Improved Spatial Transcriptomics Clustering with Nested Graph Neural Networks

We introduce a novel approach, STING, for spatial transcriptomic clustering analysis. Unlike existing state-of-the-art techniques that use graph-based neural networks (GNNs) trained on graphs generated from the spatial proximity of tissue locations (or spots), STING incorporates spot-specific related genes. This feature allows STING to better distinguish between clusters and identify meaningful gene-gene relations for knowledge discovery. It is a nested GNN framework that simultaneously models gene-gene and spatial relations. Using the gene expression, we generate a spot-specific gene-gene co-expression graph. We implement an inner GNN for these graphs to generate embeddings for each location. Next, we utilize these embeddings as features in a sample-wide graph generated using spatial information. We implement an outer GNN for this graph to reconstruct the original gene expression data. Finally, STING is trained end-to-end to generate embeddings that capture gene-gene and spatial information, which we input to a clustering algorithm to produce the spatial clusters. Experiments for 26 samples across 7 datasets and 5 spatial sequencing technologies show that STING outperforms the existing state-of-the-art techniques with a 1.58% to 4.07% improvement in the clustering evaluation metric, thus confirming that integrating gene-gene relation information with the clustering task leads to more informative embeddings and better clusters. Furthermore, experiments on a human breast cancer dataset show that STING identifies relevant genes and gene-gene relations, enabling biological hypothesis generation.