Deep learning cell type classification using nuclear DNA patterns

Abstract

Multicellular organisms comprise various types of cells, which are characterized by gene expression through interactions between chromosomal DNA and nuclear proteins. Many cutting-edge methods have been developed to reveal the three-dimensional organization of chromosomes. The detailed analyses of whole chromosomes have begun to uncover structural features specific to several cell types. Here, we show that cell types are instantly and highly accurately classified using conventional DNA staining and a convolutional neural network (CNN). A high-resolution single slice image of the nucleus is sufficient for the accurate classification of both live and fixed cells, including neurons and non-neural cells. These findings suggest that there may be cell-type-specific features decipherable by deep learning in a thin two-dimensional slice of the nucleus.