Interpretable factor models of single-cell RNA-seq via variational autoencoders

Abstract

Motivation Single-cell RNA-seq makes possible the investigation of variability in gene expression among cells, and dependence of variation on cell type. Statistical inference methods for such analyses must be scalable, and ideally interpretable. Results We present an approach based on a modification of a recently published highly scalable variational autoencoder framework that provides interpretability without sacrificing much accuracy. We demonstrate that our approach enables identification of gene programs in massive datasets. Our strategy, namely the learning of factor models with the auto-encoding variational Bayes framework, is not domain specific and may be useful for other applications. Availability and implementation The factor model is available in the scVI package hosted at https://github.com/YosefLab/scVI/. Contact v@nxn.se Supplementary information Supplementary data are available at Bioinformatics online.

pipeComp, a general framework for the evaluation of computational pipelines, reveals performant single cell RNA-seq preprocessing tools

“…GLM-PCA tended to increase the average silhouette width of already well-defined subpopulations, but Seurat's PCA procedure however proved superior on all metrics. Like GLM-PCA, scVI's Linearly Decoded (LD) data [41] and latent space do not explicitly rely on normalized counts. Their performance was however worse than Seurat's PCA in view of the silhouette width (Additional File 1: Figure S15).…”

Enhancing scientific discoveries in molecular biology with deep generative models

“…While this approach is interesting, we note that it comes with no theoretical guarantees and may produce spurious correlations. In Svensson et al (), a VAE with a linear decoder (LDVAE) is proposed for scRNA‐seq data. The LDVAE trades model fit for more interpretability since the decoder now relates directly latent variables to the expression of individual genes.…”

Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells

“…We used the mean of z n under the approximate posterior as an integrated, joint view of cell state, which we then used as input to clustering and visualization algorithms. Additionally, we modeled z n with a logistic normal distribution, meaning each z n resides in a convex polytope amenable to archetypal analysis [26,27]. Such an analysis connects each latent dimension to the expression of genes and proteins and aids with the interpretation of the model.…”

“…For example, for autoencoder-based single-cell methods like scVI or DCA [13,24], there is no straightforward way to determine which expression programs are associated with each dimension of the latent space. This is in contrast to linear methods like PCA, GLM-PCA or LDVAE [27,55], where each latent dimension is associated with a loading vector that describes the contribution made by each feature and thus enables direct interpretation. The interpretability of linear methods, however, comes at the expense of reduced capacity to fit complex data such as that obtained by scRNA-seq [27].…”

“…This is in contrast to linear methods like PCA, GLM-PCA or LDVAE [27,55], where each latent dimension is associated with a loading vector that describes the contribution made by each feature and thus enables direct interpretation. The interpretability of linear methods, however, comes at the expense of reduced capacity to fit complex data such as that obtained by scRNA-seq [27].…”