Should we discard a variant if it pass many filters and fails one?

Question

[Sharon]

Hi Everyone

Just as a general rule, if we have a very interesting variant, we have many metrics to judge the variant, from mapping quality to strand bias to coverage ,..etc. If the variant passes most filters on such metrics and fails one, so simply discard just because it fails one? I feel okay to discard if the variant has low mapping quality, even if high coverage and no strand bias, .etc.

But for other metrics, if it fails any, just discard? Could we lose interesting variants then? What do you do?

Thanks

[ATpoint]

Can you give some details on: which variant caller, which species, what kind of sequencing (exome, wgs, amplicon), control sample available, what kind of variant (SNP, Indel, complex, CNV), and what filter it failed.

[Sharon]

I am not talking about specific variant, as i have so many, some of them exome and some of them rnaseq. So, I am asking in general, if a variant fails one metric, do we discard it,, like should it be all good on all metrics? For example high coverage, good mapping quality but QD is low or there is strand bias , etc ?

[WouterDeCoster]

That's a tradeoff between sensitivity and specificity and entirely depends on the application.

Answer 1

[Kevin Blighe]

There really is no answer. There are so many NGS analysis pipelines out there and so many customised FILTER flags. However, this issue of deciding what to keep or filter out is not a unique issue to you, ATpoint, or I ... clinical scientists face these burning questions each and every day as they mostly manually curate each and every variant in their target gene of interest.

From what I know so far from working with people, if a variant fails even one aspect of quality and has a FILTER flag other than PASS, then it should be eliminated (i.e. in clinical settings). So, the cautious approach is to include variants that pass all aspects of quality.

This of course means that you should be aware of which thresholds are being used for your particular data and whether or not you agree with them. These are usually encoded in the header. Based on my background, I always push for greater quality and do not sacrifice that (unless ordered).

[Sharon]

Thanks Kevin, got it.