Accounting for Spatial Correlation in Graphical Analysis of SpatialTranscriptomics Data

Abstract

Co-expression analysis is key for understanding disease mechanisms and gene regulatory and functional relationships. In spatial transcriptomics, estimating gene correlation is challenging due to correlation among cells, which can lead to spurious associations that obscure true biological associations. To address this, we propose SpaceDecorr, a method that adjusts gene expression for technical artifacts and spatial dependencies by modeling each gene independently using a Negative Binomial Generalized Additive Model (NB-GAM) with spatial splines. Co-expression is then estimated from the Pearson residuals, yielding decorrelated expression values suitable for downstream analysis. This method targets cell-intrinsic coordination, rather than clustering genes by shared spatial patterns, and supports multi-sample analysis trough independent per-sample adjustment. Across simulations and real datasets, it consistently reduces false-positive correlations and improves the functional coherence of co-expression modules.