It is notorious that single-cell RNA sequencing (scRNA-seq) data contain a significant number of missing values due to technical variability. The issue of missing values presents a major challenge in scRNA-seq analysis, especially, complicating the identification of cell types via clustering. To address this issue, various methods have been developed to impute the missing data in scRNA-seq clustering. Most methods first impute missing expression values and then cluster scRNA-seq data. However, these approaches often fail to fully exploit the biologically meaningful cluster structures while imputing missing values. In this study, we propose DIC, a deep neural network with the Y-structure that collaboratively imputes and clusters scRNA-seq data. The Y-structure of DIC is formed by an autoencoder with an extra branch attached to its code layer. Therefore, DIC is divided into three modules: a base module (encoder), an imputation module (decoder) and a clustering module (extra branch). The imputation module and the clustering module work together to perform missing data imputation and cell clustering using deeply learned features from the base module. During the model training process, the cluster structure information is used for missing data imputation while the imputation module enhances the clustering performance by generating more accurately recovered missing data. Our experimental results illustrate that DIC is effective in both imputing missing data and identifying cell types.
Jiang et al. (Wed,) studied this question.