W219 - Powerful and Accurate Case-control Analysis of Spatially Resolved Molecular Data Using Deep Learning-defined Tissue Microniches

Abstract Text: As spatial molecular data grow in scope and resolution, there is a pressing need to identify key spatial structures associated with disease. However, current approaches often rely on hand-crafted tissue features that may overlook informative signals. We introduce a method, variational inference-based microniche analysis (VIMA), that uses deep learning and principled statistics to discover disease-associated spatial structures with greater flexibility and precision.

VIMA uses a variational autoencoder to define a large number of “microniches” – small groups of tissue patches with highly similar biology spanning multiple samples – and identifies microniches whose abundance correlates significantly with disease. We show using simulations with real spatial transcriptomics datasets that VIMA is more powerful and accurate than simpler approaches.

We applied VIMA to an N=27 immunohistochemistry dataset in rheumatoid arthritis (RA) and an N=75 spatial transcriptomics dataset of 140 genes in Alzheimer’s Disease. In the RA dataset, VIMA identified multiple spatial structures differentially enriched among disease subtypes previously defined using scRNA-seq (P=6.6e-4), including macrophages in contact with sublining fibroblasts, perivascular cellular infiltrates, highly vascularized regions, and CLIC5+ lining fibroblast-rich regions. In the Alzheimer’s dataset, VIMA revealed a previously undescribed positive association between dementia and an oligodendrocyte-rich niche in cortical layer 6, while recapitulating a known negative association between dementia and the thickness of cortical layers 2/3 (P=2.2e-3). These spatial features strongly separated dementia cases from controls, which was not possible using non-spatial sc-RNAseq. Our findings show the promise of spatial technologies for unraveling the complex tissue processes that typify autoimmune disease.