(B-58) Interpretable and scalable integration of single-cell transcriptomics across conditions with PARAFAC2

Introduction:: Analyzing how populations of single cells respond to multiple conditions or perturbations is a challenging task for current computational methods. To address this issue, we propose using PARAFAC2 (Pf2), a tensor decomposition method that can reduce the complexity of single-cell data in multi-condition and perturbational studies (Fig. A). By summarizing the data in a reduced space that explicitly separates alignment of cells from condition-specific effects, we can analyze how experimental conditions cause similar changes across single-cell experiments and how cells vary within experimental conditions. This version of irregular tensor decomposition (TD) is compatible with diverse single-cell methods and offers a significant improvement in the versatility of single-cell analyses. By harnessing the power of irregular TD to track patterns of cellular response, our study represents a major step forward in single-cell applications.

Materials and Methods:: We used a scRNA-seq dataset in which hPBMCs were treated with 40 commercially available, FDA-approved immune-modulatory drugs. PARAFAC2 (Pf2) was used to analyze the data across conditions. Pf2 uses Orthogonal Procrustes to identify similar cell states between conditions, while the transcriptional data, projected into shared cell states, is factored by canonical polyadic decomposition as a genes by conditions by cell state tensor. These results were compared to principal components analysis (PCA) as the dimensionality reduction. Uniform Manifold Approximation and Projection (UMAP) was used for visualization.

Results, Conclusions, and Discussions:: Each component of Pf2 will have an associated weight for each drug, gene, and cell state (Figs. B-D). We can map correlations across each mode in our tensor, describing how genes and single cells alter with respect to drug treatments. The Pf2 projection matrices led to better cell type isolation and alignment than with PCA (Figs. E/F). PCA led to separate clusters for several of the drugs with large effects (Figs. G/H). Through the factor matrices describing each condition, Pf2 was able to identify that glucocorticoid anti-inflammatory agents, including betamethasone valerate, triamcinolone acetonide, and budesonide, specifically perturb expression of genes including CD163, ADORA3, and MS4A6A, within macrophages and monocytes. These changes were effectively separated from the cell population identification, and could be localized to specific cell populations.



By separating the alignment of cells to certain cell states and the representation of these states across conditions, Pf2 enables straight-forward interpretation of perturbation effects, assignment of perturbation effects to specific cells, and improved resolution in the alignment of cells across conditions. We propose that Pf2 provides generally improved dimensionality reduction of single cell studies across conditions.


Acknowledgements (Optional): :

References (Optional): :