calculateAitchison - Githubissues

antagomir commented 3 years ago

I propose to add a function calculateAitchison.

This would be similar to calculateUnifrac but it will not require tree and it will return Aitchison distances.

These are defined as Euclidean distances between clr-transformed values, by default calculated from the assay="counts" or assay="relabundances" (which should lead to same result, consider including a unit test for that..).

The user should be able to choose between standard (default; with clr) and robust (optional; with rclr) transformation as part of the distance calculation.

The advantage is to standardize this calculation, it has become a relatively common choice and fluent access is needed for comparison purposes.

The default names given for the distances would be "Aitchison distance" and "Robust Aitchison distance"

As an alternative to implementing mia::calculateAitchison function, one can use method "aitchison" through coda.base::dist function because in most cases a user can just call a specific external distance function for runMDS and other ordinations: d <- coda.base::dist(X, method = 'aitchison'). The limitation with this is that the "Robust Aitchison" option is not available in standard distance functions as far as I know, this would require PR to vegan::vegdist or coda.base::dist or some other distance function, to add both Aitchison and Robust Aitchison. In principle, that would be perhaps even better solution than implementing our own versions of some rather general distances.

antagomir commented 3 years ago

@FelixErnst - wondering if we should have a more clear default policy here - how low barrier making PRs to external packages instead or expanding mia. Here we clearly have two options.

FelixErnst commented 3 years ago

coda.base seems to a single function package only, but it is using a Cpp implementation. The one fixed criteria for runMDS is that it must use the same inputs as dist. If that is the case, I would try to go the route of suggesting the function be added and optimized in these packages.

Maybe implement it yourself and do a benchmark to make sure, where the right place for it is.

The general idea for keeping the dependency tree small is that, if you are only importing one function an entry in suggest and a soft dependency is OK. If it is more than one function or a class, you have to import anyway. In this case I would keep the package in suggest, because of the single function aspect and the rare chance someone is using that function on their first encounter with mia

antagomir commented 3 years ago

Ok I opened an issue to coda.base and to vegan to discuss this.

antagomir commented 3 years ago

Given, that adding these features in the above mentioned package may take considerable time (or perhaps not?), it might be useful to have, meanwhile, our own implementation calculateAitchison (with "standard" and "robust" options).

This can be implemented easily without extra dependencies, since we already have internal implementations .calc_clr and .calc_rclr in mia. Data (counts or relabundances assay) has to be run transformCounts (or transformSamples) with the (r)clr and then just through Euclidean distance which is readily available through the standard stats::dist and vegan::vegdist functions.

We already have calculateUniFrac and calculateJSD, this would operate with the same logic.

antagomir commented 3 years ago

It seems, however, that this could be added in vegan soon, see vegan issue #433. So anyone working on this should check that before further action.

FelixErnst commented 2 years ago

Any news on this?