Multi-levels 3D Chromatin Interactions Prediction Using Epigenomic Profiles

Identification of the higher-order genome organization has become a critical issue for better understanding of how one dimensional genomic information is being translated into biological functions. In this study, we present a supervised approach based on Random Forest classifier to predict genome-wide three-dimensional chromatin interactions in human cell lines using 1D epigenomics profiles. At the first level of our in silico procedure we build a large collection of machine learning predictors, each one targets single topologically associating domain (TAD). The results are collected and genome-wide prediction is performed at the second level of multi-scale statistical learning model. Initial tests show promising results confirming the previously reported studies. Results were compared with Hi-C and ChIA-PET experimental data to evaluate the quality of the predictors. The system achieved 0.9 for the area under ROC curve, and 0.86--0.89 for accuracy, sensitivity and specificity.