Figure 4

Bivariate motif combinations that best predict S100a9 coexpression in 30 mouse cell types.

For each cell type, we identified a set of TF motifs for which the number of binding sites in the region 2 KB upstream of transcription start sites was a significant predictor of S100a9 coexpression (P < 0.05). Among these motifs, all possible pairwise combinations were evaluated as predictors of S100a9 coexpression within logistic regression models. The right margin lists the best bivariate model identified for each cell type (Akaike information criterion). For motifs shown in red font, increased motif occurrence was associated with increased probability of S100a9 coexpression (Z > 0), while conversely, for motifs shown in green font, decreased motif occurrence in was associated with increased probability of S100a9 coexpression (Z < 0). 10000 cross-validation simulations were performed to assess the ability of each model to predict S100a9 coexpression. The average AUC among all simulations is plotted in the figure (filled circles), with error bars spanning ± one standard deviation. Yellow boxes for each cell type outline the range of AUC statistics obtained for a univariate null model in which unmasked sequence length was the only predictor variable (mean AUC ± 1 standard deviation). Blue symbols represent cases for which AUC statistic distributions for the null and full models do not overlap, indicating that the combined frequencies of the two motifs yielded a sensitive and specific model for prediction of S100a9 coexpression.