scZIGVAE: A Variational Graph Attention Autoencoder Based on the Zero-Inflated Negative Binomial Distribution for Clustering Single-cell RNA-Seq Data

Background: Single-cell RNA sequencing (scRNA-seq) technology has opened new horizons in studying cellular diversity, helping researchers distinguish the gene expression patterns of each cell, identify rare cell types, and explore the dynamics of gene expression in specific cells under different environments. Clustering plays a central role in revealing unknown cell types and downstream analysis of scRNA-seq. However, the high dimensionality, high noise, and common data missing issues in scRNA-seq data significantly limit the performance of clustering. Traditional embedding algorithms often ignore the characteristics of the underlying distribution when dealing with scRNA-seq data. Aims: In this study, we aim to achieve clustering analysis of single-cell RNA sequencing (scRNAseq) data by developing and applying a variational graph attention autoencoder model based on the zero-inflated negative binomial (ZINB) distribution. Methods: Therefore, we propose a scRNA-seq data clustering analysis method, scZIGVAE, which integrates the zero-inflated negative binomial (ZINB) model and variational graph attention autoencoder. It enhances the learning of complex topological structures between cells while modeling missing events. By jointly optimizing the ZINB loss and cell graph reconstruction loss to estimate missing data, scZIGVAE generates cell representations that are more suitable for clustering. Furthermore, through the method of self-optimizing embedded clustering, the clustering centers are iteratively updated to fine-tune the clustering effect of the model further. Results: Extensive testing on twelve datasets from different single-cell RNA sequencing platforms has demonstrated that the scZIGVAE method outperforms current sota clustering techniques. Conclusion: In summary, our research findings demonstrate that by incorporating the Zero-Inflated Negative Binomial (ZINB) distribution strategy into the Variational Graph Autoencoder (VGAE) architecture, we are able to achieve better estimation of missing values during decoding. Furthermore, the utilization of multiple loss constraints on the generated latent representations renders them more conducive to downstream analyses.