A model-based constrained deep learning clustering approach for spatially resolved single-cell data

Abstract

Spatial-resolved scRNA-seq (sp-scRNA-seq) technologies provide the potential to comprehensively profile the gene expression pattern in the tissue context. However, the development of computational methods does not catch up with the fast advances of technologies and fails to fully fulfill their potential. In this study, we develop a deep learning approach for clustering sp-scRNA-seq data, named Deep Spatial-constrained Single-cell Clustering (DSSC). In this model, we integrate the spatial information of cells into the clustering process by two steps: 1) the spatial information is encoded by using a graphical neural network model; 2) cell-to-cell constraints are built based on the spatially expression pattern of the marker genes and added in the model to guide the clustering process. Then, a deep embedding clustering is performed on the bottle-neck layer of autoencoder by Kullback-Leibler (KL) divergence along with the learning of feature representation. DSSC is the first model which can utilize the information from both the spatial coordinates and the marker genes to guide the cell/spot clustering. Extensive experiments on both simulated and real datasets demonstrate that DSSC boosts clustering performance significantly compared to the state-of-art methods. It has a robust performance over different datasets with various cell type/tissue organization and/or cell type/tissue spatial dependency. We conclude that DSSC is a promising tool for clustering sp-scRNA-seq data.

HyperGCN: an effective deep representation learning framework for the integrative analysis of spatial transcriptomics data

“…Generally, high ARI and NMI values indicate good performance. Acknowledge that the spot/cell annotated in the original publications may not be fully accurate, we also used a variant of Moran’s Index to evaluate the clustering performance [ 39 ]. Moran’s Index does not require true labels and is defined as: …”

Systematic evaluation with practical guidelines for single-cell and spatially resolved transcriptomics data simulation under multiple scenarios

“…For the simulated SRT data, the reliability of spatially variable genes (SVGs) was also assessed by the approaches used for the DEGs in simulated scRNA-seq data. Furthermore, we evaluated the simulated SVGs using Moran’s I statistics which are commonly used to quantify the degree of autocorrelation of gene expression in space [ 71 – 73 ]. The expression patterns of SVGs are supposed to exhibit high spatial autocorrelation if the expression values of spots have a strong relationship with those spots near them.…”