Consider adding QC Flags to abundance estimates

[gordonkoehn]

In production, we tend to verify results manually quite frequently with lollipop deconvolute. This is suboptimal. Ideally, results should be trusted unless Lollipop explicitly flags low confidence results, requiring human review.

We could think of the following simple QC checks:

• CI Width Check: 95% CIs width >10% indicate low confidence. --> WARNING!
• Shift Detection: >20% abundance shifts between time points --> WARNING!
• Spike Detection: Label points with Z-score >3 from rolling mean as “Spike” --> WARNING!

Where the WARNING should trigger a human review before release, and could be captured from the logs.

[gordonkoehn]

@dr-david, what do you think ?

[dr-david]

Yes great idea. We should also look into integrating the "pre-deconvolution checks" which look at the properties of the variant definition + coverage, i.e. output a warning when variant definitions being to close + dropouts could hinder successful deconvolution.

[gordonkoehn]

Yes, that would be great as well! Anything brutally simple on both sides would be great to start with. Let's scope the simplest solution possible first - happy to implement it. We just need some hard numbers to start.

May I propose:

1. Variants need at least X unique mutations
2. No confidence intervals may exceed the YZ% range

Failing 1) triggered abort, or 1) & 2) trigger a warning as a warning in the command line, that one can parse.

Would you know what to put for X and ZY here ? Are these good, brutally simple criteria?