Supervised spatial inference of dissociated single-cell data with SageNet

Abstract

Spatially-resolved transcriptomics uncovers patterns of gene expression at supercellular, cellular, or subcellular resolution, providing insights into spatially variable cellular functions, diffusible morphogens, and cell-cell interactions. However, for practical reasons, multiplexed single cell RNA-sequencing remains the most widely used technology for profiling transcriptomes of single cells, especially in the context of large-scale anatomical atlassing. Devising techniques to accurately predict the latent physical positions as well as the latent cell-cell proximities of such dissociated cells, represents an exciting and new challenge. Most of the current approaches rely on an ‘autocorrelation’ assumption, i.e., cells with similar transcriptomic profiles are located close to each other in physical space and vice versa. However, this is not always the case in native biological contexts due to complex morphological and functional patterning. To address this challenge, we developed SageNet, a graph neural network approach that spatially reconstructs dissociated single cell data using one or more spatial references. SageNet first estimates a gene-gene interaction network from a reference spatial dataset. This informs the structure of the graph on which the graph neural network is trained to predict the region of dissociated cells. Finally, SageNet produces a low-dimensional embedding of the query dataset, corresponding to the reconstructed spatial coordinates of the dissociated tissue. Furthermore, SageNet reveals spatially informative genes by extracting the most important features from the neural network model. We demonstrate the utility and robust performance of SageNet using molecule-resolved seqFISH and spot-based Spatial Transcriptomics reference datasets as well as dissociated single-cell data, across multiple biological contexts. SageNet is provided as an open-source python software package at https://github.com/MarioniLab/SageNet.

Quantitative characterization of cell niches in spatial atlases

“…To assess NicheCompass’ performance, we benchmarked it against other spatial embedding methods 51,52,54,61 across diverse datasets and compared the obtained cellular representations and niche labels. In the first phase of this process, we applied all methods on a SlideSeqV2 dataset of the mouse hippocampus ( Methods ), whose spatial tissue organization closely mirrors anatomical structure 1 .…”

“…Building upon previous work 19–22 , we introduce two cell type annotation-based (CAS, CLISIS) and two unsupervised (MLAMI, GCS) metrics to evaluate spatial conservation of the learned latent representations and identified cell niches at global and local scale. All four metrics are applicable in both single sample and sample integration scenarios.…”

“…The cell-level CLISI captures the degree of cell mixing in a local neighborhood around a given cell. We then divide the latent cell-level CLISIs by the physical ones and compute the logarithm to get relative local heterogeneity scores 22 . Subsequently, we normalize these scores by the maximum possible score, compute the median of the absolute values across all cells, and subtract from 1 to arrive at the CLISIS.…”

“…Previous work has adapted existing and designed new methods to address common tasks that have evolved in analyzing spatially resolved omics data as a result 45 ; this includes integrating multiple tissue samples to construct "spatial atlases" (data integration) [46][47][48][49][50][51][52][53] , identifying spatial domains or cell niches in the tissue [46][47][48][49][50][51][53][54][55][56][57][58][59] , and inferring cell-cell communication 60,61 . Currently available methods provide important contributions to the field but are often designed for only a subset of these tasks, cannot efficiently integrate multiple samples, are not scalable, lack inherent interpretability, or only use a single modality.…”