This document describes the resource datasets and arguments that we recommend for use in the two steps of VQSR (i.e. the successive application of VariantRecalibrator and ApplyRecalibration), based on our work with human genomes, to comply with the GATK Best Practices. The recommendations detailed in this document take precedence over any others you may see elsewhere in our documentation (e.g. in Tutorial articles, which are only meant to illustrate usage, or in past presentations, which may be out of date).

The document covers:

• Explanation of resource datasets
• Important notes about annotations
• Important notes about exome experiments
• Argument recommendations for VariantRecalibrator
• Argument recommendations for ApplyRecalibration
• These recommendations are valid for use with calls generated by both the UnifiedGenotyper and HaplotypeCaller. In the past we made a distinction in how we processed the calls from these two callers, but now we treat them the same way. These recommendations will probably not work properly on calls generated by other (non-GATK) callers.

Note that VQSR must be run twice in succession in order to build a separate error model for SNPs and INDELs (see the VQSR documentation for more details).

Explanation of resource datasets

The human genome training, truth and known resource datasets mentioned in this document are all available from our resource bundle.

If you are working with non-human genomes, you will need to find or generate at least truth and training resource datasets with properties corresponding to those described below. To generate your own resource set, one idea is to first do an initial round of SNP calling and only use those SNPs which have the highest quality scores. These sites which have the most confidence are probably real and could be used as truth data to help disambiguate the rest of the variants in the call set. Another idea is to try using several SNP callers in addition to the UnifiedGenotyper or HaplotypeCaller, and use those sites which are concordant between the different methods as truth data. In either case, you'll need to assign your set a prior likelihood that reflects your confidence in how reliable it is as a truth set. We recommend Q10 as a starting value, which you can then experiment with to find the most appropriate value empirically. There are many possible avenues of research here. Hopefully the model reporting plots that are generated by the recalibration tools will help facilitate this experimentation.

Resources for SNPs

• True sites training resource: HapMap - This resource is a SNP call set that has been validated to a very high degree of confidence. The program will consider that the variants in this resource are representative of true sites (truth=true), and will use them to train the recalibration model (training=true). We will also use these sites later on to choose a threshold for filtering variants based on sensitivity to truth sites. The prior likelihood we assign to these variants is Q15 (96.84%).

• True sites training resource: Omni - This resource is a set of polymorphic SNP sites produced by the Omni geno- typing array. The program will consider that the variants in this resource are representative of true sites (truth=true), and will use them to train the recalibration model (training=true). The prior likelihood we assign to these variants is Q12 (93.69%).

• Non-true sites training resource: 1000G - This resource is a set of high-confidence SNP sites produced by the 1000 Genomes Project. The program will consider that the variants in this re- source may contain true variants as well as false positives (truth=false), and will use them to train the recalibration model (training=true). The prior likelihood we assign to these variants is Q10 (90%).

• Known sites resource, not used in training: dbSNP - This resource is a call set that has not been validated to a high degree of confidence (truth=false). The program will not use the variants in this resource to train the recalibration model (training=false). However, the program will use these to stratify output metrics such as Ti/Tv ratio by whether variants are present in dbsnp or not (known=true). The prior likelihood we assign to these variants is Q2 (36.90%).

Resources for Indels

• True sites training resource: Mills - This resource is an indel call set that has been validated to a high degree of confidence. The program will consider that the variants in this resource are representative of true sites (truth=true), and will use them to train the recalibration model (training=true). The prior likelihood we assign to these variants is Q12 (93.69%).

• Known sites resource, not used in training: dbSNP - This resource is a call set that has not been validated to a high degree of confidence (truth=false). The program will not use the variants in this resource to train the recalibration model (training=false). However, the program will use these to stratify output metrics such as Ti/Tv ratio by whether variants are present in dbsnp or not (known=true). The prior likelihood we assign to these variants is Q2 (36.90%).

• Non-true sites training resource: Axiom (1000G) - This resource is an indel callset based on the Affymetrix Axiom array on 1000 Genomes Project samples. VQSR will consider that the variants in this resource may contain true variants as well as false positives (truth=false). It will use them to train the recalibration model (training=true). The prior likelihood we assign to these variants is Q10 (90%).

Important notes about annotations

Some of the annotations included in the recommendations given below might not be the best for your particular dataset. In particular, the following caveats apply:

• Depth of coverage (the DP annotation) should not be used when working with exome datasets since there is extreme variation in the depth to which targets are captured! In whole genome experiments this variation is indicative of error but that is not the case in capture experiments.

• You may have seen HaplotypeScore mentioned in older documents. That is a statistic produced by UnifiedGenotyper that should only be used if you called your variants with UG. This statistic isn't produced by the HaplotypeCaller because that mathematics is already built into the likelihood function itself when calling full haplotypes with HC.

• The InbreedingCoeff is a population level statistic that requires at least 10 samples in order to be computed. For projects with fewer samples, or that includes many closely related samples (such as a family) please omit this annotation from the command line.

Important notes for exome capture experiments

In our testing we've found that in order to achieve the best exome results one needs to use an exome SNP and/or indel callset with at least 30 samples. For users with experiments containing fewer exome samples there are several options to explore:

• Add additional samples for variant calling, either by sequencing additional samples or using publicly available exome bams from the 1000 Genomes Project (this option is used by the Broad exome production pipeline). Be aware that you cannot simply add VCFs from the 1000 Genomes Project. You must either call variants from the original BAMs jointly with your own samples, or (better) use the reference model workflow to generate GVCFs from the original BAMs, and perform joint genotyping on those GVCFs along with your own samples' GVCFs with GenotypeGVCFs.

• You can also try using the VQSR with the smaller variant callset, but experiment with argument settings (try adding '--maxGaussians 4' to your command line, for example). You should only do this if you are working with a non-model organism for which there are no available genomes or exomes that you can use to supplement your own cohort.

Argument recommendations for VariantRecalibrator

The variant quality score recalibrator builds an adaptive error model using known variant sites and then applies this model to estimate the probability that each variant is a true genetic variant or a machine artifact. One major improvement from previous recommended protocols is that hand filters do not need to be applied at any point in the process now. All filtering criteria are learned from the data itself.

Common, base command line for VariantRecalibrator

This is the first part of the VariantRecalibrator command line, to which you need to add either the SNP-specific recommendations or the indel-specific recommendations given further below.

gatk VariantRecalibrator \
   -R Homo_sapiens_assembly38.fasta \
   -V input.vcf.gz \

SNP specific recommendations for VariantRecalibrator

For SNPs we use both HapMap v3.3 and the Omni chip array from the 1000 Genomes Project as training data. In addition we take the highest confidence SNPs from the project's callset. These datasets are available in the GATK resource bundle.

   -resource:hapmap,known=false,training=true,truth=true,prior=15.0 hapmap_3.3.b37.vcf \
   -resource:omni,known=false,training=true,truth=true,prior=12.0 1000G_omni2.5.b37.vcf \
   -resource:1000G,known=false,training=true,truth=false,prior=10.0 1000G_phase1.snps.high_confidence.b37.vcf \
   -resource:dbsnp,known=true,training=false,truth=false,prior=2.0 dbsnp_138.b37.vcf \
   -an QD -an MQ -an MQRankSum -an ReadPosRankSum -an FS -an SOR -an DP -an InbreedingCoeff \
   -mode SNP \

Please note that these recommendations are formulated for whole-genome datasets. For exomes, we do not recommend using DP for variant recalibration (see below for details of why).

Note also that, for the above to work, the input vcf needs to be annotated with the corresponding values (QD, FS, DP, etc.). If any of these values are somehow missing, then VariantAnnotator needs to be run first so that VariantRecalibration can run properly.

Also, using the provided sites-only truth data files is important here as parsing the genotypes for VCF files with many samples increases the runtime of the tool significantly.

You may notice that these recommendations no longer include the --numBadVariants argument. That is because we have removed this argument from the tool, as the VariantRecalibrator now determines the number of variants to use for modeling "bad" variants internally based on the data.

Indel specific recommendations for VariantRecalibrator

When modeling indels with the VQSR we use a training dataset that was created at the Broad by strictly curating the (Mills, Devine, Genome Research, 2011) dataset as as well as adding in very high confidence indels from the 1000 Genomes Project. This dataset is available in the GATK resource bundle.

   --maxGaussians 4 \
   -resource:mills,known=false,training=true,truth=true,prior=12.0 Mills_and_1000G_gold_standard.indels.b37.vcf  \
   -resource:dbsnp,known=true,training=false,truth=false,prior=2.0 dbsnp_138.b37.vcf\
   -an QD -an DP -an FS -an SOR -an ReadPosRankSum -an MQRankSum -an InbreedingCoeff \
   -mode INDEL \

Note that indels use a different set of annotations than SNPs. Most annotations related to mapping quality have been removed since there is a conflation with the length of an indel in a read and the degradation in mapping quality that is assigned to the read by the aligner. This covariation is not necessarily indicative of being an error in the same way that it is for SNPs.

You may notice that these recommendations no longer include the --numBadVariants argument. That is because we have removed this argument from the tool, as the VariantRecalibrator now determines the number of variants to use for modeling "bad" variants internally based on the data.

Argument recommendations for ApplyRecalibration

The power of the VQSR is that it assigns a calibrated probability to every putative mutation in the callset. The user is then able to decide at what point on the theoretical ROC curve their project wants to live. Some projects, for example, are interested in finding every possible mutation and can tolerate a higher false positive rate. On the other hand, some projects want to generate a ranked list of mutations that they are very certain are real and well supported by the underlying data. The VQSR provides the necessary statistical machinery to effectively apply this sensitivity/specificity tradeoff.

### SNP specific recommendations for ApplyRecalibration

As previously described, we used HapMap 3.3 and the Omni 2.5M chip as our SNP truth set. We typically seek to achieve 99.5% sensitivity to the accessible truth sites, but this is by no means universally applicable: you will need to experiment to find out what tranche cutoff is right for your data. Generally speaking, projects involving a higher degree of diversity in terms of world populations can expect to achieve a higher truth sensitivity than projects with a smaller scope.

   --ts_filter_level 99.5 \
   -mode SNP \

Indel specific recommendations for ApplyRecalibration

For indels we use the Mills / 1000 Genomes indel truth set described above. We typically seek to achieve 99.0% sensitivity to the accessible truth sites, but this is by no means universally applicable: you will need to experiment to find out what tranche cutoff is right for your data. Generally speaking, projects involving a higher degree of diversity in terms of world populations can expect to achieve a higher truth sensitivity than projects with a smaller scope.

   --ts_filter_level 99.0 \
   -mode INDEL \