Population-scale co-phasing of SVs & SNPs

[tuannguyen8390]

Hi teams,

As per the title, I was wondering what is a good way to do population co-phasing.

Say, I have 3 individuals, then I do it as per suggestion in your README.

Predict SVs for individual A,B,C with Sniffles Predict SNPs for individual A,B,C with Clair3

Cat A,B,C, force-calling the genotype. Then I merge SNPs & SVs of A,B,C into a multi samples VCF. Then run with longphase? Is this a good way to go and is Longphase capable of doing that?

Many thanks,

Tuan Nguyen

[ythuang0522]

Hi Tuan, LongPhase cannot phase multiple samples simultaneously (i.e., multi-sample VCF) as it's a read-based phasing algorithm. There are two possibilities. The first is phasing each individual independently (with it's own VCF and BAM file). e.g., longphase -s A_SNP.vcf --sv-file A_SV.vcf -b A_alignment.bam ..., longphase -s B_SNP.vcf --sv-file B_SV.vcf -b B_alignment.bam .... Then merge the three (phased) VCFs into a multisample VCF as you want.

However, I am not sure how existing tools (e.g., vcftools) merge phased VCFs. If vcftools can't properly retain the phased info, the alternative is generating a multi-sample VCF as you did, and then re-generate each individual VCF (yet expanded with all population SNPs) by e.g., cat multisample_VCF | cut -f 1-9,10 > A_expanded.vcf, where column 10 should be replaced with the sample column of interest.

Then you can run LongPhase for each individual with its own expanded VCF and bam files. longphase -s A_expansded_SNP.vcf --sv-file A_expanded_SV.vcf -b A_alignment.bam ...

Yao-Ting

[GuillaumeHolley]

In general, I think it would make a lot of sense to first call SVs with the population-calling mode of Sniffles 2 and GVCF merging for the SNPs (so call each sample with Clair3 but with GVCF output and then merge the GVCF into a multi-sample VCF with GLnexus). If your 3 individuals are a trio and your reads are PacBio HiFi, you can also consider DeepTrio. Then you can separate the multi-sample VCF into 3 individual VCFs with bcftools, phase them independently with LongPhase and re-merged them with bcftools merge (this should keep the phasing).

[tuannguyen8390]

Interesting, we are currently tried out the population-calling mode with *.snf file with sniffles2. But that seems won't retain read name info (?) - So I'm unsure if that would work with Longphase.

Our samples are indeed a trios but sequenced with ONT.

[ythuang0522]

From the sniffles2 README, read names can be stored in *.snf, unless they are thrown away after merging.

To output read names in SNF and VCF files, the --output-rnames option is required.

[tuannguyen8390]

From the sniffles2 README, read names can be stored in *.snf, unless they are thrown away after merging.

To output read names in SNF and VCF files, the --output-rnames option is required.

@ythuang0522 Thanks so much for letting me know, I only used --output-rnames with the single individual calling, didn't know that it would be needed at the merging step! I will test this out and update the result.

In general, I think it would make a lot of sense to first call SVs with the population-calling mode of Sniffles 2 and GVCF merging for the SNPs (so call each sample with Clair3 but with GVCF output and then merge the GVCF into a multi-sample VCF with GLnexus). If your 3 individuals are a trio and your reads are PacBio HiFi, you can also consider DeepTrio. Then you can separate the multi-sample VCF into 3 individual VCFs with bcftools, phase them independently with LongPhase and re-merged them with bcftools merge (this should keep the phasing).

@GuillaumeHolley thanks for the suggestion, I will try this out as well. I haven't used GLNexus or GVCF before so sorry if I ask a stupid question 🔢: could you please explain what is the difference between output Clair3 with GVCF vs VCF there. Because I did think about merging VCF with bcftools merge at that particular step?

[GuillaumeHolley]

@tuannguyen8390 A GVCF file follows the VCF format specification but contains records (lines) which do not correspond to a variant call but instead, a reference call or a no call. The idea is to have some information such as read coverage for sites which are not heterozygous nor homozygous alt. Merging GVCFs over VCF has (at least) 2 key advantages among others:

• It distinguishes no calls (./.) from homozygous reference calls (0/0)
• This one is a little bit controversial but if I am not mistaken, GVCF merging enables to change borderline calls. Say you have a reference call in sample A which was borderline (could have been heterozygous for example) but the call is clearly heterozygous alt in sample B, the GVCF merging tool can decide based on the reference call info in A to change the borderline ref call to a borderline heterozygous call.