parameters setting for somatic/germline variant calling

Dear team, may I know if there is any different parameters setting between somatic variant calling and germline variant calling? The reason why I posted this question is that I found one heterozygous variant called by deepvariant but seems homozygous supported in IGV. I’m wondering if my setting of deepvariant is too loose for this variant? (ps, I just run deepvariant by default pacbio data setting) Here is the result in VCF. The VAF is quite high, and1,20 AD means only 1 read supported the widetype read? chr4 3079267 . G T 36.1 PASS . GT:GQ:DP:AD:VAF:PL 0/1:4:21:1,20:0.952381:33,0,1 Here is the screenshot of IGV Thanks!!

[AndrewCarroll]

Hi @Qianwangwoo

We don’t do additional filtering beyond the probabilities from the classifier. In this case, DeepVariant does not have a high confidence in the correct genotype between HET and HOM-ALT (a GQ of 4 corresponds to a ~60% confidence in a correct genotype call). The QUAL value of 36.1 suggests that DeepVariant is at least pretty confident that the position is not REF

A few other points to keep in mind - first, are you using the two-pass DeepVariant-WhatsHap-DeepVariant method? If so, then DeepVariant may be using additional information about the phasing from longer range.

Second, this variant is at a junction between homopolymers (poly-T and poly-G) This represents the dominant error mode for PacBio HiFi, so it may nit be straightforward for a human to assess the probability of a G->T variant here as opposed to a sequencing error of Insertion T and deletion G.

If you want to for sure have a higher precision, you can additionally filter for GQ value (e.g. 10 for a 90% confidence in the genotype call). However, if you do so, you will lose variant positions like this which are very likely not reference, but difficult to genotype.

[AndrewCarroll]

Hi @Qianwangwoo

Yes, the two-pass method generally improves accuracy with PacBio small variant calling, especially for Indels. Whether it is likely to improve this call, I am not sure. Note that we anticipate a future release of DeepVariant for PacBio in the near future which will have comparable accuracy with a single pass of variant calling, so you may prefer to keep your current workflow and wait for that version if you don’t mind updating.