AndrewCarroll
commented
2 years ago Hi @Phillip-a-richmond
Thank you for the report. You are correct, that the behavior out-of-the box for DeepTrio can currently be sub-optimal for the sex chromosomes in male samples. We've benchmarked some strategies for dealing with this.
In the short term, our recommendation is to run separate calling on the non-PAR regions of ChrX and ChrY, where only the mother sample is provided as the parent for calling of the son, and (less importantly as it is unclear whether this is an issue with chrY) only the father sample is provided for calling chrY on the son.
In the documentation, this is expressed in the following statement: "Also please note: for the non-PAR regions of the sex chromosomes (X and Y), we recommend running these providing only the parent who contributed the child's chromosome (e.g. for chromosomeX, only the mother and son samples and for chromosomeY only the father and son samples)."
The DeepTrio manuscript has benchmarks for this strategy in the following section (near the end of results):
The Genome in a Bottle truth sets do not contain chromosomeX or chromosomeY variants in a male individual. As a result, DeepTrio has never been trained with hemizygous sites. Because we train DeepTrio to perform duo calling, it is likely that DeepTrio would call variants on chromosomeX similar to how it would call a duo sample. To assess this, we ran DeepVariant and DeepTrio on chromosomeX of the son (HG002) and measured the number of heterozygous variant calls in the non-PAR regions of chromosomeX.
For DeepVariant, 4.45% (455/101866) of calls in non-PAR regions of chromosomeX are heterozygous. In DeepTrio, 24.5% (21633/88314) are heterozygous. This substantial difference suggests that applying DeepTrio directly to chromosomeX in male samples is problematic. Since chromosomeX in males is inherited from the mother, we performed calling on chromosomeX with only the mother provided as the parent. This reduced heterozygous calls to 3.37% (3518/104427), which is better than in the DeepVariant case. For male samples, this recommends that variant calling should be run with both parents on the autosomal and PAR regions using a BED file to restrict location, and additional variant calling should be performed using only the mother’s file provided as parent for the non-PAR regions of chromosomeX, and only the father’s provided for the non-PAR regions of chromosomeY.
This experiment indicates that allowing the model to infer a hemizygous chromosome through coverage and explicitly training for hemizygous variants is an opportunity for improvement, both for DeepVariant and DeepTrio.
Over the long term:
We anticipate that the T2T consortium will be shortly be making available complete assemblies of ChrX and ChrY for the HG002 sample. This should allow us to make training labels for ChrX and ChrY in the hemizygous case, and we have a few internal ideas around how to use this to make models that will generally take this into account.
I would be curious in your feedback about whether the short term solution is acceptable, whether you find it insufficient, or if you would recommend other approaches for us to address the issue.
Thank you, Andrew
Phillip-a-richmond
pichuan
nurmians
pgrosu
Hello,
Noticed this issue with your tool DeepTrio regarding the representation of hemizygous variants in the non-pseudoautosomal (PAR) X-chromosome. This may be fixed now in 1.3? If so ignore this, but if not this is what I noticed.
Note, this is a simulated pathogenic variant from bamsurgeon, but the VCF representation is the focus of this problem.
Let's start with an IGV snapshot of the variant:
Now, I'll go into the representation from the DeepVariant --> GVCF --> GLnexus pipeline:
DeepVariant Pipeline:
Variant:
VCF header:
Variant line
DeepTrio
Now, with the DeepTrio -> GVCF -> GLNexus pipeline: Pipeline
Variant
Header
Variant line
Here the male proband, fully hemizygous for the variant, is represented as 0/1: 0/1:18:0,18:4:33,0,1:..
Why it matters
Now, this is a BIG problem, because downstream tools like Exomiser will treat the output here for variant prioritization in rare disease cases. So take a look at this.
Exomiser HTML for DeepVariant, where variant is represented as 1/1. This gene is the top ranked hit for this simulated rare disease case. Good:
Exomiser HTML output for DeepTrio, where variant is represented as 0/1. Notice that now the gene is not even in the top 10 candidates, as Exomiser has interpreted this variant incorrectly (likely assuming it was a PAR region on X). The variant score dropped to 0.00 because the assumption on the GT string was violated:
I can send the HTML outputs, VCFs, or even CRAMs if you are interested by email or other file transfer.
Anyways, going to try with 1.3 while I wait for an answer here.
Cheers, Phil