Abstract 2130: Uncovering the spatial landscape of tumor-immune interactions using latent spaces from spatial transcriptomics

Recent advances in spatial transcriptomics (ST) enable us to measure gene expression from cancer tissues while retaining their spatial context. We present a novel bioinformatics pipeline to infer molecular changes from tumor and immune cell interactions in the tumor microenvironment (TME) from ST data. Latent space methods enable inference of biological patterns from ST without the need for spot deconvolution into cell-based spatial features. While linear latent space methods yield interpretable biological patterns, interactions between tumor and immune cells can be nonlinear. To enable comprehensive inference of the pathways in the TME, we developed novel algorithms to characterize biological patterns from ST data using linear latent space methods and further nonlinear effects from their interactions. For any given set of genes, the patternSpotter tool visualizes the spatial variation in the relative contribution of individual patterns to the aggregate expression at each location in the tumor sample. Application of this tool to latent features identified using CoGAPS non-negative matrix factorization on a Visium ST (10x Genomics) data from a lymph node with pancreatic cancer metastasis confirms its known immune cell architecture. Furthermore, we develop a patternMarker algorithm to identify sets of coexpressed genes associated with the patterns, which help us to pinpoint the underlying biological processes and cell types. Further analyzing a breast cancer sample with invasive carcinoma and multiple precursor lesions demonstrates that this approach can uncover tumor and immune regions without prior reliance on pathology annotations from H&E imaging. In this case, an ensemble-based factorization of multiple dimensions enhances our resolution of intra-tumor heterogeneity and identifies distinct hormone receptor pathways in different precursor lesions with the patternMarker algorithm. Additional latent features are associated with immune cells, revealing further heterogeneity in immune infiltration between the invasive carcinoma and distinct precursor lesions. Still, the molecular interactions resulting from this infiltration induce a further non-linear alteration to transcription not captured through the inferred latent spaces. To resolve this, we develop a further interactionMarker statistic to identify regions of inter-pattern interaction and the associated genes. We apply this approach to detect additional intra-tumor heterogeneity in immune signaling fro