Filtering compatible overlapping truth variants about nanovarbench

[lachlancoin]

My preference would be to remove them.

[rrwick]

I'm happy to leave them in, but I think the tricky issue might be how to assess these variants.

Using your second example, if the reference sequence is CGTAGT but the correct variant-applied sequence is CGTAACAACACTTGGAAT, a variant caller could represent this in different ways.

It could do it in two variants, like you showed above:

chromosome   3   60f6a01c   T     TAACAACACTTGG
chromosome   5   a1f621d4   G     A

Or it could do it in one variant:

chromosome   3   ab49ec54   TAG   TAACAACACTTGGAA

Both are correct, so if you're assessing a variant caller, you'd need to accept both answers.

This problem applies not just to 'overlapping' variants but any close variants. And it could get even more complex with groups of close variants. E.g. if you had 3 variants near each other, you could group them in four ways: a,b,c or ab,c or a,bc or abc. And the number of possible groupings will grow enormously with more near variants.

So while I don't think you need to filter out variants because they are 'overlapping', you might want to filter out too-close variants, if that makes assessment easier.

[mbhall88]

I think the complication you've mentioned here is actually more in the realm of variant assessment. vcfdist (described in this paper) handles this by standardising the variant representation between the truth and query VCFs (See Figure 1b and 3d and Suppl. Fig 2 for good illustrations).

So while I don't think you need to filter out variants because they are 'overlapping', you might want to filter out too-close variants, if that makes assessment easier.

I'd like to keep close variants in there as this will likely be a good separating factor between the variant callers. And we don't want to make it too easy for them 😁

However, I am beginning to think that excluding these valid overlapping variants might be a good idea. I've just found some more complicated examples where bcftools +remove-overlaps doesn't think they're overlaps, but bcftools consensus does.

chromosome_2    1561582 0cb9eac6        A       ATTTCTTTTGATAAGAAAGTATTAAGTG    .       PASS    .       GT      1/1
chromosome_2    1561582 4bc9851a        A       AT      .       PASS    .       GT      1/1

[mbhall88]

So essentially, these can be removed with bcftools norm -d indels, indicating they're kind of the same variant.

I think this changes my mind and I now think we should remove all of these types of overlaps.