Build a recalibration model to score variant quality for filtering purposes
The purpose of variant recalibration is to assign a well-calibrated probability to each variant call in a call set. You can then create highly accurate call sets by filtering based on this single estimate for the accuracy of each call. The approach taken by variant quality score recalibration is to develop a continuous, covarying estimate of the relationship between SNP call annotations (such as QD, MQ, and ReadPosRankSum, for example) and the probability that a SNP is a true genetic variant versus a sequencing or data processing artifact. This model is determined adaptively based on "true sites" provided as input, typically HapMap 3 sites and those sites found to be polymorphic on the Omni 2.5M SNP chip array (in humans). This adaptive error model can then be applied to both known and novel variation discovered in the call set of interest to evaluate the probability that each call is real. The score that gets added to the INFO field of each variant is called the VQSLOD. It is the log odds of being a true variant versus being false under the trained Gaussian mixture model.
This tool performs the first pass in a two-stage process called VQSR; the second pass is performed by the ApplyRecalibration tool. In brief, the first pass consists of creating a Gaussian mixture model by looking at the distribution of annotation values over a high quality subset of the input call set, and then scoring all input variants according to the model. The second pass consists of filtering variants based on score cutoffs identified in the first pass.
VQSR is probably the hardest part of the Best Practices to get right, so be sure to read the method documentation, parameter recommendations and tutorial to really understand what these tools and how to use them for best results on your own data.
Recalibrating SNPs in exome data:
java -Xmx4g -jar GenomeAnalysisTK.jar \ -T VariantRecalibrator \ -R reference.fasta \ -input raw_variants.vcf \ -resource:hapmap,known=false,training=true,truth=true,prior=15.0 hapmap_3.3.b37.sites.vcf \ -resource:omni,known=false,training=true,truth=false,prior=12.0 1000G_omni2.5.b37.sites.vcf \ -resource:1000G,known=false,training=true,truth=false,prior=10.0 1000G_phase1.snps.high_confidence.vcf -resource:dbsnp,known=true,training=false,truth=false,prior=2.0 dbsnp_135.b37.vcf \ -an QD -an MQ -an MQRankSum -an ReadPosRankSum -an FS -an SOR -an InbreedingCoeff \ -mode SNP \ -recalFile output.recal \ -tranchesFile output.tranches \ -rscriptFile output.plots.R
java -Xmx4g -jar GenomeAnalysisTK.jar \ -T VariantRecalibrator \ -R reference.fasta \ -input raw_variants.withASannotations.vcf \ -AS \ -resource:hapmap,known=false,training=true,truth=true,prior=15.0 hapmap_3.3.b37.sites.vcf \ -resource:omni,known=false,training=true,truth=false,prior=12.0 1000G_omni2.5.b37.sites.vcf \ -resource:1000G,known=false,training=true,truth=false,prior=10.0 1000G_phase1.snps.high_confidence.vcf -resource:dbsnp,known=true,training=false,truth=false,prior=2.0 dbsnp_135.b37.vcf \ -an QD -an MQ -an MQRankSum -an ReadPosRankSum -an FS -an SOR -an InbreedingCoeff \ -mode SNP \ -recalFile output.AS.recal \ -tranchesFile output.AS.tranches \ -rscriptFile output.plots.AS.RThe input VCF must have been produced using allele-specific annotations in HaplotypeCaller. Note that each allele will have a separate line in the output .recal file with its own VQSLOD and culprit that will be transferred to the final VCF in ApplyRecalibration.
These Read Filters are automatically applied to the data by the Engine before processing by VariantRecalibrator.
This tool can be run in multi-threaded mode using this option.
All tools inherit arguments from the GATK Engine' "CommandLineGATK" argument collection, which can be used to modify various aspects of the tool's function. For example, the -L argument directs the GATK engine to restrict processing to specific genomic intervals; or the -rf argument allows you to apply certain read filters to exclude some of the data from the analysis.
This table summarizes the command-line arguments that are specific to this tool. For more details on each argument, see the list further down below the table or click on an argument name to jump directly to that entry in the list.
|Argument name(s)||Default value||Summary|
||NA||One or more VCFs of raw input variants to be recalibrated|
||||A list of sites for which to apply a prior probability of being correct but which aren't used by the algorithm (training and truth sets are required to run)|
|NA||The output recal file used by ApplyRecalibration|
|NA||The output tranches file used by ApplyRecalibration|
||SNP||Recalibration mode to employ|
|||The names of the annotations which should used for calculations|
||NA||Additional raw input variants to be used in building the model|
|stdout||A GATKReport containing the positive and negative model fits|
|NA||The output rscript file generated by the VQSR to aid in visualization of the input data and learned model|
|||If specified, the variant recalibrator will also use variants marked as filtered by the specified filter name in the input VCF file|
|2.15||The expected novel Ti/Tv ratio to use when calculating FDR tranches and for display on the optimization curve output figures. (approx 2.15 for whole genome experiments). ONLY USED FOR PLOTTING PURPOSES!|
|[100.0, 99.9, 99.0, 90.0]||The levels of truth sensitivity at which to slice the data. (in percent, that is 1.0 for 1 percent)|
|false||If specified, the variant recalibrator will ignore all input filters. Useful to rerun the VQSR from a filtered output file.|
|false||If specified, the variant recalibrator will output the VQSR model fit to the file specified by -modelFile or to stdout|
|false||If specified, the variant recalibrator will attempt to use the allele-specific versions of the specified annotations.|
||-5.0||LOD score cutoff for selecting bad variants|
||0.001||The dirichlet parameter in the variational Bayes algorithm.|
||1||Number of attempts to build a model before failing|
|8||Max number of Gaussians for the positive model|
|150||Maximum number of VBEM iterations|
|2||Max number of Gaussians for the negative model|
||2500000||Maximum number of training data|
|1000||Minimum number of bad variants|
|0||Apply logit transform and jitter to MQ values|
|100||Number of k-means iterations|
||20.0||The number of prior counts to use in the variational Bayes algorithm.|
||1.0||The shrinkage parameter in the variational Bayes algorithm.|
|10.0||Annotation value divergence threshold (number of standard deviations from the means)|
|false||Trust that all the input training sets' unfiltered records contain only polymorphic sites to drastically speed up the computation.|
Arguments in this list are specific to this tool. Keep in mind that other arguments are available that are shared with other tools (e.g. command-line GATK arguments); see Inherited arguments above.
Additional raw input variants to be used in building the model
These additional calls should be unfiltered and annotated with the error covariates that are intended to be used for modeling.
This argument supports reference-ordered data (ROD) files in the following formats: BCF2, VCF, VCF3
LOD score cutoff for selecting bad variants
Variants scoring lower than this threshold will be used to build the Gaussian model of bad variants.
double -5.0 [ [ -∞ ∞ ] ]
The dirichlet parameter in the variational Bayes algorithm.
double 0.001 [ [ -∞ ∞ ] ]
If specified, the variant recalibrator will ignore all input filters. Useful to rerun the VQSR from a filtered output file.
If specified, the variant recalibrator will also use variants marked as filtered by the specified filter name in the input VCF file
For this to work properly, the -ignoreFilter argument should also be applied to the ApplyRecalibration command.
One or more VCFs of raw input variants to be recalibrated
These variant calls must be annotated with the annotations that will be used for modeling. If the calls come from multiple samples, they must have been obtained by joint calling the samples, either directly (running HaplotypeCaller on all samples together) or via the GVCF workflow (HaplotypeCaller with -ERC GVCF per-sample then GenotypeGVCFs on the resulting gVCFs) which is more scalable. Note that the ability to pass multiple input files is only intended to facilitate scatter-gather parallelism (to enable e.g. running on VCFs generated per-chromosome), not to combine different callsets. The variant calls in the separate input files should not overlap.
R List[RodBindingCollection[VariantContext]] NA
Number of attempts to build a model before failing
The statistical model being built by this tool may fail due to simple statistical sampling issues. Rather than dying immediately when the initial model fails, this argument allows the tool to restart with a different random seed and try to build the model again. The first successfully built model will be kept. Note that the most common underlying cause of model building failure is that there is insufficient data to build a really robust model. This argument provides a workaround for that issue but it is preferable to provide this tool with more data (typically by including more samples or more territory) in order to generate a more robust model.
int 1 [ [ -∞ ∞ ] ]
Max number of Gaussians for the positive model
This parameter determines the maximum number of Gaussians that should be used when building a positive model using the variational Bayes algorithm.
int 8 [ [ -∞ ∞ ] ]
Maximum number of VBEM iterations
This parameter determines the maximum number of VBEM iterations to be performed in the variational Bayes algorithm. The procedure will normally end when convergence is detected.
int 150 [ [ -∞ ∞ ] ]
Max number of Gaussians for the negative model
This parameter determines the maximum number of Gaussians that should be used when building a negative model using the variational Bayes algorithm. The actual maximum used is the smaller value between the mG and mNG arguments, meaning that if -mG is smaller than -mNG, -mG will be used for both. Note that this number should be small (e.g. 4) to achieve the best results.
int 2 [ [ -∞ ∞ ] ]
Maximum number of training data
The number of variants to use in building the Gaussian mixture model. Training sets larger than this will be randomly downsampled.
int 2500000 [ [ -∞ ∞ ] ]
Minimum number of bad variants
This parameter determines the minimum number of variants that will be selected from the list of worst scoring variants to use for building the Gaussian mixture model of bad variants.
int 1000 [ [ -∞ ∞ ] ]
Recalibration mode to employ
Use either SNP for recalibrating only SNPs (emitting indels untouched in the output VCF) or INDEL for indels (emitting SNPs untouched in the output VCF). There is also a BOTH option for recalibrating both SNPs and indels simultaneously, but this is meant for testing purposes only and should not be used in actual analyses.
The --mode argument is an enumerated type (Mode), which can have one of the following values:
R Mode SNP
A GATKReport containing the positive and negative model fits
Apply logit transform and jitter to MQ values
MQ is capped at a "max" value (60 for bwa-mem) when the alignment is considered perfect. Typically, a huge proportion of the reads in a dataset are perfectly mapped, which yields a distribution of MQ values with a blob below the max value and a huge peak at the max value. This does not conform to the expectations of the Gaussian mixture model of VQSR and has been observed to yield a ROC curve with a jump. This argument aims to mitigate this problem. Using MQCap = X has 2 effects: (1) MQs are transformed by a scaled logit on [0,X] (+ epsilon to avoid division by zero) to make the blob more Gaussian-like and (2) the transformed MQ=X are jittered to break the peak into a narrow Gaussian. Beware that IndelRealigner, if used, adds 10 to MQ for successfully realigned indels. We recommend to either use --read-filter ReassignOriginalMQAfterIndelRealignment with HaplotypeCaller or use a MQCap=max+10 to take that into account. If this option is not used, or if MQCap is set to 0, MQ will not be transformed.
int 0 [ [ -∞ ∞ ] ]
Number of k-means iterations
This parameter determines the number of k-means iterations to perform in order to initialize the means of the Gaussians in the Gaussian mixture model.
int 100 [ [ -∞ ∞ ] ]
If specified, the variant recalibrator will output the VQSR model fit to the file specified by -modelFile or to stdout
This GATKReport gives information to describe the VQSR model fit. Normalized means for the positive model are concatenated as one table and negative model normalized means as another table. Covariances are also concatenated for positive and negative models, respectively. Tables of annotation means and standard deviations are provided to help describe the normalization. The model fit report can be read in with our R gsalib package. Individual model Gaussians can be subset by the value in the "Gaussian" column if desired.
The number of prior counts to use in the variational Bayes algorithm.
double 20.0 [ [ -∞ ∞ ] ]
The output recal file used by ApplyRecalibration
R VariantContextWriter NA
A list of sites for which to apply a prior probability of being correct but which aren't used by the algorithm (training and truth sets are required to run)
Any set of VCF files to use as lists of training, truth, or known sites. Training - The program builds the Gaussian mixture model using input variants that overlap with these training sites. Truth - The program uses these truth sites to determine where to set the cutoff in VQSLOD sensitivity. Known - The program only uses known sites for reporting purposes (to indicate whether variants are already known or novel). They are not used in any calculations by the algorithm itself. Bad - A database of known bad variants can be used to supplement the set of worst ranked variants (compared to the Gaussian mixture model) that the program selects from the data to model "bad" variants.
This argument supports reference-ordered data (ROD) files in the following formats: BCF2, VCF, VCF3
R List[RodBinding[VariantContext]] 
The output rscript file generated by the VQSR to aid in visualization of the input data and learned model
The shrinkage parameter in the variational Bayes algorithm.
double 1.0 [ [ -∞ ∞ ] ]
Annotation value divergence threshold (number of standard deviations from the means)
If a variant has annotations more than -std standard deviations away from mean, it won't be used for building the Gaussian mixture model.
double 10.0 [ [ -∞ ∞ ] ]
The expected novel Ti/Tv ratio to use when calculating FDR tranches and for display on the optimization curve output figures. (approx 2.15 for whole genome experiments). ONLY USED FOR PLOTTING PURPOSES!
The expected transition / transversion ratio of true novel variants in your targeted region (whole genome, exome, specific genes), which varies greatly by the CpG and GC content of the region. See expected Ti/Tv ratios section of the GATK best practices documentation (http://www.broadinstitute.org/gatk/guide/best-practices) for more information. Normal values are 2.15 for human whole genome values and 3.2 for human whole exomes. Note that this parameter is used for display purposes only and isn't used anywhere in the algorithm!
double 2.15 [ [ -∞ ∞ ] ]
The output tranches file used by ApplyRecalibration
R File NA
Trust that all the input training sets' unfiltered records contain only polymorphic sites to drastically speed up the computation.
The levels of truth sensitivity at which to slice the data. (in percent, that is 1.0 for 1 percent)
Add truth sensitivity slices through the call set at the given values. The default values are 100.0, 99.9, 99.0, and 90.0 which will result in 4 estimated tranches in the final call set: the full set of calls (100% sensitivity at the accessible sites in the truth set), a 99.9% truth sensitivity tranche, along with progressively smaller tranches at 99% and 90%.
List[Double] [100.0, 99.9, 99.0, 90.0]
The names of the annotations which should used for calculations
See the input VCF file's INFO field for a list of all available annotations.
R List[String] 
If specified, the variant recalibrator will attempt to use the allele-specific versions of the specified annotations.
Generate a VQSR model using per-allele data instead of the default per-site data, assuming that the input VCF contains allele-specific annotations. Annotations should be specified using their full names with AS_ prefix. Non-allele-specific (scalar) annotations will be applied to all alleles.
GATK version 3.7-0-gcfedb67 built at 2017/02/09 12:35:06.