scNiche: New Tool Maps Cell Niches in Tissues, Reveals Tumor Microenvironment Heterogeneity

A new computational tool called scNiche [ess-see-nitch] is designed to identify and characterize cell niches [nitch-ez] within tissues, offering insights into the tumor microenvironment and other biological systems. The tool integrates multi-view features of cells, including their molecular profiles, the molecular profiles of their neighborhoods, and the cellular compositions of their neighborhoods. It uses a neural network architecture of the multiple graph autoencoder (M-GAE) [em-jee-ay-ee] coupled with a graph fusion network (GFN) [jee-ef-en] to integrate these features into a joint representation. scNiche's performance was benchmarked against existing methods using simulated and biological datasets. It was applied to spatial omics [oh-micks] datasets from human triple-negative breast cancer (TNBC) [tee-en-bee-see] across two archetypical subtypes (mixed and compartmentalized) and mouse liver under normal and early-onset liver failure states. The tool identified patient- or disease-specific cell niches and provided characterization and interpretation of these niches from both the cellular composition and molecular expression perspectives. In simulated data, scNiche outperformed ten existing methods in accurately identifying cell niches, even when data quality was degraded by gene expression dropout or cell annotation dropout. Ablation studies showed that features from all three views (molecular profiles of the cell, molecular profiles of its neighborhoods, and the cellular compositions of its neighborhoods) contribute to accurate identification of cell niches. In real spatial omics data from mouse spleen and human upper tract urothelial carcinoma (UTUC) [yoo-tack], scNiche demonstrated superior overall performance compared with other methods. It also performed well on mouse brain single-cell spatial transcriptomics datasets. A modified version of scNiche was applied to human DLPFC [dee-ell-pee-ef-see] 10X Visium [viz-ee-um] data, a lower resolution spatial transcriptomics platform, and performed comparably to some state-of-the-art methods. scNiche was also tested on a large mouse whole brain MERFISH [mer-fish] dataset with over 3 million cells, identifying 14 cell niches aligned across sequential tissue sections. The niches identified by scNiche accurately corresponded to different structures in the mouse brain. Application of scNiche to a human triple-negative breast cancer (TNBC) dataset identified 13 cell niches, broadly categorized as tumor-enriched niches and immune-enriched niches. The tumor-enriched niches were predominantly enriched in the mixed subtype samples, while other immune-niches were more prevalent in the compartmentalized subtype samples. The 6 immune-enriched niches showed differential cellular composition, corresponding to distinct microenvironments, including tertiary lymphoid structure (TLS) [tee-ell-ess] and stromal microenvironment in tumor. scNiche was also applied to a mouse liver spatial transcriptomics dataset, identifying 15 cell niches, with the majority showing specific enrichment in either normal or TD (Tsc1/Depdc5) [tee-dee] livers. The 7 cell niches enriched in normal livers exhibited spatial continuity, encompassing the zonation patterns from the central vein to the portal node. The scNiche results revealed three unique niches in TD livers: Niche 4, Niche 9, and Niche 7. These niches were spatially distributed from the core to the periphery of the injury and inflammation sites, and were characterized by the enrichment of a series of emerging cell populations, including inflamed macrophages, hepatic progenitor cells (HPC) [aitch-pee-see], activated hepatic stellate cells (HSC-A) [aitch-ess-see-ay], and injured hepatocytes.