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

Abstract

The paired measurement of RNA and surface protein abundance in single cells with CITE-seq is a promising approach to connect transcriptional variation with cell phenotypes and functions. However, each data modality exhibits unique technical biases, making it challenging to conduct a joint analysis and combine these two views into a unified representation of cell state. Here we present Total Variational Inference (totalVI), a framework for the joint probabilistic analysis of paired RNA and protein data from single cells. totalVI probabilistically represents the data as a composite of biological and technical factors such as limited sensitivity of the RNA data, background in the protein data, and batch effects. To evaluate totalVI, we performed CITE-seq on immune cells from murine spleen and lymph nodes with biological replicates and with different antibody panels measuring over 100 surface proteins. With this dataset, we demonstrate that totalVI provides a cohesive solution for common analysis tasks like the integration of datasets with matched or unmatched protein panels, dimensionality reduction, clustering, evaluation of correlations between molecules, and differential expression testing. totalVI enables scalable, end-to-end analysis of paired RNA and protein data from single cells and is available as open-source software.

scMM: Mixture-of-experts multimodal deep generative model for single-cell multiomics data analysis

“…We modeled x n,m with probability distributions that capture the characteristics of data distributions for each modality. For transcriptome and surface protein data, negative binomial (NB) distribution was selected to explain non-negative counts with overdispersion [14]. In addition, chromatin accessibility data is non-negative count data; however, it exhibits extreme sparsity due to low signal (only two locus exist for each diploid cell), limited coverage, and closed chromatin.…”

“…One powerful approach to capture nonlinear latent structures is to use expressive variational autoencoders (VAEs), which consist of a pair of neural networks wherein one encodes data into the latent space, and the other decodes them to reconstruct the data distribution [7]. scMVAE and totalVI are the currently available VAE-based methods for single-cell multimodal data analysis [13, 14]. Nevertheless, scMVAE requires a simplified conversion of chromatin accessibility to transcriptome before training, which is known to lead to the nonnegligible loss of epigenetic information [15].…”

BABEL enables cross-modality translation between multiomic profiles at single-cell resolution

“…Deep learning methods for multiomic data have also been studied, but these prior works generally did not have access to large-scale paired measurements, which motivated complex techniques to align latent representations ( 15 – 18 ) or constrained these works to bulk measurements ( 19 ). More recently, new experimental techniques for generating paired single-cell data have enabled more streamlined multimodal modeling of protein epitopes and transcriptomics ( 20 ) as well as of physiological profiles and transcriptomics ( 21 ). BABEL builds off these prior works while introducing strategies for more efficient model architectures and latent space learning.…”

Maximizing statistical power to detect clinically associated cell states with scPOST

“…scPOST currently operates using a low-dimensional PCA embedding of cells. With multimodal technologies such as CITE-seq 5 becoming more available, analyses of these new data types may include dimensionality reduction with alternative methods, such as canonical correlation analysis (CCA) 10,14,34 and nonlinear embeddings 35,36 . Simulating new types of data in the context of these alternative tools, such as simulating canonical variate coordinates instead of PC coordinates, represents a possible extension of scPOST.…”