In greenelab/miQC: Flexible, probabilistic metrics for quality control of scRNA-seq data

Installation

To install the package, please use the following.

if (!requireNamespace("BiocManager"))
    install.packages("BiocManager")
BiocManager::install("miQC")

knitr::opts_chunk$set(cache=FALSE, error=FALSE, message=FALSE, warning=FALSE)

Introduction

This vignette provides a basic example of how to run miQC, which allows users to perform cell-wise filtering of single-cell RNA-seq data for quality control. Single-cell RNA-seq data is very sensitive to tissue quality and choice of experimental workflow; it’s critical to ensure compromised cells and failed cell libraries are removed. A high proportion of reads mapping to mitochondrial DNA is one sign of a damaged cell, so most analyses will remove cells with mtRNA over a certain threshold, but those thresholds can be arbitrary and/or detrimentally stringent, especially for archived tumor tissues. miQC jointly models both the proportion of reads mapping to mtDNA genes and the number of detected genes with mixture models in a probabilistic framework to identify the low-quality cells in a given dataset.

For more information about the statistical background of miQC and for biological use cases, consult the miQC paper [@hippen_miqc_2021].

You'll need the following packages installed to run this tutorial:

suppressPackageStartupMessages({
    library(SingleCellExperiment)
    library(scRNAseq)
    library(scater)
    library(flexmix)
    library(splines)
    library(miQC)
})

Example data

To demonstrate how to run miQC on a single-cell RNA-seq dataset, we'll use data from mouse brain cells, originating from an experiment by Zeisel et al [@zeisel_brain_2015], and available through the Bioconductor package scRNAseq.

sce <- ZeiselBrainData()
sce

Scater preprocessing

In order to calculate the percent of reads in a cell that map to mitochondrial genes, we first need to establish which genes are mitochondrial. For genes listed as HGNC symbols, this is as simple as searching for genes starting with mt-. For other IDs, we recommend using a biomaRt query to map to chromosomal location and identify all mitochondrial genes.

mt_genes <- grepl("^mt-",  rownames(sce))
feature_ctrls <- list(mito = rownames(sce)[mt_genes])

feature_ctrls

miQC is designed to be run with the Bioconductor package scater, which has a built-in function addPerCellQC to calculate basic QC metrics like number of unique genes detected per cell and total number of reads. When we pass in our list of mitochondrial genes, it will also calculate percent mitochondrial reads.

sce <- addPerCellQC(sce, subsets = feature_ctrls)
head(colData(sce))

miQC

Basic usage

Using the QC metrics generated via scater, we can use the plotMetrics function to visually inspect the quality of our dataset.

plotMetrics(sce)

We can see that most cells have a fairly low proportion of mitochondrial reads, given that the graph is much denser at the bottom. We likely have many cells that are intact and biologically meaningful. There are also a few cells that have almost half of their reads mapping to mitochondrial genes, which are likely broken or otherwise compromised and we will want to exclude from our downstream analysis. However, it's not clear what boundaries to draw to separate the two groups of cells. With that in mind, we'll generate a linear mixture model using the mixtureModel function.

model <- mixtureModel(sce)

This function is a wrapper for flexmix, which fits a mixture model on our data and returns the parameters of the two lines that best fit the data, as well as the posterior probability of each cell being derived from each distribution.

We can look at the parameters and posterior values directly with the functions

parameters(model)
head(posterior(model))

Or we can visualize the model results using the plotModel function:

plotModel(sce, model)

As expected, the cells at the very top of the graph are almost certainly compromised, most likely to have been derived from the distribution with fewer unique genes and higher baseline mitochondrial expression.

We can use these posterior probabilities to choose which cells to keep, and visualize the consequences of this filtering with the plotFiltering function.

plotFiltering(sce, model)

To actually perform the filtering and remove the indicated cells from our SingleCellExperiment object, we can run the filterCells parameter.

sce <- filterCells(sce, model)
sce

Other model types

In most cases, a linear mixture model will be satisfactory as well as simpler, but miQC also supports some non-linear mixture models: currently polynomials and b-splines. A user should only need to change the model_type parameter when making the model, and all visualization and filtering functions will work the same as with a linear model.

sce <- ZeiselBrainData()
sce <- addPerCellQC(sce, subsets = feature_ctrls)

model2 <- mixtureModel(sce, model_type = "spline")
plotModel(sce, model2)
plotFiltering(sce, model2)

model3 <- mixtureModel(sce, model_type = "polynomial")
plotModel(sce, model3)
plotFiltering(sce, model3)

miQC is designed to combine information about mitochondrial percentage and library complexity (number of genes discovered) to make filtering decisions, but if an even simpler model is preferred, miQC can make a model based only on mitochondrial information. This can be done by changing the model_type parameter to "one_dimensional", which runs a 1D gaussian mixture model. When library size is not added to the model, it is possible to calculate a single mitochondrial threshold to apply, which can be directly calculated using the get1DCutoff function.

model4 <- mixtureModel(sce, model_type = "one_dimensional")
plotModel(sce, model4)
plotFiltering(sce, model4)
get1DCutoff(sce, model4)

Extras

Changing posterior cutoff

miQC defaults to removing any cell with 75% or greater posterior probability of being compromised, but if we want to be more or less stringent, we can alter the posterior_cutoff parameter, like so:

plotFiltering(sce, model4, posterior_cutoff = 0.9)

Note that when performing miQC multiple times on different samples for the same experiment, it's recommended to select the same posterior_cutoff for all, to give consistency in addition to the flexibility of sample-specific models.

Preventing exclusion of low-mito cells

miQC includes two parameters to accomodate unusual and undesired behavior in the linear distributions. These issues are especially visible in some cancer datasets with a stringent posterior cutoff. We've included a bare-bones version of QC data for a high-grade serous ovarian tumor (full version of the data is available at GEO, accession GSM4816047).

data("hgsoc_metrics")
sce <- SingleCellExperiment(colData = metrics)
model <- mixtureModel(sce)
plotFiltering(sce, model, posterior_cutoff = 0.6, enforce_left_cutoff = FALSE,
                keep_all_below_boundary = FALSE)

The first issue is the group of cells at the bottom of the distribution getting marked for removal. These cells happen to be near the x-intercept of the compromised cell line, which increases their posterior probability of being compromised. But since they have decent library complexity and a low mitochondrial percentage, so it doesn't make biological sense to exclude them. When the keep_all_below_boundary parameter is set to True, as is the default, any cells below the intact cell line are kept:

plotFiltering(sce, model, posterior_cutoff = 0.6, enforce_left_cutoff = FALSE,
                keep_all_below_boundary = TRUE)

Preventing U-shaped boundary

The second issue is the U-shape in the boundary between kept and discarded cells. When this occurs, there will be cells at the bottom of the trough that are discarded, but some cells with less library complexity (farther left) and higher percentage of mitochondrial reads (higher) -- meaning they are worse in both of our QC metrics -- will be kept. To avoid this happening, the parameter enforce_left_cutoff will identify the cell marked for removal with the lowest mitochondrial percentage, determine its library complexity, and discard all cells with both lower complexity and higher mitochondrial percentage:

plotFiltering(sce, model, posterior_cutoff = 0.6, enforce_left_cutoff = TRUE,
                keep_all_below_boundary = TRUE)

This will make a de facto mitochondrial percentage cutoff for all cells with low library complexity, but will be more permissive for cells with high library complexity and high mitochondrial percentage, which are more likely to be intact cells with a biological reason for high mitochondrial expression than their low library complexity counterparts.

When not to use miQC

The miQC model is based on the assumption that there are a non-trivial number of compromised cells in the dataset, which is not true in all datasets. We recommend running plotMetrics on a dataset before running miQC to see if the two-distribution model is appropriate. Look for the distinctive triangular shape where cells have a wide variety of mitochondrial percentages at lower gene counts and taper off to lower mitochondrial percentage at higher gene counts, as can be seen in the Zeisel data.

For example of a dataset where there's not a significant number of compromised cells, so the two-distribution assumption is not met, look at another dataset from the scRNAseq package, mouse data from Buettner et al [@buettner_computational_2015].

sce <- BuettnerESCData()
mt_genes <- grepl("^mt-", rowData(sce)$AssociatedGeneName)
feature_ctrls <- list(mito = rownames(sce)[mt_genes])
sce <- addPerCellQC(sce, subsets = feature_ctrls)

plotMetrics(sce)

The mixtureModel function will throw a warning if only one distribution is found, in which case no cells would be filtered. In these cases, we recommend using other filtering methods, such as a cutoff on mitochondrial percentage or identifying outliers using median absolute deviation (MAD).

Session Information

## Session info
options(width = 120)
sessionInfo()

References

greenelab/miQC documentation built on Jan. 6, 2023, 12:16 a.m.