Predicting CTCF cell type active binding sites in human genome

Abstract

The CCCTC-binding factor (CTCF) is pivotal in orchestrating diverse biological functions across the human genome, yet the mechanisms driving its cell type-active DNA binding affinity remain underexplored. Here, we collected ChIP-seq data from 67 cell lines in ENCODE, constructed a unique dataset of cell type-active CTCF binding sites (CBS), and trained convolutional neural networks (CNN) to dissect the patterns of CTCF binding activity. Our analysis reveals that transcription factors RAD21/SMC3 and chromatin accessibility are more predictive compared to sequence motifs and histone modifications. Integrating them together achieved AUPRC values consistently above 0.868, highlighting their utility in deciphering CTCF transcription factor binding dynamics. This study provides a deeper understanding of the regulatory functions of CTCF via machine learning framework.

Keywords: CTCF binding site; Chromatin accessibility; Convolutional neural networks; RAD21; SMC3.

Fig. 1

The overview of cell type-active CBSs prediction model. (A) The construction of the positive and negative sets of CTCF cell type-active binding site. (B) The feature extraction process, detailing both single and combined features used for model training. (C) The classifier is constructed based on motif and epigenetic signals applying the CNN methodology.

Fig. 2

Chromatin accessibility signals is predictive for CBS. (A) Chromatin accessibility profiles (DNase-seq) in 21 cell lines, showing the distribution of DNase I hypersensitivity signals around CTCF binding sites (CBSs). (B) AUPRC values for predicting CBSs using DNase I signal features across 59 cell lines. The bar graph displays the performance of the predictive model, with the black line indicating the proportion of CTCF binding sites located within open chromatin regions.

Fig. 3

Analysis of histone modification signals in CBS prediction. (A) Distribution of 12 histone modification signals in the K562 cell line. The profiles show the enrichment of histone modifications around CBSs. (B) Distribution of H2AFZ signals in 12 different cell lines. (C) AUPRC values for predicting CBSs using 12 histone modifications across 13 cell lines. The bar graph depicts the performance of the predictive model for each histone modification, demonstrating the effectiveness of different histone marks in identifying CBSs.

Fig. 4

Analysis of RAD21 and SMC3 in CBS prediction. (A) The enrichment of RAD21 binding signals around CBSs in 9 cell lines. (B) Distribution of SMC3 binding signals in 4 cell lines. (C) The AUPRC values of prediction with the binding signal of RAD21 and SMC3.

Fig. 5

Analysis of CTCF motif in CBS prediction. (A) The motifs sourced from JASPAR and HOCOMOCO. (B) The common motif derived from peaks shared across 67 cell lines (abbreviated as Com_motif). (C) The motifs derived from peaks unique to one cell line (abbreviated as Uni_motif). (D) The AUPRC values of prediction with the motif scores in 33 cell lines.

Fig. 6

Accurately predicted peaks (N₁, green), mis-predicted peaks (N₂, blue), and inherent unpredictable peaks (N₀, yellow) in 8 cell lines.