We've moved!
This site is now read-only. You can find our new documentation site and support forum for posting questions here.
Be sure to read our welcome blog!

QCing IndelRealigner on low-coverage (10x) mouse WGS data


I've run the IndelRealigner on my mouse WGS *bam files with known site data from the Sanger MGP, and now I'm trying to figure out how "well" it worked.

The list created by RealignerTargetCreator contains 6547185 intervals

Parsing the output realigned.bam file for reads that had an "OC" tag added (as suggested in http://www.broadinstitute.org/gatk/events/3391/GATKw1310-BP-2-Realignment.pdf) shows that 1648299 reads were actually realigned.

I used the default settings, which means that

1) -model was USE_READS - and from what I've read, this is the correct option to use, given that Smith-Waterman modelling doesn't give greatly improved results;

2) -LOD was 5.0 - but for my data, which is mouse whole-genome sequence at average 10x coverage, this may be too high and I might be losing true positives.

I've tried randomly picking out candidate intervals from the intervals and OC-tagged reads from the realigned.bam file to check, but I was wondering if there's a more empirical way of checking how good the realignment was (I realise there's "no formal measure" as per the presentation but I'm finding it hard to make a judgement call!).

My feeling from looking at the intervals or realigned reads is that the low coverage is a major issue in terms of identifying "true" indels, so preferably we'd go for specificity over sensitivity.

Thanks for any advice/suggestions in advance!

Best Answer


  • mfletchermfletcher DEMember

    Hi @Geraldine_VdAuwera‌,

    Yes, I'm using HaplotypeCaller to call variants. It makes sense that local haplotype reconstruction means that indel realignment has less of an effect (I guess that realignment would have a much greater effect for UnifiedGenotyper, too).

    I've managed to run BQSR using the Mouse Genome Project known SNPs and there's very definitely an effect of BQSR; I have variant calls from the un-processed bam file so I'll be able to make that fully pre-processed vs original file comparison soon.

    My main concern is that that the IndelRealigner LOD threshold was set too high given the relatively low coverage (10x) of my data. However, given we have no set of known indels that we could use to validate any realignments, I think proceeding to variant calling using the pre-processed data is the best course of action.

    Thanks very much for your suggestions!

Sign In or Register to comment.