Title: Spatially Aware Pathway Enrichment for Spatial Transcriptomics
Presenter Lingzhu Shen, MSE, is a PhD Student in the Biomedical and Health Informatics program.
Spatial transcriptomics (ST) technologies extend single-cell RNA sequencing (scRNA-seq) by preserving the spatial context of gene expression within intact tissues. The integration of molecular profiles with tissue organization enables the study of cellular heterogeneity, local microenvironments, and complex biological processes that cannot be fully resolved with dissociated single-cell assays. While both ST and scRNA-seq generate high-dimensional and noisy expression data, ST introduces additional challenges including severe sparsity, spatial dependence among spots, and the integration of transcriptomic profiles with histological images. These challenges make robust gene- and pathway-level analyses particularly difficult.
In practice, there are no dedicated tools for pathway analysis in ST data. Most studies adopt workflows developed for scRNA-seq. Yet, these methods are not ideally suited for ST: they treat spatially adjacent spots as independent units and neglect histological context.Combined with the instability introduced by zero inflation, these limitations make pathway findings highly sensitive to preprocessing procedures and difficult to reproduce.
To address these issues, we introduce a pathway enrichment method specifically designed for ST data that preserves spatial structure and remains robust under zero inflation. Each spot is first represented by a low-dimensional latent embedding derived from existing graph neural network (GNN) models. Multimodal neighborhoods for each spot are then defined by integrating three complementary sources of similarity: (i) Euclidean distance in the tissue plane to capture spatial proximity, (ii) cosine similarity in latent space to capture transcriptional similarity, and (iii)optional morphological similarity derived from histology images. Within these neighborhoods, statistical models are applied to borrow expression information from similar spots, thereby enhancing spot-level expression estimates while preserving local heterogeneity. Next, from these enhanced profiles, we compute scores for selected gene sets at each spot, generating spatial enrichment heat maps. To evaluate statistical significance, we plan to implement a spatial block bootstrap procedure. By resampling spatially contiguous blocks of spots rather than individual spots, this approach would maintain local spatial correlation while enabling robust p-value estimation for pathway comparisons across domains. We will apply this method to both mouse and human brain datasets to illustrate its performance.
If unable to attend in person in Biomedical Research Building room 105, you may join via Zoom at
Meeting ID: 958 2937 2435
Passcode: 087450