Eligible studies were identified by searching the database of PubMed, ISI Web of Knowledge, Elsevier ScienceDirect, and HuGE Navigator Web server http://www.hugenavigator.net27 for relevant reports in English published before May 2009 using the following search terms: "MTHFR" or "methylenetetrahydrofolate reductase" and "liver cancer" or "hepatocellular carcinoma." The reports and dissection database published in the Chinese Biomedical Database (CBM), China National Knowledge Infrastructure (CNKI), and Wan Fang (Chinese) database were also searched to collect articles of case-control studies or cohort studies on associations between MTHFR polymorphisms and susceptibility to HCC before May 2009. Studies were selected if there were available data for MTHFR C677T polymorphism with the risk of HCC using a case-control or cohort design. The reference lists of retrieved articles were also reviewed to identify additional articles missed by the above search. Studies that determined the distribution of the C677T genotype in cases with HCC diagnosed by histopathological biopsy or by elevated α-fetoprotein (AFP) and distinct iconography changes (CT, MRI, and B ultrasonography), and in controls free of cancer were eligible for inclusion in the meta-analysis. Review articles, case-only articles, and repeated literatures were excluded.

Seven articles (which included eight studies) 19202122242526 and one unpublished Chinese research used extra blood samples to extract genome DNA, of which seven studies used polymerase chain reaction-restriction fragment length polymorphism (PCR-RFLP) to genotype, and one article 25 (which contained two studies) used fluorogenic 5'-nuclease assay (TaqMan Assay) to genotype. One study 23 published in Chinese extracted DNA from HCC tumor tissue and peritumoral tissue for genotyping using the PCR-RFLP method.

The following information was extracted from each study: first author's surname, year of publication, ethnicity of study population, country where the study was conducted, genotyping method, and the number of cases and controls for each C677T genotype. When specific results were not directly reported, available tabular data were used to calculate them.

In order to compare the odds ratio (OR) on the same baseline, crude OR was used for the meta-analysis. Following the genetic model analysis methods suggested by Thakkinstian A et al. 28, the wild-type allele was first set as A and the variant allele as B. Then meta-analysis examined the association for the allele contrast BB vs. AA (OR1), AB vs. AA (OR2), and BB vs. AB (OR3). Next, the best genetic model was determined according to the relation of the values of OR1, OR2, and OR3. If OR1 = OR3 ≠ 1 and OR2 = 1, then a recessive model is suggested. If OR1 = OR2 ≠ 1 and OR3 = 1, then a dominant model is suggested. If OR1 = 1, OR2 = 1, and OR3 ≠ 1, then a complete overdominant model is suggested. If OR1>R2>1 and OR1>OR3>1 (or OR1<OR2<1 and OR1<OR3<1), then a codominant model is suggested. The effect of association was indicated as crude OR with the corresponding 95%CI. The pooled OR was estimated using the fixed-effect (FE) model or the random-effect (RE) (DerSimonian and Laird) model 29. The heterogeneity between studies was tested using Q statistic 30. If p < 0.10, the heterogeneity was considered statistically significant, and then the RE model was used. Heterogeneity was quantified using the I²metric, which is independent of the number of studies in the meta-analysis (I²<25%, no heterogeneity; I² = 25-50%, moderate heterogeneity; I²>50%, large or extreme heterogeneity) 31. A cumulative meta-analysis was carried out to evaluate the trend of pooled OR. The potential publication bias was tested using the Egger regression test asymmetry and Begg's test for funnel plot. The distribution of the genotypes in the control group was tested for Hardy-Weinberg equilibrium using a goodness-of-fit chi-square test.

All analyses above were conducted using STATA, version 10.0, software (Stata Corp., College Station, Texas). All p values were two sided. A p value less than 0.05 was considered statistically significant.

The false positive report probability (FPRP), the probability of no true association between a genetic variant and disease given a statistically significant finding, depends not only on the observed p value but also on both the prior probability that the association between the genetic variant and the disease is real and the statistical power of the test. An Excel spreadsheet to calculate FPRP is included with the online material (see http://jncicancerspectrum.oupjournals.org/jnci/content/vol96/issue6) 32.

The statistical power was calculated using the PS software http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/PowerSampleSize3334. Given that the gene mutation was regarded as causal, we used population-attributable risk (PAR) to refer to the proportion of disease risk in a population that can be attributed to the causal effects of the risk gene. PAR can be assessed by using the formula 35: PAR (%) = p(OR-1)/(p(OR-1)+1) × 100%, where p is the proportion of the individuals exposed to risk gene in the general population, and OR is the pooled OR when cases and controls were compared in the risk model.

All together, eight articles (which included nine studies) published in English and one Chinese graduate paper (not in the list of references) met the inclusion criteria. One article 25 contained two different ethnic populations, so it was regarded as two studies. In total, ten studies (1814 cases/2862 controls) examined C677T polymorphism were included in the meta-analysis (Table 1).

All studies were published between 2004 and April 2009. In all studies, the cases were histologically confirmed or diagnosed by elevated AFP and distinct iconography changes (CT, MRI, and B ultrasonography). The controls were free of cancer. All ten studies reported the use of healthy controls or non-liver disease, non-cancer patients as controls. Four studies 19242636, on the other hand, included chronic liver disease (CLD) patients as controls. CLD patients were mainly liver cirrhosis (LC) patients. Liver cirrhosis was either histologically proved or diagnosed on concordant clinical, biological, and morphological criteria.

Studies were conducted in various populations of different ethnicity: Five studies were conducted in populations of Asian ethnicity 20222325 (one Chinese paper not in the list of references), and another five studies involved Europeans or Americans 1921242526. The genotype distributions of the control groups in two studies 2526 did not conform to the HWE equilibrium (p = 0.03 and 0.001, respectively).

In all 10 studies, the T allele frequency of C677T was 0.37 in the control group (2106/5724) and was 0.35 in the case group (1255/3628). The T allele was more likely found in the control groups (p = 0.03). The T frequency in the European controls of five studies was 0.38 (981/2588), and that in the Asian controls of five studies was 0.36 (1125/3136). This showed that there is no inter-ethnicity difference in minor allele frequency (p = 0.11).

The genetic model analysis of 10 studies (including all controls) showed pooled OR1 (TT/CC) = 1, pooled OR2 (TC/CC) = 1, and pooled OR3 (TT/CT)>1, indicating that a complete overdominant model was applicable. FE OR3 reached 1.20 (95%CI: 1.00-1.45, p for heterogeneity = 0.21) (Figure 1), showing that an overdominantly acting protective allele was applicable, that is, heterozygotes are at a lesser risk of HCC than either homozygotes (CC or TT). However, when two studies 2526 that were not in agreement with the HWE equilibrium were excluded, the association was no longer significant (pooled OR3 = 1.18, 95%CI: 0.97-1.43, p for heterogeneity = 0.39), suggesting that the result was not stable.

The meta-analysis results under three conditions, namely, dominant model, recessive model, and allele frequency contrast (T vs. C), including 10 studies, all failed to reach statistic significance. To illustrate, RE OR reached 1.02 (95%CI: 0.85-1.23, p for heterogeneity = 0.07) for the dominant model, 1.22 (95%CI: 0.92-1.61, p for heterogeneity = 0.02) for the recessive model, and 1.08 (95%CI: 0.92-1.27, p for heterogeneity = 0.00) for the T vs. C case (Table 2). The meta-regression and sensitivity analyses found that two studies 2126 were the main sources of heterogeneity. Therefore, heterogeneity was reduced significantly after excluding these two studies, without substantially altering the meta-analysis results.

Stratified analyses were conducted according to the control collection (CLD patients or non-liver disease controls) and regions (Europe or Asia). The meta-analysis of four studies (including 329 HCC cases and 625 CLD patients) showed a positive association of 677TT with HCC in CLD patients: pooled FE OR3 (TT/CT) reached 1.81 (95%CI: 1.22-2.71, p for heterogeneity = 0.25) (Figure 2), indicating that a complete overdominant model was still applicable. Meanwhile, the recessive model RE OR of four studies reached 1.85 (95%CI: 1.00-3.42, p for heterogeneity = 0.06). FE OR of four studies reached 0.75 (95%CI: 0.57-0.99, p for heterogeneity = 0.92) as CT heterozygotes was compared with (CC+TT) between cases and CLD patients of four studies. Table 3 showed the results of the meta-analysis of four studies when CLD patients were compared with the cases. It is worth noting that all four studies including CLD patients as controls were from European populations, so the result was only observed from European CLD patients.

Subgroup meta-analysis of (1) the comparison between the healthy controls and HCC cases of 10 studies; (2) five studies of European populations and (3) five studies of Asian populations all showed that 677T allele had a non-significantly increased risk of HCC and had a non-significant trend of overdominant model to HCC occurrence (detailed data not shown).

Publication bias analyses of 10 studies including all controls and four subgroup studies that used CLD controls both showed no publication bias: Egger's p reached 0.25-0.84 under various models.

However, the sensitivity analysis showed that the association between C677T and HCC (TT vs. CT) of 10 studies (including all cases and controls) was vulnerable: when someone studies was omitted at a time, the 95%CI of the OR3 (TT/CT) model would include 1.0 (Figure 3). Further, the sensitivity analysis showed that the association between C677T and HCC (TT vs. CT) was also vulnerable when CLD patients were used as controls (Figure 4). This is the same as the result in the recessive model (figure not shown).

When meta-analysis was performed in four studies using CLD patients as controls, the cumulative meta-analysis of the recessive model and the OR3 (TT/CT) of C677T showed a stable increased risk trend of HCC occurrence as evidences accumulated (figure not shown).

Because most HCC occurred in LC patients, MTHFR C677T polymorphism may have important clinical values. We further analyzed the reliability of the result in other ways below.

In four studies including CLD patients (mainly LC patients) as controls, there were 329 HCC cases and 625 patients. The proportion of CLD patients with TT or CC genotype was 0.53, and FE OR (TT+CC vs. CT) was 1.33. Then, the statistical power calculated by using the PS software was 0.55.

A statistically significant association (p < 0.05) can be further evaluated by estimating the false-positive report probability (FPRP) 32 at a given prior probability and statistical power. If the prior probability (incidence of HCC among LC patients) was set to 0.30, and OR = 1.33 (95%CI: 1.01-1.76) (TT+CC vs. CT), then FPRP equaled 0.18(<0.20), indicating that the association was noteworthy.

In CLD patients, PAR can be assessed by using the formula 35: PAR (%) = p(OR-1)/(p(OR-1)+1) × 100%, where p is the proportion of the individuals exposed to risk gene (proportion of CC and TT genotype was 0.53 in the CLD patients), and OR is the pooled OR when cases and controls were compared in the risk model (OR = 1.33). Then, PAR was 14.88% (95%CI: 1.22%-28.54%).