Leveraging Graph Information for Spatially Informed Patient Data Analysis with GIST

Patient data such as tissue samples analyzed through spatial transcriptomics have transformed our ability to study cellular subpopulations within their native microenvironments, providing unprecedented insights into tissue architecture and cellular interactions. However, accurately identifying spatial domains remains a computational challenge. Despite notable progress, no gold standard currently exists for spatial domain identification, and significant opportunities remain for further improvement.In this study, we introduce Graph Information for Spatial Transcriptomics (GIST), a Graph Neural Network (GNN)-based framework that integrates gene expression data with spatial coordinates to construct a biologically meaningful graph representation of tissue architecture. By explicitly modeling spatial dependencies and leveraging contrastive learning to optimize node embeddings, GIST substantially improves spatial domain identification. It outperforms existing methods on key clustering metrics such as the Adjusted Rand Index (ARI) demonstrating its effectiveness in capturing the true structure of spatial transcriptomic data.Furthermore, we introduce the Silhouette Spatial Score (SSS)—an extension of the traditional Silhouette Score that incorporates spatial neighborhood information. SSS enables more accurate evaluation of both transcriptomic similarity and spatial contiguity within identified domains. GIST outperforms existing methods in SSS, highlighting its ability to identify domains that are not only transcriptomically meaningful but also spatially contiguous.