Mapping the topography of spatial gene expression with interpretable deep learning

Abstract

Spatially resolved transcriptomics technologies provide high-throughput measurements of gene expression in a tissue slice, but the sparsity of this data complicates the analysis of spatial gene expression patterns such as gene expression gradients. We address these issues by deriving atopographic mapof a tissue slice—analogous to a map of elevation in a landscape—using a novel quantity called theisodepth. Contours of constant isodepth enclose spatial domains with distinct cell type composition, while gradients of the isodepth indicate spatial directions of maximum change in gene expression. We develop GASTON, an unsupervised and interpretable deep learning algorithm that simultaneously learns the isodepth, spatial gene expression gradients, and piecewise linear functions of the isodepth that model both continuous gradients and discontinuous spatial variation in the expression of individual genes. We validate GASTON by showing that it accurately identifies spatial domains and marker genes across several biological systems. In SRT data from the brain, GASTON reveals gradients of neuronal differentiation and firing, and in SRT data from a tumor sample, GASTON infers gradients of metabolic activity and epithelial-mesenchymal transition (EMT)-related gene expression in the tumor microenvironment.

Interpretable Spatial Gradient Analysis for Spatial Transcriptomics Data

“…Noticeably, several other methods 30–32 also aimed to detect gradated signals in ST data. While these methods focus largely on inference of global spatiotemporal trends from continuous gene expression data, LSGI focuses on detecting interpretable, phenotypically salient gradients factorizable by NMF.…”

Hyperlink prediction via local random walks and Jensen–Shannon divergence

“…We suggest three future extensions. First, there are several definitions of random walks on hypergraphs [14,24] and it is interesting to compare them under the hyperlink prediction task, which can give us more insight into their differences and applicability. Second, the LRW-JS and LRW-GJS perform well for most datasets except for iAF1260b, which is mainly due to the presence of a few hubs and many low-degree vertices.…”

SpottedPy quantifies relationships between spatial transcriptomic hotspots and uncovers new environmental cues of epithelial-mesenchymal plasticity in cancer

“…To delineate biologically meaningful tissue areas using spatial transcriptomics, some of the current analytical methods, such as SpaGCN 14 and BayesSpace 15 focus on unsupervised clustering of gene expression. Approaches like NeST 16 or GASTON 17 take this one step further and incorporate a nested structure or topography metrics to outline hierarchically organised co-expression hotspots aligning with tissue histology. Given that similar cells often cluster together 18,19 , methods that can reliably detect statistically significant clusters are important in reinforcing the accuracy of cell states determined from continuous signatures.…”