CNVfilteR: Remove false positives of CNV calling tools by using SNV calls

knitr::opts_chunk$set(fig.width = 8)
knitr::opts_chunk$set(fig.height = 5)


Many tools for germline copy number variant (CNV) detection from NGS data have been developed. Usually, these tools were designed for different input data like WGS, WES or panel data, and their performance may depend on the CNV size. Available benchmarks show that all these tools produce false positives, sometimes reaching a very high number of them.

With the aim of reducing the number of false positives, r BiocStyle::Biocpkg("CNVfilteR") identifies those germline CNVs that can be discarded. This task is performed by using the germline single nucleotide variant (SNV) calls that are usually obtained in common NGS pipelines. As VCF field interpretation is key when working with these files, r BiocStyle::Biocpkg("CNVfilteR") specifically supports VCFs produced by VarScan2, Strelka/Strelka2, freeBayes, HaplotypeCaller (GATK), and UnifiedGenotyper (GATK). Additionally, results can be plotted using the functions provided by the R/Bioconductor packages karyoploteR and CopyNumberPlots.


r BiocStyle::Biocpkg("CNVfilteR") is a Bioconductor package and to install it we have to use r BiocStyle::Biocpkg("BiocManager").

  if (!requireNamespace("BiocManager", quietly = TRUE))

We can also install the package from github to get the latest devel version.

```` {r, eval = FALSE} BiocManager::install("jpuntomarcos/CNVfilteR")

# Quick Start

Below we show a full example that covers the usual steps: CNVs data loading,
SNVs loading, identifying false postives and plotting the results.

First, we can load some CNV tool results:

cnvs.file <- system.file("extdata", "DECoN.CNVcalls.csv", package = "CNVfilteR", mustWork = TRUE) <- loadCNVcalls(cnvs.file = cnvs.file, chr.column = "Chromosome", start.column = "Start", end.column = "End", cnv.column = "CNV.type", sample.column = "Sample", genome = "hg19")

Then, we load the SNVs stored in a couple of VCF files.

vcf.files <- c(system.file("extdata", "variants.sample1.vcf.gz", package = "CNVfilteR", mustWork = TRUE),
               system.file("extdata", "variants.sample2.vcf.gz", package = "CNVfilteR", mustWork = TRUE))
vcfs <- loadVCFs(vcf.files, =

We observe that the function recognized VarScan2 as the source, so fields were selected and allele frequency consequently. Now we can call filterCNVs() to identify those CNVs that can be discarded.

results <- filterCNVs(, vcfs)

And we can check those CNVs that can be filtered out:

results$cnvs[results$cnvs$filter == TRUE]

As an example, we can observe that the CNV with contains 4 variants: 2 in favor of discarding it, two against discarding it. If we want to know more about the variants falling in a certain CNV we can do:


Two variants are close to the default expected heterozygous frequency, 0.5, so they obtain a positive score. The other two variants are not so clearly close to the default expected duplication value, 0.33, so they obtain a low negative score. All these default values and others can be modified in the filterCNVs() function.

Finally, we may be interested in plotting the results. For example, we can plot easily the duplication CNV with and all the variants falling in it.

plotVariantsForCNV(results, "3")

We can do the same to plot the deletion CNV with, where all variants discard the CNV except one homozygous variant that does not give us any information for supporting or discarding the CNV:

plotVariantsForCNV(results, "19")

On the opposite side, we can observe those CNVs that cannot be discarded:

results$cnvs[results$cnvs$filter != TRUE]

For example, the CNV with contains one variant. If we get the variant info, we see that the variant has an allele frequency very close to the default expected duplication value, 0.66.


So, no evidence was found for discarding the CNV. We can also plot the CNV and the variant:

plotVariantsForCNV(results, "3")

Loading Copy-Number Data

r BiocStyle::Biocpkg("CNVfilteR") functions expect germline CNVs calls to be a GRanges object with a few specificic metadata columns:

You can create this object yourself, but r BiocStyle::Biocpkg("CNVfilter") provides the proper function to do this, loadCNVcalls(). This function can interpret any tsv o csv file by indicating which columns store the information. For example, in the following code, the chr.column column is stored in the "Chromosome" column of the cnvs.file.

cnvs.file <- system.file("extdata", "DECoN.CNVcalls.csv", package = "CNVfilteR", mustWork = TRUE) <- loadCNVcalls(cnvs.file = cnvs.file, chr.column = "Chromosome", start.column = "Start", end.column = "End", cnv.column = "CNV.type", sample.column = "Sample", genome = "hg19")

loadCNVcalls() can interpret different types of CNVs. Among other options, separator can be selected using the sep parameter (defaults to \t), and first lines can be skipped using the skip parameter (defaults to 0). Also, the value used in cnv.column to store the CNV type can be modified using the deletion and duplication parameters (defaults to "deletion" and "duplication", respectively). If, for example, our cnv.column uses "CN1" and "CN3" for deletion and duplication (respectively), we should indicate the function to use these values: <- loadCNVcalls(cnvs.file = cnvs.file.2, deletion = "CN1", duplication = "CN3", chr.column = "Chromosome", start.column = "Start", end.column = "End", cnv.column = "CNV.type", sample.column = "Sample")

Some CNV tools generate results where the CNV location is stored in a single column with the format chr:start-end (i.e. 1:538001-540000). In this case, we can call loadCNVcalls() using the coord.column instead of the chr.column, start.column and end.column columns.

Loading Variants Data

Common NGS pipelines produce germline variant calling (SNVs or INDELs) in a VCF file. However, VCF interpretation is challenging due to the flexibility provided by the VCF format definition. It is not straightforward to interpret correctly the fields in the VCF file and compute the allele frequency. loadVCFs() interprets automatically VCFs produced by VarScan2, Strelka/Strelka2, freeBayes, HaplotypeCaller (GATK), and UnifiedGenotyper (GATK).

In the following example the function recognizes VarScan2 as the source.

vcf.files <- c(system.file("extdata", "variants.sample1.vcf.gz", package = "CNVfilteR", mustWork = TRUE),
               system.file("extdata", "variants.sample2.vcf.gz", package = "CNVfilteR", mustWork = TRUE))
vcfs <- loadVCFs(vcf.files, =

We can also load the VCF file spicifying how to interpret it, which can be useful if the VCF was generated by a caller not supported by r BiocStyle::Biocpkg("CNVfilteR"). For example we can specify the ref/alt fields:

vcfs <- loadVCFs(vcf.files, =, vcf.source = "MyCaller", = "RD", = "AD")

Alternatively, we can set the parameter so that field will be loaded assuming that it is a list in this order: reference allele, alternative allele. As an example:

vcf.file3 <- c(system.file("extdata", "variants.sample3.vcf", package = "CNVfilteR", mustWork = TRUE))
vcfs3 <- loadVCFs(vcf.file3, =, vcf.source = "MyCaller", = "AD")

VCF free of artifacts

CNVfilteR uses SNVs to identify false-positive CNV calls. Therefore, its performance depends on the SNV calls quality. We recommend using VCF files free of false-positive SNVs (as possible) to improve CNVfilteR accuracy. Some considerations can be followed in order to provide reliable SNVs to CNVfilteR.

Minimun total depth

Use the parameter to discard SNVs with low depth coverage in the loadVCFs function. The default value is 10, which may be appropriate in many WGS samples, but this value should be adapted to your experiment conditions. For example, we used a of 30 when using CNVfilteR on panel (targeted-enrinched) samples with high coverage and VarScan2 as SNV caller.

Regions to exclude

Low complexity and repetitive regions are genome areas where SNV callers (also CNV callers) perform poorly. If possible, ignore these regions when using CNVfilteR. We can exclude those complex regions that have already known alignement artifacts with the parameter In this example, we exclude PMS2, PRSS1, and FANCD2 genes because they are pseudogenes with alignments artifacts: <- GRanges(seqnames = c("chr3","chr7", "chr7"), ranges = IRanges(c(10068098, 6012870, 142457319), c(10143614, 6048756, 142460923)))
vcfs4 <- loadVCFs(vcf.files, =, =

INDELs excluded by default

Also, the parameter exclude.indels indicates whether to exclude INDELs when loading the variants. TRUE is the default and recommended value given that INDELs allele frequency varies differently than SNVs. Including INDELs may allow the algorithm to identify more CNVs to discard with a greater risk of identifying them wrongly. Additionally, any SNV overlapping an INDEL will be ignored because the SNV allele frequency may be affected in that region.

Other settings

The function loadVCFs() also adapts to different needs. If sample.names parameter is not provided, the sample names included in the VCF itself will be used. Both single-sample and multi-sample VCFs are accepted, but when multi-sample VCFs are used, sample.names parameter must be NULL.

If VCF is not compressed with bgzip, the function compresses it and generates the .gz file. If .tbi file does not exist for a given VCF file, the function also generates it. All files are generated in a temporary folder.

See loadVCFs() documentation to see other parameters info.

Identifying false positives

The task of identifying false positives is performed by the filterCNVs() function. It checks all the variants (SNVs and optionally INDELs) falling in each CNV present in to identify those CNVs that can be filtered out. It returns an S3 object with 3 elements: cnvs, variantsForEachCNV and filterParameters:

results <- filterCNVs(, vcfs)

Observe that those CNVs that can be filtered out have the value TRUE in the column filter. r BiocStyle::Biocpkg("CNVfilteR") employs two different strategies for identifying those CNVs:

Scoring model for duplication CNVs

The scoring model for determining whether a certain duplication CNV can be discarded is based on the allele frequency for each heterozygous variant falling in that CNV:

Although we can expect that most of the variants are close to the expected means (33.3%, 50%, and 66.6%), many of them can be far from any expected mean. The scoring model implemented in the filterCNVs() function measures the evidence - for discarding a certain CNV - using a scoring model. The scoring model is based on the fuzzy logic, where elements can have any value between 1 (True) and 0 (False). Following this idea, each variant will be scored with a value between 0 and 1 depending on how close is the allele frequency to the nearest expected mean.

The total score is computed among all the variants falling in the CNV. If the score is greater than the dup.threshold.score, the CNV can be discarded.

A common way of applying the fuzzy logic is using the sigmoid function. CNVfilteR uses the sigmoid function implemented in the pracma package, which is defined as $$ \begin{aligned} y = 1/(1 + e^{-c1(x−c2)}) \end{aligned} $$

The scoring model is built on 6 sigmoids defined on 6 different intervals. The c1 parameter is 2 by default (sigmoid.c1), and the c2 parameter is defined for the 6 sigmoids (sigmoid.c2.vector).

Where sigmoid.int1 is the mean between and, and sigmoid.int2 is the mean between and

The scoring model can be plotted using the plotScoringModel() function.

p <- results$filterParameters
plotScoringModel( = p$, 
        = p$,
        = p$,
                 sigmoid.c1 = p$sigmoid.c1, 
                 sigmoid.c2.vector = p$sigmoid.c2.vector)

And the scoring model can be modified when calling the filterCNVs() function. Let's see how the model changes when we modify the parameter sigmoid.c1 (1 instead of 2):

plotScoringModel( = p$, 
        = p$,
        = p$,
                 sigmoid.c1 = 1, 
                 sigmoid.c2.vector = p$sigmoid.c2.vector)

We can also modify the sigmoid.c2.vector parameter for each sigmoid function. For example, to make the central sigmoids narrower:

plotScoringModel( = p$, 
        = p$,
        = p$,
                 sigmoid.c1 = p$sigmoid.c1, 
                 sigmoid.c2.vector = c(28, 38.3, 46.7, 53.3, 61.3, 71.3))

The margin.pct parameter

Many CNV callers produce inaccurate CNV calls. These inaccurate CNV calls are more likely to be true (to overlap the real CNV) in the middle of the CNV than in the extremes. So, the margin.pct parameter defines the percentage of CNV (from each CNV limit) where SNVs will be ignored. By default, only 10% of SNVs from each CNV extreme will be ignored. This margin.pct parameter can be modified to better adapt it to your CNV caller. For example, we observed that DECoN produced very accurate CNV calls in our genes panel dataset, so margin.pct value was updated to 0 in this context.

Summarizing, variants in the CNV call but close to the ends of the CNV call will be ignored. margin.pct defines the percentage of CNV length, located at each CNV limit, where variants will be ignored. For example, for a CNV chr1:1000-2000 and a margin.pct value of 10, variants within chr1:1000-1100 and chr1:1900-2000 will be ignored.

Plotting results

We can plot easily a certain CNV and the variants in it. For example, the duplication CNV with can be plotted as follows:

plotVariantsForCNV(results, "16")

Some parameters can be customized, like points.cex and points.pch. It is also possible to plot all CNVs in a global schema where all the chromosomes are plotted:

cnvs.file <- system.file("extdata", "DECoN.CNVcalls.2.csv",
                         package = "CNVfilteR", mustWork = TRUE) <- loadCNVcalls(cnvs.file = cnvs.file, chr.column = "Chromosome",
                          start.column = "Start", end.column = "End", 
                          cnv.column = "CNV.type", sample.column = "Sample",
                          genome = "hg19")

Note that if a CNV is too small, it will not be visible when calling plotAllCNVs().

Session Info


Try the CNVfilteR package in your browser

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

CNVfilteR documentation built on April 17, 2021, 6:06 p.m.