Comparative Analysis of Supervised Cell Type Detection in Single-Cell RNA-seq Data

Recent studies on Single-cell RNA sequencing (scRNA-seq) technology have been widely applied in biological research and drug discovery. Before in-depth investigations of the functionality of single cells for pathological goals, identification of cell types is an essential step. Recently, several unsupervised learning methods have been developed to identify cell types. However, annotating clusters with the correct cell types require considerable efforts using marker genes. Due to the lack of enough annotated datasets, supervised techniques have not been commonly used in scRNA-seq studies. On the other hand, classification methods use feature selection algorithms to improve the prediction accuracy by finding the most informative features among many in high-dimensional datasets. Hence, to automating the process of annotation of clusters of cell types, we can take advantage of classification models. This article evaluated the performance of three state-of-the-art supervised classification methods, namely support vector machine, k-nearest neighbor, and random forest combined with three feature selection methods, namely Chi-squared, information gain, and ANOVA F-value. The results of applying nine combinations of these methods on three standard scRNA-seq datasets show that support vector machine combined with information gain outperforms other combinations of techniques. Moreover, we investigated reference gene sets and found 11 out of 20 highly variable genes in two different Pancreas gene sets to validate our findings. This article sheds some light on the potential use of identifying marker genes to improve the automatic identification of cell types.