Fig. 1

Accuracy and specificity on mRNA promoters. a Violin plots of the area under the ROC curves obtained in the 409 ChIP-seq. Best hit (red), TRAP (blue), DNAshape (green), TFcoop with no expression control (purple), and TFcoop with expression control (orange). ROC curves for Best hit, TRAP and DNAshape were computed in the non expression-controlled case. b Comparison of AUC achieved by TFcoop and DeepSea approach [45]. Comparison was done on 214 ChIP-seq experiments for which the DeepSea server provides predictions. c Intersection between pairs of ChIP-seq experiments associated with TFs identified as cooperating in promoters. These violin plots report the distribution of Jaccard indexes computed between different pairs of Chip-seq experiments. Red, olive and green: for each TF A, we measured the Jaccard index between promoters bound by A and promoters bound by a TF B whose PWM has been selected in the TFcoop model learned for A (cases B = A were not considered). λ_min, λ_1se and λ_3se denote three inclusive sets of parameters of increasing importance (see Methods for details). Purple: for each TF A, we measured the Jaccard index between promoters bound by A and promoters bound by TFs whose PWMs have not been selected in the A model. Blue: for each TF pair A-B identified in [13] (Additional file 1: Figure S1), we measured the Jaccard index between promoters bound by A and promoters bound by B. Samples red, olive, green, blue and purple have been computed from 2796, 1723, 1037, 282 and 14,529 pairs, respectively. d–e Distribution of AUC differences obtained when using a model learned on a first ChIP-seq experiment to predict the outcome of a second ChIP-seq experiment. Different pairs of ChIP-seq experiments were used: experiments on the same TF and same cell type (red), experiments on the same TF but different cell types (yellow), experiments on different TFs but same cell type (light blue), and experiments on different TFs and different cell types (blue). For each pair of ChIP-seq experiment A-B, we measured the difference between the AUC achieved on A using the model learned on A, and the AUC achieved on A using the model learned on B. AUC differences were measured on the non expression-controlled case (d) and on the expression-controlled case (e)