Fig. 1

Scan for promoter features associated with gene expression noise. a Gene expression noise is measured using the squared coefficient of variation (CV²) for each gene (subscript i). The relationship between the mean and CV², illustrated by the smoothed scatter plot (blue cloud), is accounted for by calculating an expected value for the gene expression noise for each gene (E[CV²]), shown by the orange line. The residual coefficient of variation rCV², or the mean-adjusted gene expression noise, is calculated as the absolute deviation of the observed CV² from its expected value, shown in the kernel density plot (right panel). The influence of genomic features is tested by fitting a robust linear model to prevent outlier points biasing our results, using the rCV² as the dependent variable. Each feature is regressed on rCV² individually (univariate model: X is a vector of values representing the genomic feature; multivariate model: X denotes a matrix where each column is a genomic feature and the rows are genes). The statistical significance is determined by testing the null hypothesis that the genomic feature regression coefficient, β, is equal to 0 using a t-test. Univariate and multivariate model fitting results are visualised side-by-side for all genomic features, as in the toy example (bottom right panel). Points above 0 (orange) are associated with greater noise, whilst those below the line are associated with lower noise (purple). b Static promoter features were regressed on expression noise (rCV²) in mESCs and mouse Cd4+ T cells (Additional file 2: Figure S1). Each point represents a genomic element in either a univariate (diamond) or multivariate (circle) robust linear regression model. Grey points denote features in which there is insufficient statistical evidence to reject the null hypothesis of no association with expression noise (p<0.05). c Examples of how gene expression noise differs between CGI and non-CGI promoters in both mouse embryonic stem cells and Cd4+ T cells. The y-axes in the plot of rCV² are truncated for clarity. d Genomic features across multiple cell-type lineages and between mouse and human that are consistently associated with transcriptional noise in a multivariate robust linear model. Cell types are denoted by different symbols; human- and mouse-derived cells are delineated by colour (blue for human and red for mouse). Transparent points are those where there is insufficient statistical evidence to reject the null hypothesis (p<0.05). The y-axis ranges in (b) and (d) are truncated for clarity. mESC mouse embryonic stem cell