Abstract

Spatial transcriptomics has transformed our ability to study gene expression within in-tact tissues, revealing how cellular organization shapes biological function. However, fully realizing its potential requires rigorous statistical and computational modeling. In this talk, I will present some of our recent method developments in modeling spatial transcriptomics data, focusing on three topics. First, I will introduce new developments in reference-free spatial deconvolution, which infers cellular composition from multicellular-resolution data under a latent Dirichlet allocation (LDA) framework. Second, I will describe methods for identifying cell type-specific spatially variable genes (ctSVGs) and temporally-informed SVGs (TSVGs) using kernel mixed-effects models, enabling the discovery of context-dependent transcriptional patterns. Third, I will present a robust nonparametric batch correction-free approach for SVG detection with multi-sample integration. Together, these advances demonstrate how principled statistical modeling can translate experimental complexity into meaningful biological insight. I will conclude by highlighting future opportunities in spatial transcriptomics driven by continued innovations in biotechnology.