SLICE: spatial latent inference of common expression profiles

Spatial transcriptomics has revolutionized our ability to measure gene expression within intact tissues at single-cell and subcellular resolution. In practice, gene expression patterns may occur at different resolutions: some genes follow broad spatial trends across the tissue, whereas others show more localized patterns associated with specific regions, cell populations, or microenvironments. In this talk, I will present SLICE, a Bayesian model that jointly represents spatial and latent sources of variation. The model combines a multiscale spline representation of spatial coordinates with a low-rank latent component and a Bayesian hierarchy for recurrent gene-specific spatial profiles. For scalability, inference uses stochastic variational methods, with automatic differentiation and GPU acceleration. This is joint work with Alex Cecchetto and Francesco Denti.