Association of sulfotransferase SULT1A1 with breast cancer risk: a meta-analysis of case-control studies with subgroups of ethnic and menopausal statue

Two investigators (Yiwei Jang and Liheng Zhou) independently obtained relevant articles through searches of PubMed, EBSCO and Web of Science databases using the following words: 'sulfotransferase or SULT', 'polymorphism' and 'breast cancer'. Studies had been case-control design and based on SULT1A1 Arg213His polymorphism either alone or in combination with other genes and the language of publication was restricted to English. All of the studies required study design, publication, breast cancer cases, controls selection and genotyping methods. We excluded articles on only breast cancer patients or on healthy persons and one case-series study. In the end, 10362 breast cancer patients and 14250 controls from 16 case-control studies were selected for this meta-analysis.

The following data were collected from each included studies: first authors, year of publications, study population (categorized as Asian, Caucasian, African and others), sources of controls, menopausal status and the number of different genotype in all subjects.

The risk (odds ratio, OR) was used to estimate the association between SULT1A1 polymorphism and breast cancer risk, using Arg/Arg vs His/Arg, Arg/Arg vs His/His, dominant model (Arg/His+His/His vs Arg/Arg) and recessive model (His/His vs Arg/Arg+Arg/His). For each study, the between-study heterogeneity was assessed across by the chi-square based Q statistics and I-square test. Heterogeneity was considered at either a P-value of < 0.50 or I-square > 50% [13]. All of the data from each study use either fixed-effects (Mantel-Haenszel's method) or random-effects (DerSimonian and Laird's method) model according to the heterogeneity result. If there is no between-study heterogeneity, the two methods provide similar results. Funnel plots and Egger's test were used to test the possible publication bias. Sensitivity analyses were performed to estimate the influence of individual studies on the summary effect. For the possible publication bias, we used trim and fill method and fail-safe number to evaluate the influence to the result. In the ethnic population analysis, statistical analysis was performed in Asian, Caucasian, African and other populations. For menopausal status, studies were divided into postmenopausal and premenopausal status. All of the analyses were performed by Stata 10.0 software (Stata Corporation, College Station, TX, USA) and Comprehensive Meta-Analysis software program (version 2.2.034, USA, 2006), using two-sided P values.

Based on the search strategy, 16 studies were selected. There are 8 studies focused on the menopausal status. All of the studies were divided into four ethnic categories: Asian, Caucasian, African and others. The study details are shown in the table 1. The genotype distribution is consistent with Hardy-Weinberg equilibrium but four studies [14‐30]. All of the studies were published from January 2000 to January 2010.

The details of the study characteristics and the ORs we calculated were listed in Table 2. In the dominant model (Arg/His+His/His vs Arg/Arg), there was between-study heterogeneity in the odds ratios (ORs) of the studies (Heterogeneity chi-squared = 30.09 (d.f. = 15), I-squared = 50.2%, P = 0.012), so we used the random-effect model to analyze the data and found that there was no relationship between Arg/His+His/His genotype and the risk of breast cancer (OR = 1.07, 95% CI: 0.97-1.17, P = 0.164). In the recessive model (His/His vs Arg/Arg+ Arg/His), there was no between-study heterogeneity in the odds ratios (ORs) of the studies (Heterogeneity chi-squared = 18.25 (d.f. = 12) I-squared = 34.3%, P = 0.108). Through the fixed-effect model we found that it was no relationship with breast cancer risk (OR = 1.07, 95% CI: 0.97-1.17, P = 0.169). We used random-effect model (Heterogeneity chi-squared = 31.11 (d.f. = 14) I-squared = 55.0%, P = 0.005) to analyze Arg/Arg vs Arg/His (OR = 1.06, 95%CI: 0.95-1.18, P = 0.291) (Fig. 1) and fixed-effect model (Heterogeneity chi-squared = 15.21 (d.f. = 12) I-squared = 21.1%, P = 0.230) to analyze Arg/Arg vs His/His (OR = 1.07, 95%CI: 0.97-1.18, P = 0.197) (Fig. 2), there was no relationship between SULT1A1 and breast cancer risk either. Meanwhile, we analyzed the subgroups of the studies and found that genotype Arg213His increased the risk of breast cancer among postmenopausal women (OR = 1.28, 95% CI: 1.04-1.58, P = 0.019) but not in the premenopausal women (OR = 1.06, 95% CI: 0.88-1.27, P = 0.537) by both M-H method and D-L method. Because of the different heterogeneity results for postmenopausal women (Heterogeneity chi-squared = 20.01 (d.f. = 6) I-squared = 70%, P = 0.003) and premenopausal women (Heterogeneity chi-squared = 0.73 (d.f. = 3) I-squared = 0.0%, P = 0.866), we used both M-H method and D-L method. For all the studies included in the menses subgroup (Heterogeneity chi-squared = 20.74 (d.f. = 10) I-squared = 51.8%, P = 0.023), there was also statistical significance (OR = 1.19, 95% CI: 1.03-1.36, P = 0.017) (Fig. 3). As for the ethnic subgroups, we used fixed-effects to analyze the studies. We found that racial difference influenced the relationship between the polymorphism and the breast cancer risk, especially in Asian women (M-H method, Heterogeneity chi-squared = 0.95 (d.f. = 2) I-squared = 0.0%, P = 0.621, OR = 2.03, 95% CI: 1.00-4.14, P = 0.051) but not Caucasian women (M-H method, Heterogeneity chi-squared = 10.12 (d.f. = 6) I-squared = 40.7%, P = 0.120, OR = 1.02, 95% CI: 0.92-1.13, P = 0.678) (Fig. 4).

We performed the funnel plots and Egger's test to assess the publication bias. As a result there was no publication bias in recessive model (t = 0.16, P = 0.875), Arg/Arg vs His/His model (t = 1.09, P = 0.299), subgroup for population (t = 0.02, P = 0.985) (Fig. 5). But there was publication bias for all population in dominant model (t = 2.82, P = 0.014) (Fig. 6) and Arg/Arg vs Arg/His model (t = 3.21, P = 0.007). This might be a limitation for our analysis because studies with null findings, especially those with small sample size, are less likely to be published. Also there was a publication bias (for postmenopausal women: t = 5.96, P = 0.002) as the result suggested. By using the trim and fill method, we showed that, if the publication bias was the only source of the funnel plot asymmetry, it needed two more studies to be symmetrical. The value of Log OR did not change too much after the adjustment (Fig. 7). Beside that, the fail-safe number of missing studies that would bring the P-value changed was 17. The influence of individual studies on the summary effect estimate was performed by sensitivity analyses on the overall OR (Fig. 8). No individual study affected the overall OR, since omission of any single study made no materially huge difference.