Fig. 2.

HAc data improve motif scanning-based TFBS predictions. (A) Prediction performance (area under the sensitivity versus FPR curve, or ‘ROC’ curve) for models with motif scanning and one additional feature, and a motifs-only reference model (data for models with three features are shown in Supplementary Fig. S4). Larger bar values correspond to better cross-validation-average performance on the test dataset. The performance for the reference model is shown in the blue bar (and vertical dotted line), and a random model is shown as a negative control (black bar). The motifs-only model outperformed the random model, ∼27-fold. Each green bar represents a model that used motif information plus a specific sequence-based feature (GC content, etc.). Each cyan bar represents a model that used motif information plus a HAc ChIP-Seq-based feature (Supplementary Table S3). Each error bar represents the cross-validation-wide SD of the performance difference between the indicated model and the reference model (Section 3). *P < 0.05; ^***P < 0.001. For the cyan bars, a dashed border indicates that HAc data are from LPS-stimulated cells; a solid border means the HAc data were from unstimulated cells. In the top two bar labels, ‘VS’ stands for the ‘valley score’ for local minima in the HAc ChIP-Seq signal. (B) ROC curves, for predictions by the models shown in (A) (see Supplementary Fig. S4 for the complete FPR range). The model with HAc VS (from stimulated cells; gray curve) outperforms the other models. ROC curves were obtained by varying the prediction score cutoff (Section 3). The lack of improvement for the nucleosome occupancy-based model is consistent with the very weak association between this feature and TF binding (Supplementary Fig. S6).