Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology.
Chu T, Wang Z, Pe'er D, Danko CG. · 2022
Read Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology. by Chu T, Wang Z, Pe'er D, Danko CG. (2022) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc9046084.txt.
Introduction
Chu and co-authors present BayesPrism, a method that unmixes bulk tumor gene data into cell types. Bulk sequencing blends signals from cancer and non-cancer cells, hiding who expresses what. BayesPrism uses single-cell data from patients as a guide and infers both cell shares and cell-specific expression. The team applied it to glioblastoma, head and neck, and skin cancers.
The method refined cancer subtypes once non-cancer cells were accounted for. It also linked genes in malignant cells to the arrival of macrophages, T cells, and other cell types. The authors offer it as a new lens for large tumor cohorts.
Catalog record
| CREATOR | Chu T, Wang Z, Pe'er D, Danko CG. |
|---|---|
| YEAR | 2022 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 89 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology By Tinyi Chu, Zhong Wang, Dana Pe’er, Charles G. Danko Published in: Nature Cancer DOI: 10.1038/s43018-022-00356-3 ABSTRACT Inferring single-cell compositions and their contributions to global gene expression changes from bulk RNA sequencing (RNA-seq) datasets is a major challenge in oncology. Here we develop Bayesian cell proportion reconstruction inferred using statistical marginalization (BayesPrism), a Bayesian method to predict cellular composition and gene expression in individual cell types from bulk RNA-seq, using patient-derived, scRNA-seq as prior information. We conduct integrative analyses in primary glioblastoma, head and neck squamous cell carcinoma and skin cutaneous melanoma to correlate cell type composition with clinical outcomes across tumor types, and explore spatial heterogeneity in malignant and nonmalignant cell states. We refine current cancer subtypes using gene expression annotation after exclusion of confounding nonmalignant cells. Finally, we identify genes whose expression in malignant cells correlates with macrophage infiltration, T cells, fibroblasts and endothelial cells across multiple tumor types. Our work introduces a new lens to accurately infer cellular composition and expression in large cohorts of bulk RNA-seq data. MAIN Cell–cell interactions are highly complex and can strongly impact cell behavior in biological contexts, often with medical ramifications. A quintessential example is that between malignant cells and diverse nonmalignant cell types within the tumor microenvironment (TME)1,2. Numerous studies over the past two decades have revealed interactions between cells in the TME that promote diverse functions, including angiogenesis3, metastasis4 and immunosuppression5. Nonmalignant cells can differ markedly among patients and tumor types6, and certain nonmalignant cell populations are used as clinical biomarkers7 and therapeutic targets8. These studies motivate the direct measurement of cell types within tissues. Two layers of information are critical for understanding tumor composition: (1) the proportion of each cell type and (2) the levels of gene expression in each cell type. The rise of single-cell RNA sequencing (scRNA-seq) technologies has recently enabled direct, genome-wide measurement of the transcriptome in individual cells within the TME and characterization of their heterogeneity. However, the cost of scRNA-seq and requirements for high-quality tissue limit the number of patient samples that can be assayed9. Moreover, scRNA-seq is susceptible to technical biases in cell capture9, which confound the recovery of cell type composition. As an alternative, cell type abundance can be inferred from bulk RNA-seq data using regression on a reference expression matrix constructed from a set of arbitrarily defined marker genes10–14. Pioneering methods for cell type deconvolution have demonstrated that it is possible to infer the abundance of multiple cell types in the TME. However, existing deconvolution methods make restrictive assumptions about the difference in distribution between the reference and bulk sample. These assumptions are often violated by both technical (for example, different platforms used for reference and bulk sequencing) and biological (for example, heterogeneity in gene expression within constituent cell types) differences between bulk and reference data. Critically, cell type deconvolution methods have not fully supported prediction of gene expression in a heterogeneous population of tumor cells. Thus, existing methods fail to address these key questions: how do malignant cells affect the composition of nonmalignant cells in the TME? Which genes are correlated with these interactions? To answer these questions, we need a model that can accurately represent cell type fraction and cell-type-specific expression profiles in each bulk sample, and can accommodate differences between the single-cell reference and bulk. Here, we present BayesPrism, a Bayesian model that jointly infers the posterior distribution of cell type fractions and gene expression from bulk RNA-seq data using an scRNA-seq reference as prior information. By explicitly modeling and marginalizing out the differences in gene expression between single-cell reference and bulk data, BayesPrism substantially outperforms leading methods in the inference of cell type fractions in both tumor and nontumor settings. We demonstrate the utility of our approach on a large dataset of 1,412 bulk RNA-seq and 85 scRNA-seq samples in glioblastoma (GBM), head and neck squamous cell carcinoma (HNSCC) and skin cutaneous melanoma (SKCM). Our work introduces a powerful new tool for integrative analysis of bulk and scRNA-seq data. RESULTS Bayesian inference of cell type fraction and gene expression BayesPrism uses an scRNA-seq reference to infer two statistics from each bulk RNA-seq sample: (1) the proportion of reads derived from each cell type, which we assume is proportional to the fraction of that cell type; and (2) the expression level of genes in each cell type (Fig. 1a,b, Extended Data Fig. 1 and Supplementary Note 1). The most challenging aspect of cellular deconvolution is accounting for various sources of uncertainty, including technical and biological batch variation, in gene expression between bulk and scRNA-seq reference data. To account for these uncertainties, BayesPrism adopts a Bayesian strategy that models prior distribution using scRNA-seq, and infers a joint posterior distribution of cell type proportion and gene expression in each cell type and bulk sample conditional on each observed bulk. As a result, the uncertainty in each estimate can be marginalized out from the joint posterior.Fig. 1BayesPrism algorithm flow and performance validation.a, Algorithmist flow of the deconvolution module of BayesPrism. b, Variables and their dimensions shown in a. c–f, Boxplots showing the cell-type-level Pearson’s correlation coefficient (c,e) and MSE (d,f) for deconvolution of GBM28 pseudo-bulks using refGBM8 (c,d), and bulk RNA-seq human whole blood samples with ground truth measured by flow cytometry (e,f). Boxes mark the 25th percentile (bottom), median (central bar) and 75th percentile (top). Whiskers represent extreme values within 1.5-fold of interquartile range. One-sided P values are shown for cell type fractions inferred by BayesPrism (updated θ using marker-free mode) and those by the second-best methods ranked by median value. T-test was used for MSE, and z-test was performed on Fisher’s z-transformed cell-type-level correlation coefficients (Methods). c,d, Statistics were computed for 1,350 pseudo-bulk RNA-seq samples simulated using scRNA-seq from 27 patients with GBM having more than ten malignant cells for all methods, except CIBERSORTx. For CIBERSORTx and its comparison with BayesPrism, statistics were computed using 270 downsampled pseudo-bulks across 27 patients. e,f, Statistics were computed using 12 bulk RNA-seq samples from independent healthy adults. g, Uniform manifold approximation and projection (UMAP) visualization showing expression of individual cells in GBM28. The expression profiles of nonmalignant cells before (gray) and after information pooling (black) were projected onto the UMAP manifold of scRNA-seq (left). Malignant cells in patients with more than ten malignant cells (n = 27) are visualized on the zoomed-in UMAP (right) and are colored by patient. The inferred expression profile, shown as △, and the averaged expression profile from scRNA-seq for each patient, shown as ○, are projected onto the UMAP manifold. h, Scatter plot showing Pearson’s correlation between average expression of malignant cells in pseudo-bulk and that estimated by BayesPrism (red) and CIBERSORTx group mode (orange) or the undeconvolved simulated bulk (blue), as a function of the fraction of malignant cells in a subsampled set (n = 270).Source data BayesPrism accommodates multiple gene expression subtypes of the same cell type (hereafter referred to as ‘cell states’). Internally, BayesPrism treats cell types and states in the same manner, and returns the posterior sum over different cell states defined by the scRNA-seq dataset to represent the fraction and expression