Deep learning to predict chromatin interactions from RNA-Seq data

Abstract

Chromatin interactions play important roles in gene regulation and expression. Computational methods have been developed to predict chromatin interactions due to the limitations of high-throughput techniques. The availability of large cohorts of RNA-Seq data provides an alternative data source for the prediction of chromatin interactions. We develop a deep learning model, Encoder Chromatin Interaction Neural Network (EnChINN) which predicts chromatin interactions using solely RNA-Seq gene expression information. Gene expression of both chromosome anchors in interest is first extracted from the RNA-Seq data. We then use one-dimensional convolution and transformer encoder to extract relevant features to be used for classification. The results based on four cell lines shows that EnChINN achieves satisfactory performance in predicting chromatin interactions. EnChINN also demonstrates its high generalisability based on its satisfactory across-sample performances and performance based on validation method of chromosome split. Chromatin interactions predicted by EnChINN are able to differentiate AML cancer cell samples from normal cell samples.