scPI: A Scalable Framework for Probabilistic Inference in Single-Cell RNA-Sequencing Data Analysis

The technique of single-cell RNA-sequencing (scRNA-seq) has provided an unprecedented opportunity to investigate the cellular heterogeneity of complex tissues. As large-scale scRNA-seq datasets are becoming more available and affordable, there is a growing demand for computational scalable methods to analyze scRNA-seq data. Here, we propose a scalable framework, scPI, to infer the latent low-dimensional representations of the scRNA-seq data to facilitate downstream analysis. Our method scPI makes use of the amortized variational inference, where the posterior mean and variance of the latent variable are parameterized by a nonlinear neural network. This inference structure combined with stochastic optimization enables its computational efficiency and scalability. Through the analysis of two real datasets, we demonstrate that the scPI framework can be effectively applied to several probabilistic models for scRNA-seq data, in terms of its scalability, missing value imputation and cell type clustering. The codes for reproducing the real data analysis results are available at https://github.com/YangLabHKUST/scPI.