Hypergraph Neural Networks Reveal Spatial Domains from Single-cell Transcriptomics Data

TL;DR

This paper introduces a hypergraph neural network approach for spatial transcriptomics data that captures complex cell relationships better than traditional GNNs, leading to improved clustering and cell type diversity detection.

Contribution

The novel use of hypergraph neural networks combined with autoencoders for unsupervised spatial domain detection in transcriptomics data.

Findings

Achieved highest iLISI score of 1.843, indicating diverse cell type identification.

Outperformed other methods in clustering with ARI of 0.51.

Outperformed other methods with Leiden score of 0.60.

Abstract

The task of spatial clustering of transcriptomics data is of paramount importance. It enables the classification of tissue samples into diverse subpopulations of cells, which, in turn, facilitates the analysis of the biological functions of clusters, tissue reconstruction, and cell-cell interactions. Many approaches leverage gene expressions, spatial locations, and histological images to detect spatial domains; however, Graph Neural Networks (GNNs) as state of the art models suffer from a limitation in the assumption of pairwise connections between nodes. In the case of domain detection in spatial transcriptomics, some cells are found to be not directly related. Still, they are grouped as the same domain, which shows the incapability of GNNs for capturing implicit connections among the cells. While graph edges connect only two nodes, hyperedges connect an arbitrary number of nodes along their edges, which lets Hypergraph Neural Networks (HGNNs) capture and utilize richer and more complex structural information than traditional GNNs. We use autoencoders to address the limitation of not having the actual labels, which are well-suited for unsupervised learning. Our model has demonstrated exceptional performance, achieving the highest iLISI score of 1.843 compared to other methods. This score indicates the greatest diversity of cell types identified by our method. Furthermore, our model outperforms other methods in downstream clustering, achieving the highest ARI values of 0.51 and Leiden score of 0.60.