Quality issues with DRAGEN joint variant callset

Ruebena Ellen Dawes

Hello,

I've been working with the DRAGEN joint variant call release of the 500k WGS data since it was released. We've been doing some extensive QC to run association testing, and have across these contiguous blocks where genotype quality calls drop substantially. When we set genotype calls with GQ < 20 to missing, these blocks end up with 30-100% of genotype calls missing! The blocks are around 5kb. This is highly unusual, and is not the case in the Graphtyper joint variant call set. Has anyone else found this issue when working with the DRAGEN dataset? We are considering swapping over to using Graphtyper for the project, which will be a very costly and time-consuming setback a year into using the data. On the future data releases page it says there is a ML-corrected DRAGEN population level dataset slated to be released in Q1 2025- is this ‘ML-corrected’ because there are known issues with the dataset? Any information would be much appreciated!

Kind Regards,

Ruby 

Comments

1 comment

Date Votes
  • Comment author
    Ole Benjamin Schulz-trieglaff

    Hi Ruby,

    Thanks for sharing this — and for the detailed QC work you’ve done. Sorry for the very late response!

    I would advice applying a global threshold on genotypes qualities (FORMAT/GQ) - the GQ is meant to reflect uncertainty in calling but not a hard-call failure. Also, in our experience the hom-ref and variant genotypes qualities will follow different distributions (the underlying models are different). Hom-ref genotypes are also far more numerous than variants.

    I realize that this response might be too later but before switching pipelines, it may be worth exploring:

    • using the ML‑corrected DRAGEN population‑level dataset, which applies cohort‑level machine‑learning filtering to improve genotype consistency and reduce downstream missingness, and
    • applying different QC thresholds for hom‑ref and variant genotypes, rather than a single global GQ cutoff.

    The ML‑corrected release is intended as an improved starting point for downstream analyses such as association testing. We would highly recommend using that for most downstream analyses.

    If you continue to see strong block‑level effects after those steps, it would be interesting to compare notes on specific regions and QC strategies.

    I hope this is useful, please do get in touch if we can help with anything else.

    Best regards,


    Ole

    1

Please sign in to leave a comment.