Cell Classification Based on Stacked Autoencoder for Single-Cell RNA Sequencing

Abstract

Single-cell transcriptome sequencing (scRNA-seq) allows researchers to investigate cellular heterogeneity in gene expression profiles, identify cell types and predict cell fate at the single-cell level. Analysis of scRNA-seq data can effectively extract unknown heterogeneity and functional diversity of cell populations. Cell classification is one of the most important tasks in scRNA-seq data analysis, which contains cell clustering and classification of cell subtypes. Before assigning a cell type to each cluster, the unsupervised clustering methods look for marker genes for each cluster. These approaches are susceptible to a number of drawbacks in terms of sources of variation, technology, etc. Meanwhile, as more cell subtypes are gradually discovered, cluster-based cell type identification methods have been gradually leaning towards classification-based cell type identification. In this paper, we proposed a new cell classification method based on stacked autoencoder for representation learning (scSAERLs), which enhances the accuracy of classification by learning the feature representation of the data through deep network models. The stacked autoencoder-based classification model employed an unsupervised greedy pre-training learning procedure, which was followed by supervised label-based fine-tuning of the entire classification model. We tested the model in the intra and inter datasets, evaluated its performance with a standard classification metric. Experimental results showed that scSAERLs outperformed other commonly used classification methods in terms of classification accuracy and F1-score.