Association of PEDF polymorphisms with age-related macular degeneration and polypoidal choroidal vasculopathy: a systematic review and meta-analysis

This study assesses the association of the pigment epithelium-derived factor (PEDF) gene with age-related macular degeneration (AMD) and polypoidal choroidal vasculopathy (PCV). Publications in MEDLINE and EMBASE up to 21/08/2014 were searched for case-control association studies of PEDF with AMD and/or PCV. Reported studies giving adequate genotype and/or allele information were included. Pooled odds ratios (OR) and 95% confidence intervals (CI) of each polymorphism were estimated. Our literature search yielded 297 records. After excluding duplicates and reports with incomplete information, 8 studies were eligible for meta-analysis, involving 2284 AMD patients versus 3416 controls, and 317 PCV patients versus 371 controls. Four PEDF polymorphisms were meta-analyzed: rs1136287, rs12150053, rs12948385 and rs9913583, but none was significantly associated with AMD or PCV. The most-investigated polymorphism, rs1136287, had a pooled-OR of 1.02 (95% CI: 0.94–1.11, P = 0.64) for AMD. In subgroup analysis by ethnicity, no significant association was identified. Polymorphisms present in single report showed no association. Therefore, existing data in literature does not support the role of PEDF in the genetic susceptibility of AMD and PCV, although replication in specific populations is warranted. Since the pooled-sample size for PCV was small, there is a need of PEDF genotyping in larger samples of PCV.

progress of CNV in AMD 14,15 . The vascular endothelial growth factor A (VEGF-A) plays an important role in inducing the growth of choroidal new vessels [16][17][18] , which triggers proliferation and migration of vascular endothelial cells 19 . Anti-VEGF therapies have been widely used to treat both nAMD and PCV. Also, variants in the VEGF-A gene have been implicated in the genetic mechanism of AMD and PCV 20,21 . A recent meta-analysis of 9 articles with 2281 AMD cases versus 2820 controls revealed that VEGF-A rs1413711 and rs833061 increased the risk of AMD 21 .
In contrast to VEGF, the PEDF, a member of the serine proteinase inhibitor family, potently inhibits angiogenesis and regulates choroidal neovascularization in humans [22][23][24][25] . PEDF has been detected in the aqueous humor, vitreous, retina and choroid 26,27 . In the retina, PEDF inhibits the proliferation and migration of retinal endothelial cells and vascular permeability induced by VEGF, promotes the apoptosis of endothelial cells and down-regulates the pro-angiogenic factors 23,28,29 . PEDF is also a highly effective inhibitor of angiogenesis in cell culture and animal models 25,[30][31][32] . Decreased vitreous level of PEDF had been associated with the CNV in AMD 33 . PEDF as a potential therapeutic agent has been investigated in animal models of CNV, and it was found that periocular or intravitreal introduction of PEDF could inhibit CNV 22,[34][35][36] . Moreover, patients with nAMD were found to have reduced CNV size after a single intravitreal injection of PEDF-expressing adenoviral vector in a phase I clinical trial 37 . These studies altogether suggest that PEDF is an important factor for CNV, and hypothetically the PEDF gene is an excellent candidate gene for nAMD.
In 2005, Yamagishi et al. hypothesized that a single nucleotide polymorphism (SNP) in PEDF, rs1136287 (c.C215T, p.Met72Thr), might be a genetic marker for AMD 38 . This amino acid substitution is located at the end of a helix domain of the PEDF protein and leads to a BsstSI restriction site 39 , suggesting it could have functional impact. Later, rs1136287 was found to be associated with AMD in the Taiwanese and Korean populations 40,41 , providing initial evidences to support PEDF as a susceptibility gene for AMD. Subsequent studies in Caucasians and other Asian cohorts, including Japanese and Chinese, did not identify a significant association between rs1136287 and AMD, but the effects of the SNP, represented by the odds ratio (OR), were variable across study cohorts [42][43][44][45][46] . As such, whether PEDF rs1136287 is a genuine genetic marker for AMD remains inconclusive. Also, whether there is population-specific association of this SNP with AMD needs further confirmation. Moreover, SNPs other than rs1136287, such as rs12150053, rs12948385 and rs9913583 in PEDF had also been reported in AMD or PCV 41,[43][44][45][46] , but their associations remains inconclusive.
Since PEDF is functionally important in AMD pathogenesis and could be a new target for AMD treatment, the identification of disease-associated gene variants could provide useful targets for studying the roles of PEDF in AMD pathogenesis and pharmacogenetics. To confirm the role of PEDF as a candidate gene for AMD and PCV, we conducted a systematic review and meta-analysis to evaluate the associations of all reported PEDF SNPs with AMD and PCV. This report is about the results of the meta-analysis.

Results
Characteristics of eligible studies on PEDF in AMD and/or PCV. Figure 1 showed the study inclusion of this meta-analysis. A total of 297 articles published between January 1, 1980 and August 21, 2014 were identified in the EMBASE and MEDLINE databases. By excluding duplicated and unrelated records, we got 13 relevant reports for further assessment. However, 6 of them were excluded because 2 articles are about PEDF and the treatment response in AMD and PCV 47,48 , 1 is not a case-control study 38 , and the other 3 are reviews [49][50][51] . By searching the reference lists of relevant studies and genome-wide association studies (GWAS) of AMD, we identified another 2 relevant articles 52,53 . However, one GWA study did not provide genotype or allele data in controls and was excluded from further analysis 52 . Therefore, a total of 8 case-control studies were finally included for the meta-analysis [40][41][42][43][44][45][46]53 , involving 2284 AMD cases versus 3416 controls, and 317 PCV cases versus 371 controls. The main characteristics of the included studies were summarized in Table 1. The sample sizes of AMD groups ranged from 109 to 893, PCV groups 140 to 177, and control groups 90 to 2199. The mean age ranged from 70.9 to 78.6 years in the AMD groups, 65 to 73 years in the PCV groups, and 44 to 77.4 years in the control groups. The gender ratios (male/female) varied from 0.51  Risk of bias assessment in eligible studies. As shown in Table 2, risk of ascertainment bias in the diagnostic criteria for AMD and PCV was present in 3 studies, in which subjects were not defined based on fundus fluorescein angiography (FFA) and indocyanine green angiography (ICGA) 40,41,53 . One paper did not mention the ascertainment of control subjects 41 . Risk of population stratification presented in 1 study, which used spouses and friends as control subjects 53 . There was risk of confounding bias in 5 studies, in which results were not adjusted for confounding variables 40,41,45,46,53 . All the studies reported Hardy-Weinberg equilibrium (HWE) in controls.
Meta-analysis of PEDF polymorphisms in AMD. Totally 9 SNPs had been tested in AMD and/or PCV in the literature. However, only 4 SNPs (rs1136287, rs12150053, rs12948385 and rs9913583) in AMD and 1 SNP (rs1136287) in PCV were reported in at least two studies and thus eligible for meta-analysis. Summary of the allelic associations of these four PEDF polymorphisms are shown in Table 3. The other 5 SNPs, i.e., rs11658342, rs1894286, rs2269344, rs9889773 and rs3786136, appeared in single report but showed no association with AMD 45,54 . We also searched for SNP IDs that were merged with PEDF SNPs and found that rs1804144, rs3199567, rs16951641, rs17352972, rs17845405, rs17858264 and rs58553017 have been merged into rs1136287 in the dbSNP database. However, none of these SNP IDs except rs1136287 was adopted in the studies. SNP rs1136287 is the most-investigated SNP in AMD, with a total of 2284 cases and 3416 controls included for the meta-analysis [40][41][42][43][44][45][46]53 .
Regarding the other three SNPs, rs12150053 (c.-5736T.C), rs12948385 (c.-5304G.A) and rs9913583 (c.-86C.A), which are located in the promoter region or 59-untranslated region of PEDF, the pooled ORs were not statistically significant in AMD in the allelic (Table 3  Data extracted from a genome-wide association study; for the control subjects only age range was given in the study. nAMD: neovascular AMD; nnAMD: non-neovascular AMD; PCV: polypoidal choroidal vasculopathy; NA: not available from the initial report; SD: standard deviation. www.nature.com/scientificreports Table S1). Leave-one-out sensitivity analysis indicated that there was no substantial change after excluding any individual study, which manifested stable results (P . 0.1). Also, there was no publication bias for any SNPs according to the funnel plots and Egger's test (P . 0.1).

Discussion
In this systematic review and meta-analysis, we have, for the first time, summarized the association profiles of PEDF in AMD and PCV, and we found no significant association between reported PEDF SNPs and AMD/PCV in overall samples or different ethnic subgroups. Therefore, current literatures do not support the role of PEDF as a susceptibility gene for AMD and PCV, despite the PEDF protein is functionally important in choroidal neovascularization. PEDF, an antiangiogenic and neurotrophic factor, involves in the neuronal survival and differentiation in the eye and brain. The PEDF gene, also known as SERPINF1, is located on chromosomal region 17p13.1. It contains 8 exons, encoding a polypeptide of 418 amino acids. PEDF was considered as a disease gene for AMD because (1) the PEDF protein is one of the most effective endogenous inhibitors of angiogenesis and neovascularization 28,55 ; (2) decreased PEDF level was found in the vitreous of AMD patients than controls, suggesting that the deficiency of PEDF in the eye could play a role in the pathogenesis of AMD 18 ; (3) intravitreal injection of PEDF-expressing adenoviral vector reduced CNV size in AMD 37 ; and (4) there were studies reporting significant association of PEDF SNPs with AMD and/or PCV.
In the studies of Lin et al., rs1136287 was found to be associated with nAMD in Taiwanese, with an odds ratio of 2.21 (95% CI: 1.43-3.42, P 5 0.005) for the T allele, but not with atrophic AMD (OR 5 0.94, 95% CI: 0.61-1.45, P 5 0.87) 40 . Also rs1136287 was associated with nAMD in a Korean population (OR 5 1.39, 95% CI: 1.01-1.93, P 5 0.045) 41 . However, as summarized in our meta-analysis, in  53 . Therefore, together with the stratification analysis in our meta-analysis, we conclude that SNP rs1136287 is not a susceptibility genetic marker for AMD in Caucasians. This leads to a question that whether the association of rs1136287 with AMD is population-or disease subtypespecific. Notably, the criteria for AMD classification for the subjects in all the included studies were consistent, except the Korean cohort, in which no explicit criteria for AMD were described 44 . Thus, subtypespecific association is less likely, especially for the Taiwanese cohort. Also, in our meta-analysis, we found no significant association of rs1136287 with all forms of AMD in Caucasians or Asians, with the ORs towards different directions in different populations and small heterogeneity among studies (Fig. 2). This suggests the associations detected in the Taiwanese and Korean cohorts might be chance findings from a limited sample size, or there could be population-specific association of this SNP, confirmation of which requires replication among Taiwanese and Korean populations. In recent years, GWAS, a hypothesis-free approach, has become a predominant method for identifying associated loci for AMD. However, findings across different GWAS had been variable 4,7,53,56,57 . This could be due to the variation in sample sizes and ethnicities, leading to variable association signals. For example, the association signal at the 4q12 gene cluster, first identified in the Japanese population 56 , was not detected in a recent large-scale AMD GWAS in Caucasians 57 , suggesting population-specific effect. Moreover, in GWAS only SNPs reaching a predefined threshold are replicated, thus some disease-associated SNPs might have been missed, resulting in false negatives. Therefore, on the one hand, results from mul-tiple cohorts should be incorporated to validate the initial signals by methods such as meta-analysis; whereas on the other hand, methods other than GWAS should play a role in gene identification. The candidate gene approach, which is hypothesis-based, has been an important method for mapping AMD genes. For example in the complement pathway, while the CFH gene was identified for AMD by GWAS, the complement components 2 and 3 (C2 and C3), and complement factor B (CFB) were identified under the hypothesis that variation in genes encoding proteins of the same pathway with CFH could be associated with AMD 58,59 . These genes have later been confirmed in genotyping studies 57 or meta-analysis 60 on AMD and PCV 61 , suggesting the importance of candidate gene analysis in the discovery of AMD genes. In our present meta-analysis, we aimed to confirm the role of an excellent candidate gene for AMD -PEDF, but we found no significant association of a major SNP rs1136287 with AMD or PCV. Nevertheless, our observation that no replication study has yet been available in Taiwanese and Korean populations may arouse follow-up studies on this SNP in the two populations.
Apart from rs1136287, we also found that another 3 PEDF SNPs, rs12150053, rs12948385 and rs9913583, were not associated with AMD. Notably among the included studies, only one involved haplotype-tagging SNPs in the association analysis 45 . In the study, 4 tagging SNPs (rs11658342, rs1894286, rs2269344 and rs9889773) with r 2 cutoff of 1.0 and minor allele frequency (MAF) cutoff of 5%, and 3 reported SNPs (rs1136287, rs12150053 and rs12948385) were selected. The CT genotype of rs1136287 was found to be associated with AMD (OR 5 0.59, 95% CI: 0.36-0.95, P 5 0.03), although the allelic associations of all SNPs were non-significant 45 . Since a r 2 cutoff of 1.0 is likely to limit the number of SNPs that capture the major proportion of gene information, we searched for the tagging SNPs of PEDF using a MAF cutoff of 5% and r 2 cutoff of 0.8 in the HapMap Genome Browser (release No. 27, http://hapmap.ncbi.nlm. nih.gov/; accessed Oct 27, 2014), and we identified more SNPs (Supplementary Table S3). Among the included studies, one tagging SNP rs1136287 in the CHB (Han Chinese in Beijing) population  was included in 3 Chinese cohorts 40,45,46 , while one tagging SNP rs9913583 in the JPT (Japanese in Tokyo) population was studied in 1 Japanese cohort 44 . Since haplotype-tagging SNP analysis is useful for identifying the responsible SNP in a genetic locus 62 , further studies using the tagging SNPs would be warranted to provide a comprehensive evaluation of PEDF in AMD and PCV.
Upon identifying a lack of association of the PEDF SNPs with AMD and PCV, we searched for possible association of PEDF with other ophthalmic and non-ophthalmic diseases with a view to better understand the role of PEDF. Previously, based on linkage analysis, Koenekoop et al. suggested PEDF as a causative gene for Leber's congenital amaurosis 63 . In the study of Miyake et al., PEDF rs12603825 was found to have marginal association with myopic CNV, while rs1136287 was found of no association with CNV in Japanese myopic patients 64 . A meta-analysis is not possible because only one article was found for each SNP. Of note, PEDF was evaluated in diabetic retinopathy in three studies 39,65,66 . However, three SNPs, rs1136287, rs12150053 and rs12948385, showed no significant association with DR by meta-analysis ( Supplementary Fig. S2).
The current systematic review and meta-analysis is an overview of published genetic reports on PEDF in AMD and PCV. Our study revealed several limitations in this topic. First, AMD and PCV are multifactorial diseases involving both environmental and genetic factors. Our results were based on unadjusted assessment and indicated discrepancy even in the same ethnicity, suggesting environmental factors may be involved. Thus, a more rigorous analysis should be performed by stratifying other risk factors if data was available. Second, we found a limited number of studies for this meta-analysis and most cohorts were Asians. In particular, only 2 studies with relatively small sample sizes reported the association of PEDF with PCV. Therefore, a comprehensive association of PEDF with AMD and PCV should be elucidated in more study cohorts. Also, no replication study was available in the Taiwanese and Korean populations, in which significant association of rs1136287 with AMD was detected. Therefore, further studies should be performed in these 2 populations. Third, significant heterogeneity was detected for some SNPs (e.g., rs9913583 in AMD), thus the random-effect model was applied to yield more conservative odds ratio. Fourth, some studies got low score of quality assessment 40,41 or did not give the classification of AMD 44 , which may cause imprecise results when AMD or PCV was evaluated separately. However, since our metaanalysis showed no association of PEDF SNPs with either AMD or PCV, the lack of classification in the initial studies would have no major impact to the final conclusion.
In conclusion, this systematic review and meta-analysis has, for the first time, provided an overview of reported PEDF SNPs in AMD and PCV, and the results suggest that PEDF is not a major susceptibility gene for the diseases in the overall population. However, further studies should be warranted to confirm the association of PEDF SNP rs1136287 with AMD in specific populations such as Taiwanese and Korean. Moreover, since the pooled sample size for PCV was small, further studies of PEDF in large PCV samples are warranted.

Methods
Literature search. Literature search was done in the MEDLINE and EMBASE databases for genetic studies on PEDF in AMD and/or PCV. We used MeSH terms and free words: (Pigment epithelium-derived factor or PEDF or SERPINF1 or serpin peptidase inhibitor, clade F or OI6 or OI12 or EPC-1 or PIG35) and (age-related macular degeneration or AMD or ARMD or age-related macular disease or agerelated maculopathy or ARM or PCV or polypoidal choroidal vasculopathy). All related articles published before August 21, 2014 were retrieved without language restriction. To find more papers, we manually screened the reference lists of all eligible articles. Moreover, to maximize the usable data we also searched all reported genomewide association studies of AMD including their supplementary materials. Details of search strategy were illuminated in Supplementary Table S4.
Inclusion and exclusion criteria. A study was included if it fulfills all the following criteria: (1) case-control study, cohort study or population-based study investigating the association between PEDF and AMD and/or PCV; (2) data of genotype and/or allele counts or frequencies were presented in the papers; (3) unrelated control subjects were free of AMD or PCV; (4) for articles published by the same group of authors on the same gene or markers, only the latest study or the one with the largest sample size was included. Animal researches, case report, reviews, conference report, editorial comment and reports without sufficient data were excluded (Supplementary  Table S5).
Literature review and data extraction. Two reviewers (L.M. and S.M.T.) independently reviewed and extracted data from studies on the association between PEDF SNPs and AMD/PCV. Any discrepancies were resolved by another two reviewers (S.S.R. and L.J.C.) after thorough discussion. The following information was extracted from each record: the name of first author, publication year, ethnicity of the study population, study design, sample size, disease subtype, gender composition, mean age, allele and genotype distribution in cases and controls, Hardy-Weinberg equilibrium (HWE) test results in controls. The records were combined into one group if listed allele and/or genotype distribution was stratified by AMD classification 40 .
Risk of bias assessment. The reviewers appraised the qualities of retrieved records by a modified risk-of-bias score for genetic association studies, based on traditional epidemiologic and genetic considerations (Supplementary Table S6) [67][68][69] . The assessment consists of 3 domains: (1) information bias: evaluation of diagnostic criteria for AMD, PCV and controls; (2) confounding bias: assessment of population bias and other confounding variables; (3) consideration of HWE in individual study. Each domain contains 3 answers: yes, no or unclear, which presents low risk of bias, high risk of bias and unclear if the included study was assessed based on insufficient information.
Statistical analysis. Meta-analysis for each polymorphism was performed if it had been reported in $2 studies or cohorts. The association was assessed using different genetic models, including allelic, dominant, recessive, heterozygous and homozygous models. Pooled odds ratios and 95% confidence intervals of each SNP were estimated for the strength of association, using the fixed-effect (I 2 # 50%) or random-effect (I 2 . 50%) model based on the heterogeneity test 70 . The I 2 test was used to assess heterogeneity among studies. The I 2 value was explained as of no (0-25%), low (25-50%), moderate (50-75%) or high heterogeneity (75-100%) 71 .
Sensitivity analysis was carried out to examine the influence by removing one study each time 69 . The potential publication bias was evaluated with funnel plots and the Egger's test 72,73 . When the p value of the Egger test was ,0.05, publication bias was expected to exist. Statistical analyses were conducted using the software Review Manager (RevMan, version 5.2, The Cochrane Collaboration, Copenhagen, Denmark), and Egger's test was performed in R (version 2.15.0, http://cran.r-project. org/). A pooled P value of less than 0.05 was considered statistically significant.