Nature Biomedical EngineeringSource publication:
Loading content…
Spatial transcriptomics analysis has long clustered at a single scale, so cells, neighboring cell groups and tissue regions were handled separately. SpatialZoomer applies a set of spectral graph low-pass filters to extract spatial molecular features at the single-cell, niche and domain scales from one dataset, and locates critical scales by partitioning a cross-scale similarity map. In Alzheimer's disease and tumor microenvironment data it identified disease-progression signals at specific scales and spatial context-dependent cell subtypes, matching or exceeding existing methods on single-scale tasks with lower compute and hardware demands.
One spatial transcriptomics dataset can now yield spatial molecular features at the single-cell, niche and tissue-domain scales at once, instead of building a separate model for each scale. Prior analyses typically fixed the task at a single scale, treating cell typing, niche characterization and domain segmentation as separate problems. In Alzheimer's disease data, disease-progression signals appeared at a specific scale; in tumor microenvironment data, spatial context-dependent cell subtypes were identified, and spatially heterogeneous niches were uncovered in cancer.
The multi-scale features come from applying a family of low-pass filters to a cell graph, letting smoothly varying spatial molecular signal through while suppressing sharply fluctuating components. These spectral graph filters take the form of a heat kernel, whose scale parameter sets diffusion time on the graph and thus the effective diffusion distance. Critical scales are determined automatically by partitioning a cross-scale similarity map, with analyses across Xenium, CosMx SMI, MERFISH, STARmap, Stereo-seq and Visium HD mouse brain and tumor datasets.
On single-scale tasks the performance is comparable or superior to existing methods, with higher computational efficiency and lower hardware requirements. Multi-scale analysis usually carries a heavier compute and memory burden. Runtime and peak memory were measured on nine datasets ranging from about 50,000 to over one million cells, where one dataset exceeding 1.1 million cells failed on a laptop due to memory overflow.
The report covers multi-scale feature extraction and critical-scale identification for spatial transcriptomics, aimed at tissue-section data from single-cell-resolution imaging and sequencing platforms. Demonstrated settings include an Alzheimer's disease model, the tumor microenvironment and several mouse brain datasets, with comparisons to existing methods on single-scale tasks. The method is released as a Python package with tutorials, and code and processed data are available via Zenodo.
Automatic critical-scale identification depends on block partitioning of the cross-scale similarity map, with the number of blocks chosen by the elbow method. On a dataset exceeding 1.1 million cells, a laptop failed due to memory overflow. The comparison with single-scale methods is comparable or better, not uniformly leading across all tasks.