Unraveling the timeline of gene expression: A pseudotemporal trajectory analysis of single-cell RNA sequencing data

Version Changes

Revised. Amendments from Version 1

In this revised version, we have incorporated additional information and provided more comprehensive explanations regarding the individual scRNA-seq analysis, the integration analysis, and the downstream edgeR time-course analysis. We have also revised the design matrix used in the downstream pseudo-bulk time-course analysis. In particular, we have introduced a sample effect into the design, which enables the incorporation of differences between the five biological samples. This substantially increases the statistical power of the time-course analysis, resulting in the detection of more genes significantly associated with pseudotime. Figure 6 has been split into two separate Figures (Figure 6 and 7). This is because users need to first generate Figure 6 to decide the starting node before they proceed to the cell ordering and pseudotime calculation. The visualization of pseudotime (Figure 7) comes after that. Consequently, all the Figure numbers after Figure 7 will change accordingly.

Abstract

Background

Single-cell RNA sequencing (scRNA-seq) technologies have rapidly developed in recent years. The droplet-based single cell platforms enable the profiling of gene expression in tens of thousands of cells per sample. The goal of a typical scRNA-seq analysis is to identify different cell subpopulations and their respective marker genes. Additionally, trajectory analysis can be used to infer the developmental or differentiation trajectories of cells.

Methods

This article demonstrates a comprehensive workflow for performing trajectory inference and time course analysis on a multi-sample single-cell RNA-seq experiment of the mouse mammary gland. The workflow uses open-source R software packages and covers all steps of the analysis pipeline, including quality control, doublet prediction, normalization, integration, dimension reduction, cell clustering, trajectory inference, and pseudo-bulk time course analysis. Sample integration and cell clustering follows the Seurat pipeline while the trajectory inference is conducted using the monocle3 package. The pseudo-bulk time course analysis uses the quasi-likelihood framework of edgeR.

Results

Cells are ordered and positioned along a pseudotime trajectory that represented a biological process of cell differentiation and development. The study successfully identified genes that were significantly associated with pseudotime in the mouse mammary gland.

Conclusions

The demonstrated workflow provides a valuable resource for researchers conducting scRNA-seq analysis using open-source software packages. The study successfully demonstrated the usefulness of trajectory analysis for understanding the developmental or differentiation trajectories of cells. This analysis can be applied to various biological processes such as cell development or disease progression, and can help identify potential biomarkers or therapeutic targets.

Introduction

Single-cell RNA sequencing (scRNA-seq) has emerged as a popular technique for transcriptomic profiling of samples at the single-cell level. With droplet-based methods, thousands of cells can be sequenced in parallel using next-generation sequencing platforms. ^{1 , 2} One of the most widely used droplet-based scRNA-seq technologies is the 10x Genomics Chromium which enables profiling transcriptomes of tens of thousands of cells per sample. ³ A common goal of a scRNA-seq analysis is to investigate cell types and states in heterogeneous tissues. To achieve this, various pipelines have been developed, such as Seurat ⁴ and the Bioconductor’s OSCA pipeline, ⁵ and scanpy. ⁶ A typical scRNA-seq data analysis pipeline involves quality control, normalization, dimension reduction, cell clustering, and differential expression analysis.

With the advent of single-cell multiplexing technologies, the per sample cost of scRNA-seq experiments has significantly decreased. This makes it feasible and more affordable to conduct single-cell RNA-seq profiling across a variety of biological samples within a given experimental study. In a multiple sample single-cell experiment, an integration method is required to investigate all cells across all samples simultaneously. This ensures that sample and batch effects are appropriately considered in visualizing and clustering cells. Popular integration methods include the Seurat’s anchor-based integration method, ⁴ Harmony, ⁷ and the MNN. ⁸

After integration and cell clustering, differential expression analysis is often performed to identify marker genes for each cell cluster. Various methods have been developed at the single-cell level for finding marker genes. ^{9 , 10} Recently, the pseudo-bulk method has become increasingly popular due to its superior computational efficiency and its ability to consider biological variation between replicate samples. ¹¹ Under this approach, pseudo-bulk expression profiles are formed by aggregating read counts for all cells within the same group (e.g., cluster, cell type) and from the same sample.

Trajectory inference is another popular downstream analysis that aims to study cell differentiation or cell type development. Popular software tools to perform trajectory analysis include monocle3 ¹² and slingshot. ¹³ These methods learn trajectories based on the change of gene expression and order cells along a trajectory to obtain pseudotime. ^{14 , 15} This allows for pseudotime-based time course analysis in single-cell experiments, which is extremely useful for investigating specific biological questions of interest.

Here we present a new single-cell workflow that integrates trajectory analysis and pseudo-bulking to execute a single-cell pseudo time course analysis. The inputs for this workflow are single-cell count matrices, such as those generated by 10x Genomic’s cellranger. The methods involved open source packages in R. The single-cell QC, clustering and integration analyses are performed in Seurat, whereas the trajectory analysis is conducted using monocle3. Once the pseudo-bulk samples are created and assigned pseudotime, a time course analysis is conducted in edgeR. ¹⁶ The analysis pipeline presented in this article can be used for examining dynamic cellular changes along a specific trajectory in any single-cell RNA-seq experiment with replicate samples.

Description of the biological experiment

The scRNA-seq data used to demonstrate this workflow consists of five mouse mammary epithelium samples at five different stages: embryonic, early postnatal, pre-puberty, puberty and adult. The puberty sample is from the study in Pal et al. 2017, ¹⁷ whereas the other samples are from Pal et al. 2021. ¹⁸ These studies examined the stage-specific single-cell profiles in order to gain insight into the early developmental stages of mammary gland epithelial lineage. The cellranger count matrix outputs of these five samples are available on the GEO repository as series GSE103275 and GSE164017.

Data preparation

Downloading the data

The cellranger output of each sample consists of three key files: a count matrix in mtx.gz format, barcode information in tsv.gz format and feature (or gene) information in tsv.gz format.

The outputs of the mouse mammary epithelium at embryonic stage (E18.5), post-natal 5 days (P5), 2.5 weeks (Pre-puberty), and 10 weeks (Adult) can be downloaded from GSE164017, ¹⁸ whereas the output of mouse mammary epithelium at 5 weeks (Puberty) can be downloaded from GSE103275. ¹⁷

We first create a data directory to store all the data files.

> data_dir <- "data"
> if(!dir.exists(data_dir)){dir.create(data_dir, recursive=TRUE)}

We then download the barcode and count matrix files of the five samples.

> accessions <-c("GSM4994960","GSM4994962","GSM4994963","GSM2759554","GSM4994967")
> stages <- c("E18-ME", "Pre-D5-BL6", "Pre-BL6", "5wk-1", "Adult-BL6")
> file_suffixes <- c("barcodes.tsv.gz", "matrix.mtx.gz")
> for ( i in 1:length(accessions) ) {
+  for (file_suffix in file_suffixes) {
+    filename <- paste0(accessions[i],"_",stages[i],"-",file_suffix)
+    url <- paste0("http://www.ncbi.nlm.nih.gov/geo/download/?acc=",
+            accessions[i],"&","format=file&","file=",filename)
+    download.file(url=url,destfile=paste0(data_dir,"/",filename))
+  }
+ }

Since the five samples in this workflow are from two separate studies and were processed using different cellranger references but built from the same mouse genome (mm10), the feature information is slightly different between the two runs. In general, the same cellranger reference build is preferred for the sake of consistency, although the effect on the downstream analysis is negligible. Here, we download the feature information of both runs. The GSM2759554_5wk-1-genes.tsv.gz file contains the feature information for the 5wk-1 sample, whereas GSE164017_features.tsv.gz contains the feature information for the other four samples.

> GSE <- c("GSE164017", "GSM2759554")
> feature_filenames <- c("GSE164017_features.tsv.gz",
+               "GSM2759554_5wk-1-genes.tsv.gz")
> for (i in 1:length(GSE) ) {
+  url <- paste0("http://www.ncbi.nlm.nih.gov/geo/download/?acc=",
+           GSE[i],"&","format=file&","file=",feature_filenames[i])
+  download.file(url=url,destfile=paste0(data_dir,"/",feature_filenames[i]))
+ }

A target information file is created to store all the sample and file information.

> samples <- c("E18.5-epi", "P5", "Pre-puberty", "Puberty", "Adult")
> targets <- data.frame(
+   samples=samples,
+   stages=stages,
+   accessions=accessions,
+   matrix.file = paste0("data/",accessions[1:5],"_",stages[1:5],"-","matrix.mtx.gz"),
+   barcode.file = paste0("data/",accessions[1:5],"_",stages[1:5],"-","barcodes.tsv.gz"),
+   feature.file = paste0("data/",feature_filenames[c(1,1,1,2,1)]))
> targets

      samples     stages accessions                              matrix.file
1   E18.5-epi     E18-ME GSM4994960     data/GSM4994960_E18-ME-matrix.mtx.gz
2          P5 Pre-D5-BL6 GSM4994962 data/GSM4994962_Pre-D5-BL6-matrix.mtx.gz
3 Pre-puberty    Pre-BL6 GSM4994963    data/GSM4994963_Pre-BL6-matrix.mtx.gz
4     Puberty      5wk-1 GSM2759554      data/GSM2759554_5wk-1-matrix.mtx.gz
5       Adult  Adult-BL6 GSM4994967  data/GSM4994967_Adult-BL6-matrix.mtx.gz
                                barcode.file                       feature.file
1     data/GSM4994960_E18-ME-barcodes.tsv.gz     data/GSE164017_features.tsv.gz
2 data/GSM4994962_Pre-D5-BL6-barcodes.tsv.gz     data/GSE164017_features.tsv.gz
3    data/GSM4994963_Pre-BL6-barcodes.tsv.gz     data/GSE164017_features.tsv.gz
4      data/GSM2759554_5wk-1-barcodes.tsv.gz data/GSM2759554_5wk-1-genes.tsv.gz
5  data/GSM4994967_Adult-BL6-barcodes.tsv.gz     data/GSE164017_features.tsv.gz

Reading in the data

The downloaded cellranger outputs of all the samples can be read in one-by-one using the read10X function in the edgeR package. First, a DGElist object is created for each sample, which is then consolidated into a single DGElist object by merging them altogether.

> library(edgeR)
> dge_all <- list()
> for ( i in 1:5 ) {
+   y <- read10X(mtx = targets$matrix.file[i],
+     barcodes = targets$barcode.file[i], genes = targets$feature.file[i])
+   y$samples$group <- targets$samples[i]
+   colnames(y) <- paste0(targets$accessions[i],"-",y$samples$Barcode)
+   y$genes$Ensembl_geneid <- rownames(y)
+   y$genes <- y$genes[,c("Ensembl_geneid","Symbol")]
+   y <- y[!duplicated(y$genes$Symbol),]
+   rownames(y) <- y$genes$Symbol
+   dge_all[[i]] <- y
+ }
> rm(y)
> common.genes <- Reduce(intersect, lapply(dge_all, rownames))
> for(i in 1:5) dge_all[[i]] <- dge_all[[i]][common.genes, ]
> dge_merged <- do.call("cbind", dge_all)

The levels of group in the sample information data frame are reordered and renamed from the early embryonic stage to the late adult stage.

> dge_merged$samples$group <- factor(dge_merged$samples$group, levels=samples)

The number of genes, the total number of cells, and the number of cells in each sample are shown below.

> dim(dge_merged)

[1] 26589 33735

> table(dge_merged$samples$group)

  E18.5-epi         P5 Pre-puberty      Puberty      Adult
       6969       3886        4183         5428      13269

Single-cell RNA-seq analysis

Quality control

Quality control (QC) is essential for single-cell RNA-seq data analysis. Common choices of QC metrics include number of expressed genes or features, total number of reads, and proportion of reads mapped to mitochondrial genes in each cell. The number of expressed genes and mitochondria read percentage in each cell can be calculated as follows.

> dge_merged$samples$num_exp_gene <- colSums(dge_merged$counts>0)
> mito_genes <- rownames(dge_merged)[grep("^mt-",rownames(dge_merged))]
> dge_merged$samples$mito_percentage <-
+  colSums(dge_merged$counts[mito_genes,])/
+  colSums(dge_merged$counts)*100

These QC metrics can be visualized in the following scatter plots ( Figure 1).

> library(ggplot2)
> my_theme_ggplot <- theme_classic() +
+  theme(axis.text=element_text(size=12),
+      axis.title=element_text(size=15,face="bold"),
+      plot.title=element_text(size=15,face="bold",hjust=0.5),
+      plot.margin=margin(0.5, 0.5, 0.5, 0.5, "cm"))
> my_theme_facet <-
+  theme(strip.background=element_rect(colour="white",fill="white"),
+      strip.text=element_text(size=15, face="bold",color="black"))
> my_colors_15 <- c("cornflowerblue", "darkorchid1", "firebrick1", "gold",
+            "greenyellow", "mediumspringgreen", "mediumturquoise",
+            "orange1", "pink", "deeppink3", "violet", "magenta",
+            "goldenrod4", "cyan", "gray90")

> p1 <- ggplot(data = dge_merged$samples,
+         aes(x=num_exp_gene, y=lib.size, color = group ) ) +
+     geom_point(size=0.5, show.legend=FALSE) +
+     facet_wrap(group~., ncol=1) +
+     scale_color_manual(values=my_colors_15 ) +
+     labs(x="Number of genes", y="Number of reads") +
+     my_theme_ggplot + my_theme_facet
> p2 <- ggplot(data = dge_merged$samples,
+         aes(x = mito_percentage, y=lib.size, color = group ) ) +
+     geom_point(size = 0.5, show.legend = FALSE) +
+     facet_wrap(group~., ncol=1) +
+     scale_color_manual(values=my_colors_15) +
+     labs(x="Mito-percentage", y="Number of reads") +
+     my_theme_ggplot + my_theme_facet
> patchwork::wrap_plots(p1, p2, ncol=2)

Figure 1. Scatter plots of quality control metrics across all the samples.

Each dot represents a cell. The plots on the left show number of reads vs number of genes detected, whereas those on the right show number of reads vs mitochondria read percentage.

Cells with a very low number of genes (<500) are considered of low quality and hence are removed from the analysis. Cells with high mitochondria read percentage (>10%) are also removed as high expression level of mitochondrial genes indicate damaged or dead cells. In general, these QC thresholds are dependent on the study data and hence should be considered carefully. For example, quiescent cells may normally have low RNA expression levels, and metabolically active cells may have a high mitochondrial content. Cells expressing a large number of genes are also removed as they are likely to be doublets. Even though a separate doublet detection analysis is performed later on, we notice from our own practice that the combination of both doublet detection and the removal of cells with large counts works the best. Different thresholds are selected for different samples based on the distribution of the number of genes expressed. Here, we choose 5000, 6000, 6000, 3000, 4000 for E18.5-epi, P5, pre-puberty, puberty and adult samples, respectively. In this workflow, most of the single-cell analysis is conducted using the Seurat package. We begin by reading in the scRNA-seq data from the five samples together with an initial QC process. Specifically, we filter out genes expressed in fewer than 3 cells and cells expressing fewer than 200 genes for each sample. Then the abovementioned QC thresholds are applied in order to further remove cells of low quality in each sample. The data after QC are stored as a list of five Seurat objects.

> library(Seurat)
> n_genes_max <- c(5000, 6000, 6000, 3000, 4000)
> data_seurat <- list()
> for (i in 1:5) {
+  sel <- dge_merged$samples$group == samples[i]
+  y <- dge_merged[, sel]
+  data_seurat[[i]] <- CreateSeuratObject( counts=y$counts,
+     meta.data=y$samples, min.cells=3, min.features=200,
+     project=samples[i] )
+  data_seurat[[i]] <- subset( data_seurat[[i]],
+     subset = (nFeature_RNA > 500) & (nFeature_RNA < n_genes_max[i]) &
+           (mito_percentage < 10) )
+ }
> names(data_seurat) <- samples

Standard Seurat analysis of individual sample

A standard Seurat analysis is performed for each individual sample. This would provide us some general information on how each individual sample looks like and what cell types present within them. More details on how to perform a scRNA-seq analysis can be found in Seurat online vignettes.

For each individual sample analysis, the default log normalization method in NormalizeData is applied to each sample. The top 2000 highly variable genes (HVGs) are identified by FindVariableFeatures. The normalized data of the 2000 highly variable genes are scaled by ScaleData to have a mean of 0 and a variance of 1. The principal component analysis (PCA) dimension reduction is performed on the highly variable genes by RunPCA. Uniform manifold approximation and projection (UMAP) dimension reduction is performed on the first 30 PCs by RunUMAP. Here we use the first 30 PCs to be consistent with the analysis in Pal et al. 2021. ¹⁸ Based on the Seurat vignette and our own practice, the number of PCs chosen would not change the results dramatically if it is large enough (> 10). Cell clustering is performed individually for each sample by FindNeighbors and FindClusters, which by default uses the Louvain algorithm. Cell clustering resolution is carefully chosen for each sample so that distinct cell types are grouped into separate clusters. For this dataset, the cell clustering resolution is set at 0.1, 0.1, 0.2, 0.2 and 0.2 for E18.5-epi, P5, pre-puberty, puberty and adult, respectively.

> data_seurat <- lapply(data_seurat, NormalizeData)
> data_seurat <- lapply(data_seurat, FindVariableFeatures, nfeatures=2000)
> data_seurat <- lapply(data_seurat, ScaleData)
> data_seurat <- lapply(data_seurat, RunPCA, verbose = FALSE)
> data_seurat <- lapply(data_seurat, RunUMAP, reduction = "pca", dims = 1:30)
> data_seurat <- lapply(data_seurat, FindNeighbors, reduction="pca", dims=1:30)
> resolutions <- c(0.1, 0.1, 0.2, 0.2, 0.2)
> for(i in 1:5)
+   data_seurat[[i]] <- FindClusters(data_seurat[[i]],
+      resolution=resolutions[i], verbose=FALSE)

Removing potential doublets and non-epithelial cells

Although high-throughput droplet-based single-cell technologies can accurately capture individual cells, there are instances where a single droplet may contain two or more cells, which are known as doublets or multiplets. Here we use the scDblFinder package ¹⁹ to further remove potential doublets. To do that, each Seurat object in the list is first converted into a SingleCellExperiment object using the as.SingleCellExperiment function in Seurat. Then the scDblFinder function in the scDblFinder package is called to predict potential doublets on each SingleCellExperiment object. The scDblFinder output for each sample is stored in the corresponding Seurat object.

> library(scDblFinder)
> for (i in 1:5) {
+  sce <- as.SingleCellExperiment(DietSeurat(data_seurat[[i]],
+     graphs=c("pca","umap")) )
+  set.seed(42)
+  sce <- scDblFinder(sce)
+  data_seurat[[i]]$db_score <- sce$scDblFinder.score
+  data_seurat[[i]]$db_type <- factor( sce$scDblFinder.class,
+     levels=c("singlet", "doublet") )
+ }

The main object of this single-cell experiment is to examine the early developmental stages of the mouse epithelial mammary gland. Hence, for the rest of the analysis we will mainly focus on the epithelial cell population which is typically marked by the Epcam gene. The cell clustering, the expression level of Epcam and doublet prediction results of each sample are shown below ( Figure 2).

> p1 <- lapply(data_seurat,function(x){DimPlot(x, pt.size=0.1, cols=my_colors_15) +
+          ggtitle(x$group[1]) + theme(plot.title=element_text(hjust=0.5))})
> p2 <- lapply(data_seurat, FeaturePlot, feature="Epcam", pt.size=0.1)
> p3 <- lapply(data_seurat, DimPlot, group.by="db_type", pt.size=0.1,
+         cols=c("gray90", "firebrick1"))
> patchwork::wrap_plots(c(p1,p2,p3), nrow=5, byrow=FALSE)

Figure 2. UMAP visualization of each individual samples.

The UMAP plots, in sequence from the top row to the bottom row, correspond to E18.5-epi, P5, Pre-puberty, Puberty, and Adult, respectively. In each row, cells are coloured by cluster on the left, by Epcam expression level in the middle, and by doublet prediction on the right.

By examining the expression level of the Epcam gene, together with some other known marker genes of basal, LP and ML, we select the following clusters in each sample as the epithelial cell population.

> epi_clusters <- list(
+  "E18.5-epi" = 0,
+  "P5" = c(1,3),
+  "Pre-puberty" = c(0:2, 5),
+  "Puberty" = 0:6,
+  "Adult" = 0:3
+ )

Cells that are non-epithelial and those identified as potential doublets by scDblFinder are excluded from the subsequent analysis. The cellular barcodes of the remaining epithelial cells from each sample are stored in the list object called epi_cells. The respective number of epithelial cells that are retained for each sample is shown below.

> epi_cells <- list()
> for (i in samples) {
+  epi_cells[[i]] <- rownames(
+   subset(data_seurat[[i]]@meta.data,
+    (db_type == "singlet") & (seurat_clusters %in% epi_clusters[[i]])))
+ }
> do.call(c, lapply(epi_cells, length))

  E18.5-epi          P5 Pre-puberty      Puberty      Adult
       4343        1140        2546         4706       9341

Data integration

Integrating epithelial cells of five samples

Since we have five individual scRNA-seq samples, conducting an integration analysis is necessary to explore all cells across these samples simultaneously. In this workflow, we use the default anchor-based integration method of the Seurat package. Depending on the single-cell analysis workflow, users are free to use other integration methods they may prefer (e.g., Harmony and MNN). A Seurat object is first created from the merged DGEList object of epithelial cells using CreateSeuratObject function without filtering any cells ( min.features is set to 0). Lowly expressed genes are removed as they are not of any biological interest here. Here we keep genes expressed in at least 3 cells in each sample ( min.cells is set to 3) although different thresholds can be adopted in general depending on the data.

> epi_cells <- do.call(c, epi_cells)
> dge_merged_epi <- dge_merged[, epi_cells]
> seurat_merged <- CreateSeuratObject(counts = dge_merged_epi$counts,
+              meta.data = dge_merged_epi$samples,
+              min.cells = 3, min.features = 0, project = "mammary_epi")

Then the Seurat object is split into a list of five Seurat objects, where each object corresponds to one of the five samples. For each sample, the log normalization method is applied to normalize the raw count by NormalizeData, and highly variable genes are identified by FindVariableFeatures.

> seurat_epi <- SplitObject(seurat_merged, split.by = "group")
> seurat_epi <- lapply(seurat_epi, NormalizeData)
> seurat_epi <- lapply(seurat_epi, FindVariableFeatures, nfeatures = 2000)

The feature genes used for integration are chosen by SelectIntegrationFeatures, and these genes are used to identify anchors for integration by FindIntegrationAnchors. The integration process is performed by IntegrateData based on the identified anchors. Please note that the integration step is computationally intensive and might take a substantial amount of time to complete (20-40 minutes depending on the computational resource).

> anchor_features <- SelectIntegrationFeatures(seurat_epi,
+              nfeatures = 2000, verbose = FALSE)
> anchors <- FindIntegrationAnchors(seurat_epi, verbose = FALSE,
+              anchor.features = anchor_features)
> seurat_int <- IntegrateData(anchors, verbose = FALSE)

The integrated data are then scaled to have a mean of 0 and a variance of 1 by ScaleData. PCA is performed on the scaled data using RunPCA, followed by UMAP using RunUMAP. Same as before, we use 30 PCs for the sake of consistency and the results would not change dramatically provided a good amount of PCs (>10) are used. Cell clusters of the integrated data are identified by using FindNeighbors and FindClusters. We choose 0.2 as the cell clustering resolution after experimenting with different resolution parameters. This is because under this resolution the three major epithelial subpopulations, two intermediate cell clusters, and a small group of stroma cells can be clearly separated in distinct cell clusters.

> DefaultAssay(seurat_int) <- "integrated"
> seurat_int <- ScaleData(seurat_int, verbose = FALSE)
> seurat_int <- RunPCA(seurat_int, npcs = 30, verbose = FALSE)
> seurat_int <- RunUMAP(seurat_int, reduction = "pca",
+           dims = 1:30, verbose = FALSE)
> seurat_int <- FindNeighbors(seurat_int, dims = 1:30, verbose = FALSE)
> seurat_int <- FindClusters(seurat_int, resolution = 0.2, verbose = FALSE)

UMAP plots are generated to visualize the integration and cell clustering results ( Figure 3). The UMAP plot indicates the presence of three major cell clusters (cluster 0, 1, and 2), which are bridged by intermediate clusters located in between them. Cells at the later stages largely dominate the three major cell clusters, while cells at the earlier stages are predominantly present in the intermediate clusters in the middle.

> seurat_int$group <- factor(seurat_int$group, levels = samples)
> p1 <- DimPlot(seurat_int, pt.size = 0.1, cols = my_colors_15)
> p2 <- DimPlot(seurat_int, pt.size = 0.1, group.by = "group",
+       shuffle = TRUE, cols = my_colors_15) + labs(title="")
> p1 | p2

Figure 3. UMAP visualization of the integrated data.

Cells are coloured by cluster on the left and by original sample on the right.

Cell type identification

The mammary gland epithelium consists of three major cell types: basal myoepithelial cells, luminal progenitor (LP) cells and mature luminal (ML) cells. These three major epithelial cell populations have been well studied in the literature. By examining the classic marker genes of the three cell types, we are able to identify basal, LP and ML cell populations in the integrated data ( Figure 4). Here we use Krt14 and Acta2 for basal, Csn3 and Elf5 for LP, and Prlr and Areg for ML. We also examine the expression level of Hmgb2 and Mki67 as they are typical markers for cycling cells and the expression level of Igfbp7 and Fabp4 as they are marker genes for stromal cells.

> markers <- c("Krt14", "Acta2", "Csn3","Elf5", "Prlr","Areg",
+              "Hmgb2", "Mki67", "Igfbp7","Fabp4")
> DefaultAssay(seurat_int) <- "RNA"
> FeaturePlot(seurat_int, order = TRUE, pt.size = 0.1, features = markers, ncol = 2)

Figure 4. Feature plots of the integrated data.

Genes from the top row to the bottom rows are the markers of basal, LP, ML, cycling, and stromal cells, respectively.

Based on the feature plots, cluster 1, cluster 2 and cluster 0 represent the basal, LP and ML cell populations, respectively. Cluster 4 mainly consists of cycling cells, whereas cluster 3 seems to be a luminal intermediate cell cluster expressing both LP and ML markers. Cluster 5 consists of a few non-epithelial (stromal) cells that have not been filtered out previously.

The number of cells in each cluster for each sample is shown below.

> tab_number <- table(seurat_int$group, seurat_int$seurat_clusters)
> tab_number

                0    1    2    3    4   5
 E18.5-epi   171  878   29 2341  921   3
 P5          272  347   47  351  120   3
 Pre-puberty 381 1590  482   64   23   6
 Puberty    1894  986 1535   20  265   6
 Adult      4281 2495 2362  171   32   0

The proportion of cells in each cluster is calculated for each sample to compare the variation in cell composition across different stages.

> tab_ratio <- round(100*tab_number/rowSums(tab_number), 2)
> tab_ratio <- as.data.frame.matrix(tab_ratio)
> tab_ratio

                0    1     2     3     4    5
E18.5-epi    3.94 20.2  0.67 53.90 21.21 0.07
P5          23.86 30.4  4.12 30.79 10.53 0.26
Pre-puberty 14.96 62.5 18.93  2.51  0.90 0.24
Puberty     40.25 20.9 32.62  0.42  5.63 0.13
Adult       45.83 26.7 25.29  1.83  0.34 0.00

The bar plot ( Figure 5) shows the proportion of different cell types in samples at different developmental stages. Specifically, the proportion of basal cells (purple) demonstrates an ascending trend from E18.5 to pre-puberty stage, after which it declines towards adult stage. The LP cell proportion (red) rises from E18.5 to puberty stage, followed by a slight dip at adult stage. Although the proportion of ML cells (blue) is higher at P5 than pre-puberty stage, it shows an overall increasing trend. Cycling cells (green) constitute the highest proportion at E18.5 stage, but decrease to a smaller proportion at pre-puberty stage, with a slight increase at puberty stage, and subsequently, they reduce to a negligible proportion at adult stage. The augmented cycling cell proportion at puberty stage aligns with the ductal morphogenesis characteristics of the mammary gland. The luminal intermediate cell proportion (yellow) displays a decreasing trend from E18.5 stage to adult stage.

> par(mar=c(5, 7, 1, 7), xpd=TRUE)
> barplot(t(tab_ratio), col=my_colors_15, xlab="Cell proportion (%)",
+   horiz = TRUE, las=1)
> legend("right", inset = c(-0.3,0), legend = 0:5, pch = 15,
+   col=my_colors_15, title="Cluster")

Figure 5. Bar plot of cell proportion of each cluster in each sample.

Trajectory analysis with monocle3

Constructing trajectories and pseudotime

Many biological processes manifest as a dynamic sequence of alterations in the cellular state, which can be estimated through a “trajectory” analysis. Such analysis is instrumental in detecting the shifts between different cell identities and modeling gene expression dynamics. By treating single-cell data as a snapshot of an uninterrupted process, the analysis establishes the sequence of cellular states that forms the process trajectory. The arrangement of cells along these trajectories can be interpreted as pseudotime.

Here, we use the monocle3 package to infer the development trajectory in the mouse mammary gland epithelial cell population. The Seurat object of the integrated data is first converted into a cell_data_set object to be used in monocle3.

> library(monocle3)
> cds_obj <- SeuratWrappers::as.cell_data_set(seurat_int)

monocle3 re-clusters cells to assign them to specific clusters and partitions, which are subsequently leveraged to construct trajectories. If multiple partitions are used, each partition will represent a distinct trajectory. The calculation of pseudotime, which indicates the distance between a cell and the starting cell in a trajectory, is conducted during the trajectory learning process. These are done using the cluster_cells and learn_graph functions. To obtain a single trajectory and avoid a loop structure, both use_partition and close_loop are turned off in learn_graph.

> set.seed(42)
> cds_obj <- cluster_cells(cds_obj)
> cds_obj <- learn_graph(cds_obj, use_partition=FALSE, close_loop=FALSE)

Visualizing trajectories and pseudotime

The plot_cells function of monocle3 is used to generate a trajectory plot that superimposes the trajectory information onto the UMAP representation of the integrated data. By adjusting the label_principal_points parameter, the names of roots, leaves, and branch points can be displayed. Cells in the trajectory UMAP plot ( Figure 6) are coloured by cell cluster identified in the previous Seurat integration analysis.

Figure 6. UMAP visualization of trajectory inferred by monocle3.

Cells are coloured by cluster.

> p1 <- plot_cells(cds_obj, color_cells_by="seurat_clusters",
+           group_label_size=4, graph_label_size=3,
+           label_cell_groups=FALSE, label_principal_points=TRUE,
+           label_groups_by_cluster=FALSE) +
+    scale_color_manual(values = my_colors_15)
> p1

Along the monocle3 trajectory analysis, several nodes are identified and marked with black circular dots on the resulting plot, representing principal nodes along the trajectories. To establish the order of cells and calculate their corresponding pseudotime, it is necessary to select a starting node among the identified principal nodes. For this analysis, node “Y_65” in the basal population (cluster 1) was selected as the starting node, as mammary stem cells are known to be enriched in the basal population and give rise to LP and ML cells in the epithelial lineage. ²⁰ It should be noted that node numbers may vary depending on the version of monocle3 used.

> cds_obj <- order_cells(cds_obj, root_pr_nodes="Y_65")

The cells are then ordered and assigned pseudotime values by the order_cells function in monocle3. The resulting pseudotime information can be visualized on the UMAP plot by using the plot_cells function, as demonstrated in the UMAP plot ( Figure 7).

> p2 <- plot_cells(cds_obj, color_cells_by="pseudotime",
+           label_groups_by_cluster=FALSE, label_leaves=FALSE,
+           label_branch_points=FALSE)
> p2

Figure 7. UMAP visualization of pseudotime computed by monocle3.

Cells are coloured by pseudotime.

The pseudotime function in monocle3 allows users to extract the pseudotime values of the cells from a cell_data_set object. This information can then be stored in the metadata of the Seurat object for further analysis.

> seurat_int$pseudotime <- pseudotime(cds_obj)

Pseudo-bulk time course analysis with edgeR

Constructing pseudo-bulk profiles

After obtaining the pseudotime of each cell, we proceed to a time course analysis to identify genes that change significantly along the pseudotime. Our approach involves creating pseudo-bulk samples using a pseudo-bulking approach and performing an edgeR-style time course analysis.

To create the pseudo-bulk samples, read counts are aggregated for all cells with the same combination of sample and cluster. The number of cells used to construct each pseudo-bulk sample is added to the sample metadata. For simplicity, the average pseudotime of all cells in each pseudo-bulk sample is used as the pseudotime for that sample. One could also use the median of the cellwise pseudotime instead of the mean, but the results will not change dramatically.

> y <- dge_merged[, colnames(seurat_int)]
> y$samples <- cbind(y$samples[, 1:3],
+   seurat_int@meta.data[, c("seurat_clusters", "pseudotime")])
> sample_cluster <- paste0(y$samples$group, "_C", y$samples$seurat_clusters)
> avg_pseudotime <- tapply(y$samples$pseudotime, sample_cluster, mean)
> cell_number <- table(sample_cluster)
> y <- sumTechReps(y, ID = sample_cluster)
> y$samples$pseudotime <- avg_pseudotime[colnames(y)]
> y$samples$cell_number <- cell_number[colnames(y)]

The Entrez gene IDs are added to the gene information. Genes with no valid Entrez gene IDs are removed from the downstream analysis.

> library(org.Mm.eg.db)
> entrez_id <- select(org.Mm.eg.db, keys = y$genes$Symbol,
+         columns = c("ENTREZID", "SYMBOL"), keytype = "SYMBOL")
> y$genes$ENTREZID <- entrez_id$ENTREZID
> y <- y[!is.na(y$genes$ENTREZID), ]

The samples are ordered by average pseudotime for the following analysis.

> y <- y[, order(y$samples$pseudotime)]

Filtering and normalization

We now proceed to the standard edgeR analysis pipeline, which starts with filtering and normalization. The sample information, such as library sizes, average pseudotime and cell numbers, are shown below.

> y$samples[, c("lib.size", "pseudotime", "cell_number")]

               lib.size pseudotime cell_number
Pre-puberty_C1 11886898       4.65        1590
Adult_C1        9285265       4.77        2495
P5_C1           2680089       6.41         347
Puberty_C1      3112796       6.48         986
E18.5-epi_C1    8084434      10.16         878
E18.5-epi_C5       5179      15.61           3
P5_C5             24491      15.61           3
Pre-puberty_C5    57834      15.61           6
Puberty_C5        12278      15.61           6
Adult_C4         204212      19.25          32
E18.5-epi_C4   11725731      19.31         921
P5_C4           1917228      19.68         120
Puberty_C4      1860770      19.72         265
Pre-puberty_C4   289370      22.10          23
E18.5-epi_C3   15806768      28.03        2341
Puberty_C2      5841364      28.62        1535
P5_C3           3862152      28.94         351
E18.5-epi_C2     167242      29.14          29
Adult_C2        8320630      29.66        2362
P5_C2            347365      29.67          47
Pre-puberty_C2  4625160      30.51         482
Puberty_C3        63722      31.73          20
Pre-puberty_C3  1432336      32.64          64
Adult_C3         997670      33.99         171
Pre-puberty_C0  6024872      38.90         381
E18.5-epi_C0     990670      39.64         171
Adult_C0       27621462      40.44        4281
P5_C0           3113053      40.68         272
Puberty_C0      8806924      41.09        1894

To ensure the reliability of the analysis, it is recommended to remove pseudo-bulk samples that are constructed from a small number of cells. We suggest each pseudo-bulk sample should contain at least 30 cells. In this analysis, we identified seven pseudo-bulk samples that were constructed with less than 30 cells and removed them form the analysis.

> keep_samples <- y$samples$cell_number > 30
> y <- y[, keep_samples]

Genes with very low count number are also removed from the analysis. This is performed by the filterByExpr function in edgeR.

> keep_genes <- filterByExpr(y)
> y <- y[keep_genes, , keep.lib.sizes=FALSE]

The number of genes and samples after filtering are shown below.

> dim(y)
[1] 11550   22

Normalization is performed by the trimmed mean of M values (TMM) method ²¹ implemented in the calcNormFactors function in edgeR.

> y <- calcNormFactors(y)

A Multi-dimensional scaling (MDS) plot serves as a valuable diagnostic tool for investigating the relationship among samples. MDS plots are produced using the plotMDS function in edgeR ( Figure 8).

> par(mar = c(5.1, 5.1, 2.1, 2.1), mfrow=c(1,2))
> cluster <- y$samples$seurat_clusters
> group <- y$samples$group
> plotMDS(y, labels = round(y$samples$pseudotime, 2),
+   xlim=c(-6,4), ylim=c(-3,3), col=my_colors_15[cluster])
> legend("topleft", legend=levels(cluster), col=my_colors_15, pch=16)
> plotMDS(y, labels = round(y$samples$pseudotime, 2),
+   xlim=c(-6,4), ylim=c(-3,3), col=my_colors_15[group])
> legend("topleft", legend=levels(group), col=my_colors_15, pch=16)

Figure 8. Multi-dimensional scaling (MDS) plot of the pseudo-bulk samples labelled by pseudotime.

Samples are coloured by original cell cluster on the left and by developmental stage on the right.

On the MDS plot, pseudo-bulk samples derived from the same cell cluster are close to each other. The samples are positioned in ascending order of pseudotime from left to right, suggesting a continuous shift in the gene expression profile throughout the pseudotime.

Design matrix

The aim of a time course experiment is to examine the relationship between gene abundances and time points. Assuming gene expression changes smoothly over time, we use a natural cubic spline with degrees of freedom of 3 to model gene expression along the pseudotime. In general, any degrees of freedom in range of 3 to 5 is reasonable provided there are sufficient time points for the degrees of freedom of the residuals.

The spline design matrix is generated by ns function in splines. The three spline coefficients of the design matrix (i.e., Z1, Z2 and Z3) do not have any particular meaning in general. However, we can re-parametrize the design matrix using QR decomposition so that the first coefficient Z1 represents the linear trend in pseudotime.

> t1 <- y$samples$pseudotime
> X <- splines::ns(as.numeric(t1),df = 3)
> A <- cbind(1,t1,X)
> QR <- qr(A)
> r <- QR$rank
> R_rank <- QR$qr[1:r,1:r]
> Z <- t(backsolve(R_rank,t(A),transpose=TRUE))
> Z <- Z[,-1]
> design <- model.matrix(~ Z)

Since the five samples are from different timepoints, the pseudo-bulk samples derived from these five samples are not independent replicates. The sample effect at the pseudo-bulk level can also be seen from the MDS plot ( Figure 8 right). Hence, we add the sample effect to the design in addition to the re-parametrized spline coefficients. The full design matrix is shown below.

> group <- y$samples$group
> design <- model.matrix(~ Z + group)
> colnames(design) <- gsub("group", "", colnames(design))
> design
   (Intercept)      Z1      Z2      Z3 P5 Pre-puberty Puberty Adult
1            1 -0.3593 -0.0550 -0.3206  0           1       0     0
2            1 -0.3572 -0.0572 -0.3116  0           0       0     1
3            1 -0.3285 -0.0887 -0.1837  1           0       0     0
4            1 -0.3271 -0.0901 -0.1780  0           0       1     0
5            1 -0.2626 -0.1489  0.0887  0           0       0     0
6            1 -0.1034 -0.0918  0.4047  0           0       0     1
7            1 -0.1024 -0.0900  0.4042  0           0       0     0
8            1 -0.0958 -0.0775  0.4002  1           0       0     0
9            1 -0.0951 -0.0761  0.3997  0           0       1     0
10           1  0.0505  0.2625  0.0541  0           0       0     0
11           1  0.0609  0.2746  0.0262  0           0       1     0
12           1  0.0666  0.2796  0.0118  1           0       0     0
13           1  0.0790  0.2862 -0.0180  0           0       0     1
14           1  0.0792  0.2863 -0.0185  1           0       0     0
15           1  0.0940  0.2854 -0.0491  0           1       0     0
16           1  0.1313  0.2389 -0.1021  0           1       0     0
17           1  0.1551  0.1804 -0.1195  0           0       0     1
18           1  0.2410 -0.1584 -0.1106  0           1       0     0
19           1  0.2540 -0.2199 -0.1034  0           0       0     0
20           1  0.2682 -0.2885 -0.0947  0           0       0     1
21           1  0.2722 -0.3084 -0.0922  1           0       0     0
22           1  0.2794 -0.3435 -0.0876  0           0       1     0
attr(,"assign")
[1] 0 1 1 1 2 2 2 2
attr(,"contrasts")
attr(,"contrasts")$group
[1] "contr.treatment"

Dispersion estimation

The edgeR package uses negative binomial (NB) distribution to model read counts of each gene across all the sample. The NB dispersions are estimated by the estimateDisp function. The estimated common, trended and gene-specific dispersions can be visualized by plotBCV ( Figure 9).

> y <- estimateDisp(y, design)
> sqrt(y$common.dispersion)

[1] 0.588

> plotBCV(y)

Figure 9. A scatter plot of the biological coefficient of variation (BCV) against the average abundance of each gene in log2 count-per-million (CPM).

The square-root estimates of the common, trended and gene-wise NB dispersions are shown.

The NB model can be extended with quasi-likelihood (QL) methods to account for gene-specific variability from both biological and technical sources. ^{22 , 23} Note that only the trended NB dispersion is used in the QL method. The gene-specific variability is captured by the QL dispersion, which is the dispersion parameter of the negative binomial QL generalized linear model.

The glmQLFit function is used to fit a QL model and estimate QL dispersions. The QL dispersion estimates can be visualized by plotQLDisp ( Figure 10).

> fit <- glmQLFit(y, design, robust=TRUE)
> plotQLDisp(fit)

Figure 10. A scatter plot of the quarter-root QL dispersion against the average abundance of each gene in log2 count-per-million (CPM).

Estimates are shown for the raw, trended and squeezed dispersions.

Time course trend analysis

The QL F-tests are performed by glmQLFTest in edgeR to identify genes that change significantly along the pseudotime. The tests are conducted on all three covariates of the spline model matrix. This is because the significance of any of the three coefficients would indicate a strong correlation between gene expression and pseudotime.

> res <- glmQLFTest(fit, coef=2:4)

The number of genes significantly associated with pseudotime (FDR < 0.05) are shown below.

> summary(decideTests(res))

       Z3-Z2-Z1
NotSig     4843
Sig        6707

Top significant genes can be viewed by topTags.

> topTags(res, n=10L)

Coefficient:  Z1 Z2 Z3
            Ensembl_geneid  Symbol ENTREZID logFC.Z1 logFC.Z2 logFC.Z3 logCPM
Fhod3   ENSMUSG00000034295   Fhod3   225288   -13.88   1.1400    1.201   4.73
Mlph    ENSMUSG00000026303    Mlph   171531    10.83   1.5074    2.449   5.98
Luzp2   ENSMUSG00000063297   Luzp2   233271   -13.93   0.0755    2.317   2.16
Ptpre   ENSMUSG00000041836   Ptpre    19267    -8.28   1.9119    0.113   5.45
Aoc1    ENSMUSG00000029811    Aoc1    76507    16.04  -3.3051  -18.932   5.02
Col27a1 ENSMUSG00000045672 Col27a1   373864    -8.71   1.0071   -2.231   4.02
Jph2    ENSMUSG00000017817    Jph2    59091   -15.64  -5.8376    3.242   2.20
Popdc2  ENSMUSG00000022803  Popdc2    64082   -17.44  -7.3846    3.814   2.32
Myh11   ENSMUSG00000018830   Myh11    17880   -16.38  -3.4680    2.343   7.94
Tns1    ENSMUSG00000055322    Tns1    21961    -9.56   1.1794   -3.453   3.87
            F   PValue      FDR
Fhod3   331.3 4.26e-16 4.92e-12
Mlph    136.9 1.07e-12 6.17e-09
Luzp2   121.9 2.92e-12 9.08e-09
Ptpre   118.6 3.72e-12 9.08e-09
Aoc1    469.7 4.27e-12 9.08e-09
Col27a1 115.4 4.72e-12 9.08e-09
Jph2    104.5 1.11e-11 1.83e-08
Popdc2  101.0 1.49e-11 2.15e-08
Myh11   121.6 2.03e-11 2.60e-08
Tns1     90.7 3.72e-11 4.25e-08

The logFC.Z1, logFC.Z2, and logFC.Z3 values in the table above denote the estimated coefficients of Z1, Z2, and Z3 for each gene. It should be noted that these values do not carry the same interpretation as log-fold changes in traditional RNA-seq differential expression analysis. For each gene, the sign of the coefficient logFC.Z1 indicates whether the expression level of that gene increases or decreases along pseudotime in general. The top increasing and the top decreasing genes are listed below.

> tab <- topTags(res, n=Inf)$table
> tab$trend <- ifelse(tab$logFC.Z1 > 0, "Up", "Down")
> tab.up <- tab[tab$trend == "Up", ]
> tab.down <- tab[tab$trend == "Down", ]
> head(tab.up)

           Ensembl_geneid Symbol ENTREZID logFC.Z1 logFC.Z2 logFC.Z3 logCPM     F
Mlph   ENSMUSG00000026303   Mlph   171531    10.83     1.51    2.449   5.98 136.9
Aoc1   ENSMUSG00000029811   Aoc1    76507    16.04    -3.31  -18.932   5.02 469.7
Mpzl3  ENSMUSG00000070305  Mpzl3   319742     5.37     0.92    0.723   4.90  89.8
Prr15l ENSMUSG00000047040 Prr15l   217138    13.35     2.39    3.568   4.63  87.7
Elf5   ENSMUSG00000027186   Elf5    13711     5.90     8.90    7.192   6.26  97.5
Lrrc26 ENSMUSG00000026961 Lrrc26   227618    13.99     3.05    3.276   4.68  82.9
         PValue      FDR trend
Mlph   1.07e-12 6.17e-09    Up
Aoc1   4.27e-12 9.08e-09    Up
Mpzl3  4.04e-11 4.25e-08    Up
Prr15l 4.97e-11 4.78e-08    Up
Elf5   5.80e-11 5.06e-08    Up
Lrrc26 8.00e-11 5.44e-08    Up

> head(tab.down)

            Ensembl_geneid  Symbol ENTREZID logFC.Z1 logFC.Z2 logFC.Z3 logCPM   F
Fhod3   ENSMUSG00000034295   Fhod3   225288   -13.88   1.1400    1.201   4.73 331
Luzp2   ENSMUSG00000063297   Luzp2   233271   -13.93   0.0755    2.317   2.16 122
Ptpre   ENSMUSG00000041836   Ptpre    19267    -8.28   1.9119    0.113   5.45 119
Col27a1 ENSMUSG00000045672 Col27a1   373864    -8.71   1.0071   -2.231   4.02 115
Jph2    ENSMUSG00000017817    Jph2    59091   -15.64  -5.8376    3.242   2.20 104
Popdc2  ENSMUSG00000022803  Popdc2    64082   -17.44  -7.3846    3.814   2.32 101
          PValue      FDR trend
Fhod3   4.26e-16 4.92e-12  Down
Luzp2   2.92e-12 9.08e-09  Down
Ptpre   3.72e-12 9.08e-09  Down
Col27a1 4.72e-12 9.08e-09  Down
Jph2    1.11e-11 1.83e-08  Down
Popdc2  1.49e-11 2.15e-08  Down

Line graphs are produced to visualize the relationship between gene expression level and pseudotime for the top 6 increasing and the top 6 decreasing genes ( Figure 11).

Figure 11. Line graphs of expression level of top genes along pseudotime.

The red line represents the predicted expression level in log2-CPM along pseudotime.

For each gene, the expression levels (in log2-CPM) are averaged across five samples, and the line is smoothed using its predicted expression level at 100 evenly spaced pseudotime points within the pseudotime range. The smooth curves for the first 6 genes exhibit a generally increasing trend in gene expression over pseudotime, while the curves for the last 6 genes show a general decreasing trend.

> design2 <- model.matrix(~ X + group)
> fit2 <- glmQLFit(y, design2, robust=TRUE)
> pt <- y$samples$pseudotime
> pt_new <- round(seq(min(pt), max(pt), length.out=100), 2)
> X_new <- predict(X, newx=pt_new)
> topGenes <- c(rownames(tab.up)[1:6], rownames(tab.down)[1:6])
> par(mfrow=c(4,3))
> for(i in 1:12) {
+   Symbol <- topGenes [i]
+   beta <- coef(fit2)[Symbol,]
+   AverageIntercept <- beta[1] + mean(c(0,beta[5:8]))
+   Trend <- AverageIntercept + X_new %*% beta[2:4]
+   Trend <- (Trend + log(1e6))/log(2)
+   plot(pt_new, Trend,type="l", frame=FALSE, col="red", lwd=2,
         xlab="Pseudotime", ylab="Log2CPM", main=Symbol)
+}

A heatmap is generated to examine the top 20 up and top 20 down genes collectively ( Figure 12). In the heatmap, pseudo-bulk samples are arranged in increasing pseudotime from left to right. The up genes are on the top half of the heatmap whereas the down genes are on the bottom half. The heatmap shows a gradual increase in expression levels of the up genes from left to right, while the down genes display the opposite trend.

> logCPM.obs <- edgeR::cpm(y, log=TRUE, prior.count=fit$prior.count)
> topGenes <- c(rownames(tab.up)[1:20], rownames(tab.down)[1:20])
> z <- logCPM.obs[topGenes, ]
> z <- t(scale(t(z)))
> ComplexHeatmap::Heatmap(z, name = "Z score",
+   cluster_rows = FALSE,cluster_columns = FALSE)

Figure 12. Heatmap of top 20 up and top 20 down genes.

Rows are genes and columns are pseudo-bulk samples.

Time course functional enrichment analysis

Gene ontology analysis

To interpret the results of the time course analysis at the functional level, we perform gene set enrichment analysis. Gene ontology (GO) is one of the commonly used databases for this purpose. The GO terms in the GO databases are categorized into three classes: biological process (BP), cellular component (CC) and molecular function (MF). In a GO analysis, we are interested in finding GO terms that are over-represented or enriched with significant genes.

GO analysis is usually directional. For simplicity, we re-perform the QL F-test on the Z1 coefficient to identify genes that exhibit a general linear increase or decrease along pseudotime. The numbers of genes with a significant increasing or decreasing linear trend are shown below.

> res_2 <- glmQLFTest(fit, coef=2)
> summary(decideTests(res_2))

         Z1
Down   2742
NotSig 6273
Up     2535

To perform a GO analysis, we apply the goana function to the above test results. Note that Entrez gene IDs are required for goana, which has been added to the ENTREZID column in the gene annotation. The top enriched GO terms can be viewed using topGO function.

> go <- goana(res_2, geneid="ENTREZID", species="Mm")
> topGO(go, truncate.term = 30, n=15)

                                     Term Ont    N  Up Down    P. Up   P.Down
GO:0071944                 cell periphery  CC 2928 695 1029 0.003837 4.73e-60
GO:0009653 anatomical structure morpho...  BP 1828 345  669 0.999801 6.61e-42
GO:0005576           extracellular region  CC  995 212  415 0.707860 1.19e-39
GO:0030312 external encapsulating stru...  CC  297  43  174 0.999620 9.63e-39
GO:0031012           extracellular matrix  CC  297  43  174 0.999620 9.63e-39
GO:0005886                plasma membrane  CC 2650 646  877 0.000352 2.21e-36
GO:0062023 collagen-containing extrace...  CC  240  34  145 0.999223 1.69e-34
GO:0007155                  cell adhesion  BP  833 169  350 0.894200 5.69e-34
GO:0040011                     locomotion  BP 1186 245  458 0.879526 6.97e-34
GO:0006928 movement of cell or subcell...  BP 1361 282  508 0.885421 2.59e-33
GO:0032501 multicellular organismal pr...  BP 4101 838 1238 0.998415 4.06e-33
GO:0048731             system development  BP 2741 536  886 0.999787 2.18e-32
GO:0032502          developmental process  BP 3917 776 1188 0.999971 3.33e-32
GO:0048856 anatomical structure develo...  BP 3657 724 1118 0.999941 3.45e-31
GO:0007275 multicellular organism deve...  BP 3188 615  995 0.999992 1.74e-30

It can be seen that most of the top GO terms are down-regulated. Here, we choose the top 10 down-regulated terms for each GO category and show the results in a bar plot ( Figure 13).

> top_go <- rbind.data.frame(topGO(go, ont =c("BP"), sort="Down",n=10),
+                            topGO(go, ont =c("CC"), sort="Down",n=10),
+                            topGO(go, ont =c("MF"), sort="Down",n=10))
> d <- transform(top_go, P_DE = P.Down, neg_log10_P = -log10(P.Down))
> d$Term <- factor(d$Term,levels = d$Term)
> ggplot(data = d, aes(x = neg_log10_P, y = Term, fill = Ont) ) +
+  geom_bar(stat = "identity", show.legend = TRUE) +
+  labs(x="-log10 (P value)", y="", title = "Down") +
+  facet_grid(Ont~.,scales = "free",space = "free") +
+  my_theme_ggplot + my_theme_facet +
+  scale_fill_manual(values = my_colors_15[-2]) +
+  theme(strip.text = ggplot2::element_blank())

Figure 13. Bar plot of −log ₁₀ p-values of the top 10 down-regulated GO terms under each GO category.

KEGG pathway analysis

The Kyoto Encyclopedia of Genes and Genomes ²⁴ (KEGG) is another commonly used database for exploring signaling pathways to understand the molecular mechanism of diseases and biological processes. A KEGG analysis can be done by using kegga function.

The top enriched KEGG pathways can be viewed by using topKEGG function.

> kegg <- kegga(res_2, geneid="ENTREZID", species="Mm")
> topKEGG(kegg, truncate.path=40, n=15)

                                                Pathway    N   Up Down     P.Up    P.Down
mmu03010                                 Ribosome  127  79    3 1.07e-22 1.00e+00
mmu05171           Coronavirus disease - COVID-19  161  92   15 1.91e-22 1.00e+00
mmu04510                           Focal adhesion  157  19   81 1.00e+00 2.79e-14
mmu04512                 ECM-receptor interaction   56   5   40 9.97e-01 5.48e-14
mmu04974         Protein digestion and absorption   49   7   35 9.36e-01 2.11e-12
mmu04015                   Rap1 signaling pathway  150  15   71 1.00e+00 1.92e-10
mmu04921               Oxytocin signaling pathway   93   9   48 9.99e-01 4.92e-09
mmu04151               PI3K-Akt signaling pathway  243  32   98 1.00e+00 4.99e-09
mmu04020                Calcium signaling pathway  115  18   54 9.64e-01 3.98e-08
mmu05200                       Pathways in cancer  373  55  134 1.00e+00 5.16e-08
mmu05414                   Dilated cardiomyopathy   53   4   31 9.99e-01 6.08e-08
mmu01100                       Metabolic pathways 1062 302  219 1.10e-07 9.95e-01
mmu04360                            Axon guidance  143  22   62 9.81e-01 1.60e-07
mmu04024                   cAMP signaling pathway  119  19   54 9.59e-01 1.65e-07
mmu05412 Arrhythmogenic right ventricular card...   44   3   26 9.98e-01 5.38e-07

The results show that most of the top enriched KEGG pathways are down-regulated. Here, we select the top 15 down-regulated KEGG pathways and visualize their significance in a bar plot ( Figure 14).

> top_path <- topKEGG(kegg,sort="Down",n=15)
> data_for_barplot <- transform(top_path, P_DE=P.Down, neg_log10_P=-log10(P.Down))
> data_for_barplot$Pathway <- factor(data_for_barplot$Pathway,
+                                    levels=data_for_barplot$Pathway)
> ggplot(data=data_for_barplot,aes(x=neg_log10_P, y=Pathway) ) +
+   geom_bar(stat="identity", show.legend=FALSE, fill=my_colors_15[1]) +
+   labs(x="-log10 (P value)", y="", title="Down" ) +
+   my_theme_ggplot

Figure 14. Bar plot of −log ₁₀ p-values of the top 15 down-regulated KEGG pathways.

Among the top down-regulated pathways, the PI3K-Akt signaling pathway is noteworthy as it is typically involved in cell proliferation and plays a crucial role in mammary gland development.

To assess the overall expression level of the PI3K-Akt signaling pathway across pseudotime, a plot is generated by plotting the average expression level of all the genes in the pathway against pseudotime. The information of all the genes in the pathway can be obtained by getGeneKEGGLinks and getKEGGPathwayNames.

> kegg_links <- getGeneKEGGLinks("mmu")
> p_names <- getKEGGPathwayNames("mmu")
> p1 <- p_names[grep("PI3K", p_names$Description), ]
> p1_GeneIDs <- subset(kegg_links, PathwayID == p1$PathwayID)$GeneID
> tab_p1 <- tab[tab$ENTREZID %in% p1_GeneIDs, ]
> d <- logCPM.obs[tab_p1$Symbol,]
> d <- apply(d, 2, mean)
> d <- data.frame(avg_logCPM = d, avg_pseudotime = y$samples$pseudotime)
> head(d)

               avg_logCPM avg_pseudotime
Pre-puberty_C1       4.43           4.65
Adult_C1             4.69           4.77
P5_C1                4.70           6.41
Puberty_C1           4.18           6.48
E18.5-epi_C1         4.46          10.16
Adult_C4             3.53          19.25

The plot below clearly illustrates a significant down-regulation of the PI3K-Akt pathway along pseudotime ( Figure 15).

> ggplot(data = d,aes(x = avg_pseudotime, y = avg_logCPM) ) +
+   geom_smooth(color=my_colors_15[1],se = FALSE) +
+   labs(x="Pseudotime", y="Average log-CPM",
+      title = "PI3K-Akt signaling pathway" ) +
+   my_theme_ggplot

Figure 15. A smooth curve of PI3K-Akt signaling pathway expression level against pseudotime.

Discussion

In this article, we demonstrated a complete workflow of a pseudo-temporal trajectory analysis of scRNA-seq data. This workflow takes single-cell count matrices as input and leverages the Seurat pipeline for standard scRNA-seq analysis, including quality control, normalization, and integration. The scDblFinder package is utilized for doublet prediction. Trajectory inference is conducted with monocle3, while the edgeR QL framework with a pseudo-bulking strategy is applied for pseudo-time course analysis. Alternative methods and packages can be used interchangeably with the ones implemented in this study, as long as they perform equivalent functions. For instance, the Bioconductor workflow may be substituted for the Seurat pipeline in scRNA-seq analysis, whereas the slingshot package may replace monocle3 for performing trajectory analysis.

This workflow article utilized 10x scRNA-seq data from five distinct stages of mouse mammary gland development, with a focus on the lineage progression of epithelial cells. By performing a time course analysis based on pseudotime along the developmental trajectory, we successfully identified genes and pathways that exhibit differential expression patterns over the course of pseudotime. The results of this extensive analysis not only confirm previous findings in the literature regarding the mouse mammary gland epithelium, but also reveal genes and pathways that exhibit continuous changes along the epithelial lineage. The analytical framework presented here can be utilized for any single-cell experiments aimed at studying dynamic changes along a specific path, whether it involves cell differentiation or the development of cell types.

Packages used

This workflow depends on various packages from the Bioconductor project version 3.15 and the Comprehensive R Archive Network (CRAN), running on R version 4.2.1 or higher. The complete list of the packages used for this workflow are shown below:

> sessionInfo()

R version 4.2.1 (2022-06-23)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: CentOS Linux 7 (Core)

Matrix products: default
BLAS:   /stornext/System/data/apps/R/R-4.2.1/lib64/R/lib/libRblas.so
LAPACK: /stornext/System/data/apps/R/R-4.2.1/lib64/R/lib/libRlapack.so

locale:
 [1] LC_CTYPE=en_AU.UTF-8       LC_NUMERIC=C
 [3] LC_TIME=en_AU.UTF-8        LC_COLLATE=en_AU.UTF-8
 [5] LC_MONETARY=en_AU.UTF-8    LC_MESSAGES=en_AU.UTF-8
 [7] LC_PAPER=en_AU.UTF-8       LC_NAME=C
 [9] LC_ADDRESS=C               LC_TELEPHONE=C
[11] LC_MEASUREMENT=en_AU.UTF-8 LC_IDENTIFICATION=C

attached base packages:
 [1] stats4  stats   graphics  grDevices utils  datasets methods  base

other attached packages:
 [1] org.Mm.eg.db_3.15.0         AnnotationDbi_1.58.0
 [3] monocle3_1.2.9              SingleCellExperiment_1.18.0
 [5] SummarizedExperiment_1.26.1 GenomicRanges_1.48.0
 [7] GenomeInfoDb_1.32.2         IRanges_2.30.0
 [9] S4Vectors_0.34.0            MatrixGenerics_1.8.0
[11] matrixStats_0.62.0          Biobase_2.56.0
[13] BiocGenerics_0.42.0         scDblFinder_1.10.0
[15] sp_1.5-0                    SeuratObject_4.1.0
[17] Seurat_4.1.1                ggplot2_3.3.6
[19] edgeR_3.38.1                limma_3.55.5

loaded via a namespace (and not attached):
 [1] utf8_1.2.2              R.utils_2.11.0             reticulate_1.25
 [4] lme4_1.1-29             tidyselect_1.1.2           RSQLite_2.2.14
 [7] htmlwidgets_1.5.4       grid_4.2.1                 BiocParallel_1.30.3
[10] Rtsne_0.16              munsell_0.5.0              ScaledMatrix_1.4.0
[13] codetools_0.2-18        ica_1.0-2                  statmod_1.4.36
[16] scran_1.24.0            xgboost_1.6.0.1            future_1.26.1
[19] miniUI_0.1.1.1          withr_2.5.0                spatstat.random_2.2-0
[22] colorspace_2.0-3        progressr_0.10.1           highr_0.9
[25] knitr_1.39              ROCR_1.0-11                tensor_1.5
[28] listenv_0.8.0           labeling_0.4.2             GenomeInfoDbData_1.2.8
[31] polyclip_1.10-0         bit64_4.0.5                farver_2.1.0
[34] parallelly_1.32.0       vctrs_0.4.1                generics_0.1.2
[37] xfun_0.31               doParallel_1.0.17          R6_2.5.1
[40] clue_0.3-61             ggbeeswarm_0.6.0           rsvd_1.0.5
[43] locfit_1.5-9.5          cachem_1.0.6               bitops_1.0-7
[46] spatstat.utils_2.3-1    DelayedArray_0.22.0        assertthat_0.2.1
[49] promises_1.2.0.1        BiocIO_1.6.0               scales_1.2.0
[52] rgeos_0.5-9             beeswarm_0.4.0             gtable_0.3.0
[55] beachmat_2.12.0         Cairo_1.5-15               globals_0.15.0
[58] goftest_1.2-3           rlang_1.0.2                GlobalOptions_0.1.2
[61] splines_4.2.1           rtracklayer_1.56.0         lazyeval_0.2.2
[64] spatstat.geom_2.4-0     BiocManager_1.30.18        yaml_2.3.5
[67] reshape2_1.4.4          abind_1.4-5                httpuv_1.6.5
[70] tools_4.2.1             ellipsis_0.3.2             spatstat.core_2.4-4
[73] RColorBrewer_1.1-3      proxy_0.4-27               ggridges_0.5.3
[76] Rcpp_1.0.8.3            plyr_1.8.7                 sparseMatrixStats_1.8.0
[79] zlibbioc_1.42.0         purrr_0.3.4                RCurl_1.98-1.7
[82] rpart_4.1.16            deldir_1.0-6               GetoptLong_1.0.5
[85] pbapply_1.5-0           viridis_0.6.2              cowplot_1.1.1
[88] zoo_1.8-10              ggrepel_0.9.1              cluster_2.1.3
[91] magrittr_2.0.3          data.table_1.14.2          RSpectra_0.16-1
[94] scattermore_0.8         circlize_0.4.15            lmtest_0.9-40
[97] RANN_2.6.1              fitdistrplus_1.1-8         patchwork_1.1.1
[100] mime_0.12              evaluate_0.15              xtable_1.8-4
[103] XML_3.99-0.10          shape_1.4.6                gridExtra_2.3
[106] compiler_4.2.1         scater_1.24.0              tibble_3.1.7
[109] KernSmooth_2.23-20     crayon_1.5.1               R.oo_1.25.0
[112] minqa_1.2.4            htmltools_0.5.2            mgcv_1.8-40
[115] later_1.3.0            tidyr_1.2.0                DBI_1.1.3
[118] ComplexHeatmap_2.12.0  MASS_7.3-57                boot_1.3-28
[121] leidenbase_0.1.11      Matrix_1.5-3               cli_3.3.0
[124] R.methodsS3_1.8.2      parallel_4.2.1             metapod_1.4.0
[127] igraph_1.3.2           pkgconfig_2.0.3            GenomicAlignments_1.32.0
[130] terra_1.5-34           plotly_4.10.0              scuttle_1.6.2
[133] spatstat.sparse_2.1-1  foreach_1.5.2              vipor_0.4.5
[136] dqrng_0.3.0            XVector_0.36.0             stringr_1.4.0
[139] digest_0.6.29          sctransform_0.3.3          RcppAnnoy_0.0.19
[142] spatstat.data_2.2-0    Biostrings_2.64.0          leiden_0.4.2
[145] uwot_0.1.11            DelayedMatrixStats_1.18.0  restfulr_0.0.15
[148] shiny_1.7.1            Rsamtools_2.12.0           nloptr_2.0.3
[151] rjson_0.2.21           lifecycle_1.0.1            nlme_3.1-158
[154] jsonlite_1.8.0         SeuratWrappers_0.3.0       BiocNeighbors_1.14.0
[157] viridisLite_0.4.0      fansi_1.0.3                pillar_1.7.0
[160] lattice_0.20-45        GO.db_3.15.0               KEGGREST_1.36.2
[163] fastmap_1.1.0          httr_1.4.3                 survival_3.3-1
[166] remotes_2.4.2          glue_1.6.2                 iterators_1.0.14
[169] png_0.1-7              bit_4.0.4                  bluster_1.6.0
[172] stringi_1.7.6          blob_1.2.3                 BiocSingular_1.12.0
[175] memoise_2.0.1          dplyr_1.0.9                irlba_2.3.5
[178] future.apply_1.9.0

Funding Statement

This work was supported by the Medical Research Future Fund (MRF1176199 to YC), a University of Melbourne Research Scholarship to JC, the National Health and Medical Research Council (Fellowship 1154970 to GKS) and the Chan Zuckerberg Initiative (2021-237445 to GKS and YC). The WEHI is supported by the Australian Government Independent Research Institutes Infrastructure Support Scheme and a Victorian State Government Operational Infrastructure Support Grant.

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

Underlying data

The single-cell RNA-seq datasets used in this study were obtained from the Gene Expression Omnibus (GEO) with accession numbers of GSE103275 ¹⁷ and GSE164017. ¹⁸

Software availability

Source code available from: https://github.com/jinming-cheng/TimeCoursePaperWorkflow

Archived source code at time of publication: https://doi.org/10.5281/zenodo.7879833 ²⁵

License: GNU General Public License version 3 (GPL-3.0-only)

All the packages used in this workflow are publicly available from the Bioconductor project (version 3.15) and the Comprehensive R Archive Network ( CRAN).