R software v2

Concentrations of plasma carotenoids are available in Supplementary Table S1. Plasma total carotenoid concentrations (µmol/L of plasma) were defined as the sum of α-carotene, β-carotene, β-cryptoxanthin, lutein, lycopene, and zeaxanthin concentrations. Regressions between normalized DNA methylation levels of all 472,245 CpG sites and total carotenoid concentrations adjusted for the family ID were computed using R software v2.14.1 (R Foundation for Statistical Computing; http://www.r-project.org) [36 ]. Plasma total carotenoid concentrations (independent variable) were used to predict DNA methylation levels (dependent variable). A p-value ≤ 0.05 was used to identify significant associations. Regressions were adjusted for family ID (fixed effect) to account for the familial structure. In order to maintain a more exploratory approach, the regressions were not adjusted for other confounding factors. Choices of linear model and confounding factors were made for comparison purposes with similar study by our group [37 (link)]. Weighted gene correlation network analysis (WGCNA) was performed with methylation levels of 20,687 CpG sites showing a significant association (p-value ≤ 0.05) with total carotenoid concentrations. This allowed evaluating co-methylation similarities only in CpG sites that are associated with carotenoids.

Continuous and categorical variables were compared using Mann-Whitney U and Fisher’s exact tests, respectively. Unsupervised clustering of normalized miRNA levels was performed using classical multidimensional scale (cmdscale) function. Differential analysis of miRNA expression levels was performed using the limma package [29 (link)]. MiRNA levels with fold changes of ≥ 1 (or ≤ -1) and p-values adjusted for false discovery rate (FDR) of ≤ 0.05 were considered significant. The log2(fold change)|>1|is a widely used threshold used to restrict the analysis to miRNA that are at least two fold up or down regulated to identify meaningful and functionally relevant differences in miRNA expressions. FDR-adjusted p-values were calculated using the Benjamini and Hochberg procedure for multiple testing corrections. Heatmaps were generated using the ‘heatmap’ function, where samples and genes were clustered by the hclust function that uses Euclidian distance. Statistical analysis and data visualization were performed using R software V2.15.0 (http://www.R-project.org) and Bioconductor packages.

Continuous variables were initially explored for normal distribution using the Shapiro-Wilk test. As a normal distribution could not be confirmed for most variables, the variables were expressed as median and IQR. Differences between the groups were compared by the t-test or Mann-Whitney U test. Categorical variables were compared using the χ² test, Yates’ correction test or Fisher's exact test. The collinearity was assessed by examining the variance inflation factors, and no collinearity was found between the variables. Survival curves were estimated by the Kaplan–Meier method.
Cumulative incidences of recurrence and ICC-related death were estimated while taking into account the competing risk of non-ICC-related death using a competing risk analysis as defined by Fine and Gray [31 ]. Predictive analysis of variables associated with the cause-specific hazard of recurrence and ICC-related death was done using the univariate and multivariate Cox proportional hazard model [31 ].
All statistical analyses were performed using SPSS 19.0 for Windows (SPSS, Chicago, IL) and R software v. 2.15.3 (R Foundation for Statistical Computing, Vienna, Austria; www.r-project.org). All reported p values were two-sided, and p < 0.05 was considered statistically significant.

Time-to-event data (e.g., OS, PFS) were analyzed using the hazard ratios (HRs) with a value less than one indicating favorable outcomes in elderly patients with methylated MGMT promoter or those with TMZ-containing therapies. If the HR was not directly reported, the value was calculated using the Kaplan-Meier survival curves or the methods reported by Tierney et al [17] (link). For proportions, data were computed using the logit transformation formula.[18] (link) The inverse-variance method was used, and the application of either fixed- or random-effect model was based on between-study heterogeneity. The heterogeneity was estimated using Chi² test and I² statistic (the percentage of the variability in effect estimates that is due to heterogeneity rather than random error), with P_{heterogeneity}<0.1 or I²>40% considered to be statistically significant.
Publication bias was assessed by visual impression, and was confirmed by analytic methods such as Egger’s test [19] (link).
Both a subgroups analysis of studies with similar treatment and an interaction analysis between treatment and MGMT promoter status were implemented. A sensitivity analysis based on the assessment of risk of bias was also performed.
All analyses were done in R software v2.15.3 (R Foundation for Statistical Computing, Vienna, Austria) and Review Manager v5.2 (the Cochrane Collaboration, Software Update, Oxford, UK).

All statistical analyses were performed in R software (V 2.15.3) (R Foundation for Statistical Computing, Vienna, Austria) and IBM SPSS 20.0 (IBM Corporation, USA). Analysis of variance (ANOVA) and multiple comparison analysis was applied to calculate the mean and standard deviation and for statistical tests. The differences of parameter variance were estimated by ANOVA and the Kruskal Wallis rank-sum test based on the distribution of parameter statistics. For post-hoc comparison, Tukey’s honest significant differences tests and Wilcoxon matched-pairs signed-rank test were applied. R (V 3.2.2) was employed to perform heat map, PCoA, and hierarchical clustering. The R language did a comparison of microorganisms through the ‘adonis’ function from R’s ‘vegan’ package. The multiple null hypothesis testing of the prior group was achieved using the ‘anosim’ or ‘adonis’ function from R’s ‘vegan’ package. According to the method of Franzosa et al. [52 (link)], we clustered the differentially abundant features via a custom approach. Then, the correlation analysis of the differentially abundant chemical components with the differentially abundant microorganisms was undertaken (For all clustering analyses, we applied Spearman’s rank correlation as a similarity measure with a threshold of r = 0.7). Unless otherwise stated, the significance level was set at p < 0.05.