VQSR vs Hard Filtering

Follow

@Anna

• October 22, 2021 12:52

Hi, I’m analyzing >200 human exomes following GATK Best Practices. I’ve tried to read all the GATK documentation about this, but I’d like to ask you something just to be sure. I understand that it is recommended using VQSR instead of hard filtering. I used the option --truth-sensitivity-filter-level 99.0 in ApplyVQSR, and in the end, I kept only the variants that had the PASS filter. I would like to confirm if I should apply any more filters (like QD<2) to make sure I don’t have any false positives. Thank you in advance

0

• 

  Pamela Bretscher

  • October 22, 2021 14:50

  Hi Anna,

  The filters you decide to use really just depend on your data and the sensitivity and precision that you would like to achieve with filtering. You're correct that VQSR is generally recommended for filtering and the filtering you have already done should be sufficient. If you wish, you could use VariantFiltration to hard filter your variants using filters like QD<2, QUAL<30, MQRankSum, etc., and compare your results. However, I think using VQSR should give you high enough accuracy. I hope this helps answer your question.

  Kind regards,

  Pamela

  0

  Comment actions Permalink
• 

  @Anna

  • October 25, 2021 11:59

  Hi Pamela,

  Thank you so much for your explanatory answer. I have below two variants as an example of my doubt, hope that’s ok (I just copied 3 representative samples, as I have >200, all with DP<10) .

  QUAL=974

  PASS

  AC=14;AF=0.35;AN=40;DP=43;ExcessHet=0;FS=0;InbreedingCoeff=0.3492;MLEAC=120;MLEAF=1;MQ=48.52;POSITIVE_TRAIN_SITE;QD=25.51;RAW_MQ=61200;SOR=1.061;VQSLOD=3.14;culprit=FS

  GT:AD:DP:GQ:PL

  1/1:0,2:2:6:49,6,0

  ./.:0,0:0:.:0,0,0

  ./.:0,0:0:.:0,0,0

  QUAL=1129

  PASS

  AC=17;AF=0.149;AN=114;BaseQRankSum=0;DP=157;ExcessHet=0;FS=0;InbreedingCoeff=0.3537;MLEAC=58;MLEAF=0.509;MQ=47.61;MQRankSum=0;POSITIVE_TRAIN_SITE;QD=34.23;RAW_MQ=122400;ReadPosRankSum=1.44;SOR=0.859;VQSLOD=5.87;culprit=DP

  GT:AD:DP:GQ:PGT:PID:PL

  1/1:0,1:1:3:.:.:45,3,0

  0/0:4,0:4:12:.:.:0,12,99./.:0,0:0:.:.:.:0,0,0

  ./.:0,0:0:.:.:.:0,0,0

  So, should I consider them as true variants, even though both have really low DP across all samples?

  Or should I filter variants that have at least two samples with DP>10, for example?

  Best regards,
  Anna

  0

  Comment actions Permalink
• 

  Pamela Bretscher

  • October 25, 2021 13:21

  Hi Anna,

  I understand your concern with these variants. However, it is actually not recommended to use the DP annotation when working with exome samples because of the really high variation in depth. The variation is seen as an error by the filtering tools when working with whole-genome data, but it isn't necessarily indicative of error when working with exome samples, which you mentioned that you are. Given that you have a large number of samples, VQSR should still be suitable for filtering, but the DP filter shouldn't be used with exomes. I hope this is helpful.

  Kind regards,

  Pamela

  0

  Comment actions Permalink
• 

  @Anna

  • October 25, 2021 20:07
  • Edited

  Hi Pamela,

  OK, I understand it. Again, thanks a lot for explaining this. It is truly helpful for me.

  Best regards

  0

  Comment actions Permalink

Please sign in to leave a comment.