sequencing artifacts passing GATK's ApplyRecalibration filters

[efoss]

I have five human exome data sets that I am putting through the GATK "Best Practices" pipeline. After running things through UnifiedGenotyper, I separated the SNPs from the indels and then ran VariantRecalibrator on the SNPs:

java -Xmx36g -jar /home/efoss/sequencing/GenomeAnalysisTK-1.1-31-gdc8398e/GenomeAnalysisTK.jar -T VariantRecalibrator -R /mnt/fred/home/efoss/gene_databases/human_genome/human_g1k_v37/human_g1k_v37.fasta --percentBadVariants 0.05 --maxGaussians 6 -B:input,VCF /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.raw.vcf -B:hapmap,VCF,known=false,training=true,truth=true,prior=15.0 /mnt/fred/home/efoss/gene_databases/GATK_resource_files/hapmap_3.3.b37.sites.vcf -Bmni,VCF,known=false,training=true,truth=false,prior=12.0 /mnt/fred/home/efoss/gene_databases/GATK_resource_files/1000G_omni2.5.b37.sites.vcf -B:dbsnp,VCF,known=true,training=false,truth=false,prior=8.0 /mnt/fred/home/efoss/gene_databases/GATK_resource_files/dbsnp_132.b37.vcf -an QD -an HaplotypeScore -an MQRankSum -an ReadPosRankSum -an FS -an MQ -mode SNP -recalFile /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.recal -tranchesFile /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.tranches -rscriptFile /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.plots.R

I then ran ApplyRecalibration:

java -Xmx36g -jar /home/efoss/sequencing/GenomeAnalysisTK-1.1-31-gdc8398e/GenomeAnalysisTK.jar -T ApplyRecalibration -R /mnt/fred/home/efoss/gene_databases/human_genome/human_g1k_v37/human_g1k_v37.fasta -B:input,VCF /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.raw.vcf --ts_filter_level 99.0 -tranchesFile /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.tranches -recalFile /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.recal -o /mnt/fred/home/efoss/sequencing/MDS/all_five_patients_UnifiedGenotyper_output_SNPs_091711_1.recalibrated.raw.vcf


My problem is that there are plenty of bad calls getting through ApplyRecalibration. I pick up quite a few mutations that appear in 5 out of 5 (unrelated) individuals, so I can be confident that they are sequencing artifacts. However, these mutations are often listed as passing the filters. Here are two examples:

1 801943 rs7516866 C T 44026.78 PASS AC=10;AF=1.00;AN=10;BaseQRankSum=4.380;DB;DP=1449;DS;Dels=0.00;FS=0.000;HRun=1;HaplotypeScore=16.2975;MQ=47.88;MQ0=66;MQRankSum=7.651;QD=30.38;ReadPosRankSum=5.739;SB=-18779.46;VQSLOD=-4.0745;culprit=ReadPosRankSum GT:ADP:GQ:PL 1/1:10,268:278:99:7953,572,0 1/1:6,293:299:99:9124,682,0 1/1:5,268:273:99:7936,586,0 1/1:4,295:299:99:9809,761,0 1/1:10,290:300:99:8784,610,0
1 802191 rs71507475 G A 1686.49 PASS AC=5;AF=0.50;AN=10;BaseQRankSum=2.704;DB;DP=1499;DS;Dels=0.00;FS=5.855;HRun=0;HaplotypeScore=44.5371;MQ=22.55;MQ0=460;MQRankSum=-3.161;QD=1.13;ReadPosRankSum=-0.860;SB=-867.92;VQSLOD=-15469854734583906.0000;culprit=HaplotypeScore GT:ADP:GQ:PL 0/1:224,69:300:99:286,0,2161 0/1:225,71:299:99:264,0,2035 0/1:237,58:300:99:230,0,2248 0/1:217,75:300:99:525,0,1676 0/1:213,77:300:99:426,0,1708

I gathered up all of these 5/5 sequencing artifacts and checked the results of the exome analysis published in Nature Genetics and not a single one appeared in those results, so clearly I'm either running the analysis incorrectly or I'm interpreting the results i ncorrectly (or both).

I would appreciate any suggestions.

Thank you very much for the help.

Eric

[efoss]

To add a bit more information: The number of sequence information lines in my unrecalibrated file is the same as in my recalibrated file, so clearly nothing was filtered out. Furthermore, the FILTER column almost always says "PASS" and when it doesn't, it says "TruthSensitivityTranche99.00to99.90".

[gavin.oliver]

I may be wrong, but isn't the recalibrator just scoring your variants with a new metric? Therefore the filtering itself would still remain to be done based on the VQSLOD values?

[wanguan2000]

I have the same question. And I found the number of LowQual using ApplyRecalibration as the same as the LowQual in raw.vcf.
LowQual number does not change!!! and a little PASS change to TruthSensitivityTranche99.00to99.90.

should I filtering itself based on the VQSLOD values? but:

http://www.broadinstitute.org/gsa/wiki/index.php/Best_Practice_Variant_Detection_with_the_GATK_v3

don't say anything.

[petriedish]

I realise this is a very late reply, but if you look at those results there are rs numbers from dbSNP showing that those calls are common SNPs. For example, the first one, rs7516866 is seen in approximately 25% of alleles. With the huge number of calls it is not surprising that some of these very common SNPs will be shared between all of your 5 samples.