emptyDrops {DropletUtils}R Documentation

Identify empty droplets

Description

Distinguish between droplets containing cells and ambient RNA in a droplet-based single-cell RNA sequencing experiment.

Usage

testEmptyDrops(m, lower=100, niters=10000, test.ambient=FALSE,
    ignore=NULL, alpha=NULL, BPPARAM=SerialParam())

emptyDrops(m, lower=100, retain=NULL, barcode.args=list(), ...)

Arguments

m

A numeric matrix object - usually a dgTMatrix or dgCMatrix - containing droplet data prior to any filtering or cell calling. Columns represent barcoded droplets, rows represent genes.

lower

A numeric scalar specifying the lower bound on the total UMI count, at or below which all barcodes are assumed to correspond to empty droplets.

niters

An integer scalar specifying the number of iterations to use for the Monte Carlo p-value calculations.

test.ambient

A logical scalar indicating whether results should be returned for barcodes with totals less than or equal to lower.

ignore

A numeric scalar specifying the lower bound on the total UMI count, at or below which barcodes will be ignored (see Details for how this differs from lower).

alpha

A numeric scalar specifying the scaling parameter for the Dirichlet-multinomial sampling scheme.

BPPARAM

A BiocParallelParam object indicating whether parallelization should be used to compute p-values.

retain

A numeric scalar specifying the threshold for the total UMI count above which all barcodes are assumed to contain cells.

barcode.args

Further arguments to pass to barcodeRanks.

...

Further arguments to pass to testEmptyDrops.

Value

testEmptyDrops will return a DataFrame with the following components:

Total:

Integer, the total UMI count for each barcode.

LogProb:

Numeric, the log-probability of observing the barcode's count vector under the null model.

PValue:

Numeric, the Monte Carlo p-value against the null model.

Limited:

Logical, indicating whether a lower p-value could be obtained by increasing niters.

For barcodes with counts below lower, NA values are returned for all fields if test.ambient=FALSE. This is to ensure that the number of rows in the output DataFrame is identical to ncol(m).

emptyDrops will return a DataFrame like testEmptyDrops, with an additional FDR field.

The metadata of the output DataFrame will contains the ambient profile in ambient, the estimated/specified value of alpha, the specified value of lower and the number of iterations in niters. For emptyDrops, the metadata will also contain the retention threshold in retain.

Details about testEmptyDrops

The testEmptyDrops function will obtain an estimate of the composition of the ambient pool of RNA based on the barcodes with total UMI counts less than or equal to lower. This assumes that a cell-containing droplet would generally have higher total counts than empty droplets containing RNA from the ambient pool. Counts for the low-count barcodes are pooled together, and an estimate of the proportion vector for the ambient pool is calculated using goodTuringProportions. The count vector for each barcode above lower is then tested for a significant deviation from these proportions.

The null hypothesis is that transcript molecules are included into droplets by multinomial sampling from the ambient profile. For each barcode, the probability of obtaining its count vector based on the null model is computed. Then, niters count vectors are simulated from the null model. The proportion of simulated vectors with probabilities lower than the observed multinomial probability for that barcode is used to calculate the p-value. We use this Monte Carlo approach as an exact multinomial p-value is difficult to calculate.

The ignore argument can also be set to ignore barcodes with total counts less than or equal to ignore. This differs from the lower argument in that the ignored barcodes are not necessarily used to compute the ambient profile. Users can interpret ignore as the minimum total count required for a barcode to be considered as a potential cell. In contrast, lower is the maximum total count below which all barcodes are assumed to be empty droplets.

Details about emptyDrops

The emptyDrops function combines the results of testEmptyDrops with barcodeRanks to identify droplets that are likely to contain cells. Barcodes that contain more than retain total counts are always retained. This ensures that large cells with profiles that are very similar to the ambient pool are not inadvertently discarded. If retain is not specified, it is set to the total count at the knee point detected by barcodeRanks. Manual specification of retain may be useful if the knee point was not correctly identified in complex log-rank curves. Users can also set retain=Inf to disable automatic retention of barcodes with large totals.

The Benjamini-Hochberg correction is also applied to the Monte Carlo p-values to correct for multiple testing. Cells can then be defined by taking all barcodes with significantly non-ambient profiles, e.g., at a false discovery rate of 0.1%. All barcodes with total counts above K (or retain) are assigned p-values of zero during correction, reflecting our assumption that they are true positives. This ensures that their Monte Carlo p-values do not affect the correction of other genes, and also means that they will have FDR values of zero. Nonetheless, their original Monte Carlo p-values are still reported in the output.

Handling overdispersion

If alpha is set to a positive number, sampling is assumed to follow a Dirichlet-multinomial (DM) distribution. The parameter vector of the DM distribution is defined as the estimated ambient profile scaled by alpha. Smaller values of alpha model overdispersion in the counts, due to dependencies in sampling between molecules. If alpha=NULL, a maximum likelihood estimate is obtained from the count profiles for all barcodes with totals less than or equal to lower. If alpha=Inf, the sampling of molecules is modelled with a multinomial distribution.

Users can check whether the model is suitable by extracting the p-values for all barcodes with test.ambient=TRUE. Under the null hypothesis, the p-values for presumed ambient barcodes (i.e., with total counts below lower) should be uniformly distributed. Skews in the p-value distribution are indicative of an inaccuracy in the model and/or its estimates (of alpha or the ambient profile).

Author(s)

Aaron Lun

References

Lun A, Riesenfeld S, Andrews T, Dao TP, Gomes T, participants in the 1st Human Cell Atlas Jamboree, Marioni JC (2018). Distinguishing cells from empty droplets in droplet-based single-cell RNA sequencing data. biorXiv.

Phipson B, Smyth GK (2010). Permutation P-values should never be zero: calculating exact P-values when permutations are randomly drawn. Stat. Appl. Genet. Mol. Biol. 9:Article 39.

See Also

barcodeRanks, defaultDrops

Examples

# Mocking up some data:
set.seed(0)
my.counts <- DropletUtils:::simCounts()

# Identify likely cell-containing droplets.
out <- emptyDrops(my.counts)
out

is.cell <- out$FDR <= 0.01
sum(is.cell, na.rm=TRUE)

# Check if p-values are lower-bounded by 'niters'
# (increase 'niters' if any Limited==TRUE and Sig==FALSE)
table(Sig=is.cell, Limited=out$Limited)

[Package DropletUtils version 1.4.3 Index]