Introduction to the scFeatureFilter package


Package aim

One of the steps in single cell RNA-seq analysis is the filtering of features (i.e. genes or transcripts) to retain only the most expressed ones. This allows to focus on features less affected by technical variations (relatively), and often increases the statistical power of the subsequent analyses.

This step is usually done by applying one or more arbitrary thresholds (i.e. mean TPM > 1, retention of features frequently detected across cells, etc.), or by using the data from spike-in RNA. This package aims to help the threshold decision, especially when spike-in controls are not available.

Method summary

Highly expressed features are less affected by technical variability than lowly expressed ones, and thus capture biological information more successfully. We propose to use highly expressed features as a reference to to estimate the drowning of biological variation in bins of features of decreasing expression. In this scenario, it is expected that features from major transcription programs will be correlated with each other across bins, unless the technical variation component dominates the expression results.

One of the package's first steps is to select this reference bin of features (or 'top window' of features). We propose to use the 100 features with the highest mean expression across cells, and a reasonably low fraction of drop-outs (zero-expression values) as reference. Using more features will increase the noise in the biological variation reflected in this reference bin (as well as increase computation time), while using less features might result in insufficient capture of the biological complexity of the samples. Subsequent bins of features of decreasing mean expression across cells are then defined (by default, 1000 features per bin). Note that the size of the subsequent feature bins matters less than the size of the reference bin.

For comparison, several negative control bins are defined by randomly shuffling values of the top bin expression matrix, building a window that is expected to be completely uncorrelated. Then, every feature in each bin is correlated to every feature in the reference bin and the control bins to obtain correlation coefficient distributions. Ultimately, the variation of the distributions of correlation coefficients can be used to inform the threshold decision.


The stable version of scFeatureFilter can be installed from Bioconductor:

## try http:// if https:// URLs are not supported
if (!requireNamespace("BiocManager", quietly=TRUE))

Once installed, the package is ready to be loaded:


library(cowplot) # multipanel figures + nice theme

Integrated usage

A single-function call option is available to the user, in which the entire package functionality is executed. This will filter the expression matrix and keep only the most informative highly expressed features, using default options:

# example dataset included with the package:

# filtering of the dataset with a single function call:

Detailed usage

Loading data

scFeatureFilter uses an expression matrix (either as a matrix, a data.frame, a tibble or a SingleCellExperiment R object) as input. Support of Bioconductor's expression set objects will arrive shortly. Note that features should be in rows, while cells should be in columns.

We recommend providing normalized expression values where at least library size is accounted for, such as TPM or FPKM, rather than raw counts.

An example dataset is supplied with the package:


The example dataset was processed by Mantsoki et al. (2016), data from human Embryonic Stem cells was generated by Yan et al. (2013).

Binning data

Firstly, a mean expression and a coefficient of variation column is added to the expression matrix provided using the calculate_cvs function:


We can explore the mean - variance relationship of the dataset with the plot_mean_variance function:

library(magrittr) # to use the pipe %>%

calculate_cvs(scData_hESC) %>%
    plot_mean_variance(colourByBin = FALSE)

The top window (or reference bin) is then defined with the define_top_genes function, while the rest of bins are created with the bin_scdata function:

scData_hESC %>%
    calculate_cvs %>%
    define_top_genes(window_size = 100) %>%
    bin_scdata(window_size = 1000)

Then, we can plot the resulting bin definition using the same plot_mean_variance function, setting the colourByBin argument to TRUE:

myPlot <- scData_hESC %>%
    calculate_cvs %>%
    define_top_genes(window_size = 100) %>%
    bin_scdata(window_size = 1000) %>%
    plot_mean_variance(colourByBin = TRUE, density_color = "blue")


Note that the plotting function outputs are ggplot2 objects that can be further customized. See an example here:

myPlot + annotation_logticks(sides = "l")

Correlation of the bins

After binning the data, correlations (by default, Pearson's) can be calculated. Every feature in each window is correlated to every other feature in the reference bin (as well as in the randomized control bins) using the function correlate_windows:

corDistrib <- scData_hESC %>%
    calculate_cvs %>%
    define_top_genes(window_size = 100) %>%
    bin_scdata(window_size = 1000) %>%
    correlate_windows(n_random = 3)

For visualization of the correlation values, probability distributions can be computed using and plotted using the correlations_to_densities and plot_correlations_distributions functions:

corDens <- correlations_to_densities(corDistrib, absolute_cc = TRUE)
plot_correlations_distributions(corDens, facet_ncol = 5) +
    scale_x_continuous(breaks = c(0, 0.5, 1), labels = c("0", "0.5", "1"))

The coloured lines indicate the correlation coefficient distributions for each bin compared to the reference bin (bin 1 shows the autocorrelation of the reference bin). The thicker blue lines with grey area (barely visible) represents the correlation coefficient distributions for each bin compared to the randomized control bins.

Notice that the correlation distributions shift from high to low values as the bin number increases, i.e. as the mean expression level of the features decreases, indicating a higher influence of technical variability in the expression results.

Note: the absolute_cc = TRUE option, which is the default, instructs the package to work with the absolute value of the correlation coefficient. Therefore, highly negatively correlated features will count as highly correlated. Setting this option as TRUE also produces clearer plots, and reduces the emphasis on non-symmetrical, near 0, shifts of correlation values that are not interpretable.

We can quantify the steady decrease of correlation values as feature expression goes down using the get_mean_median function, improving interpretability of the extent of technical variability in the data:

metrics <- get_mean_median(corDistrib)
plot_correlations_distributions(corDens, metrics = metrics, facet_ncol = 5) +
    scale_x_continuous(breaks = c(0, 0.5, 1), labels = c("0", "0.5", "1"))

The added dashed lines are the mean values of the distributions. The plot_metric function focalized on these means of distributions:

plot_metric(metrics, show_ctrl = FALSE, show_threshold = FALSE)

The bar represents the mean value of the correlation coefficient distribution of each bin against the reference bin, while the dot and whiskers are against the randomized control windows. In our example dataset, the mean of the correlation distributions decreases steadily, and reaches a plateau after bin 11.

Choosing a threshold

On the previous plot, a background reference can be added as a dashed line, by taking the mean value of the correlations of every bin to the randomized control bins:

plot_metric(metrics, show_ctrl = TRUE, show_threshold = FALSE)

Then, the user can decide to plot only bins for which the mean correlation value against the reference window is higher than, for example, twice the background:

plot_metric(metrics, show_ctrl = TRUE, show_threshold = TRUE, threshold = 2)

To further assist thresholding, the function determine_bin_cutoff will return the number of the first bin of features that falls below the set threshold:

determine_bin_cutoff(metrics, threshold = 2)

Finally, using this information, the function filter_expression_table can be used to obtain a filtered expression matrix from the original data:

binned_data <- scData_hESC %>%
    calculate_cvs %>%
    define_top_genes(window_size = 100) %>%
    bin_scdata(window_size = 1000)
metrics <- correlate_windows(binned_data, n_random = 3) %>%

filtered_data <- filter_expression_table(
    bin_cutoff = determine_bin_cutoff(metrics)


Note: most of the features are filtered at the calculate_cvs step because hey have a too high fraction of drop-outs or zero-expression. The maximum fraction of 0 expression tolerated by scFeatureFilter is 75%, by default.

Working with SingleCellExperiment objects

This package accpets SingleCellExperiment as inputs for sc_feature_filter and calculate_cvs functions. However, it does not retun such objects at the moment. One can retrieve a filtered SingleCellExperiment objects like this:

library(scRNAseq) # example datasets

sce_allen <- ReprocessedAllenData()

# sce_allen is an SingleCellExperiment object

filtered_allen <- sc_feature_filter(sce_allen, sce_assay = "rsem_tpm")
is.matrix(filtered_allen) # filtered_allen is a tibble

sce_filtered_allen <- sce_allen[rownames(filtered_allen), ]

Parameter values

Several parameters are used in sc_feature_filter or in step-by-step functions. Their default values are set using mouse an human data and should work for eukaryote species with multi-exonic genes (metazoans & plants).


The max_zeros parameter (default value 0.75) is used as a first filtering step. Features that present zero-expression in a proportion of cells higher than the set threshold will be removed. Zero expression can reflect both an absence of expression in the cell of origin, or a drop-out event occuring during library preparation. Ocurrence of these technical drop-outs can be quite high in scRNAseq protocols. More stringent values for this parameter could be 0.5 or 0.25. This parameter will mostly impact the total number of features, and therefore the number of bins that the data is divided into. This filtering step will remove low-expression, highly affected by noise features before the method is applied.


The top_window_size parameter (default to 100) can have a significant impact on the rest of the pipeline. It sets the number of features with the highest mean expression that will be used as a reference bin. If the reference window is too small (i.e. 10), it will not capture enough of the biological variation. On the contrary, a too large top window (i.e. 2000) might include features mildly affected by technical variation, diluting the biological component of the variation too much to be of any use. We therefore suggest keeping it at 100 or near that value.

The function plot_top_window_autocor can be used to visually explore the mean auto-correlation value of features in the top bin with themselves depending on the number of features in the top bin:



The other_window_size parameter (default to 1000) is the number of features included in the bins of features (outside the top window). It is a more flexible parameter, and can safely be decreased to 500 or 250, (which will result in more bins) without significantly altering the proposed filtering threshold:

metrics_bigBins <- scData_hESC %>%
    calculate_cvs %>%
    define_top_genes(window_size = 100) %>%
    bin_scdata(window_size = 1000) %>%
    correlate_windows(n_random = 3) %>%

metrics_smallBins <- scData_hESC %>%
    calculate_cvs %>%
    define_top_genes(window_size = 100) %>%
    bin_scdata(window_size = 500) %>%
    correlate_windows(n_random = 3) %>%

    plot_metric(metrics_bigBins) +
        labs(title = "1000 genes per bin"),
    plot_metric(metrics_smallBins) +
        labs(title = "500 genes per bin")


The background correlation values used to determine a threshold are calculated by randomly shuffling the expression values of features in the top window. n_random (default to 3) sets the number of control bins that will be generated. Note that increasing this value will also increase the computation time. and given that results differ only marginally among the randomised windows, there is not much need to increase that value. Results differ only marginally between the different random windows, therefore there is not much need to generate more than 3 control windows.


One can decide which bins of features to keep by looking at the plot_metric or plot_correlations_distributions plots, by visually exploring their correlation to the top bin of features. We also provide an automatic threshold decision method, which keeps only windows of features whose mean correlation value to the top window is higher than threshold (default to 2) times the background level. The background level of correlation is estimated using the randomised control bins. Depending on the intended downstream analyses, one can increase the threshold value to be more stringent in the feature selection, keeping less features, but ensuring that they will contain a sufficiently high amount of biological information.

Try the scFeatureFilter package in your browser

Any scripts or data that you put into this service are public.

scFeatureFilter documentation built on Nov. 8, 2020, 7:49 p.m.