Abstract
Motivation The Hi-C technology has become an efficient tool to measure the spatial organization of the genome. With the recent advance of 1Kb resolution Hi-C experiment, some of the essential regulatory features have been uncovered. However, most available Hi-C datasets are in coarse-resolution due to the extremely high cost for generating high-resolution data. Therefore, a computational method to maximum the usage of the current available Hi-C data is urgently desired.
Results Inspired by the super-resolution image technique, we develop a computational approach to impute the high-resolution Hi-C data from low-resolution Hi-C data using the deep convolutional neural network. We hypothesize that the Hi-C interaction heatmap contains the repeating features, and develop an end-to-end framework to map these features from low-resolution Hi-C heatmap to high-resolution Hi-C heatmap at the feature level. Our approach successfully reconstructs the high-resolution Hi-C interaction map from the low-resolution counterpart, which also proves that the Hi-C interaction matrix is a combination of the regional features. Besides, our approach is highly expandable, and we can also increase prediction accuracy by incorporating ChIA-PET data.
Availability Source code is publicly available at https://github.com/zhangyan32/HiCPlus
Contact jtang{at}cse.sc.edu, fyue{at}hmc.psu.edu