Generating Synthetic Single Cell Data from Bulk RNA-seq Using a Pretrained Variational Autoencoder

Abstract

Single cell RNA sequencing (scRNA-seq) is a powerful approach which generates genome-wide gene expression profiles at single cell resolution. Among its many applications, it enables determination of the transcriptional states of distinct cell types in complex tissues, thereby allowing the precise cell type and set of genes driving a disease to be identified. However, scRNA-seq remains costly, and there are extremely limited samples generated in even the most extensive human disease studies. In sharp contrast, there is a wealth of publicly available bulk RNA-seq data, in which single cell and cell type information are effectively averaged. To further leverage this wealth of RNA-seq data, methods have been developed to infer the fraction of cell types from bulk RNA-seq data using single cell data to train models. Additionally, generative AI models have been developed to generate more of an existing scRNA-seq dataset. In this study, we develop an innovative framework that takes full advantage of powerful generative AI approaches and existing scRNA-seq data to generate representative scRNA-seq data from bulk RNA-seq. Our bulk to single cell variational autoencoder-based model, termed bulk2sc, is trained to deconvolve pseudo-bulk RNA-seq datasets back into their constituent single-cell transcriptomes by learning the specific distributions and proportions related to each cell type. We assess the performance of bulk2sc by comparing synthetically generated scRNA-seq to actual scRNA-seq data. Application of bulk2sc to large-scale bulk RNA-seq human disease datasets could yield single cell level insights into disease processes and suggest targeted scRNA-seq experiments.