GATK MuTect2 duplication filter

Question

[umn_bist]

So I have a set of tumor:matched normal samples. I have them deduped with picard for PCR contamination. Afterwards I use MuTect2 to call somatic variants against dbSNP, COSMIC coding mutation, COSMIC noncoding mutation. And for some reason about 10% of my reads are being filtered out as duplicates.

I suspect that these "duplicates" are not contaminants and was wondering what may be going on. Could it be rRNA that were not trimmed during pre-processing QC?

Answer 1

[DG]

picard de-duplication is based almost entirely on reads mapping to identical start/end points. A 10% duplication rate is high, but wouldn't be totally out of the ordinary either. You have your post tagged as RNA-Seq and mention rRNA but your workflow with Mutect2 reads more like a DNA-Seq alignment and processing workflow. Are you calling variants from your RNA-Seq data? Given how Picard deduplicates data it tends to grossly overestimate duplicated reads when dealing with RNA-Seq data so it is usually skipped. Some clarification about the source of your data and the type of experiment might clear things up further.

[umn_bist]

Of course. I am working with RNA-seq data from TCGA for calling somatic variant calling using Mutect2. I have both tumor and its matched normals.

From my understanding, deduping was encouraged (according to Broad/GATK) to remove PCR contaminants. I am not trimming but simply marking my duplicates.

Could it be that the duplicates that are being filtered by MuTect2 are actually my marked duplicates?

[DG]

Unless you specify not to MuTect2 automatically filters marked duplicates. That's the whole point of marking duplicates is so they aren't considered by downstream variant callers and other tools.The GATK/Broad best-practices documents is primarily geared towards working with DNA sequencing data. Many of the steps have not been validated when working with RNA. RNA-based variant calling has always been considered a little bit more problematic than DNA-based results as the underlying error rate is higher for individual nucleotides.