This meta-analysis was registered with the PROSPERO international prospective register of systematic reviews (registration number CRD42021226402). We searched medical literature for relevant studies using PubMed, EMBASE, the Cochrane Library, Web of Science, WanFang Database, and China National Knowledge Infrastructure (CNKI) from their date of establishment to December 2020. A total of 463 records were identified using electronic search strategies; in addition, one relevant record was obtained from the reference lists of the included studies. We used the search terms: “Steroid 17-alpha-Hydroxylase or CYP17,” “aromatase or CYP19,” “sex hormone binding globulin or SHBG,” “Single nucleotide polymorphism (SNP) or polymorphisms or genotype or genetic or mutation or variant,” and “Polycystic ovary syndrome or PCOS.” We limited the publication type to case-control studies, and there were no language or location restrictions. We also tried to search in grey literature, but no new relevant cohort studies were found.

To be included in this meta-analysis, the studies needed to meet the following inclusion criteria: (a) originated from a case-control study design; (b) reported the cases of PCOS patients diagnosed with one of the following three diagnostic criteria: National Institutes of Health (NIH) 1999, Rotterdam 2003, and AE-PCOS Society 2006; (c) evaluated the association between SHBG or CYP17 or CYP19 gene polymorphisms and PCOS; (d) reported odds ratios (ORs) and corresponding 95% confidence intervals (CIs) or provided the distribution of sufficient genotypic and allelic data for estimation in cases and controls.

Exclusion criteria included the following: (a) insufficient data on genotyping; (b) no control population; (c) study included disorders other than PCOS; (d) case reports, case series, abstracts, reviews, meta-analysis, comments, editorial articles, or letters without original data; (e) duplicate data; (f) the SNP was reported in less than 5 case-control study.

For published articles, two investigators (CX and HZ) independently extracted data and assessed methodological quality. Data were extracted from the included studies to collect the following necessary information: first author, year of publication, country of origin, ethnicity, number of cases and controls, PCOS diagnostic criteria, genotype method, genotype data, and evidence of Hardy-Weinberg equilibrium (HWE). Genotype data of case and control studies were extracted to calculate the OR with 95% CI and p-value of HWE in the control group. As described elsewhere, quality, internal validity, and risk of bias of the included studies were assessed using the validated quality of genetic association studies checklist (Q-Genie) (Santana et al., 2020). The Q-genie tool consists of 11 questions, which address the following aspects of study methodologies: study rationale, outcome, comparability, exposure, bias, sample size, analyses, statistical methods and control for confounding, inferences for genetic analyses, and inferences drawn from results. Each question has seven possible answers as follows: “1 (poor),” “2,” “3 (good),” “4,” “5 (very good),” “6,” “7 (excellent).” The overall quality of studies is classified as “poor quality” if score is <35, a score of 36–45 indicates “moderate quality” and a score of >45 indicates “good quality.” Discrepancies were discussed or adjudicated with the third author (H.B.), until a satisfactory consensus was reached.

Pooled ORs with corresponding 95% CIs were calculated to identify the potential association between susceptibility to PCOS and gene polymorphisms for the following genotypic models: dominant model (mtmt + wtmt vs. wtwt), recessive model (mtmt vs. wtwt + wtmt), co-dominant model (mtmt vs. wtwt) or (wtmt vs. wtwt), and complete overdominant model (mtmt + wtwt vs. wtmt) for rs743572 of CYP17 gene polymorphisms, rs2414096 of CYP19 gene polymorphisms, rs6529 of SHBG gene polymorphisms. In order to synthesize the best genetic model, we refer to the method of screening the best genetic model in the study of Thakkinstian et al. (2005). Data from each study is extracted as the number of subjects with each genotype (AA, Aa, and aa) in the case and control groups. The gene effects of each study are defined as OR, which are OR1, OR2, and OR3 for AA versus aa, Aa versus aa, and AA versus Aa, respectively. The pooled ORs are calculated by the inverse variance method. To determine the overall gene effect, the model that includes gene is compared with the model without gene. If the overall gene effect is statistically significant, further comparisons of OR1, OR2, and OR3 are explored. These comparisons should be performed using residual variance obtained from the regression model. These pairwise differences can be used to indicate the most appropriate genetic model, as outlined below (assuming that the risk allele is A): (a) If OR1 = OR3 = 1 and OR2 = 1, then a recessive model is suggested; (b) If OR1 = OR2 = 1 and OR3 = 1, then a dominant model is suggested; (c) If OR2 = 1 = OR3 = 1 and OR1 = 1, then a complete overdominant model is suggested; (d) If OR1 > OR2 > 1 and OR1 > OR3 > 1 (or OR1 < OR2 < 1 and OR1 < OR3 < 1), then a co-dominant model is suggested. The heterogeneity between the included studies was estimated using the Q test and the I-squared statistic. Values of p < 0.1 and I ² > 50% were considered to indicate significant heterogeneity (Higgins et al., 2003), using of the quality effects models. Subgroup analysis was performed for gene with more than five different studies according to PCOS, control, and ethnicity. LFK index and Doi plots were used to statistically assess publication bias (Furuya-Kanamori et al., 2018). For all tests, p < 0.05 was considered statistically significant. Stata version 16.0 (Stata Corp, College Station, TX) was used to perform the statistical analysis.

The initial search strategy retrieved a total of 464 studies from all electronic databases. After reading the titles, abstracts, and full texts, we selected 26 studies (Diamanti-Kandarakis et al., 1999; Marszalek et al., 2001; Cousin et al., 2004; Kahsarmiller et al., 2004; Tan et al., 2005; Bendlová et al., 2007; Ferk et al., 2007; Echiburú et al., 2008; Park et al., 2008; Jin et al., 2009; Unsal et al., 2009; Martinez-Garcia et al., 2012; Dasgupta et al., 2014; Dai et al., 2015; Li et al., 2015; Abu-Hijleh et al., 2016; Banerjee et al., 2016; Mehdizadeh et al., 2017; Wu et al., 2017; Jiao et al., 2018; Kaur et al., 2018; Bhatnager et al., 2019; Liu et al., 2019; Rahimi and Mohammadi, 2019; Munawar Lone et al., 2020; Ashraf et al., 2021), which included a total of 8,903 women, for inclusion in the analysis. The remaining studies were excluded: 84 were duplicates, 315 were of apparent irrelevance based on a review of abstracts and titles, four were not case-control studies, 19 lacked sufficient genotyping data, 12 included other SNPs without gene locus, two were combined with other diseases, and two had SNPs reported in less than 5 case-control study. The literature screening process and results are shown in Figure 1, followed the PRISMA flowchart.

Table 1 presents the characteristics of included studies. During the search no language restriction was applied. Trials were performed in European, American, and Asian countries and were published from 1999 to 2020. Participants in these trials were recruited from outpatient clinics, hospitals, or medical centers and were diagnosed with PCOS according to either the NIH or Rotterdam criteria. Participants aged 18–49 years were included in all 26 studies exploring the association between CYP17A1 (Diamanti-Kandarakis et al., 1999; Marszalek et al., 2001; Kahsarmiller et al., 2004; Tan et al., 2005; Echiburú et al., 2008; Park et al., 2008; Unsal et al., 2009; Dasgupta et al., 2014; Li et al., 2015; Banerjee et al., 2016; Wu et al., 2017; Kaur et al., 2018; Rahimi and Mohammadi, 2019; Munawar Lone et al., 2020; Ashraf et al., 2021), CYP19A1 (Jin et al., 2009; Mehdizadeh et al., 2017; Jiao et al., 2018; Kaur et al., 2018; Munawar Lone et al., 2020), and SHBG (Cousin et al., 2004; Bendlová et al., 2007; Ferk et al., 2007; Martinez-Garcia et al., 2012; Dai et al., 2015; Abu-Hijleh et al., 2016; Bhatnager et al., 2019; Liu et al., 2019) gene polymorphisms and PCOS risk. There were 15 studies of CYP17A1 gene polymorphisms, including 15 studies of CYP17A1 rs743572 and one study of CYP17A1 loci -600, CYP17A1 loci +723, CYP17A1 loci +2612, CYP17A1 loci +4259, CYP17A1 loci +4444 and CYP17A1 loci +4994; there were six studies of CYP19A1 gene polymorphisms, including five studies of CYP19A1 rs2414096, one study of CYP19A1 rs2236722, two studies of CYP19A1 rs700519 and one study of CYP19A1 rs700518; there were eight studies of SHBG gene polymorphisms, including eight studies of SHBG rs6529, three studies of SHBG rs727428, one study of SHBG rs13894, one study of SHBG rs858521, one study of SHBG rs1799941 and one study of SHBG rs6527. There were 17 studies of Asian patients (Tan et al., 2005; Park et al., 2008; Jin et al., 2009; Unsal et al., 2009; Dasgupta et al., 2014; Dai et al., 2015; Li et al., 2015; Banerjee et al., 2016; Mehdizadeh et al., 2017; Wu et al., 2017; Jiao et al., 2018; Kaur et al., 2018; Bhatnager et al., 2019; Liu et al., 2019; Rahimi and Mohammadi, 2019; Munawar Lone et al., 2020; Ashraf et al., 2021), eight studies of Caucasian patients (Diamanti-Kandarakis et al., 1999; Marszalek et al., 2001; Cousin et al., 2004; Kahsarmiller et al., 2004; Bendlová et al., 2007; Ferk et al., 2007; Martinez-Garcia et al., 2012; Abu-Hijleh et al., 2016), and one study of mixed patients (Echiburú et al., 2008), which included 4860 PCOS cases and 4043 controls overall. All studies extracted DNA from peripheral blood, and PCR, PCR-RFLP, PCR-LDR, and TaqMan assays were used. The genotype distributions among the controls of all studies were consistent with HWE in most studies. With regard to the quality of the included studies, the Q-Genie scores were of medium quality, with an average Q-Genie score of 40.88 (range 38–44).

For a quantitative analysis, at least two studies are required for each polymorphism. For there were few cohort studies on specific loci polymorphisms (such as only one study on CYP17A1 loci -600, CYP17A1 loci +723, CYP17A1 loci +2612, CYP17A1 loci +4259, CYP17A1 loci +4444, CYP17A1 loci +4994, CYP19A1 rs2236722, CYP19A1 rs700518, SHBG rs13894, SHBG rs858521, SHBG rs1799941 and SHBG rs6527; two studies on CYP19A1 rs700519; three studies on SHBG rs727428), to make the conclusions more reliable, we only included loci with no less than five cohort studies for quantitative analysis.

In the current meta-analysis, 15 case-control studies involving 2429 PCOS patients and 2050 control women were included to estimate the relationship between the CYP17A1 (rs743572) polymorphisms and PCOS risk. As shown in Table 2, the CYP17A1 rs743572 polymorphisms was found to be negatively associated with PCOS risk under dominant model (p = 0.017, OR = 0.83, 95%CI 0.72–0.97, I ² = 74.80%, P _{heterogeneity} = 0.000) in the general population. In subgroup analyses by ethnicity, however, no significant association between the CYP17A1 rs743572 polymorphisms and the PCOS susceptibility was found in the Asian (p = 0.054, OR = 0.83, 95%CI 0.68–1.00), Caucasian (p = 0.252, OR = 0.87, 95%CI 0.72–1.09) and mixed (p = 0.148, OR = 0.80, 95%CI 0.60–1.08) population. We evaluated the overall potential publication biases with LFK index = 0.37 and Doi plots showing no asymmetry (Figure 2A).

In the current meta-analysis, five case-control studies involving 1260 PCOS patients and 1030 control women were included to estimate the relationship between the CYP19A1 (rs2414096) polymorphisms and PCOS risk. All of the five studies were conducted in the Asian population (Table 2), no significant association between the CYP19A1 rs2414096 polymorphisms and the PCOS susceptibility was found under dominant model (p = 0.578, OR = 0.87, 95%CI 0.54–1.41, I ² = 95.90%, P _{heterogeneity} = 0.000). We evaluated the overall potential publication biases with LFK index = -0.37 and Doi plots showing no asymmetry (Figure 2B).

In the current meta-analysis, we explored the relationship between the SHBG (rs6529) polymorphisms and PCOS risk. As shown in Table 2, for SHBG rs6529 polymorphisms and the PCOS susceptibility, eight case-control studies involving 1625 PCOS patients and 1313 control women were included, with no significant association found under dominant model (p = 0.752, OR = 0.99, 95%CI 0.94–1.05, I ² = 60.90%, P _{heterogeneity} = 0.012) in the general population. In subgroup analyses by ethnicity, no significant association between the SHBG rs6529 polymorphisms and the risk of PCOS susceptibility was found in the Asian (p = 0.794, OR = 0.98, 95%CI 0.81–1.18) and Caucasian (p = 0.704, OR = 1.01, 95%CI 0.97–1.05) population. The overall potential publication biases were evaluated with LFK index = -0.60 and Doi plots showing no asymmetry (Figure 2C).