Tensor decomposition reveals coordinated multicellular patterns of transcriptional variation that distinguish and stratify disease individuals

Abstract

Tissue- and organism-level biological processes often involve coordinated action of multiple distinct cell types. Current computational methods for the analysis of single-cell RNA-sequencing (scRNA-seq) data, however, are not designed to capture co-variation of cell states across samples, in part due to the low number of biological samples in most scRNA-seq datasets. Recent advances in sample multiplexing have enabled population-scale scRNA-seq measurements of tens to hundreds of samples. To take advantage of such datasets, here we introduce a computational approach called single-cell Interpretable Tensor Decomposition (scITD). This method extracts multicellular gene expression patterns that vary across different biological samples. These patterns capture how changes in one cell type are connected to changes in other cell types. The multicellular patterns can be further associated with known covariates (e.g., disease, treatment, or technical batch effects) and used to stratify heterogeneous samples. We first validated the performance of scITD using in vitro experimental data and simulations. We then applied scITD to scRNA-seq data on peripheral blood mononuclear cells (PBMCs) from 115 patients with systemic lupus erythematosus and 56 healthy controls. We recapitulated a well-established pan-cell-type signature of interferon-signaling that was associated with the presence of anti-dsDNA autoantibodies and a disease activity index. We further identified a novel multicellular pattern that appears to potentiate renal involvement for patients with anti-dsDNA autoantibodies. This pattern was characterized by an expansion of activated memory B cells along with helper T cell activation and is predicted to be mediated by an increase in ICOSLG-ICOS interaction between monocytes and helper T cells. Finally, we applied scITD to two PBMC datasets from patients with COVID-19 and identified reproducible multicellular patterns that stratify patients by disease severity. Overall, scITD is a flexible method for exploring co-variation of cell states in multi-sample single-cell datasets, which can yield new insights into complex non-cell-autonomous dependencies that define and stratify disease.

Precise identification of cell states altered in disease with healthy single-cell references

“…While here we applied a naive approach to define sample similarity (suppl. figure 7A), novel methods to model sample-level heterogeneity in scRNA-seq data are being explored (Chen et al ., 2020; Boyeau et al ., 2022; Mitchel et al ., 2022). These could improve the matching of disease samples to optimal controls, and provide new insights into which technical and demographic variables are likely to affect disease-to-healthy comparisons.…”

Multicellular factor analysis of single-cell data for a tissue-centric understanding of disease

“…Here, we show that multi-omics integration methods, such as Multi-Omics Factor Analysis (MOFA) (Argelaguet et al, 2018(Argelaguet et al, , 2020, can be repurposed in a straightforward manner to perform similar tissue-centric analyses as the ones performed by the aforementioned methods, since MOFA treats similar multi-view data representations and model objectives to create latent spaces. Moreover, MOFA is a flexible statistical framework that overcomes the limitation of data completeness that some tissue-centric methods enforce (Armingol et al, 2022;Mitchel et al, 2022), where all samples must contain information in all cell-type views and all cell-type views must contain the same features. In contrast to the aforementioned methods, it also provides the possibility of jointly analyzing independent groups of samples with various classes of cell-type views.…”

“…A set of novel tissue-centric computational methods have emerged that are helpful in the definition of multicellular programs associated with clinical covariates of interest (Jerby-Arnon & Regev, 2022), and the unsupervised analysis of samples from cross-condition single-cell atlases (Armingol et al , 2022; Mitchel et al , 2022). These methods are extensions of matrix factorization that aim to reduce the dimensionality of the data while retaining most of the variability.…”

Deep generative modeling of sample-level heterogeneity in single-cell genomics

“…A common approach for quantifying differences between single-cell samples relies on clustering cells into groups representing cell states/types and then quantifying sample-specific differences in the relative abundance of each group. This approach can be used to evaluate the distance between any pair of samples, and thus enables an exploratory analysis [6, 7, 12, 13, 14]. However, it also oversimplifies the task by reducing the high-dimensional omic information of every cell to a single, discrete label.…”