Figure 1.

Meta fold-change obtained from the inverse variance method plotted against -log₁₀(p_{meta, Fisher}) for each gene, counts smoothed with hexbin [27]. Panel A shows the Fisher method meta p-values obtained from a one sided t-test checking against lowered expression of the gene. Panel B shows the Fisher method applied to p-values from a one sided test against a hypothesis of increased expressed. The third panel, C, shows the minimum of the two Fisher method obtained values obtained for each gene plotted against the meta fold-change. Note that although a meta p-value for a lowered expression is indeed associated with a lower meta-fold change, and vice versa, overall extreme values in meta fold-change are not associated with extremely significant p-values. Panel C which combines the two tests does not show the extreme, marked forked structure that is characteristic of a "volcano plot". One of the key features of a nonparametric approach like Fisher's method is that it allows small changes in fold change that are consistent across experiments to rise to statistical significance across studies. It is important to note however, that these meta analysis results show that meta fold-change and meta p-value can be more decoupled than they are in a single study, and that the two methods can give different rank prioritizations of genes that differentially expressed across studies.