version 1.1 from the Haplotype Reference Consortium (http://www.haplotype-reference-consortium.org) as reference panel. For an overview of our pipeline see Figure 1.
Figure 1Pipeline and results. Relevant phenotypes, genotypes and fundus images were collected from the UKBB, OphtalmoLaus and SKIPOGH. After quality control, the images were processed by deep learning, classifying arteries and veins. A range of tortuosity measures were then calculated, which provided the phenotypes for the GWASs. The primary results were 173 novel genetic trait loci. These associations include signals which were shared between retinal tortuosity and several diseases (metabolic syndrome and CVD). Their aggregation on annotated gene-sets identified relevant pathways and GO terms. Tissue-wide expression analysis revealed expression in the arteries and heart. Correlation analysis revealed associations between retinal tortuosity and cardiometabolic diseases.
]. The exclusion criteria was based on upper and lower thresholds on the total length of the vasculature, and on the number of vessels (see Supplemental Text 1). Roughly two out of three images passed this strict quality control (116 639 out of 173 837 in the UKBB). Based on ARIA's vessel annotations, we calculated a tortuosity measure known as the distance factor (DF) [
where the total vessel length, , is divided by the Euclidean distance between the vessel segment endpoints, . DF is referred to in a recent review as the arc over chord ratio [
Abdalla M, Hunter A, Al-Diri B. Quantifying retinal blood vessels’ tortuosity — Review. 2015 Science and Information Conference (SAI). 2015. doi:10.1109/sai.2015.7237216
]. In addition to DF, we also calculated six other tortuosity phenotypes based on alternative measures using integrals over the curvature along the vessel (see Supplemental Text 2).
We phenotyped each individual by calculating median retinal tortuosities, then averaging the values derived from one image of the left and one from the right eye, when available. If only one retinal image was available we used the value of this image. In the few cases where multiple images were available for the same eye, we only considered one image from the earliest time point (for the resulting distribution, refer to Supplemental Text 3).
Deep Learning classification of arteries and veins
We calculated pixel-wise artery and vein classifications using the Deep Learning algorithm Little W-Net [
Adrian Galdran, André Anjos, José Dolz, Hadi Chakor, Hervé Lombaert, Ismail Ben Ayed. The Little W-Net That Could: State-of-the-Art Retinal Vessel Segmentation with Minimalistic Models. arXiv. 2020. doi:The Little W-Net That Could: State-of-the-Art Retinal Vessel Segmentation with Minimalistic Models
]. For each vessel segment recognized by ARIA, we used the difference between pixels classified as arterial and venous as a score that was required to be positive or negative for the segment to be annotated as artery or vein, respectively. On a set of 44 images, manually annotated by an ophthalmologist, we obtained an area under the curve of 0.93 and an accuracy of 0.88. Thus, we performed vessel type classification for the entire set of retinal fundus images, computing artery- and vein-specific tortuosity values (see Supplemental Text 4).
Genome-wide association analyses
We ran genetic association studies on tortuosity of arteries, of veins, and combining both vessel types (from UKBB CFIs). We used BGENIE [
], applying linear regression to confounder-corrected, quantile-quantile normalized, retinal vessel tortuosity on the genotypes of the matching subjects imputed to a panel of approximately 15M genetic variants. In order to account for confounding effects [
], the following variables were provided as covariates, as usual in GWAS: age, sex, PC of the genotypes (we considered only PCs with a significant correlation to tortuosity, namely 1, 2, 5, 6, 7, 8, 16, 17 and 18). A sensitivity analysis controlling for additional covariates, including age-squared, spherical power, smoking, hypertension, diabetes, eye-related diseases and conditions, assessment-center and genotyping array, indicated only minor impact on the significant association p-values (Supplemental Text 14). We considered SNPs to be nominally significant if their p-value was below the classical Bonferroni threshold of 5×10-8 (i.e. correcting for an estimated 1M of independent SNPs). A list of independent SNP was obtained by performing linkage disequilibrium (LD) pruning using the LDpair function of the R package LDlinkR [
] and the kinship matrix in the model to account for family structure. We also included the recruitment center as a covariable. For the GWAS on the OphtalmoLaus cohort, we used the same parameters and tools as for the discovery cohort. Results from SKIPOGH and OphtalmoLaus were meta-analyzed using an inverse-variance weighting scheme for the respective effect sizes. Due to the small sample size of the replication cohort, we only attempted replication for the SNPs and genes that were significant in the discovery cohort.
LD Hub: a centralized database and web interface to perform LD score regression that maximizes the potential of summary level GWAS data for SNP heritability and genetic correlation analysis.
] tool, to aggregate SNP-wise summary statistics into gene scores using a sum of χ2 statistics: PascalX takes into account LD by effectively transforming the sum of χ2 from all SNPs within the gene window into a new basis of independent “Eigen-SNPs” corresponding to a weighted sum of χ2 statistics. Using multiple-precision arithmetics, PascalX computes the corresponding null cumulative probability distribution to essentially arbitrary precision, while other tools usually only approximate the underlying distribution. We thus computed p-values up to a precision of 10-100, allowing for accurate scoring of genes with contributions from extremely significant SNPs, which become increasingly frequent in highly powered GWASs such as this one.
We used the following configurations: We computed gene scores from SNPs lying within a window of 50kb before the transcription start site and 50kb after the transcript end. The annotation of the gene positions was based on the Genome Reference Consortium Human genome build 37 (GRCh37/hg19) downloaded from the Ensembl biomart [
Kinsella RJ, Kähäri A, Haider S, Zamora J, Proctor G, Spudich G, et al. Ensembl BioMarts: a hub for data retrieval across taxonomic space. Database . 2011;2011: bar030.
] was used to estimate the SNP-SNP correlations (LD effects). PascalX uncovered 265 significant genes (after Bonferroni correction for 25 489 gene-based tests p < 0.05 / 25489 2.0×10-6).
] to compute gene set enrichment scores based on ranking derived from the gene-based tests. As a large number of genes have inflated p-values in highly powered GWASs, this ranking approach was more conservative. We first computed scores for 2 868 canonical pathways (BioCarta, KEGG, PID, Reactome, and WikiPathways), then extended our analysis to the 31 120 pathways in MSigDB (v7.2) [
]. To adjust for statistical dependence and co-expression, genes that are less than 100kb apart were “fused” (i.e. considered as single entities termed “fusion genes” [
] (v8) dataset, comprising 54 tissues. We defined gene sets based on the significant genes from each of our three GWAS on DF tortuosity (artery, vein and combined). PascalX was used to perform an enrichment analysis that indicated whether these sets were over-expressed in any particular tissue. PascalX corrected for the co-expression of gene sub-clusters within each gene set by merging nearby genes to fusion genes. We computed the fusion genes expression values in transcripts per kilobase million from the raw read counts. These values values were made uniform via ranking, transformed to χ2-distributed random variables, summed, and tested against a χ2 distribution with as many degrees of freedom as there were “fusion genes” in each set. We applied a Bonferroni threshold: p = 0.05 / 54 = 9.2×10-4.
Shared genetic signal with disease
We computed the overlap between DF tortuosity SNPs (from the combined-vessel GWAS) and disease-related SNPs. To this end, we first identified which of the independent SNPs in the combined-vessel GWAS were listed in the GWAS Catalog [
] to search for evidence of causal effects between DF tortuosity (from the combined-vessel GWAS) and the following traits: body mass index (BMI), coronary artery disease (CAD), systolic blood pressure (SBP), and lipid traits, namely high-density lipoprotein, low-density lipoprotein (LDL), total cholesterol, and triglycerides. For each trait, we used independent (r2 < 0.01) significant (P < 5×10-8) SNPs as instrumental variables. All summary statistics (estimated univariate effect size and standard error) originated from the most recent meta-analyses (not including UKBB individuals) and were downloaded from the publicly available NIH Genome-wide Repository of Associations Between SNPs and Phenotypes [
Genome-wide Repository of Associations Between SNPs and Phenotypes. In: National Institutes of Health (NIH) [Internet]. [cited Feb 2021]. Available: https://grasp.nhlbi.nih.gov/
], which allowed us to estimate the LD among these SNPs and prune them. We removed strand ambiguous SNPs. Causal estimates were based on the inverse variance weighted method [
The UK Biobank has obtained Research Tissue Bank (RTB) approval from its ethics committee that covers our use of the Resource. The UK Biobank Research Ethics Committee (REC) approval number is 16/NW/0274. OphtalmoLaus obtained ethics approval from La Commission cantonale d'éthique de la recherche sur l'être humain (project PB_2019-00168). The same commission approved ethics for SKIPOGH (Protocols 92/07 and 303/12). All three studies adhere to the Declaration of Helsinki, and obtained informed consent from all subjects.
RESULTS
Baseline characteristics and tortuosity quantification
Following quality control measures, we analyzed 116 639 images from 62 751 subjects of the UKBB (mean±SD age = 56±8 years; 35 098 females at birth [54%]; 4 618 smokers [7%]). We analyzed 1 352 images from 379 subjects of the SKIPOGH cohort (mean±SD age = 48±16 years; 211 females [53%]; 107 smokers [27%]). We analyzed 7 254 images from 512 subjects of the OphtalmoLaus cohort (mean±SD age = 51±10 years; 270 females [53%]). Baseline characteristics and disease prevalence are presented in Supplemental Text 6. For an overview of our pipeline see Figure 1. Note that we did not explicitly exclude subjects with retinal diseases or other ocular conditions from the dataset, but that images from such subjects often did not pass our quality control standards (see Supplemental Text 1).
The distributions of DF tortuosity were similar across cohorts: long-tailed, left-skewed, with means ranging from 1.030 (UKBB) to 1.034 (OphtalmoLaus). DF was higher in the elderly population (Cohen’s d = 0.49, p = 1×10-195), and in women (Cohen’s d = 0.049, p = 9×10-10). Overall, DF was higher in veins (Cohen’s d = 0.13, p = 9×10-142). For details about the stratified analysis of the DF phenotype in the UKBB see Supplemental Text 3.
We extracted six additional tortuosity measures based on alternative mathematical definitions. Correlations analysis and dimensionality reduction in terms of principle components showed that the DF is most similar to the path integral of the squared curvature (3) and least similar to the path integral of the curvature (2). The other alternative measures (4-7) were similar to each other, very different from 2 and of intermediate similarity to the DF and 3 (see Supplemental Text 2).
Vessel tortuosity correlates with disease status
We found that the DF tortuosity of arteries was associated with hypertension (beta = 0.19, p = 3×10-56) and angina (beta = 0.09, p = 6×10-4), but not with myocardial infarction, stroke or DVT. In the case of veins, the DF was significantly associated with hypertension (beta = 0.25, p = 7×10-99), angina (beta = 0.18, p = 2×10-10), myocardial infarction (beta = 0.12, p = 2×10-4), stroke (beta = 0.16, p = 5×10-5) and DVT (beta = 0.11, p = 5×10-4). For predictive power over disease status, see Supplemental Text 7.
Vessel tortuosity GWASs identify 173 novel loci
We identified 7 072 significantly associated SNPs in the combined-vessel GWAS on DF tortuosity in the UKBB (Supplemental Dataset 4A). The vessel type specific GWAS resulted in 6 563 significantly associated SNPs for arteries, and 2 896 SNPs for veins when using a Bonferroni threshold of 5×10-8 (Supplemental Dataset 4B and 4C). We applied LD pruning, identifying 128 independent loci in the combined-vessel GWAS, 116 in the artery-specific GWAS, and 63 in the vein-specific GWAS. Accounting for overlap between these sets (see Supplemental Text 9), we obtained a total of 175 independent lead SNPs (see Figure 2a-c). The top 10 SNPs are listed in Table 1, ordered by significance (for complete listings, see Supplemental Dataset 1). Among the significantly associated variants, rs1808382 and rs7991229 had been previously reported [
] (Supplemental Text 8), whereas the remaining 173 independent lead SNPs represented novel loci associated (see Supplemental Dataset 5).
Figure 2SNP p-values and effects. a, Manhattan plot of Genome-Wide Association Study (GWAS) of retinal vessel tortuosity, combining all vessel types (both arteries and veins). The red line indicates the genome-wide significance level after Bonferroni correction (p = 5×10-8). Oblique dashes on top of peaks mark extremely significant p-values that have been cropped. Squares mark the position of disease SNPs (see Table 4). The trait was corrected for phenotypic variables which showed a statistically significant association, i.e.: age, sex, and a subset of principal components of genotypes. b, Manhattan plots of the vessels-specific GWAS (artery-specific on top, vein-specific at the bottom). Confounder correction, significance level and cropping of extremely significant p-values as in the (a). c, GWAS q-q plot: arteries in red, veins in blue, combined-vessels signal in black; the genome-wide significance level is represented as a green dashed line. d, Statistically significant correlation between the measured effect sizes in the discovery cohort (UKBB, n = 62 751) and replication meta-cohort (SKIPOGH plus OphtalmoLaus, n = 911). We considered all lead (independent) SNPs in the UKBB. We tested all 136 SNPs with matching rsIDs in the replication meta-cohort except one censored outlier (rs187691758), 89 of which had the same sign of their effect size estimate in the UKBB. The resulting Pearson correlation is r = 0.36; p = 1.18×10-5. e, Benjamini-Hochberg procedure on discovery lead SNPs from the UKBB yields 4 hits in the replication cohort using FDR=0.2.
Table 1Top retinal tortuosity SNPs. The 10 most significant DF tortuosity SNPs, ordered by p-value. For full results, refer to the list of 175 independent lead SNPs in Supplemental Dataset 1. Chr: chromosome; SNP: rsIDs of the single nucleotide polymorphism; EA: effect allele; RA: reference allele; freq: allele frequency of effect allele; beta: effect size estimate; -log10p: normalized p-value in the discovery cohort; GWAS type: vessel type to which the signal applied.
Heritability of DF is larger than for other tortuosity measures
The SNP-based heritability differed substantially across tortuosity measures, with DF receiving the highest estimate (h2SNP = 0.25, SE = 0.025). This was approximately twice the heritability estimate of the six alternative curvature-based measures (0.11 ≤ h2SNP ≤ 0.13, 0.011 ≤ SE ≤ 0.012, Supplemental Text 2). We did not observe any significant genomic inflation (see Table 2). Heritability also varied depending on vessel type (h2SNP = 0.23 [SE = 0.020] for arteries, and h2SNP = 0.15 [SE = 0.021] for veins). The distribution of the DF phenotype for each vessel type is shown in Supplemental Text 3.
Table 2| SNP-based heritability. h2SNP: portion of phenotypic variance cumulatively explained by the SNPs; lambda GC: inflation, measure of the effect of confounding and polygenicity acting on the trait; intercept: LD score regression intercept (values close to 1 indicates little influence of confounders, mostly of population stratification); ratio: ratio of the proportion of the inflation in the mean Chi2 that is not due to polygenicity (a ratio close to, or smaller than, 0 is desirable as it indicates low inflation from population stratification). SE are given in parentheses.
Replication of lead SNPs and genes in a small meta-cohort
The sample size of the replication meta-cohort (n=909) is too low to replicate any of our discoveries with a fixed Bonferroni p-value threshold to correct for multiple hypotheses testing. We therefore used the well-established Benjamini–Hochberg procedure [
], which fixes a false discovery rate (FDR), corresponding to a variable threshold that is less stringent for SNPs with lower rank. With this procedure, for FDR=0.1 (so expecting one in 10 positives to be false) we replicated four SNPs (rs10788873, rs2571461, rs501943 and rs35252676, indicated in Fig 2e), and at FDR=0.5 four additional SNPs replicate. At FDR=0.05 we could not replicate any of our hits. For genes we found that 58 replicated at FDR=0.5, but none at FDR=0.1. Clearly our replication meta-cohort lacks power, but many candidate SNPs, and even more so candidate genes, have more significant p-values than expected. Consistently, we observed a Pearson correlation of r = 0.36 (p = 1.18×10-5) between the SNP effect size estimates in the two studies (see Figure 2d and Supplemental Text 5), and r = 0.13 (p = 0.02) between normalized gene ranks (Figure 3d).
Figure 3Gene p-values and replication scores. a, Gene-based Manhattan plot of retinal vessel tortuosity, combining all vessel types (both arteries and veins). 203 genes were significant in arteries, 123 in genes, and 265 when combining the vessel types. Gene-based tests were computed by PascalX [
]. The red line indicates the genome-wide significance level after Bonferroni correction (p = 5×10-8). Squares mark the position of particularly relevant genes (see corresponding Results section). b, Gene-based Manhattan plots of the vessels-specific GWAS (artery-specific on top, vein-specific at the bottom). c, q-q plot of gene p-values: arteries in red, veins in blue, combined-vessel signal in black; the genome-wide significance level is represented as a green dashed line. d, Statistically significant correlation between q-q normalized genes’ p-values in the discovery (UKBB) and in the replication meta-cohort (SKIPOGH + OphtalmoLaus). Only genes that were significant in the discovery cohort were considered. The resulting Pearson correlation is r = 0.13 (p = 0.02). e, Benjamini-Hochberg procedure replicates 58 hits at FDR=0.5 in the replication meta-cohort. We used a candidate approach, meaning only genes that were significant in the discovery cohort were considered.
Tortuosity genes and pathways affect vascular tissue remodeling and angiogenesis
Mapping the SNP-wise association signals onto genes (Methods) we identified 265 significant genes in the discovery GWAS combining vessel types, 203 in the artery-specific GWAS, and 123 in the vein-specific GWAS. Accounting for overlap between these sets (see Supplemental Text 9), we obtained a total of 312 genes (see Figure 3a-c). Top genes are reported in Table 3 (for a complete listing, see Supplemental Datasets 6A/6B/6C). Among those, we replicate the three genes in two independent loci (ACTN4/CAPN12, COL4A2) that were found in a previous GWAS study on tortuosity [
]. A large fraction of these genes carried annotations related to vessel integrity, vascular tissue remodeling and angiogenesis. Specifically, we identified a cluster of highly significant genes on chromosome 19, including ACTN4 (related to actin filament bundling), TNS1 (cross-linking of actin filaments), and CAPN12 (involved in structural integrity to blood vessel walls). This locus also included three genes involved in adhesion to the connective tissue [
]: LGALS7, LGALS7B and LGALS4. We also replicated the highly significant association of tortuosity with two type IV collagen genes, COL4A2 and COL4A1 [
], also received a highly significant association. Finally, among the artery-specific genes, we found FLT1 coding for VEGFR1, which plays a role in vessel formation and vascular biology [
]. (See Discussion for further details and interpretation of these results.)
Table 3| Top retinal tortuosity genes. The 15 most significant DF tortuosity genes, for each GWAS (combining all vessels, considering only arteries, and only veins). P-values were computed by PascalX [
Gene set enrichment (Methods) yielded 78 significant sets in total (see Figure 4), with the strongest signals arising from the combined and artery-specific analysis (see Supplemental Text 9 and Supplemental Datasets 7A/7B/7C). Similarly to genes, many of the pathways pointed to specific biological processes, cellular components, and molecular functions related to vessel integrity and remodeling. These included “human retinal fibroblasts”, “vascular smooth muscle cells” (both in the kidney and the neuroepithelium), and “epithelium development”. We also observed a pathway related to “vascular endothelial growth factors”, VEGFA-VEGFR2, which is a well-known therapeutic target for ocular diseases. We highlight several transcription factors and binding motifs for further experimentation (see Figure 4b). The role of integrity and development of blood vessels for tortuosity was supported by the enrichment of several GO terms such as “circulatory system development”, “anatomical structure morphogenesis” and “tube development”. The enriched terms “cell-substrate junction”, “anchoring junction”, “actin” and “actomyosin” revealed some of the molecular players involved. (See Discussion for more details).
Figure 4Enriched pathways and gene-sets. Arteries in red, veins in blue, combined-vessel signal in black: scores for 31 120 gene-sets in MSigDB (v7.2) [
]. Only gene-sets for which significance was reached by at least one GWAS are shown. The red dashed line indicates Bonferroni-threshold (-log10p = 5.7). The number of genes in each set is indicated in squared brackets. Gene-set names have been shortened and some redundant GO categories are not shown. For details, refer to the extended plot in Supplemental Text 13. a, Enrichment in GO categories. b, Enrichment in pathways referring to a particular molecule (typically a transcription factor) and/or binding motif. c, Enrichment in gene-set obtained from transcriptomic analysis of tissues of treated cell types.
Compared to the DF analysis, the alternative tortuosity measures had lower heritability and fewer enriched genes and pathways. However, some were unique and disease-relevant, such as a pathway related to “abnormal cardiac ventricle morphology” (see Supplemental Text 2).
Tortuosity genes are overexpressed in arteries and heart tissues
Performing enrichment analyses across expression data from 54 tissues, we found that tortuosity genes were overexpressed in three types of arteries (i.e., aorta, tibial artery and coronary artery), two heart tissues (i.e. ventricle and atrial appendage), and, less significantly, fibroblasts and muscular tissues. The profile of enrichment significance values across tissues for tortuosity genes detected by combined-vessel type GWAS analysis is more similar to that of the artery-specific GWAS than that of vein-specific one (see Figure 5), which did not result in any significant tissue associations (for a strict Bonferroni threshold of p = 0.05/54 = 9.2×10-4).
Figure 5Tissue expression results. Arteries in red, veins in blue, combined-vessel signal in black: tissue-specific gene expression analysis of GTEx (v8) [
]. We defined sets based on the significant genes from each of the three GWAS we carried out and asked whether they were over-expressed in a particular tissue. Only top tissues are shown here, for full results refer to Supplemental Figure 19.
Nine of the discovered tortuosity loci had been previously reported as disease variants that mapped to specific genes (Table 4): three loci were linked to vascular diseases (coronary heart disease, myocardial infarction, and arterial hypertension), two loci were linked to ocular diseases (glaucoma and myopia), three loci were linked to other systemic diseases (chronic lymphocytic leukemia, type 2 diabetes, and Alzheimer’s disease), and one loci was linked to digestive conditions (diverticular disease). Similarly, we identified 12 loci influencing both tortuosity and disease risk factors. We also uncovered 26 additional disease variants that have not been confidently mapped to a specific gene (see Supplemental Text 10).
Table 4| Pleiotropic disease-variants. List of variants identified in the tortuosity GWAS (combined-vessel analysis) which were found to be associated with a disease outcome or risk factor in an independent study. We report only exact variants (same rsID in both tortuosity and disease GWAS), which we could confidently map to a gene. Gene p-values were computed by PascalX [
Mehta NN. Large-scale association analysis identifies 13 new susceptibility loci for coronary artery disease. Circulation. Cardiovascular genetics. 2011. pp. 327–329.
Myocardial Infarction Genetics Consortium Genome-wide association of early-onset myocardial infarction with single nucleotide polymorphisms and copy number variants.
Novel insights into the genetics of smoking behaviour, lung function, and chronic obstructive pulmonary disease (UK BiLEVE): a genetic association study in UKBB.
The impact of donor and recipient common clinical and genetic variation on estimated glomerular filtration rate in a European renal transplant population.
We expanded our analysis of disease variants to SNPs belonging to the same LD block (Figure 6). We observe a sizable number of tortuosity-associated variants that overlap with CVD (54 SNPs). Several traits related to metabolic syndrome also stand out: blood pressure (55 SNPs for SBP, 49 for DBP, 15 for pulse pressure), blood cholesterol levels (54 SNPs), BMI (54 SNPs), blood pressure linked to alcohol intake and smoking (44 SNPs for SBP + alcohol, 27 for DBP + alcohol) and type2 diabetes (5 SNPs). In addition, other CVD risk factors share a high number of variants associated with tortuosity, such as protein levels (27 SNPs) and type1 diabetes (9 SNPs). Finally, we detected an overlap with various eye morphology traits, including optic disc morphometry (40 SNPs).
Figure 6Overlap in genetic signals with diseases and other complex traits. Arteries in red, veins in blue, combined-vessel signal in black: number of variants shared with other traits reported in the GWAS Catalog [
] (also considering SNPs in high LD with the lead SNP, r2 > 0.8). Only traits with at least 5 shared associations are included (for a full list, including rsIDs, refer to the Supplemental Dataset 3). The traits with the highest number of shared SNPs belong to metabolic syndrome (blood pressure, BMI, blood cholesterol levels) and CVD. This analysis was generated using FUMA [
Using inverse-variance weighting MR, we observed that exposure to elevated (standardized) levels of LDL reduced the tortuosity of veins by 3% (p = 0.02) and arteries by 5% (p = 0.001). Conversely, increased venous (but not arterial nor combined) tortuosity reduced BMI by 4.4% (p = 0.01) (see Supplemental Text 11).
DISCUSSION
Blood vessel tortuosity is a complex trait whose variation is induced in part during developmental angiogenesis and vascular differentiation, and in part through vessel remodeling due to pathological processes in adult life. Both sources of variation are modulated by the environment, but also genetically through gene and regulatory variants that subtly modulate these processes. In order to better understand the involved genetic architecture we conducted the largest GWAS on retinal vessel tortuosity to date, identifying 173 novel loci and pinpointing numerous genes and gene-sets enriched with these primary association signals. Leveraging the unprecedented number of hits, we performed MR that revealed the causal relationships between retinal tortuosity, BMI and blood lipids. This provides context for the considerable overlap we observed between variants associated with vessel tortuosity and cardiometabolic diseases as well as their risk factors. Our results were consistent with the overexpression of tortuosity-related genes in the aorta, tibial artery, coronary artery, and heart tissues. We found these genes to be involved in the development of blood vessels, the maintenance of vessel integrity and the remodeling as a consequence of disease processes.
Vessel integrity
Several enriched GO categories that are integral to vessel development were enriched, namely “morphogenesis of anatomical structures”, “development of circulatory system”, and “tube development”. Similarly GO categories pertinent to the structural integrity of vessels and the stability of specific tissues were highlighted: “cell-substrate junction” and “anchoring junction” which are responsible for the mechanical attachment of a cell and its cytoskeleton to the extracellular matrix. Molecularly, “actin cytoskeleton”, “actin binding”, “actin filament bundle organization”, and “positive regulation of actin filament bundle assembly” highlighted the important role of actin.
Among the top hits, we found genes directly related to vessel integrity. The product of ACTN4, contributes to cell adhesion and to assembly of the tight junction by mediating actin filament bundling. The paralogues COL4A1 and COL4A2 provide structural support and elasticity to connective tissues by forming the hetero-trimer α1α1α2, which is the most abundant collagen in the basement membrane [
J.M.B. Sand, F. Genovese, N.S. Gudmann, M.A. Karsdal. Type IV collagen. In: Karsdal MA, editor. Biochemistry of Collagens, Laminins and Elastin 2nd edition. 2019.
]. We found both COL4A2 and ACTN4 to be over-expressed in vascular tissues (see Supplemental Text 12). Two more genes with actin-related activity were also among our top hits: TNS1, which promotes cell migration and regulates angiogenesis [
]. Finally, we identified three genes coding for galectins, which are involved in adhesion to the connective tissue via modulation of cell-cell and cell-matrix interactions [
Pathological stresses such as inflammation, infection, or injury, can cause remodeling of vessels, manifesting as occlusions, kinks, tubulations, or other collateral formation of vessels. Pathway analysis identified gene sets of ECs (four sets), SMCs (2 sets), fibroblasts (1 set) and pericytes (1 set) which are the basic cell types composing vessel walls. Dysregulated response of vascular SMC can induce hypertension, and excessive proliferation of these cells contributes to CVD progression [
]. We identified a gene set for “human retinal fibroblasts'' consistent with the fact that this cell type is the most common in connective tissue and involved in maintaining the extracellular matrix. Under stress, fibroblasts proliferate resulting in the accumulation of extracellular materials that ultimately limits elasticity [
]. In addition, we found enrichment in a gene set related to “mesangial cells”, which are kidney-specific pericyte cells. Retinal capillaries are composed of endothelial cells and pericytes. These contractile cells control blood flow in capillaries [
]. Therefore dysregulation of these gene sets has the potential to induce vessel remodeling under stress.
We identified genes directly involved in vessel remodeling. In particular, FLT1 plays a role in the process of collateral vessel formation, which is a form of vascular remodeling in response to stress, such as hypoxia or hypertension [
]. In the retina, VEGFR1 is observed in ECs, SMCs, pericytes and RPE cells (which modulate fibroblast proliferation), and excess VEGFR1 contributes to vessel leakage and angiogenesis [
We detected pleiotropic effects of tortuosity loci, which we showed to be independently associated with CAD, myocardial infarction, hypertension, diabetes, chronic lymphocytic leukemia, Alzheimer’s disease, myopia and glaucoma. We also found tortuosity related genes to be involved in disease pathomechanisms. ACTN4, our top hit, was recently associated with vasorelaxation [
], a mechanism that can lead to hypertension when malfunctioning. The lead SNP in ACTN4 tortuosity (rs1808382) is also independently associated with CAD [
]. Fittingly, COL4A2 also figured among our variants with pleiotropic effects on disease risk (see Table 4). Variants in the fetal genome near FLT1 have been associated with preeclampsia [
], a condition of pregnant women presenting with hypertension and damage to the liver and kidneys, whose underlying mechanism involves abnormal formation of blood vessels in the placenta [
We elucidated causal links between tortuosity and disease risk factors by applying MR. Specifically, we established that elevated LDL exposure causally reduces arterial tortuosity. High LDL is known to cause the buildup of atherosclerotic plaque [
Low-density lipoproteins cause atherosclerotic cardiovascular disease. 1. Evidence from genetic, epidemiologic, and clinical studies. A consensus statement from the European Atherosclerosis Society Consensus Panel.
]. In fact, arteriosclerosis may make retinal arterial walls less flexible and thereby reduce their DF. We observed a negative causal effect of venous tortuosity on BMI, despite the known positive correlation between BMI and retinal tortuosity [
], suggesting that environmental factors may play a role in the relationship between BMI and vascular tortuosity.
Limitations
This study was subject to the following limitations: First, we focused on the DF as a tortuosity measure, since the corresponding GWAS revealed many more significant loci, genes and pathways, as well as a higher heritability estimate in comparison to the alternative curvature-based tortuosity measures. These measures are more sensitive to local physiological vessel features, such as aneurysms or sharp bending (“kinks”), while DF only captures the total vessel elongation. Yet, they may also be more sensitive to the vessel segmentation procedure than the DF. Interestingly, the GWAS for these measures revealed several specific genes and pathways that were not significant in the DF analysis, which may be associated with pathologies manifesting as local disruptions in the microvascular network. Further work is needed to elucidate to what extent the stronger association signals for the DF are due to its robustness as a tortuosity measure or its quality to capture total vessel elongation as the most physiologically relevant trait. Second, due to the small size of our replication meta-cohort we essentially just had sufficient power to verify an overall concordance with the discovery cohort in terms of the highly significant correlation between SNP- and gene-effect sizes, providing independent evidence that they were not driven by any artifacts specific to the UKBB [
]. Even though we could only replicate very few of our SNP-wise hits (4/136 at FDR=0.1), the situation was somewhat better at the level of genes (57/262 at FDR=0.5), underlining the usefulness of signal aggregation from SNPs to genes [
]. Our specific findings should thus be viewed as discoveries in the UK population that still need to be replicated in a much bigger cohort than our Swiss meta-cohort. Finally, we did not attempt to stratify this population by existing diseases, including retinal disorders or other ocular conditions, nor remove subjects with a retinal image from one eye only, all of which may affect our results.
This study exploits advanced automated image processing to characterize different vessel type specific retinal tortuosity measures from retinal fundus images of close to 70k subjects to conduct a high-powered GWAS on this trait. The resulting significant association signals allowed us to provide novel insights into the genetic architecture of retinal tortuosity. Specifically we identified a large number of genes, annotated gene-sets and tissues relevant for this trait, and revealed pleiotropic links with and causal effects to or from disease-related traits. Our study makes important methodological advancements in the large-scale analysis of medically relevant images, which can be applied to other retinal and non-retinal features both in fundamental and clinical research. Our findings provide a significant progress in understanding of molecular players and mechanisms modulating retinal vessel tortuosity, and their links to ocular and cardiometabolic diseases, which is fundamental for developing better tools for their diagnosis and treatment.
Author contributions
MT and SB designed the study. MT and MJB performed QC on the raw images. MT and SOV extracted tortuosity measurements from the image data (UKBB, OphalmosLaus and SKIPOGH). MJB performed classification of arteries and veins. MT carried out the median DF tortuosity GWAS, with the guidance of SB, NM and EP. SOV carried out the tortuosity GWASs based on alternative measures with input from MT. MJB and SOV carried out gene and pathway scoring with the guidance of DK. ALB performed LD Score Regression analysis. SOV evaluated the correlation between different tortuosity measures and their impact on genetic associations. MJB, MT and TC performed the replication analysis in SKIPOGH and CoLaus with input from MB. HA, LK, RS and CB provided ophthalmological expertise and manually annotated the raw image data. MT, CB and SB lead the writing of the manuscript with contributions from all other authors.
Acknowledgements
This work was conducted using data from the UKBB (application ID 43805), SKIPOGH and OphtalmoLaus. We thank Micha Hersch for inspiring this project, to the UKBB team for their support and responsiveness, and to all UKBB participants for sharing their personal data. We also thank aSciStance Ltd for their help in revising the manuscript.
Funding
This work was supported by the Swiss National Science Foundation (#FN 310030_152724/1 to SB) and by the Swiss Personalized Health Network (2018DRI13 to Thomas J. Wolfensberger). The SKIPOGH study was also supported by the Swiss National Science Foundation (#FN 33CM30-124087 to MB). The OphtalmoLaus study was supported by the Claire et Selma Kattenburg Foundation.
Additional information
Please refer to the Supplemental Publication Material and files.
Retinal vascular image analysis as a potential screening tool for cerebrovascular disease: a rationale based on homology between cerebral and retinal microvasculatures.
Sabanayagam C, Xu D, Ting DSW, Nusinovici S, Banu R, Hamzah H, et al. A deep learning algorithm to detect chronic kidney disease from retinal photographs in community-based populations. The Lancet Digital Health. 2020;2: e295–e302.
The CoLaus study: a population-based study to investigate the epidemiology and genetic determinants of cardiovascular risk factors and metabolic syndrome.
Tapp RJ, Owen CG, Barman SA, Welikala RA, Foster PJ, Whincup PH, et al. Associations of Retinal Microvascular Diameters and Tortuosity With Blood Pressure and Arterial Stiffness. Hypertension. 2019. pp. 1383–1390. doi:10.1161/hypertensionaha.119.13752
Abdalla M, Hunter A, Al-Diri B. Quantifying retinal blood vessels’ tortuosity — Review. 2015 Science and Information Conference (SAI). 2015. doi:10.1109/sai.2015.7237216
Adrian Galdran, André Anjos, José Dolz, Hadi Chakor, Hervé Lombaert, Ismail Ben Ayed. The Little W-Net That Could: State-of-the-Art Retinal Vessel Segmentation with Minimalistic Models. arXiv. 2020. doi:The Little W-Net That Could: State-of-the-Art Retinal Vessel Segmentation with Minimalistic Models
LD Hub: a centralized database and web interface to perform LD score regression that maximizes the potential of summary level GWAS data for SNP heritability and genetic correlation analysis.
Kinsella RJ, Kähäri A, Haider S, Zamora J, Proctor G, Spudich G, et al. Ensembl BioMarts: a hub for data retrieval across taxonomic space. Database . 2011;2011: bar030.
Genome-wide Repository of Associations Between SNPs and Phenotypes. In: National Institutes of Health (NIH) [Internet]. [cited Feb 2021]. Available: https://grasp.nhlbi.nih.gov/
J.M.B. Sand, F. Genovese, N.S. Gudmann, M.A. Karsdal. Type IV collagen. In: Karsdal MA, editor. Biochemistry of Collagens, Laminins and Elastin 2nd edition. 2019.
Low-density lipoproteins cause atherosclerotic cardiovascular disease. 1. Evidence from genetic, epidemiologic, and clinical studies. A consensus statement from the European Atherosclerosis Society Consensus Panel.
Mehta NN. Large-scale association analysis identifies 13 new susceptibility loci for coronary artery disease. Circulation. Cardiovascular genetics. 2011. pp. 327–329.
Novel insights into the genetics of smoking behaviour, lung function, and chronic obstructive pulmonary disease (UK BiLEVE): a genetic association study in UKBB.
The impact of donor and recipient common clinical and genetic variation on estimated glomerular filtration rate in a European renal transplant population.
We performed the largest Genome-Wide Association Study of retinal vascular tortuosity to date, identifying several loci that are shared with ocular diseases (glaucoma, myopia), cardiovascular diseases and metabolic syndrome.
Competing interests
The authors declare no competing interests.
Declaration of interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
00020-9&id=gr1.jpg){kind=link}
00020-9&id=gr2.jpg){kind=link}
00020-9&id=gr3.jpg){kind=link}
00020-9&id=gr4.jpg){kind=link}
00020-9&id=gr5.jpg){kind=link}
00020-9&id=gr6.jpg){kind=link}