Bayesian cell-type deconvolution and gene expression inference reveals tumor-microenvironment interactions

Abstract

Understanding the dynamic interactions between malignant cells and the tumor stroma is a major goal of cancer research. Here we developed a Bayesian model that jointly infers both cellular composition and gene expression in each cell type, including heterogeneous malignant cells, from bulk RNA-seq using scRNA-seq as prior information. We conducted an integrative analysis of 85 single-cell and 1,412 bulk RNA-seq datasets in primary human glioblastoma, head and neck squamous cell carcinoma, and melanoma. We identified cell types correlated with clinical outcomes and explored regional heterogeneity in tumor state and stromal composition. We redefined common molecular subtypes using gene expression in malignant cells, after excluding confounding non-malignant cell types. Finally, we identified genes whose expression in malignant cells correlated with infiltration of macrophages, T-cells, fibroblasts, and endothelial cells across multiple tumor types. Our work provides a new lens that we used to measure cellular composition and expression in a statistically powered cohort of three primary human malignancies. Results Bayesian inference of cell type composition and tumor expressionTED uses a scRNA-seq reference dataset to infer two parameters of interest from bulk RNA-seq data: (i) the proportion of cell types in the bulk population and (ii) the average expression profiles of each cell type. TED describes the proportion and expression profiles of each cell type as latent variables that it infers from the data ( Fig. 1a, Supplementary Fig. 1, Supplementary Note 1). TED makes the key simplifying assumption that each non-malignant cell type shares a common gene expression profile across patients, as observed in the cases analyzed to date 28,30,31 .Critically, each bulk RNA-seq sample is then assumed to have a unique tumor expression profile that we infer from the data.Expression in the reference and bulk RNA-seq data often differ substantially due to batch effects or tumor heterogeneity. To account for uncertainty in the reference cell type expression matrix, TED implements a fully Bayesian inference of tumor composition. First, TED uses Gibbs sampling to estimate the posterior joint distribution of cell type composition, θ0, and gene expression profiles, Z, i.e. P(θ0, Z | φ, X; α) ( Fig. 1a, red, top). Second, to account for tumor cells, or other cell types which cannot be observed in the reference dataset, TED infers a maximum likelihood estimate (MLE) for tumor expression profiles, ψtum (Fig. 1a, red, mid). During this step TED also infers the maximum a posterior estimate (MAP) of the expression profile of non-malignant stromal cells, ψstr, to correct for batch effects between bulk and single cell RNAseq platforms. Last, TED uses the updated expression profile for each patient and cell type to resample the posterior distribution of cell type composition, θ, i.e. P(θ | ψtum , ψstr ,X; α) ( Fig. 1a, red, bottom). Optionally, TED has an additional mode which can be used to learn common patterns of expression heterogeneit...

A lymphatic-stem cell interactome regulates intestinal stem cell activity

“…For this purpose, we used BayesPrism, a Bayesian statistical model that jointly deconvolves cell type composition and cell type-specific gene expression profiles within each spatial spot by using a scRNA-seq reference from matched tissue as prior information (Figure 4A) (Chu et al, 2021; McKellar et al, 2020). This method significantly outperforms other regression-based tools in bulk RNA-seq deconvolution (Chu et al, 2021; McKellar et al, 2021) (Figure S5A), and its robustness to platform batch effects, technical artefacts and noise made it particularly well suited for spatial deconvolution, treating each Visium spot as a bulk RNA sample. Moreover, our enrichment for LECs and LGR5 + ISCs in our scRNA-seq strategy allowed for an accurate inference of these rarer cell types across the intestine (Figures S5B and S5C).…”

Gene expression variability in human and chimpanzee populations share common determinants

“…Given the correlation between dispersion and the extent of cell-type-specific regulation, we sought to estimate the proportions of different cell types amongst our bulk RNA-seq samples for both chimpanzee and human. We applied BayesPrism cell type deconvolution and expression estimation ( Chu and Danko, 2020 ) to the bulk RNA-seq profiles using reference cell type profiles derived from Tabula Muris (Materials and methods). As expected, cell type proportions between chimpanzee and human hearts are qualitatively similar, although much inter-individual variation exists in both species for particular cell types, such as cardiac muscle cells and myofibroblasts ( Figure 4—figure supplement 1A ).…”

“…BayesPrism ( Chu and Danko, 2020 ) was used to estimate cell types in the bulk RNA-seq datasets used for the RNA-seq dispersion and power analyses. The intersection of genes that are one-to-one mouse-human orthologs and used in DE and bulk dispersion estimation were used to filter the cell-type-labeled mouse scRNA-seq reference gene expression matrix for deconvolution ('run.Ted' function with default parameters) of the 39 chimpanzee and 39 human bulk samples used in DE analysis.…”

Strength in numbers: Large-scale integration of single-cell transcriptomic data reveals rare, transient muscle progenitor cell states in muscle regeneration

“…Lastly, we supplemented our compendium with the first, to our knowledge, transcriptome-level spatial RNA sequencing dataset of regenerating murine skeletal muscle. We repurposed BayesPrism 43 , a recently developed algorithm for the deconvolution of bulk RNA sequencing datasets, to estimate the cell composition of each spot. We then identified putative cellular interactions within those spots that may help drive myogenesis.…”

“…Each Visium spot is 55 µm in diameter and therefore contains several cells. To deconvolve each spot we used BayesPrism, a Bayesian algorithm designed to estimate cell type composition within a bulk RNAseq dataset using a single-cell reference as prior information 43 . We treated each individual spot as a bulk RNA sequencing sample and used BayesPrism to estimate what fraction of the transcripts (theta) within that spot are derived from each cell type within a single-cell reference.…”