Does anyone have any thoughts on merging gVCFs for a single sample from different callers? Good idea, bad idea, example of how to do it?

I think we might have started doing that in #CGFUX04HZ, or is it just for the VCFs?

For what purpose?

I am leaning against it, I prefer to make results tables for my "end users", which may contain variants from different callers (and you can see which one[s]), but I keep the VCFs separate. Different variant callers have different output in the

INFO

and

FORMAT

fields, so if you merge the files, some entries may have a certain metric, while others don't. If you later try to filter the combined file on that and are not aware of that, you probably won't get the expected results. Also, it is always easier to combine things later than to try to tease them apart. The purpose is also important. E.g., I would never do some quantitative metric like tumor mutational burden on a combination of variant callers -- they are just too different in how they work and which biases they all have, and the one with the most calls (even if many are garbage) will dominate the results. I much prefer doing the calculation separately for each one and then compare any trends (and if there are no trends, you probably should distrust at least something). On the other hand, if you are looking for e.g. driver mutations, you can have a table with all callers, but I would filter the VCF of each caller based on appropriate criteria (e.g. FreeBayes does not populate

FILTER

, so you have to use something else, but you probably would consider

FILTER

for Mutect2 and Strelka, probably in combination with other things) and then combine the results in one table (with a column for which tool[s] found it, ideally more than one).

I think a lot of what you are saying makes sense to me, although I’m not sure how different a single combined VCF is to e.g. a csv with effectively the same information but manually curated, as long as analysts understand what they are being given. The use-case for these merged gVCFs is that we want to create multi-sample VCFs but we also use multiple callers currently (rare disease work, so the analysts want to capture the maximal possible calls and filter down on their choices after). Some googling suggests it could be done by merging the gVCFs then running GLNexus (which we currently use for joint-genotyping the gVCFs for each caller) but that’s where I think the issues might be messy - the merged gVCF would need to coerce the calls to look like one or other of the callers as GLNexus config is caller-based. We also use small variant calls to phase and as inputs to some of our CNV/SV callers and I am unsure on whether it is better to combine the small variant caller output (standard per-sample VCFs, not the gVCFs) or to just pick one small variant caller for this use?

... I’m not sure how different a single combined VCF is to e.g. a csv with effectively the same information ...

I use Excel sheets or tab-separated files mostly as a way to get information to the "end user", i.e. someone who might be looking for variants that meet specific criteria, or they might be looking for variants in their favorite genes, etc. If your user is an analyst, your priorities might be different, but a consideration that remains is that it is harder to separate out components of a mixed file than to combine them. Apart from GLNexus requirements, there may be some variants that might be very hard to normalize, so they look the same, no matter which caller generated them, see e.g. here for a tricky example. I would look into why GLNexus differentiates between callers, because that should tell you about important differences between the different VCFs. Would combining results from GLNexus make more sense? Not sure if that works with GLNexus and your other tools, but MAF is a file format that is fairly well standardized, and maybe you could convert VCFs to MAF and then combine MAFs? Those might be more consistent/standardized than VCFs.

... I am unsure on whether it is better to combine the small variant caller output (standard per-sample VCFs, not the gVCFs) or to just pick one small variant caller for this use?

Good question. Variant callers differ in how they call things like genotype, at least in ambiguous cases (I usually work with somatic samples, where the concept often does not make sense, so I might see this more frequently than people working with germline variants), and I would suspect some are better at this than others, so I would try to find which one(s) is/are most trustworthy. If there is more than one "best", maybe combining these might make sense if the variants they find are sufficiently non-overlapping.