Skip to content

Latest commit

 

History

History
59 lines (48 loc) · 2.89 KB

deepvariant-haploid-support.md

File metadata and controls

59 lines (48 loc) · 2.89 KB
Raw

DeepVariant support for variant calling in chromosome X and Y

Case study

A case study on how to use the parameters mentioned here are described in DeepVariant X, Y calling case study.

Haploid calling support

As DeepVariant is a diploid variant caller, it assigns genotypes as {Hom-ref, Het, Hom-alt} for each candidate allele it observes. For samples with karyotype XY, the chromosome X and Y are effectively haploid. So, we are introducing two flags to re-adjust the genotypes in regions that are considered to be haploid for those samples.

You can use --haploid_contigs and --par_regions_bed parameters to readjust the genotypes in haploid regions. For samples with XY karyotype, it is expected that users will set --haploid_contigs="chrX,chrY" for GRCh38 and --haploid_contigs="X,Y" for GRCh37. You can also provide a PAR region bed file with --par_regions_bed="/input/GRCh3X_par.bed" parameter. The regions in the PAR bed file will be skipped from genotype readjustment. You can download the PAR bed files from here: GRCh38_par.bed, GRCh37_par.bed.

How it works

The genotype re-adjustment is implemented in the postprocess_variants stage of DeepVariant. For any variant, that is in the--haploid_contigs regions and not in the --par_regions_bed regions, the genotype likelihoods of heterozygous variants are set as 0 and the genotypes are normalized again after re-adjusting the likelihoods. After that the most-likely genotype is assigned to the allele which excludes any heterozygous calls.

For example, suppose we observe an alternate allele ALT1 at a position that we consider to be haploid. So the observed alleles at that position are: Candidates: {REF, ALT1} The neural network generates likelihoods for the genotypes for this candidate as such:

Homozygous reference:   likelihood(REF,REF)
Heterozygous alternate: likelihood(REF,ALT1)
Homozygous alternaate:  likelihood(ALT1,ALT1)

So the likelihood vector looks like: L={L[(REF, REF)], L[(REF, ALT1)], L[(ALT1, ALT1)]} In the post processing step we set L[(REF, ALT1)] = 0 as that is a likelihood associated with heterozygous genotype and heterozygous calling is excluded in haploid regions. The likelihood vector becomes: L={L[(REF, REF), 0, L(ALT1, ALT1)]}. Then we normalize the likelihood vector and assign the genotype based on the adjusted values from the vector.

In DeepVariant r1.8, we added extra logic in the make_examples stage to adjust for reference blocks as well. See the discussion in #811.