In LuyiTian/scPipe: Pipeline for single cell multi-omic data pre-processing

getVal <- function(x, default = NULL) {
  ifelse(is.null(x), default, x)
}

library(scales)
library(readr)
library(ggplot2)
library(plotly)
library(DT)
library(scater)
library(scran)
library(scPipe)
library(Rtsne)
knitr::opts_chunk$set(echo = FALSE)

Data summary

The organism is r getVal(params$organism, "unknown"), and gene id type is "r getVal(params$organism, "unknown")".

Cell barcode statistics

sce <- create_sce_by_dir(
  params$outdir,
  organism = params$organism,
  gene_id_type = params$gene_id_type
)
overall_stat <- demultiplex_info(sce)
datatable(overall_stat, width=800)

Plot barcode match statistics in pie chart:

plot_demultiplex(sce) +
  theme(axis.text.x = element_text(size = 10))

Read alignment statistics

ggplotly(plot_mapping(sce, percentage=FALSE))

ggplotly(plot_mapping(sce, percentage=TRUE))

Summary and distributions of QC metrics

if (any(colSums(counts(sce)) == 0)) {
  zero_cells <- sum(colSums(counts(sce)) == 0)
  sce <- sce[, colSums(counts(sce)) > 0]
} else {
  zero_cells = 0
}

r if (zero_cells > 0){paste(zero_cells, "cells have zero read counts, remove them.")}

Datatable of all QC metrics:

sce <- calculate_QC_metrics(sce)

rounded_qc_metrics <- round(as.data.frame(QC_metrics(sce)), 2)
datatable(rounded_qc_metrics, width=800, options=list(scrollX=TRUE))

Summary of all QC metrics:

summary_table <- round(do.call(cbind, lapply(QC_metrics(sce), summary)), 2)
datatable(summary_table, width=800, options=list(scrollX=TRUE))

Number of reads mapped to exon before UMI deduplication VS number of genes detected:

ggplotly(ggplot(as.data.frame(QC_metrics(sce)), aes(x=mapped_to_exon, y=number_of_genes))+geom_point(alpha=0.8))

Quality control

Detect outlier cells

A robustified Mahalanobis Distance is calculated for each cell then outliers are detected based on the distance. However, due to the complex nature of single cell transcriptomes and protocol used, such a method can only be used to assist the quality control process. Visual inspection of the quality control metrics is still required. By default we use comp = 1 and the algorithm will try to separate the quality control metrics into two gaussian clusters.

The number of outliers:

sce_qc = detect_outlier(sce, type="low", comp=1)
table(QC_metrics(sce_qc)$outliers)

Pairwise plot for QC metrics, colored by outliers:

plot_QC_pairs(sce_qc) + theme_bw()

Plot high expression genes

Remove low quality cells and plot highest expression genes.

sce_qc <- remove_outliers(sce_qc)
sce_qc <- convert_geneid(sce_qc, returns="external_gene_name")
sce_qc <- calculateQCMetrics(sce_qc)
plotHighestExprs(sce_qc, n=20)

Remove low abundance genes

Plot the log10 average count for each gene:

ave.counts <- rowMeans(counts(sce_qc))
hist(log10(ave.counts), breaks=100, main="", col="grey80",
     xlab=expression(Log[10]~"average count"))

As a loose filter we keep genes that are expressed in at least two cells and for cells that express that gene, the average count larger than 1. This is not

keep1 <- rowMeans(counts(sce_qc)) > 1 # average count larger than one
keep2 <- rowSums(counts(sce_qc) > 0) > 2 # expressed in at least three cells

sce_qc <- sce_qc[(keep1 & keep2), ]
dim(sce_qc)

We have r nrow(sce_qc) genes left after removing low abundance genes.

Data normalization

Sample normalization

We perform normalization using scater and scran,

5-point summary of size factors:

ncells <- ncol(sce_qc)
if (ncells > 200) {
  sce_qc <- computeSumFactors(sce_qc)
} else {
  sizes <- as.integer(c(ncells/7, ncells/6, ncells/5, ncells/4, ncells/3))
}

has_spikes_ins <- any( str_detect(rownames(sce_qc), "ERCC") )
if (has_spikes_ins) {
  sce_qc <- computeSpikeFactors(sce_qc, general.use=FALSE)
}

summary(sizeFactors(sce_qc))

r if (min(sizeFactors(sce_qc)) <= 0) { print("We have negative size factors in the data. They indicate low quality cells and we have removed them. To avoid negative size factors, the best solution is to increase the stringency of the filtering.") }

# filter out samples with negative size factors
if (min(sizeFactors(sce_qc)) <= 0) {
  sce_qc <- sce_qc[, sizeFactors(sce_qc) > 0]
}

PCA plot using gene expressions as input, colored by the number of genes.

cpm(sce_qc) = calculateCPM(sce_qc, use_size_factors=TRUE)
plotPCA(sce_qc, exprs_values="cpm", colour_by="total_features")

Normalize the data using size factor and get high variable genes

The highly variable genes are chosen based on trendVar from scran with FDR > 0.05 and biological variation larger than 0.5. If the number of highly variable genes is smaller than 100 we will select the top 100 genes by biological variation. If the number is larger than 500 we will only keep top 500 genes by biological variation.

sce_qc <- normalize(sce_qc)

var.fit <- trendVar(sce_qc, method="loess", use.spikes=FALSE, span=0.2)
var.out <- decomposeVar(sce_qc, var.fit)
var.out <- var.out[order(var.out$bio, decreasing=TRUE), ]

signif.genes <- which(var.out$FDR <= 0.05 & var.out$bio >= 0.5)

if (length(signif.genes) < 100) {
  hvg.out <- var.out[1:100, ]
} else if (length(signif.genes) > 500){
  hvg.out <- var.out[1:500, ]
} else {
  hvg.out <- var.out[signif.genes, ]
}

plot(var.out$mean, var.out$total, pch=16, cex=0.6,
     xlab="Mean log-expression", ylab="Variance of log-expression")
o <- order(var.out$mean)
lines(var.out$mean[o], var.out$tech[o], col="dodgerblue", lwd=2)
points(var.out$mean[rownames(var.out) %in% rownames(hvg.out)],
       var.out$total[rownames(var.out) %in% rownames(hvg.out)],
       col="red", pch=16)

Heatmap of high variable genes

gene_exp <- exprs(sce_qc)

gene_exp <- gene_exp[rownames(hvg.out), ]

hc.rows <- hclust(dist(gene_exp))
hc.cols <- hclust(dist(t(gene_exp)))

gene_exp = gene_exp[hc.rows$order, hc.cols$order]

m = list(
  l = 100,
  r = 40,
  b = 10,
  t = 10,
  pad = 0
)

plot_ly(
  x = colnames(gene_exp), y = rownames(gene_exp),
  z = gene_exp, type = "heatmap"
) %>%
  layout(autosize = F, margin = m)

Dimensionality reduction using high variable genes

Dimensionality reduction by PCA

plotPCA(sce_qc, exprs_values="logcounts", colour_by="total_features")

Dimensionality reduction by t-SNE

set.seed(100)
if (any(duplicated(t(logcounts(sce_qc)[rownames(hvg.out), ])))) {
  sce_qc <- sce_qc[, !duplicated(t(logcounts(sce_qc)[rownames(hvg.out), ]))]
}

out5 <- plotTSNE(sce_qc, exprs_values="logcounts", perplexity=5,
                 colour_by="total_features", feature_set=rownames(hvg.out)) + 
                    ggtitle("Perplexity = 5")
out10 <- plotTSNE(sce_qc, exprs_values="logcounts", perplexity=10,
                  colour_by="total_features", feature_set=rownames(hvg.out))  + 
                    ggtitle("Perplexity = 10")
out20 <- plotTSNE(sce_qc, exprs_values="logcounts", perplexity=20,
                  colour_by="total_features", feature_set=rownames(hvg.out))  + 
                    ggtitle("Perplexity = 20")

multiplot(out5, out10, out20, cols=3)

Session information

sessionInfo()

LuyiTian/scPipe documentation built on Dec. 11, 2023, 8:21 p.m.