We’ve tried to use standard file formats where possible in CNVkit. However, in a few cases we have needed to extend the standard BED format to accommodate additional information.
All of the non-standard file formats used by CNVkit are tab-separated plain text and can be loaded in a spreadsheet program, R or other statistical analysis software for manual analysis, if desired.
BED and GATK/Picard Interval List¶
Note that BED genomic coordinates are 0-indexed, like C or Python code – for example, the first nucleotide of a 1000-basepair sequence has position 0, the last nucleotide has position 999, and the entire region is indicated by the range 0-1000.
GATK and Picard interval list coordinates are 1-indexed, like R or Matlab code. In the same example, the first nucleotide of a 1000-basepair sequence has position 1, the last nucleotide has position 1000, and the entire region is indicated by the range 1-1000. These files usually have the extension .interval_list.
In GATK4, the term “interval list” also refers to samtools-style genomic coordinate specifications of the form chromosome:start-end, e.g. chr1:1-1000. As with Picard and older GATK style interval lists, the coordinates are 1-indexed. When used with GATK4, these files usually have the extension .list or .interval.
CNVkit will load these files by automatically determining the specific format based on the file contents, not the filename extension.
See the VCF specifications.
CNVkit currently uses VCF files in two ways:
- To extract single-nucleotide variant (SNV) allele frequencies, which can be plotted in the scatter command, used to assign allele-specific copy number in the call command, or exported along with bin-level copy ratios to the “nexus-ogt” format.
- To export CNVs, describing/encoding each CNV segment as a structural variant (SV).
For the former – investigating allelic imbalance and loss of heterozygosity (LOH) – it’s most useful to perform paired calling on matched tumor/normal samples. You can use a separate SNV caller such as FreeBayes, VarDict, or MuTect to do this. For best results, ensure that:
- Both the tumor and normal samples are present in the same VCF file.
- Include both germline and somatic variants (if any) in the VCF file. (For MuTect, this means keeping the “REJECT” records.) Mark somatic variants with the “SOMATIC” flag in the INFO column.
- Add a PEDIGREE tag to the VCF header declaring the tumor sample(s) as “Derived” and the normal as “Original”. Without this tag, you’ll need to tell CNVkit which sample is which using the -i and -n options in each command.
An example VCF constructed from the 1000 Genomes samples NA12878 and NA12882 is included in CNVkit’s test suite.
Target and antitarget bin-level coverages (.cnn)¶
CNVkit saves its information in a tabular format similar to BED, but with additional columns. Each row in the file indicates an on-target or off-target (a.k.a. “antitarget”) bin. Genomic coordinates are 0-indexed, like BED. Column names are shown as the first line of the file.
In the output of the coverage command, the columns are:
- Chromosome or reference sequence name (
- Start position (
- End position (
- Gene name (
- Log2 mean coverage depth (
- Absolute-scale mean coverage depth (
Essentially the same tabular file format is used for coverages (.cnn), ratios (.cnr) and segments (.cns) emitted by CNVkit.
Copy number reference profile (.cnn)¶
In addition to the columns present in the “target” and “antitarget” .cnn files, the reference .cnn file has the columns:
- GC content of the sequence region (
- RepeatMasker-masked proportion of the sequence region (
- Statistical spread or dispersion (
The log2 coverage depth is the robust average of coverage depths, excluding extreme outliers, observed at the corresponding bin in each the sample .cnn files used to construct the reference. The spread is a similarly robust estimate of the standard deviation of normalized log2 coverages in the bin. The depth column is the robust average of absolute-scale coverage depths from the input .cnn files, but without any bias corrections.
To manually review potentially problematic targets in the built reference, you can sort the file by the spread column; bins with higher values are the noisy ones.
It is important to keep the copy number reference file consistent for the duration of a project, reusing the same reference for bias correction of all tumor samples in a cohort. If your library preparation protocol changes, it’s usually best to build a new reference file and use the new file to analyze the samples prepared under the new protocol.
Bin-level log2 ratios (.cnr)¶
In addition to the
depth columns present in .cnn files, the .cnr file includes each bin’s
proportional weight or reliability (
The weight value is derived from several sources:
- The size of the bin relative to the average bin size (for targets or antitargets, separately)
- For a paired or pooled reference, the deviation of the reference log2 value from neutral coverage (i.e. distance from 0.0)
- For a pooled reference, the inverse of the variance (i.e. square of
spreadin the reference) of normalized log2 coverage values seen among all normal samples at that bin.
This calculated value is used to weight the bin log2 ratio values during segmentation. Also, when a genomic region is plotted with CNVkit’s “scatter” command, the size of the plotted datapoints is proportional to each bin’s weight – a relatively small point indicates a less reliable bin.
Segmented log2 ratios (.cns)¶
In addition to the
weight columns present in .cnr files, the .cns file format has
the additional column
probes, indicating the number of bins covered by the
The gene column concatenates the gene names of all the bins that the segment covers. The weight column sums the bin-level weights, and the depth and log2 is the weighted mean of the input bin-level values corresponding to the segment.