Missing parental genotypes in smallvariants VCF lead to false de novo calls

[emmamiz]

In the PacBio (v3.1.0) smallvariants output, we are seeing variants that appear potentially de novo based on the VCF, but manual review in IGV shows they are clearly inherited. For these variants, the parent the variant was inherited from has all sample-level fields reported as missing in the VCF, despite clear alternate allele support in IGV. This results in a high number of false de novo candidates that require manual review.

For example, chr9:42064090 C>T was paternally inherited but was noted in the VCF as missing in the father (screenshots included below).

Can deeptrio be added for family analyses? Without that, de novo calls are highly unreliable and full of false positives.

[williamrowell]

Could you please share the FORMAT and sample columns from the GLnexus joint-called VCF for this variant? I'm especially interested in GQ and AD, to see if any trace of this variant in the father made it through joint-calling.

How are you currently identifying small de novo variants? Is it based on GT alone, some combination of GT/GQ/AD, or some tool like bcftools +trio-dnm2 that calculates based on genotype quality? slivar, the tool built into our very-basic-tertiary-analysis, has a denovo function that requires that the total parental AD be no more than 1 read. The only reason we didn't implement this in the current workflow is that we wanted the task to work for non-trios as well.

function denovo(kid, mom, dad){
  // check genotypes match expected pattern
  if(!(kid.het && mom.hom_ref && dad.hom_ref)){ return false; }
  if(!hq(kid, mom, dad)){ return false; }
  return ((mom.AD[1] + dad.AD[1]) < 2)
}

There are no immediate plans on the roadmap to integrate DeepTrio, but I can add it to the feature request list.

[emmamiz]

Image below includes the GQ and AD columns, both of which are missing for this variant in the father. I usually filter for small de novo variants using VAF.

[williamrowell]

Could you check the gVCF in the father and proband for this position as well? Thanks, just trying to understand how it missed something that is ~30%.

I suspect this variant is coming from a LINE element from some other part of the sample mapping to the closest sequence in the reference. It's possible that DeepVariant has learned something from the features of this variant that indicate that it is not trustworthy, e.g. variant almost exclusively carried on alignments with soft clipping on both sides. Which version of DeepVariant is this?

[gevro]

Just to add my thoughts that I think deeptrio would be essential to handle these situations and obtain reliable de novo mutation calls in families. Without joint family/trio calling, there is no feasible way to achieve the required specificity.

[emmamiz]

From the gVCFs:

chr9:42064090 C>T,<*> Father GQ=4 AD=39,26,0

chr9:42064090 C>T,<*> Proband GQ=21 AD=21,15,0

[williamrowell]

Thanks for sharing the gVCF outputs. Could you share one more image, please? Can you please regenerate the first image you shared, except 1) squish alignments, 2) group alignments in each track by "supplementary flag", and 3) zoom out a bit. I'm trying to understand more about why the variant is ignored in the father, because that could indicate whether DeepTrio would have a better chance of correctly genotyping it.

[emmamiz]

Here’s a slightly zoomed-out view with alignments squished and grouped by supplementary flag.

My apologies for the delayed response! Things got busy and this accidentally fell off my radar.

[williamrowell]

No problem, thanks for sharing.

This variant is near the end of an annotated LINE in the reference, and I suspect that the reads carrying the variant come from a LINE somewhere else in the patient sample that is aligning here because it's the most closely related reference sequence. The DeepVariant model can learn that features of this sort of variant (variant only carried by alignments with soft clipping on both sides) are often associated with false positives. That doesn't mean that they can't slip through from time to time. Do you have evidence that DeepTrio calls this variant differently in your trio?

[emmamiz]

Yes, DeepTrio calls this as paternally inherited.