De novo gene signature identification from single‐cell RNA ‐seq with hierarchical Poisson factorization

Abstract

Common approaches to gene signature discovery in single‐cell RNA ‐sequencing (sc RNA ‐seq) depend upon predefined structures like clusters or pseudo‐temporal order, require prior normalization, or do not account for the sparsity of single‐cell data. We present single‐cell hierarchical Poisson factorization (sc HPF ), a Bayesian factorization method that adapts hierarchical Poisson factorization (Gopalan et al , 2015 , Proceedings of the 31st Conference on Uncertainty in Artificial Intelligence , 326) for de novo discovery of both continuous and discrete expression patterns from sc RNA ‐seq. sc HPF does not require prior normalization and captures statistical properties of single‐cell data better than other methods in benchmark datasets. Applied to sc RNA ‐seq of the core and margin of a high‐grade glioma, sc HPF uncovers marked differences in the abundance of glioma subpopulations across tumor regions and regionally associated expression biases within glioma subpopulations. sc HFP revealed an expression signature that was spatially biased toward the glioma‐infiltrated margins and associated with inferior survival in glioblastoma.

Latent factor modelling of scRNA-seq data uncovers novel pathways dysregulated in cell subsets of autoimmune disease patients

“…In the RA dataset, almost all methods fail to reach an AUPR > 0.5 regardless of the dimensionality of the latent space. In contrast, for the SLE dataset, all methods see a consistent increase in predictive power as the dimensionality increases, with scCoGAPS and LDA showing the best performance at low (16)(17)(18)(19)(20)(21)(22)(23)(24) and high (26)(27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38)(39)(40) dimensionality of the latent space, respectively. Taken together, these results suggest that the dimensionality of the latent space is critical for extracting biological features related to the disease state of the cell.…”

“…Four methods were selected for evaluation: scCoGAPS 21 , LDA 29 , scHPF 22 and scVI 23 . The number of dimensions for the latent space (16 to 14 latent variables, step 2) has been selected for efficiency and for consistent evaluation across models (i.e.…”

“…For LDA we employed hyperparameters as described in a similar use case 27 . For scHPF, we employed hyperparameters as suggested by the authors 22 .…”

“…Standard matrix factorization approaches, such as principal component analysis (PCA), nonnegative matrix factorization (NMF) and independent component analysis (ICA), have been widely applied to scRNA-seq data 1,2,20 . Nevertheless, novel methods have been developed that account for the specificities of single-cell data, using meaningful prior distributions and enforcing sparsity [21][22][23][24][25][26] . A major challenge of these approaches is the determination of the number of latent dimensions to use.…”

“…To explore the potential of these methods to uncover previously unidentified pathway activities, we set up to perform a systematic comparison of four latent factor models recently proposed: scCoGAPS 21 , LDA 29 , scHPF 22 and scVI 23 . We test these models on two scRNA-seq datasets from autoimmune diseases patients.…”