Low coverage WGS data

Follow

Goran Rakocevic

• January 16, 2024 14:37

Hi All,

 

I'll be looking at calling variants on a cohort of ~15,000 low coverage genomes (1-2x). 

I searched the forums and found that the recommendation is to use the regular best practices in this case as well, but I still wonder about a few things.

 

1) Should we set --min-pruning to 1 in the HaplotypeCaller? I realise this will (I guess?) emit every sequencing error, but I'm hoping the joint calling might help there. 

 

2) In most places there won't be that many reads for HC to derive haplotypes with, so I'm wondering if it would make sense to do HC with multiple BAM files (or one multisample BAM file).

 

15,000 samples, even at this low coverage, is likely well beyond feasible, but I was thinking about doing groups of 50-100, followed by merging and joint genotyping on all 15,000.

Doing 50-100 still might blow up the number of paths in local graphs (especially if coupled to --min-pruning 1), so it might be 10-15 or whatever ends up working, I'm more interested if this is a good idea or not.

 

3) I'm getting the DRAGEN outputs, so would those be ok as starting points (either CRAMs or gVCFs, in case answers to questions 1 and 2 are No).

 

Thanks!

0

• 

  Laura Gauthier

  • January 29, 2024 19:36

  Hi Goran Rakocevic,

  Unfortunately with the low-pass WGS study design, you're going to be limited to common variants, but on the other hand seeing a variant in multiple samples is pretty good evidence. This is what we did for the first phase of the 1000 Genomes project and the final callset was pretty high quality, though definitely enriched for common variants.

  Yes, I would recommend reducing the min-pruning argument, potentially even to zero. Then you'll rely on the site-level QUAL score to ensure that the likelihood of the site being variant is higher than the (human) SNP rate of 1/1000, i.e. QUAL >= 30.  I think that ends up being at least four ALT reads across samples, but that's a very rough estimate and it will require more evidence for low base qualty, low mapping quality, and potentially indels. VQSR should take care of systematic errors like strand bias. Single-sample GVCF mode should be fine. I don't think you'll be able to provide HaplotypeCaller with 15,000 input bams.  There are strange limitations like UNIX commandline length or the number of file handles the operating system will allow open at the same time.  I don't think there's a way to create a multi-sample GVCF with multiple input bams, so I would suggest you follow the typical best practices GVCF->GenomicsDB->GenotypeGVCFs workflow with the additional pruning change for haplotypecaller.

  0

  Comment actions Permalink
• 

  Goran Rakocevic

  • January 29, 2024 20:07

  Hi Laura Gauthier

  Thanks!

  I didn't mean to shove 15,000 BAMs into a single HC call. I was thinking of maybe doing random batches of 50 or so... 

  That will cause batch effects, just wondering if that would help or hurt more.

  0

  Comment actions Permalink
• 

  Laura Gauthier

  • January 29, 2024 20:44

  HaplotypeCaller won't do multi-sample GVCFs and you're going to run into issues combining multiple multi-sample genotyped VCFs.  Most importantly, you won't have reference information to get confident allele frequencies across all samples.

  0

  Comment actions Permalink

Please sign in to leave a comment.