xcms3 as a branch or a completely rewritten package?

[sneumann]

xcms3 as a branch or a completely rewritten package?

[jorainer]

OK, so, following @sneumann's advice I'll start doing under the hood changes within xcms first (i.e. introducing the layer of do_ functions as detailed in #30).
Also, I would like to do some clean up and re-organization of the existing code:

• Fix the warnings of R CMD check
• Re-organize the R files (i.e. Classes.R, methods-...R, functions-...R)

Any objections?

[stanstrup]

I guess the answer really depends on how radical the change. The more radical the more b makes sense.
MzMine is going through a similar process of re-writing everything and as far as I understand they went for b and is slowly porting over features.

[lgatto]

A in between alternative would be the following:

Start new development in a xcmsdev package/repository to have the clean sheet advantage. Once that package has reached some basic functionality, xcms would depend on/import it and it would get the necessary wrappers/infrastructure to use it. (I believe it is possible to make use of github packages, although I haven't done it yet). Depending on maturity/stability, the xcmsdev code could even be moved to xcms and old xcms code deprecated/removed. At some point, dependency on xcmsdev could be removed too.

Although xcmsdev could be functional on its own (albeit possibly concentrating on low-level functionality), a typical user would be expected to use xcms rather than the xcmsdev.

[jorainer]

@stanstrup : the changes I would like to introduce are:

• Use MSnbase's OnDiskMSnExp object to read and represent the raw data (without actually keeping it in memory); this should be similar to the xcmsRaw object, but without all the data in memory.
• Methods (e.g. detectFeatures) that take the raw data object from above to identify the features. Dispatching to the various algorithms (e.g. centwave) would be done using a parameter class (e.g. CentWaveParam).

This will already change the xcms workflow considerably; still, I would be fine with each option.

[jorainer]

@lgatto : I like that idea, however, I don't know how github dependencies are seen in Bioconductor... based on the latest discussions at BioC2016 not so well...

[sneumann]

Hi, I am leaning to doing the changes in-place. That's what development cycles are for.
W.r.t. the timing, disruptive changes and/or those that might change the outcomes
(peak picking, alignment etc.) should be early in the BioC development cycles.

[jorainer]

I can live will all options; as detailed in #30 I am planning to separate the UI from the actual analysis functions hence making the code as modular as possible.
I am only afraid that users might get upset if some of the methods are removed at some point.

[lgatto]

I don't know how github dependencies are seen in Bioconductor... based on the latest discussions at BioC2016 not so well...

I am aware of the Remotes entry in the DESCRIPTION file: see https://github.com/hadley/devtools/blob/master/vignettes/dependencies.Rmd. Don't know how/if this work with biocLite though.

[jorainer]

Closing this as we're already on our way towards xcms3