Relation equivariant graph neural networks to explore the mosaic-like tissue architecture of kidney diseases on spatially resolved transcriptomics

Abstract

Motivation: Chronic kidney disease (CKD) and acute kidney injury (AKI) are prominent public health concerns affecting more than 15% of the global population. The ongoing development of spatially resolved transcriptomics (SRT) technologies presents a promising approach for discovering the spatial distribution patterns of gene expression within diseased tissues. However, existing computational tools are predominantly calibrated and designed on the ribbon-like structure of the brain cortex, presenting considerable computational obstacles in discerning highly heterogeneous mosaic-like tissue architectures in the kidney. Consequently, timely and cost-effective acquisition of annotation and interpretation in the kidney remains a challenge in exploring the cellular and morphological changes within renal tubules and their interstitial niches.

Results: We present an empowered graph deep learning framework, REGNN (Relation Equivariant Graph Neural Networks), designed for SRT data analyses on heterogeneous tissue structures. To increase expressive power in the SRT lattice using graph modeling, REGNN integrates equivariance to handle n-dimensional symmetries of the spatial area, while additionally leveraging Positional Encoding to strengthen relative spatial relations of the nodes uniformly distributed in the lattice. Given the limited availability of well-labeled spatial data, this framework implements both graph autoencoder and graph self-supervised learning strategies. On heterogeneous samples from different kidney conditions, REGNN outperforms existing computational tools in identifying tissue architectures within the 10× Visium platform. This framework offers a powerful graph deep learning tool for investigating tissues within highly heterogeneous expression patterns and paves the way to pinpoint underlying pathological mechanisms that contribute to the progression of complex diseases.

Availability and implementation: REGNN is publicly available at https://github.com/Mraina99/REGNN.

Figure 1.

Comparison of the tissue architecture and cell type distribution of the brain cortex vs kidney nephron. (A) Brain cortex diagram and ribbon-like cell types distribution of a brain cortex sample (Maynard et al. 2021) on 10× Visium platform. (B) Kidney nephron diagram and mosaic-like cell type distribution of a kidney sample from KPMP (de Boer et al. 2021) on 10× Visium platform. Each color represents a different cell type present in the tissue. For Fig 1A-B, the Brain and Kidney drawing was sourced from BioRender.

Figure 2.

Schema of REGNN model. Take 10× Visium platform as an example, REGNN models the SRT data as a spot-spot graph and learns the embeddings of the data, then infers tissue architecture through clustering. (A) The empowered REGNN contains: (1) Equivariance in AGG operation. (2) Positional Encoding in UPDATE operation. (B) REGNN_GAE as a graph autoencoder architecture built on REGNN, and (C) REGNN_SSL as a graph contrastive learning strategy built on REGNN.

Figure 3.

Performance comparison on ARI in all 23 samples from KPMP. For each method, the median is shown by the solid black line and the median is displayed by the dotted line. The Wilcoxon signed rank test is performed to determine the significance of REGNN_GAE’s mean compared to other competitive methods. REGNN_GAE and REGNN_SSL have no significant difference between their result means.

Figure 4.

Visualization of results from computational methods on a representative CKD sample. The gold standard annotations and calculated results of the computational methods are mapped to the original locations of CKD sample V10S14-085_XY04_21-0057.

Figure 5.

(A) Comparing ablation test results on REGNN_GAE, shown with CKD representative sample V10S14-085_XY04_21-0057. (B) Different Clustering algorithm performance on REGNN_SSL’s graph embeddings on representative CKD sample V10S14-085_XY04_21-0057. (C) REGNN_SSL performance on ARI (Y-axis) compared with the sample’s Moran’s I (X-axis) on samples used in the study within different heterogeneity. (D) All the comparative methods’ performances on ARI (Y-axis) compared with the sample’s Moran’s I (X-axis) on the 23 kidney samples used in the study. Both plots (C) and (D) are fit by linear regression to estimate the general trend across the increase in Moran’s I.