How should I interpret VCF files produced by the GATK?

1. What is VCF?

VCF stands for Variant Call Format. It is a standardized text file format for representing SNP, indel, and structural variation calls. See this page for detailed specifications.

VCF is the primary (and only well-supported) format used by the GATK for variant calls. We prefer it above all others because while it can be a bit verbose, the VCF format is very explicit about the exact type and sequence of variation as well as the genotypes of multiple samples for this variation.

That being said, this highly detailed information can be challenging to understand. The information provided by the GATK tools that infer variation from NGS data, such as the UnifiedGenotyper and the HaplotypeCaller, is especially complex. This document describes some specific features and annotations used in the VCF files output by the GATK tools.

2. Basic structure of a VCF file

The following text is a valid VCF file describing the first few SNPs found by the UG in a deep whole genome data set from our favorite test sample, NA12878:

##fileformat=VCFv4.0
##FILTER=<ID=LowQual,Description="QUAL < 50.0">
##FORMAT=<ID=AD,Number=.,Type=Integer,Description="Allelic depths for the ref and alt alleles in the order listed">
##FORMAT=<ID=DP,Number=1,Type=Integer,Description="Read Depth (only filtered reads used for calling)">
##FORMAT=<ID=GQ,Number=1,Type=Float,Description="Genotype Quality">
##FORMAT=<ID=GT,Number=1,Type=String,Description="Genotype">
##FORMAT=<ID=PL,Number=3,Type=Float,Description="Normalized, Phred-scaled likelihoods for AA,AB,BB genotypes where A=ref and B=alt; not applicable if site is not biallelic">
##INFO=<ID=AC,Number=.,Type=Integer,Description="Allele count in genotypes, for each ALT allele, in the same order as listed">
##INFO=<ID=AF,Number=.,Type=Float,Description="Allele Frequency, for each ALT allele, in the same order as listed">
##INFO=<ID=AN,Number=1,Type=Integer,Description="Total number of alleles in called genotypes">
##INFO=<ID=DB,Number=0,Type=Flag,Description="dbSNP Membership">
##INFO=<ID=DP,Number=1,Type=Integer,Description="Total Depth">
##INFO=<ID=DS,Number=0,Type=Flag,Description="Were any of the samples downsampled?">
##INFO=<ID=Dels,Number=1,Type=Float,Description="Fraction of Reads Containing Spanning Deletions">
##INFO=<ID=HRun,Number=1,Type=Integer,Description="Largest Contiguous Homopolymer Run of Variant Allele In Either Direction">
##INFO=<ID=HaplotypeScore,Number=1,Type=Float,Description="Consistency of the site with two (and only two) segregating haplotypes">
##INFO=<ID=MQ,Number=1,Type=Float,Description="RMS Mapping Quality">
##INFO=<ID=MQ0,Number=1,Type=Integer,Description="Total Mapping Quality Zero Reads">
##INFO=<ID=QD,Number=1,Type=Float,Description="Variant Confidence/Quality by Depth">
##INFO=<ID=SB,Number=1,Type=Float,Description="Strand Bias">
##INFO=<ID=VQSLOD,Number=1,Type=Float,Description="log10-scaled probability of variant being true under the trained gaussian mixture model">
##UnifiedGenotyperV2="analysis_type=UnifiedGenotyperV2 input_file=[TEXT CLIPPED FOR CLARITY]"
#CHROM  POS ID      REF ALT QUAL    FILTER  INFO    FORMAT  NA12878
chr1    873762  .       T   G   5231.78 PASS    AC=1;AF=0.50;AN=2;DP=315;Dels=0.00;HRun=2;HaplotypeScore=15.11;MQ=91.05;MQ0=15;QD=16.61;SB=-1533.02;VQSLOD=-1.5473 GT:AD:DP:GQ:PL   0/1:173,141:282:99:255,0,255
chr1    877664  rs3828047   A   G   3931.66 PASS    AC=2;AF=1.00;AN=2;DB;DP=105;Dels=0.00;HRun=1;HaplotypeScore=1.59;MQ=92.52;MQ0=4;QD=37.44;SB=-1152.13;VQSLOD= 0.1185 GT:AD:DP:GQ:PL  1/1:0,105:94:99:255,255,0
chr1    899282  rs28548431  C   T   71.77   PASS    AC=1;AF=0.50;AN=2;DB;DP=4;Dels=0.00;HRun=0;HaplotypeScore=0.00;MQ=99.00;MQ0=0;QD=17.94;SB=-46.55;VQSLOD=-1.9148 GT:AD:DP:GQ:PL  0/1:1,3:4:25.92:103,0,26
chr1    974165  rs9442391   T   C   29.84   LowQual AC=1;AF=0.50;AN=2;DB;DP=18;Dels=0.00;HRun=1;HaplotypeScore=0.16;MQ=95.26;MQ0=0;QD=1.66;SB=-0.98 GT:AD:DP:GQ:PL  0/1:14,4:14:60.91:61,0,255

It seems a bit complex, but the structure of the file is actually quite simple:

[HEADER LINES]
#CHROM  POS ID      REF ALT QUAL    FILTER  INFO          FORMAT          NA12878
chr1    873762  .       T   G   5231.78 PASS    [ANNOTATIONS] GT:AD:DP:GQ:PL  0/1:173,141:282:99:255,0,255
chr1    877664  rs3828047   A   G   3931.66 PASS    [ANNOTATIONS] GT:AD:DP:GQ:PL  1/1:0,105:94:99:255,255,0
chr1    899282  rs28548431  C   T   71.77   PASS    [ANNOTATIONS] GT:AD:DP:GQ:PL  0/1:1,3:4:25.92:103,0,26
chr1    974165  rs9442391   T   C   29.84   LowQual [ANNOTATIONS] GT:AD:DP:GQ:PL  0/1:14,4:14:60.91:61,0,255

After the header lines and the field names, each line represents a single variant, with various properties of that variant represented in the columns. Note that here everything is a SNP, but some could be indels or CNVs.

3. How variation is represented

The first 6 columns of the VCF, which represent the observed variation, are easy to understand because they have a single, well-defined meaning.

• CHROM and POS : The CHROM and POS gives the contig on which the variant occurs. For indels this is actually the base preceding the event, due to how indels are represented in a VCF.

• ID: The dbSNP rs identifier of the SNP, based on the contig and position of the call and whether a record exists at this site in dbSNP.

• REF and ALT: The reference base and alternative base that vary in the samples, or in the population in general. Note that REF and ALT are always given on the forward strand. For indels the REF and ALT bases always include at least one base each (the base before the event).

• QUAL: The Phred scaled probability that a REF/ALT polymorphism exists at this site given sequencing data. Because the Phred scale is -10 * log(1-p), a value of 10 indicates a 1 in 10 chance of error, while a 100 indicates a 1 in 10^10 chance. These values can grow very large when a large amount of NGS data is used for variant calling.

• FILTER: In a perfect world, the QUAL field would be based on a complete model for all error modes present in the data used to call. Unfortunately, we are still far from this ideal, and we have to use orthogonal approaches to determine which called sites, independent of QUAL, are machine errors and which are real SNPs. Whatever approach is used to filter the SNPs, the VCFs produced by the GATK carry both the PASSing filter records (the ones that are good have PASS in their FILTER field) as well as those that fail (the filter field is anything but PASS or a dot). If the FILTER field is a ".", then no filtering has been applied to the records, meaning that all of the records will be used for analysis but without explicitly saying that any PASS. You should avoid such a situation by always filtering raw variant calls before analysis.

For more details about these fields, please see this page.

In the excerpt shown above, here is how we interpret the line corresponding to each variant:

• chr1:873762 is a novel T/G polymorphism, found with very high confidence (QUAL = 5231.78)
• chr1:877664 is a known A/G SNP (named rs3828047), found with very high confidence (QUAL = 3931.66)
• chr1:899282 is a known C/T SNP (named rs28548431), but has a relative low confidence (QUAL = 71.77)
• chr1:974165 is a known T/C SNP but we have so little evidence for this variant in our data that although we write out a record for it (for book keeping, really) our statistical evidence is so low that we filter the record out as a bad site, as indicated by the "LowQual" annotation.

4. How genotypes are represented

The genotype fields of the VCF look more complicated but they're actually not that hard to interpret once you understand that they're just sets of tags and values. Let's take a look at three of the records shown earlier, simplified to just show the key genotype annotations:

chr1    873762  .       T   G   [CLIPPED] GT:AD:DP:GQ:PL    0/1:173,141:282:99:255,0,255
chr1    877664  rs3828047   A   G   [CLIPPED] GT:AD:DP:GQ:PL    1/1:0,105:94:99:255,255,0
chr1    899282  rs28548431  C   T   [CLIPPED] GT:AD:DP:GQ:PL    0/1:1,3:4:25.92:103,0,26

Looking at that last column, here is what the tags mean:

• GT : The genotype of this sample. For a diploid organism, the GT field indicates the two alleles carried by the sample, encoded by a 0 for the REF allele, 1 for the first ALT allele, 2 for the second ALT allele, etc. When there's a single ALT allele (by far the more common case), GT will be either:

  • 0/0 - the sample is homozygous reference
  • 0/1 - the sample is heterozygous, carrying 1 copy of each of the REF and ALT alleles
  • 1/1 - the sample is homozygous alternate In the three examples above, NA12878 is observed with the allele combinations T/G, G/G, and C/T respectively.

• GQ: The Genotype Quality, or Phred-scaled confidence that the true genotype is the one provided in GT. In the diploid case, if GT is 0/1, then GQ is really L(0/1) / (L(0/0) + L(0/1) + L(1/1)), where L is the likelihood that the sample is 0/0, 0/1/, or 1/1 under the model built for the NGS dataset.

• AD and DP: These are complementary fields that represent two important ways of thinking about the depth of the data for this sample at this site. See the Technical Documentation for details on AD (DepthPerAlleleBySample) and DP (DepthOfCoverage).

• PL: This field provides the likelihoods of the given genotypes (here, 0/0, 0/1, and 1/1). These are normalized, Phred-scaled likelihoods for each of the 0/0, 0/1, and 1/1, without priors. To be concrete, for the heterozygous case, this is L(data given that the true genotype is 0/1). The most likely genotype (given in the GT field) is scaled so that it's P = 1.0 (0 when Phred-scaled), and the other likelihoods reflect their Phred-scaled likelihoods relative to this most likely genotype.

With that out of the way, let's interpret the genotypes for NA12878 at chr1:899282.

chr1    899282  rs28548431  C   T   [CLIPPED] GT:AD:DP:GQ:PL    0/1:1,3:4:25.92:103,0,26

At this site, the called genotype is GT = 0/1, which is C/T. The confidence (GQ=25.92) isn't so good, largely because there were only a total of 4 reads at this site (DP=4), 1 of which was ref (=had the reference base) and 3 of which were alt (=had the alternate base) (AD=1,3). The lack of certainty is evident in the PL field, where PL(0/1) = 0 (the normalized value), whereas there's a serious chance that the subject is hom-var (=homozygous with the variant allele) since PL(1/1) = 26 = 10^(-2.6) = 0.25%. Either way, though, it's clear that the subject is definitely not home-ref (=homozygous with the reference allele) here since PL(0/0) = 103 = 10^(-10.3) which is a very small number.

5. Understanding annotations

Finally, variants in a VCF can be annotated with a variety of additional tags, either by the built-in tools or with others that you add yourself. The way they're formatted is similar to what we saw in the Genotype fields, except instead of being in two separate fields (tags and values, respectively) the annotation tags and values are grouped together, so tag-value pairs are written one after another.

chr1    873762  [CLIPPED] AC=1;AF=0.50;AN=2;DP=315;Dels=0.00;HRun=2;HaplotypeScore=15.11;MQ=91.05;MQ0=15;QD=16.61;SB=-1533.02;VQSLOD=-1.5473
chr1    877664  [CLIPPED] AC=2;AF=1.00;AN=2;DB;DP=105;Dels=0.00;HRun=1;HaplotypeScore=1.59;MQ=92.52;MQ0=4;QD=37.44;SB=-1152.13;VQSLOD= 0.1185
chr1    899282  [CLIPPED] AC=1;AF=0.50;AN=2;DB;DP=4;Dels=0.00;HRun=0;HaplotypeScore=0.00;MQ=99.00;MQ0=0;QD=17.94;SB=-46.55;VQSLOD=-1.9148

Here are some commonly used built-in annotations and what they mean: