SpaceDX: A Bayesian test for localized differential expression in population-level spatial transcriptomics datasets

Niklas Stotzem, Simon Chang, Na Cai, Francesco Paolo Casale

Track: Main track (up to 8 pages)

Abstract:

Spatial transcriptomics allows for the study of gene expression within its spatial context, yet current spatial methods for differential expression require the definition of specific discrete regions of interest across the analyzed sections, which limits their applicability and statistical power. To address this limitation, we introduce SpaceDX, the first framework for spatial differential expression that automatically localizes regions of interest without requiring tissue registration or manual annotations. SpaceDX employs an attention mechanism to detect tissue contexts exhibiting differential gene expression and uses a hierarchical Bayesian framework to overcome the typical challenge of low sample sizes in spatial datasets.
We first applied SpaceDX to a structured mouse brain dataset consisting of Visium sections from 38 animals, comparing stressed and control groups.
Since the brain has well-defined anatomical regions, we could benchmark SpaceDX against traditional differential expression methods that rely on predefined regions, showing a 110% increase in significant gene detection and the automatic localization of regions exhibiting these differences. Next, we tested SpaceDX on a less structured dataset, specifically using sections from patients with inflammatory skin disease, where it successfully identified regions of interest exhibiting differential gene expression, demonstrating its broad applicability.

Submission Number: 56