CMStatistics 2022: Start Registration
View Submission - CMStatistics
Title: Spectral clustering using gene expression and histology identifies disease-relevant spatial domains in SRT Authors:  Kyle Coleman - University of Pennsylvania (United States) [presenting]
Jian Hu - Emory University (United States)
Daiwei Zhang - University of Pennsylvania (United States)
Mingyao Li - University of Pennsylvania (United States)
Abstract: Spatially resolved transcriptomics (SRT) provides an unprecedented opportunity to integrate gene expression with histology and spatial location information when studying the disease. A prominent goal in SRT data analysis is the identification of spatial domains through the clustering of SRT spots. We present SpeCTrE (Spectral Clustering using Transcriptomics and H\&E Histology), a deep learning-based spectral clustering algorithm for the grouping of SRT spots into spatial domains that are distinct with respect to gene expression and histology. SpeCTrE first employs HIPT to extract spot-level histology features while capturing the long-range histological dependencies among spots. The algorithm then constructs two adjacency matrices representing the transcriptional and histological similarities of each pair of spots. The columns of the transcriptomics adjacency matrix are projected onto the top eigenvectors of the normalized Laplacian of the histology adjacency matrix to obtain a modified transcriptomics adjacency matrix containing histological information useful for clustering. Using this updated adjacency matrix and a multilayer perceptron, SpeCTrE obtains spot-level feature vectors that are used as input for the k-means clustering algorithm. Through analyses of SRT datasets from cancerous tissue sections and extensive benchmark evaluations, we show that SpeCTrE outperforms state-of-the-art spatial clustering methods in separating spots into disease-relevant spatial domains.