scGAC: a graph attentional architecture for clustering single-cell RNA-seq data

Abstract

Motivation Emerging single-cell RNA sequencing (scRNA-seq) technology empowers biological research at cellular level. One of the most crucial scRNA-seq data analyses is clustering single cells into subpopulations. However, the high variability, high sparsity, and high dimensionality of scRNA-seq data pose lots of challenges for clustering analysis. Although many single-cell clustering methods have been recently developed, few of them fully exploit latent relationship among cells, thus leading to suboptimal clustering results. Results Here, we propose a novel unsupervised clustering method, scGAC (single-cell Graph Attentional Clustering), for scRNA-seq data. scGAC firstly constructs a cell graph and refines it by network denoising. Then it learns clustering-friendly representation of cells through a graph attentional autoencoder, which propagates information across cells with different weights and captures latent relationship among cells. Finally, scGAC adopts a self-optimizing method to obtain the cell clusters. Experiments on 16 real scRNA-seq datasets show that scGAC achieves excellent performance and outperforms existing state-of-art single-cell clustering methods. Availability Python implementation of scGAC is available at Github (https://github.com/Joye9285/scGAC) and Figshare (https://figshare.com/articles/software/scGAC/19091348). Supplementary information Supplementary data are available at Bioinformatics online.

A Meta-learning based Graph-Hierarchical Clustering Method for Single Cell RNA-Seq Data

“…Here we reduced LANDER whole feature dimension with the denoising auto-encoder we pre-trained instead of PCA to the latent shape as clustering centers. [26] 0.4955 0.6476 0.6195 DCA [11] 0.5035 0.6306 0.5989 IDEC [20] 0.4708 0.5888 0.5952 scDeepCluster [18] 0.5600 0.7299 0.6433 Seurat [24] 0.5792 0.7474 0.6765 SIMLR [25] 0.5368 0.6721 0.6518 scGAC [22] 0.6319 0.7268 0.7277 scNAME [23] 0 All the ablation experiments were conveyed on same conditions, and then we measured average performance from formula (18) for 13 normal scale datasets (Table III). From Uniform Manifold Approximation and Projection visualization(UMAP, Fig.…”

“…In Fig. 2, we plot all ARI, NMI and CA [28] 0.4955 0.6476 0.6195 DCA [9] 0.5035 0.6306 0.5989 IDEC [22] 0.4708 0.5888 0.5952 scDeepCluster [20] 0.5600 0.7299 0.6433 Seurat [26] 0.5792 0.7474 0.6765 SIMLR [27] 0.5368 0.6721 0.6518 scGAC [24] 0.6319 0.7268 0.7277 scNAME [25] 0.6447 0.7192 0.7476 MeHi-SCC without meta-learning 0.6031 0.6820 0.7148 performance with violin plots. On each plot, each violin bar includes 13 points indicating 13 normal scale datasets.…”

“…To better reveal inter-cellular relationship, Zeng et al integrated cell-cell relationship with graph neuron network as extra information to optimize dual self-supervised clustering [21]. In addition, Yi et al applied graph attention network enhancement for the same goal and Hui et al brought in contrastive learning to promote robustness [22,23]. As for traditional machine learning based single cell clustering method, Seurat uses log-normalization to impute raw RNA counts and generate clustering labels based on nearest neighbors with high variant gene markers [24].…”

Bioinformatic Methods for Identifying Differentially Abundant Subpopulations in scRNA-seq Data

“…The method utilizes latent relationship information across cells to graphically obtain cell clusters. [45]…”

An optimized graph-based structure for single-cell RNA-seq cell-type classification based on non-linear dimension reduction

“…Furthermore, the large dimensions of the primary data require efficient methods for dimension reduction. In order to overcome these challenges, in recent years, several methods have been proposed for analyzing single-cell RNA sequencing data taking advantages of deep learning approaches [ 5 – 7 ]. Most of these scRNA-seq pipelines consist of three stages: 1) imputation of dropout events, 2) adoption of dimension reduction methods to identify lower-dimensional representations that explain the maximum variance, 3) Clustering of various cells with similar expressions [ 8 ].…”