CNV evaluate confidence

Question

[Daria]

Hello! I am doing a CNV calling based on the CBS algorithm using this article https://jeremy9959.net/Blog/cbs-fixed/#:~:text=Circular%20Binary%20Segmentation%20is%20an,based%20DNA%20copy %20number%20data. I have a question, how can I evaluate the quality of the CNVs found? Are there any metrics and formulas for them? So far I have found Derivative log ratio spread (DLRS), Z-score.

Answer 1

[cmdcolin]

there are some simple statistics defined by https://github.com/brentp/duphold

quote from readme

    DHFC: fold-change for the variant depth relative to the rest of the chromosome the variant was found on
    DHBFC: fold-change for the variant depth relative to bins in the genome with similar GC-content.
    DHFFC: fold-change for the variant depth relative to Flanking regions.

It also adds GCF to the INFO field indicating the fraction of G or C bases in the variant.

After annotating with duphold, a sensible way to filter to high-quality variants is:

bcftools view -i '(SVTYPE = "DEL" & FMT/DHFFC[0] < 0.7) | (SVTYPE = "DUP" & FMT/DHBFC[0] > 1.3)' $svvcf

In our evaluations, DHFFC works best for deletions and DHBFC works slightly better for duplications. For genomes/samples with more variable coverage, DHFFC should be the most reliable.