An integrative approach for fine-mapping chromatin interactions

Abstract

Motivation: Chromatin interactions play an important role in genome architecture and gene regulation. The Hi-C assay generates such interactions maps genome-wide, but at relatively low resolutions (e.g. 5-25 kb), which is substantially coarser than the resolution of transcription factor binding sites or open chromatin sites that are potential sources of such interactions.

Results: To predict the sources of Hi-C-identified interactions at a high resolution (e.g. 100 bp), we developed a computational method that integrates data from DNase-seq and ChIP-seq of TFs and histone marks. Our method, χ-CNN, uses this data to first train a convolutional neural network (CNN) to discriminate between called Hi-C interactions and non-interactions. χ-CNN then predicts the high-resolution source of each Hi-C interaction using a feature attribution method. We show these predictions recover original Hi-C peaks after extending them to be coarser. We also show χ-CNN predictions enrich for evolutionarily conserved bases, eQTLs and CTCF motifs, supporting their biological significance. χ-CNN provides an approach for analyzing important aspects of genome architecture and gene regulation at a higher resolution than previously possible.

Availability and implementation: χ-CNN software is available on GitHub (https://github.com/ernstlab/X-CNN).

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

The structure of the CNN in χ-CNN. A data matrix from an interacting locus is passed through an encoding layer, convolutional layer, global max-pooling layer, a dense layer and finally, a logistic regression layer. The encoder, convolutional and dense layers use a ReLu activation function

Fig. 2.

An example of a fine-mapped peak. The left and right sides correspond to the two sides of an interaction. The top images show tracks for H3K4me3, H3K27ac, H3K36me3, H3K9me3 and CTCF. The bottom images show χ-CNN’s fine-mapping score for each position in the region (in kilobases). There is a sharp peak on the left corresponding to a CTCF peak, and in the right region, χ-CNN assigns the highest importance score to one of three CTCF peaks

Fig. 3.

Distribution of fine-mapping predictions for different size HiCCUPs peaks. Kernel density estimation (KDE) plots showing the distribution of χ-CNN’s fine-mapping predictions within K562 peaks after extending the original peak equally in both directions to form a 25 kb peak. To generate plots, we used the ‘jointplot’ function with the KDE option in Python’s Seaborn package. (a) For 5 kb interaction peaks extended to 25 kb, fine-mapped positions are strongly concentrated around the original 5 kb peak (center blue box). Enrichment in center 5 kb bin is 8.3-fold compared to random guessing. (b) For 10 kb peaks extended to 25 kb, fine-mapped positions are concentrated in the original 10 kb peak (center blue box). Enrichment in center 5 kb bin is 4.2-fold. (c) Fine-mapped positions are not concentrated in any specific region in interactions called at 25 kb. Enrichment in center 5 kb bin is 1.6-fold. The positive direction on the axes points toward the exterior of the interactions. The mode of the 5 kb peak plot is shifted toward the positive direction, meaning that fine-mapped peaks are most likely to be approximately 1 kb further out than the center of the originally called peak. Similar plots for GM12878 can be found in Supplementary Figure S3. (Color version of this figure is available at Bioinformatics online.)

Fig. 4.

The 5 kb peak fine-mapping performance for χ-CNN and baseline methods. Fine-mapping performance using individual features is marked with points, and methods integrating multiple features are emphasized with horizontal bars. Light blue points and the dark blue bar correspond to ‘primary’ histone marks and χ-CNN trained on these marks, respectively. Similarly, light green points and the dark green bar correspond to ‘secondary’ marks. CTCF, in lavender, performs well, but χ-CNN trained on ‘secondary’ marks and CTCF performs better. Cohesion sub-units, in pink, are the best performing single marks; however, χ-CNN trained on all features, in red, shows greater enrichment than any individual mark. All other TFs, in orange, perform similarly to histone marks. Finally, a baseline method of averaging all features is marked with a brown bar. (Color version of this figure is available at Bioinformatics online.)