bcftools merge with gVCF blocks

[jordilc]

Hi,
I want to use bcftools merge with files with gVCF blocks, but if one of the files has an SNP where the other file has a gVCF block, bcftools doesn't merge them properly. An example:

File 1:
18 7447502 . C T 228 PASS DP=12;VDB=0.985771 ...
18 7447775 . T . . LowQual END=7447781 GT 0/0
18 7447783 . T . . LowQual END=7447803 GT 0/0

File 2:
18 7447472 . C A 228 PASS DP=14;VDB=0.346151 ...
18 7447777 . T G 190 PASS DP=9;VDB=0.904845 ...

Merged file:
18 7447502 . C T 228 PASS VDB=0.985771 ...
18 7447775 . T . . LowQual END=7447781;AN=2 GT 0/0 ./.
18 7447777 . T G 190 PASS VDB=0.904845 ...
18 7447783 . T . . LowQual END=7447803;AN=2 GT 0/0 ./.

I think the expected result should be split the gVCF block into two blocks, with the SNP between them:
18 7447502 . C T 228 PASS VDB=0.985771 ...
18 7447775 . T . . LowQual END=7447776;AN=2 GT 0/0 ./.
18 7447777 . T G 190 PASS VDB=0.904845 ...
18 7447778 . T . . LowQual END=7447781;AN=2 GT 0/0 ./.
18 7447783 . T . . LowQual END=7447803;AN=2 GT 0/0 ./.

[pd3]

Unfortunately, merging of gVCFs is not supported yet, partly because I am not sure about the logic of the operation. For example, when merging normal records, filters behave differently from what you proposed: a site can have FILTER=PASS only if it is not filtered in any of the files. I don't have much experience with gVCF and don't know what behaviour is expected / most useful.

[dwaggott]

Were there any ideas on this subsequent to #244 ? Doing a gvcf merge is a great poor man's synthetic joint caller.

I've got several hundred whole genome gvcf's that need merging. My first rather clunky approach is:

1. determine the union of all variable sites across the cohort
2. extract those sites from each gvcf using convert i.e. bcftools convert --gvcf2vcf -R all_variable_sites.bed
3. merge the resulting vcf's

[pd3]

There is an initial version of gVCF merging in https://github.com/pd3/bcftools. However, it is not ready for production use yet.

[mcshane]

The gVCF merging code is now on the develop branch. Would be great if people can test it for their particular use cases.