Transcription factor binding predicts histone modifications in human cell lines

Abstract

Gene expression in higher organisms is thought to be regulated by a complex network of transcription factor binding and chromatin modifications, yet the relative importance of these two factors remains a matter of debate. Here, we show that a computational approach allows surprisingly accurate prediction of histone modifications solely from knowledge of transcription factor binding both at promoters and at potential distal regulatory elements. This accuracy significantly and substantially exceeds what could be achieved by using DNA sequence as an input feature. Remarkably, we show that transcription factor binding enables strikingly accurate predictions across different cell lines. Analysis of the relative importance of specific transcription factors as predictors of specific histone marks recapitulated known interactions between transcription factors and histone modifiers. Our results demonstrate that reported associations between histone marks and gene expression may be indirect effects caused by interactions between transcription factors and histone-modifying complexes.

Keywords: epigenetics; gene regulation.

Fig. 1.

Histone modifications can be predicted from DNA sequence. (A) Representative ROC curves of the performance of k-mer LR-based classifiers for histone modifications at gene promoters in H1 cells. The AUC for each task is indicated in the legend. The ROC curves shown are for a single iteration of a 70–30 split of the data. (B) H3K4me3 profile at test set promoters in H1 cells. Shown on the Left is the mean H3K4me3 profile at promoters predicted to be positive (green) and negative (red) for H3K4me3 in a single iteration of the analysis. The cutoff used was P = 0.5. The panel on the Right shows the profile at all of the promoters in the test set ordered by their predicted probability of being marked by H3K4me3 (white, low; red, high).

Fig. 2.

Histone modifications can be predicted from TF-binding data. (A) Representative ROC curves of the performance of TF ChIP-Seq LR-based classifiers for histone modifications at gene promoters in H1 cells. The AUC for each task is indicated in the legend. The ROC curves shown are for a single iteration of a 70–30 split of the data. (B) TF-binding prediction outperforms DNA sequence. Shown is a scatter plot comparing the AUCs achieved from TF-binding LR classifiers (y axis) and DNA sequence LR classifiers (x axis). Each point represents the mean of 10 computational experiments for one histone mark in one cell line. (C) H3K9ac profile at test set promoters. Shown on the Left is the mean H3K9ac profile at test set promoters predicted to be positive (green) and negative (red) for H3K9ac. Both predictions were performed on the same set of promoters. The dashed lines are predictions from sequence, and the solid lines are predictions from TFs. Notice the higher average of TF predicted positive marks. On the Right are heat maps of H3K9ac levels (white, low; red, high) at the individual promoters ordered by predicted positive probability (increasing along the y axis) as provided by sequence LR (Center) and by TF LR (Right).

Fig. 3.

TF-binding data can predict histone marks across cell lines. Shown are ROC curves for TF-binding LR classifiers for H3K4me3. Classifiers were trained on the cell line indicated by the row and tested on each of the three cell lines (indicated by the column). The AUC is given on the plot in each case.

Fig. 4.

TF weights uncover histone modifier–TF interactions. Heat map of LR weights from the prediction of histone modifications from TF-binding data in H1 cells. Each cell represents the weight assigned to a particular TF in predicting the occurrence of a particular histone mark. Both the rows and columns were subject to hierarchical clustering.