Deconvolution of the tumor microenvironment with omnideconv

1. Introduction and dataset

In this vignette, we will use the omnideconv package to deconvolve a bulk RNA-seq dataset from 24 breast cancer patients with two different methods (DWLS and BayesPrism). The datasets are from a recent breast cancer study (Wu et al. 2021). This study provides access to a primary-tumor single cell RNA-seq (scRNA-seq) dataset from 26 breast cancer patients across three major cancer subtypes (ER+, HER2+, TNBC). The dataset includes cell-type annotation for three resolution levels. In addition, the data includes bulk RNA-seq sequencing for 24 of the patients. In this chapter, we use multi-sample, breast-cancer scRNA-seq atlas (100,064 cells) as a reference to train the methods for the deconvolution of the bulk RNA-seq samples. The single-cell and bulk RNA-seq data is deposited on GEO, under accession number: GSE176078. The single cell data comes with the author’s cell type annotations. We will need to download and unzip the datasets (GSE176078_Wu_etal_2021_BRCA_scRNASeq.tar.gz, GSE176078_Wu_etal_2021_bulkRNAseq_raw_counts.txt.gz), and store them in the working directory. For this example analysis, we will also need to retrieve the additional clinical information about the patients – although it is not required by omnideconv. This is available in the Supplementary Table 1, included with the paper supplementary materials

2. Library loading

library(Seurat)
library(tidyverse)
library(omnideconv)
library(readxl)

3. Single cell data processing

Although not strictly required by omnideconv, we suggest performing quality control and filtering of the input scRNA-seq data according to the best practice (Heumos et al. 2023), to ensure the best training conditions for deconvolution algorithms. We first pre-process the single-cell dataset to remove low-quality cells. We will use the R package Seurat (Hao et al. 2023), which allows fast and easy manipulation of single-cell data. We will create a Seurat object with the cell counts and their metadata of interest curated by the authors, which include cell-type annotation on three levels of resolution:

single.cell.data <- Seurat::ReadMtx(
  mtx = 'C:/Users/c1041161/book_chapter/Wu_etal_2021_BRCA_scRNASeq/count_matrix_sparse.mtx',
  cells = 'C:/Users/c1041161/book_chapter/Wu_etal_2021_BRCA_scRNASeq/count_matrix_barcodes.tsv',
  features = 'C:/Users/c1041161/book_chapter/Wu_etal_2021_BRCA_scRNASeq/count_matrix_genes.tsv',
  feature.column = 1
)
single.cell.metadata <- read.table('C:/Users/c1041161/book_chapter/Wu_etal_2021_BRCA_scRNASeq/metadata.csv',
                                   sep = ',',
                                   header = TRUE,
                                   row.names = 1)

single.cell.data = CreateSeuratObject(single.cell.data,
                                      project='Wu_dataset',
                                      assay='RNA',
                                      min.cells = 0,
                                      min.features = 1, meta.data = single.cell.metadata)

We can have an overview of the number of cells per cell type in the dataset:

In order to remove low quality cells, we will follow the best practices for single cell normalization (Heumos et al. 2023). We will perform quality control on each cell by considering metrics such as the number of total counts, the number of expressed features (genes), and the fraction of mitochondrial genes per cell. We will remove cells that have \(MAD = median(|X_i - median(X)|)\), where

• \(X_i\) is the respective metric of an observation \(i\) (cell), for example the number of detected genes
• \(MAD\) is the Mead Absolute Deviation, computed as \(MAD = median(|X_i - median(X)|)\)
• \(n\) is set to 5 for the number of counts and expressed features, and to 3 for the fraction of mitochondrial genes

In order to do this filtering, we will create a function to identify the outliers for each metric

is_outlier <- function(SeuratObject, metric, nmads){
  eval(parse(text = paste0("M <- SeuratObject$",metric)))
  outlier <- (M < median(M) - nmads * mad(M)) | (M > median(M) + nmads * mad(M))
  return(outlier)
}

check_outliers_nFeature <- is_outlier(single.cell.data, 'nFeature_RNA', 5)
check_outliers_nCount <- is_outlier(single.cell.data, 'nCount_RNA', 5)
check_outliers_mito <- is_outlier(single.cell.data, 'percent.mito', 3)

non_outliers_nFeature <- names(check_outliers_nFeature[!check_outliers_nFeature])
non_outliers_nCount <- names(check_outliers_nCount[!check_outliers_nCount])
non_outliers_mito <- names(check_outliers_mito[!check_outliers_mito])

We will retain only those that satisfy all three of the conditions described above.

non_outliers <- intersect(non_outliers_nFeature, non_outliers_nCount) %>%
  intersect(non_outliers_mito)

single.cell.data <- subset(single.cell.data, cells = non_outliers)

as.data.frame(table(single.cell.data$celltype_major, dnn = list("celltype_major")), responseName = "number_cells")
#>      celltype_major number_cells
#> 1           B-cells         3150
#> 2              CAFs         6154
#> 3 Cancer Epithelial        16613
#> 4       Endothelial         6991
#> 5           Myeloid         8855
#> 6 Normal Epithelial         3533
#> 7      Plasmablasts         3164
#> 8               PVL         5120
#> 9           T-cells        34991

4. Bulk data preprocessing

We will now read in the bulk sequencing data file, which consists of 24 samples.

bulk.data <- read.table('C:/Users/c1041161/book_chapter/GSE176078_Wu_etal_2021_bulkRNAseq_raw_counts.txt', skip=1)

header <- read.table('C:/Users/c1041161/book_chapter/GSE176078_Wu_etal_2021_bulkRNAseq_raw_counts.txt',
                     header = FALSE, nrows = 1, skipNul = TRUE, sep='\t')

colnames(bulk.data) <- c('Genes', gsub('A|N', '', header[2:25]))

bulk.data <- bulk.data[bulk.data$Genes != '', ]
bulk.data <- column_to_rownames(bulk.data, 'Genes')
bulk.data <- as.matrix(bulk.data)

5. Subsampling of single cell data

The various methods included in omnideconv rely on the single cell dataset that will be used to train them for the deconvolution of those specific cell types. This training involves the optimization of internal features of the methods and can happen in different ways. Some methods use the single cell data to build a ‘signature matrix’, i.e. a reduced transcriptional fingerprints of the cell types provided, while others use this data in a statistical or deep learning model. Since single cell datasets can often encompass thousands of cells, we will need to subsample it in order to be able to run the analysis on our machines. In this case we will retain a maximum of 200 cells per cell type, but this step can be costumed, or eventually skipped, depending on the computational resources available.

max_cells_per_celltype = 200

sampled.metadata <- single.cell.data@meta.data %>%
      rownames_to_column(., 'barcode') %>%
      group_by(., celltype_major) %>%
      nest() %>%
      mutate(n =  map_dbl(data, nrow)) %>%
      mutate(n = min(n, max_cells_per_celltype)) %>%
      ungroup() %>%
      mutate(samp = map2(data, n, sample_n)) %>%
      select(-data) %>%
      unnest(samp)

single.cell.data.sampled <- subset(single.cell.data, cells = sampled.metadata$barcode)


#as.data.frame(table(single.cell.data.sampled$celltype_major, dnn = list("celltype_major")), responseName = "number_cells")

6. Deconvolution of the bulk data

Each methods has different requirements, but in general to compute the deconvolution results we will need the single cell counts matrix, the cell type annotations and the information on the individual/experiment fom which the cells were retrieved (batch ID).

counts.matrix <- as.matrix(single.cell.data.sampled@assays$RNA@counts)
cell.type.annotations <- single.cell.data.sampled$celltype_major
batch.ids <- single.cell.data.sampled$orig.ident

# We need to insert the normalization as well
signature.matrix.dwls <- omnideconv::build_model(single_cell_object = counts.matrix,
                        cell_type_annotations = cell.type.annotations,
                        method = 'dwls',
                        dwls_method = 'mast_optimized')

deconvolution.results.dwls <- deconvolute(bulk_gene_expression = bulk.data,
                                          signature = signature.matrix.dwls,
                                          method='dwls',
                                          dwls_submethod = 'DampenedWLS')

# BayesPrism deconvolution

deconvolution.results.bayesprism <- deconvolute(bulk_gene_expression = bulk.data,
                                           single_cell_object = counts.matrix,
                                           cell_type_annotations = cell.type.annotations,
                                           signature=NULL,
                                           method = 'bayesprism',
                                           n_cores=12)

omnideconv::plot_deconvolution(list('bayesprism' = deconvolution.results.bayesprism), "bar", "method", "Spectral")

7. Deconvolution of the bulk data at a lower resolution

The considered single-cell breast cancer dataset includes cell-type annotations at three levels of resolution: celltype_major, celltype_minor, and celltype_subset, which distinguish 9, 29, and 58 cell types respectively. The different cell-type annotations can be accessed with:

single.cell.data$celltype_major      # Major annotation
single.cell.data$celltype_minor      # Minor annotation
single.cell.data$celltype_subset     # Subset annotation

These additional annotations provide a cell-type classification at a finer resolution. For instance, at the celltype_major level, we only have the T cell population, while at the celltype_minor level, we can distinguish between CD4+ and CD8+ T cells. In the following, we will again perform deconvolution analysis with DWLS but, this time, using the celltype_minor information. We will subsample the dataset as before, this time considering the second level of resolution for the cell types, and extract the objects needed for deconvolution.

max_cells_per_celltype = 200


sampled.metadata <- single.cell.data@meta.data %>%
      rownames_to_column(., 'barcode') %>%
      group_by(., celltype_minor) %>%
      nest() %>%
      mutate(n =  map_dbl(data, nrow)) %>%
      mutate(n = min(n, max_cells_per_celltype)) %>%
      ungroup() %>%
      mutate(samp = map2(data, n, sample_n)) %>%
      select(-data) %>%
      unnest(samp)

single.cell.data.sampled <- subset(single.cell.data, cells = sampled.metadata$barcode)

counts.matrix <- as.matrix(single.cell.data@assays$RNA@counts)
cell.type.annotations <- single.cell.data$celltype_minor
batch.ids <- single.cell.data$orig.ident

signature.matrix.dwls.minor <- omnideconv::build_model(single_cell_object = counts.matrix,
                        cell_type_annotations = cell.type.annotations,
                        method = 'dwls',
                        dwls_method = 'mast_optimized')

deconvolution.results.dwls.minor <- deconvolute(bulk_gene_expression = bulk.data,
                                          signature = signature.matrix.dwls.minor,
                                          method='dwls',
                                          dwls_submethod = 'DampenedWLS')

We can visualize the results as before:

omnideconv::plot_deconvolution(list('dwls' = deconvolution.results.dwls.minor), "bar", "method", "Spectral")

8. Comparison of cell fractions across conditions

We can consider as well the metadata provided with the original paper, which include patient’s data, cancer subtype information and treatment details. We can first harmonize the sample names, and then combine all this information with the deconvolution results in one dataframe.

patient.metadata <- read_excel("C:/Users/c1041161/book_chapter/41588_2021_911_MOESM4_ESM.xlsx", sheet = 1, skip = 3) %>%
  select(., c(1, 4, 5, 11, 12))
colnames(patient.metadata) <- c('Sample', 'Grade', 'Cancer_type', 'IHC_subtype', 'Treatment')

patient.metadata$Sample <- gsub('-', '', patient.metadata$Sample) %>%
  paste0('CID', .)

patients.results <- rownames_to_column(as.data.frame(deconvolution.results.dwls), 'Sample') %>%
  gather(., key='celltype', value='cell_fraction', -Sample) %>%
  left_join(., patient.metadata)

Each condition has a different number of samples:

We can visualize the estimated cell fractions across samples with a boxplot, and group the results either by IHC subtype or by treatment status

HER2-positive samples are enriched in T cells, while the ER-positive and triple negative breast cancer (TNBC) samples show an enrichment in cancer and normal epithelial cells, respectively. The patient metadata includes information about eventual treatments undergone by the patients. We can see that 5 out of the 24 patients were treated with Neoadjuvant and/or Paclitaxel, while the other 19 were untreated. We can therefore visualize the distribution of cell fractions across treated and untreated patients

The major differences in this case are the enrichment of the CAFs and normal epithelial cells in the treated samples, as opposed to the cancer epithelial cells and B cells which have a higher median cell fraction in the untreated samples.

As mentioned before, with the lower level of annotations we can identify additional cell types, such as CD4+/CD8+ subtypes, Fibroblasts and cancer cells. We can therefore visualize the estimated cell fractions in a boxplot, focusing in particular on the CD4+ and CD8+ T-cell subtypes.

The estimates for the CD8+ T cells seem to be very low for all samples. On the other hand, in the HER2/ER positive samples the cell fractions estimated for the CD4+ T cells seem to be significantly higher than in the other subtypes. The estimates for the CD8+ T cells are close to zero for almost every sample, with very low fractions detected for the HER2+/ER+ ones.

Similarly, we can visualize the distribution of the estimates for the cancer associated fibroblasts (CAFs), respectively the myoblastic CAFs (myCAF) and the bone marrow derived inflammatory CAFs (MSC iCAF).

Here we can notice that, while the myoblastic CAFs have a comparable median value, the MSC CAFs have higher infiltration in the samples characterized as HER2+/ER+ subtype.

Finally, we can visualize the distribution of the different cancer cells molecular subtypes that were described by the authors.

We can see that the HER2+ shows an overrepresentation of the cancer cells described as HER2-like by the authors. Similarly, the ER+ samples show higher fractions of Luminal B cancer cells. These findings on bulk RNA-seq are concordant with the subtypings that the authors described on the single cell RNA-seq (‘Results - scSubtype’).