Position filtering

[csoneson]

• [x] sequence context (done in af66b27676da764d7226f055b3acd2fc91aeb435)
• [x] coverage (currently inside the calcReadStats function) (done in 26c7e41)
• [x] collapsing/pruning repeated positions (done in fe95936)
• [x] remove positions that are NA in all reads in all samples (done in 8ad2d22)
• [x] each of these go into a separate helper function
• [x] new function filterPositions() - calls one or more of the helper functions (done in b1e79e5)
• [x] call read-NA-filtering in the end (by default turned on) to remove reads that are now all NA in the retained positions (done in 10d37e08b488f3368244fd38ad4081da3afaba8d)

[csoneson]

Some things to discuss:

• Currently, keeping positions matching sequence context TAG also retains (e.g.) NNN - don't allow this (N in the query are fine, but remove hits where the provided sequence context contains N).
• Not totally sure if the current names for these filtering functions are the most intuitive.
• Should the coverage filter be applied to the sum across samples or to individual samples (e.g. require at least one sample to pass the coverage threshold)?
• For .removeAllNAPositions(), do we allow a summary assay (and if so, check for rowSums > 0 instead of presence of any non-NA values?

[mbstadler]

Just my 2-cents to the questions above:

• N-containing sequence contexts: As discussed, Ns in rowData(se)$sequence.context may result both from Ns in the reference as well as from Ns added to sequence contexts near the boundaries that run out of a reference sequence. I agree that we probably don't want to allow this.
• The names are fine to me. An alternative would be to make them all similar, e.g. .filterPositionsBy..., but I am not sure that would be worth it, as the current names are clear and they are internal functions usually called via filterPositions().
• Coverage filter: I don't really know what makes more sense. In order to get rid of false positive calls, the global coverage (as implemented now) will probably work well. By the way, I find the use of addReadsSummary very elegant :-)
• remove NA positions: I would maybe rather implement such a function separately under a different name (.removeAllZeroPositions, and a new argument in filterPositions(..., assay.type.zero)), if we need it.

I have a question regarding the coverage filter, at this line: I guess our assays always have row names, but is it safer/required to go via rownames, or would a logical keep also work and be more general?

[csoneson]

Thanks!

I have a question regarding the coverage filter, at this line: I guess our assays always have row names, but is it safer/required to go via rownames, or would a logical keep also work and be more general?

You're right, we don't need rownames - the two objects (mat and se) should always have the same rows anyway.

remove NA positions: I would maybe rather implement such a function separately under a different name (.removeAllZeroPositions, and a new argument in filterPositions(..., assay.type.zero)), if we need it.

Also makes sense to me

[csoneson]

• Removed the rownames matching for subsetting in 2dac82e
• Added min.nbr.samples argument to .filterPositionsByCoverage in 2dac82e
• Only allow ambiguities in the pattern to be interpreted as wildcards in sequence context filtering in 58618db

[csoneson]

Included in #22