Integrating spatial transcriptomics data across different conditions, technologies and developmental stages

Abstract

With the rapid generation of spatial transcriptomics (ST) data, integrative analysis of multiple ST datasets from different conditions, technologies, and developmental stages is becoming increasingly important. However, identifying shared and specific spatial domains across ST datasets of multiple slices remains challenging. To this end, we develop a graph attention neural network STAligner for integrating and aligning ST datasets, enabling spatially-aware data integration, simultaneous spatial domain identification, and downstream comparative analysis. We apply STAligner to the integrative analysis of ST datasets of the human cortex slices from different samples, the mouse olfactory bulb slices generated by two profiling technologies, the mouse hippocampus tissue slices under normal and Alzheimer's disease conditions, and the spatiotemporal atlases of mouse organogenesis. STAligner efficiently captures the shared tissue structures across different slices, the disease-related substructures, and the dynamical changes during mouse embryonic development. Additionally, the shared spatial domain and nearest neighbor pairs identified by STAligner can be further considered as corresponding pairs to guide the threedimensional reconstruction of consecutive slices, achieving more accurate local structure-guided registration results than the existing method.

DenoiseST: A dual-channel unsupervised deep learning-based denoising method to identify spatial domains and functionally variable genes in spatial transcriptomics

“…In the 12 benchmark datasets, each comprises three batches of slices. Batch effects across different datasets may obscure genuine biological signals 51 , with noise arising from disparate times, different processing personnel, and technological platforms, resulting in significant variations or batch effects. These effects can be either linear or nonlinear, making them challenging to distinguish from biological variability.…”

Navigating the landscapes of spatial transcriptomics: How computational methods guide the way

“…This requires further methodological improvement. Using mutual nearest neighbors (MNN) as a way to determine anchors, graph deep learning techniques have shown impressive performance in solving this problem (Guo et al, 2023;Xia et al, 2023;Zhou et al, 2023). For example, STAligner (Zhou et al, 2023) integrates a graph attention autoencoder with MNN methods to achieve spatially aware data integration and remove batch effects by spot triplet learning.…”

“…Using mutual nearest neighbors (MNN) as a way to determine anchors, graph deep learning techniques have shown impressive performance in solving this problem (Guo et al, 2023;Xia et al, 2023;Zhou et al, 2023). For example, STAligner (Zhou et al, 2023) integrates a graph attention autoencoder with MNN methods to achieve spatially aware data integration and remove batch effects by spot triplet learning. Similarly, SPIRAL (Guo et al, 2023) uses GraphSAGE (Hamilton et al, 2017) and domain adaptation to learn representations of transcriptome data and align different batch coordinates using cluster-aware Gromov-Wasserstein distance.…”

Graspot: A graph attention network for spatial transcriptomics data integration with optimal transport

“…Moreover, nonlinear embedding based methods are developed, such as GPSA [13] and STAligner [20]. GPSA [13] integrates multiple ST slices into a common coordinate system using deep Gaussian processes, while STAligner [20] embeds gene expression and spatial neighbor network of spots with a graph attention network and aligns slices using the mutual nearest neighbor (MNN) originally developed for single-cell data integration. However, none of them are capable of handling the partial alignment of slices or providing the probabilistic mapping of spots in adjacent slides for downstream analysis, as achieved by PASTE and PASTE2.…”