Single-cell RNA Clustering Based on ZINB Masked Autoencoder

Due to the high-dimensional and sparse nature of single-cell RNA data, it has been challenging to find an appropriate method for clustering such data. Many approaches have attempted to leverage the fitting capability of neural networks to obtain a low-dimensional representation for subsequent unsupervised clustering but have struggled to achieve satisfactory results. In this paper, we first apply a masking matrix to shuffle the original RNA data before inputting it into a masked auto encoder. We introduce ZINB (Zero-Inflated Negative Binomial) loss on the masked autoencoder and utilize a mask predictor to enable the autoencoder to better identify the inherent noise in the raw single-cell RNA data and provide a lower-dimensional representation of the single-cell RNA data. Finally, clustering can be performed in the low-dimensional space according to the number of cells, using algorithms such as K-means or Leiden. Experiments were conducted on six datasets, comparing our method with several other clustering methods. Various clustering metrics indicate that our approach demonstrates good performance in clustering.