Bayesian Inference for a Generative Model of Transcriptome Profiles from Single-cell RNA Sequencing

By Romain Lopez, Jeffrey C. Regier, Michael Cole, Michael I. Jordan, Nir Yosef. 2018

38 citations recorded by the source corpus.

Transcriptome profiles of individual cells reflect true and often unexplored biological diversity, but are also affected by noise of biological and technical nature. This raises the need to explicitly model the resulting uncertainty and take it into account in any downstream analysis, such as dimensionality reduction, clustering, and differential expression. Here, we introduce Single-cell Variational Inference (scVI), a scalable framework for probabilistic representation and analysis of gene expression in single cells. Our model uses variational inference and stochastic optimization of deep neural networks to approximate the parameters that govern the distribution of expression values of each gene in every cell, using a non-linear mapping between the observations and a low-dimensional latent space. By doing so, scVI pools information between similar cells or genes while taking nuisance factors of variation such as batch effects and limited sensitivity into account. To evaluate scVI, we conducted a comprehensive comparative analysis to existing methods for distributional modeling and dimensionality reduction, all of which rely on generalized linear models. We first show that scVI scales to over one million cells, whereas competing algorithms can process at most tens of thousands of cells. Next, we show that scVI fits unseen data more closely and can impute missing data more accurately, both indicative of a better generalization capacity. We then utilize scVI to conduct a set of fundamental analysis tasks – including batch correction, visualization, clustering and differential expression – and demonstrate its accuracy in comparison to the state-of-the-art tools in each task. scVI is publicly available, and can be readily used as a principled and inclusive solution for multiple tasks of single-cell RNA sequencing data analysis.

Subfields: Artificial Intelligence, Molecular Biology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Molecular Biology

Citations in the corpus

Information Constraints on Auto-Encoding Variational Bayes (2018, methodology): “Experiment We compare scVI [33], a state-of-the-art model, with no observed nuisance variables (8 latent dimensions for z), and our proposed model with observed quality control metrics.”
Unsupervised removal of systematic background noise from droplet-based single-cell experiments using CellBender (2019, methodology): “Furthermore, a plurality of widely-used algorithms such as voom [55] for differential expression (DE) analysis, Seurat v3’s highly-variable gene (HVG) selection [56], and scVI [34] for latent space learning, explicitly expect integer counts as input due to employment of discrete likelihood models su”
Graph abstraction reconciles clustering with trajectory inference through a topology preserving map of single cells (2017, background): “Any other distance measure, for example, the kernel-based measure of Reference [44], or autoencoder representations [45, 46] would also be a viable option.”