How is the "weight" calculated by CNVkit ?

Question

[Hällyss]

Hello,

We search a kind of score whilch can eliminates many FP call in CNVkit results. We think that the weight can be used. We search a signification of this weight, an equation. In the manual of CNVkit, we found this extract :

A weight is assigned to each remaining bin depending on:

1. The size of the bin;
2. The deviation of the bin’s log2 value in the reference from 0;
3. The “spread” of the bin in the reference.

(The latter two only apply if at least one normal/control sample was used to build the reference.)

So, we have many questions :

• what "bin" means ? The segment ? The bin in my_target.bed ? my_antitarget.bed ?
• what is the "spread" ? the number of bins in the segment ? the length of the segment ? something else ?
• the score seems to be strongly affected by segment size and / or bin number in the segment, is this the case?
• is it possible to get an equation of the weight ?

Thank you

Alice

[WouterDeCoster]

You wrote in the tags which tool this question is about, but that would have been useful information in your post as well.

[Hällyss]

All of my apologies, thank you for your answer.

Answer 1

[Eric T.]

So:

• Bins are the unsegmented regions seen in my_target.bed, my_antitarget.bed, and the .cnn and .cnr files emitted by CNVkit. Sometimes also called "probes" in the code.
• Spread is the statistical spread of coverages in a bin observed across all of the samples in your pooled reference, similar to standard deviation but calculated differently to be more robust to outliers.
• The weight listed in the segmented .cns files is the sum of the weights of the bins/probes spanned by the segment. It correlates with segment length and number of bins, but will be a bit lower if the segment covers a region with less reliable sequencing coverage or mapping (i.e. lower-weight bins).
• The calculation is in the function cnvlib.fix.apply_weights. It is not one equation; it depends on which data sources are available, mainly whether the reference is paired, pooled, or flat.

Also see cnvkit.py segmetrics --ci and cnvkit.py call --filter ci for filtering out potential FP segments by calculating confidence intervals for each segment's mean log2 ratio.