A Bayesian Spatial Poisson-Lognormal Model with Pathway-Informed Priors for Detecting Spatially Expressed Genes

Thomas M. Menino Convention & Exhibition Center 

Spatial transcriptomics (ST) enables high-resolution mapping of gene expression across tissues, offering spatial insights into cellular organization, tissue development, disease progression, and treatment response. A key objective in ST analysis is the identification of spatially expressed (SE) genes. Most existing approaches, however, analyze genes independently and therefore fail to account for biologically meaningful gene–gene dependencies. We propose SPHERE (Spatial Poisson Hierarchical modEl with pathway-infoRmed gEne networks), a Bayesian spatial Poisson log-normal model that jointly captures spatial dependence across tissue locations and gene level dependence informed by biological pathways. To address the high dimensionality of ST data, we introduce a pathway-informed conditional autoregressive (CAR) prior that incorporates external biological knowledge to model dependencies among genes within pathways. The proposed hierarchical framework enables the simultaneous detection of clusters of SE and non-SE genes while borrowing strength across related genes. By integrating these localized gene dependencies into a hierarchical spatial framework, SPHERE improves both sensitivity and interpretability in detecting SE genes. Simulation studies and applications to real ST datasets demonstrate that SPHERE achieves higher power and accuracy than existing approaches while providing biologically meaningful insights into gene–gene relationships.

Spatially expressed

Pathway

Conditional autoregressive

Spatial transcriptomics

Dependence

Detection 

Section on Statistics in Genomics and Genetics