Supplementary Materials: Differential meta-analysis of RNA-seq data from multiple studies

In this work, we filter weakly expressed genes using the HTSFilter Bioconductor package, which implements a databased filtering procedure based on the calculation of a global Jaccard similarity index among biological replicates for read counts arising from replicated transcriptome sequencing (RNA-seq) data; see Rau et al. (2013) and the HTSFilter vignette for additional details. This technique provides an intuitive data-driven way to filter RNA-seq data and to effectively remove those genes that contribute to a peak of raw p-values close to 1, due to the discretization of p-values from conditional tests (such as the Fisher’s exact test) for small counts. This latter point is particularly important for the p-value combination methods (Inverse Normal and Fisher) investigated in the main paper, as both rely on an assumption of uniformly distributed p-values under the null hypothesis. Briefly, the HTSFilter method seeks to identify the threshold that maximizes the filtering similarity among replicates (as measured by the Jaccard similarity index), that is, one where most genes tend to either have normalized counts less than or equal to the cutoff in all samples (i.e., filtered genes) or greater than the cutoff in all samples (i.e., non-filtered genes). The data-based filter is chosen by examining the behavior of the global Jaccard index (see Supplementary Figure 10), and identifying the cutoff that corresponds to the maximum global Jaccard index. For both the real and simulated data, we note that for individual per-study analyses of differential expression (and consequently, for the p-value combination methods), data filters are applied independently to each study (e.g., the left and middle panels of Supplementary Figure 10) following estimation of library sizes and dispersion parameters, meaning that it is possible for a gene to be filtered in one study and not in another. For the DESeq approaches, both with and without a fixed study effect, a data filter is applied to all studies simultaneously (e.g., right panel of Supplementary Figure 10) following estimation of library sizes and dispersion parameters. To apply the filter, genes