HybridSVG: Ensemble Framework for Detecting Spatially Variable Genes in Spatial Transcriptomics using Fusion of Global and Local Autocorrelation

Abstract

Spatial transcriptomics enables the study of gene expression within intact tissue architecture, providing insights into cellular organization and tissue-specific molecular processes. Accurately identifying spatially variable genes is critical for understanding how gene expression patterns correspond to anatomical structures and functional domains. Here, we present HybridSVG, an ensemble framework that integrates global and local spatial autocorrelation measures to detect genes with diverse spatial expression patterns. We applied HybridSVG to publicly available mouse brain and human dorsolateral prefrontal cortex datasets generated using the 10x Genomics Visium platform. Our results demonstrate that HybridSVG captures both broad tissue-wide gradients and fine-grained spatial niches with greater precision than existing methods, including SpatialDE, scGCO, and SPARK-X. Genes identified by HybridSVG show stronger spatial autocorrelation and clearer local clustering, reflecting biologically meaningful structures such as cortical layers and subcortical domains. These findings highlight the ability of HybridSVG to robustly identify spatially variable genes across species and tissue types, providing a versatile tool for the analysis of complex spatial transcriptomics data.