Classification of genes using clustering of chromatin state segmentations in human epigenomes

Abstract

Combinatorial patterns of chromatin marks have been shown to play a significant role in gene regulation activities by changing the landscape of the DNA through chemical means. Recent work has expanded on this observation using ChIP-seq signals of chromatin marks and supervised algorithms to build gene expression prediction models based on correlation analysis. However, no approach to date has attempted to use chromatin states to identify various classes of genes outside of the high-low expression classes. This research aims to fill this void by utilizing chromatin state segmentation and RNA-seq expression datasets from the NIH Roadmap Epigenomes project. A gene classification model was built using a k-fuzzy clustering approach of chromatin state features from a subset of training genes and then applied to a larger test set of genes. The models were found to be robust and show striking correspondence between training and test sets. 8 classes of genes that represent silent, repressed, and subsets of actively transcribed genes were identified and several metrics to validate the classes were computed. The systematic analysis outlined in this research is shown to a be promising approach for gene classification and future de novo discovery of gene like regions.