`anndata.acc` based API

[flying-sheep]

What kind of feature would you like to request?

Additional function parameters / changed functionality / changed defaults?

Please describe your wishes

instead of layer or obsm, we should have in_/out or so, where out=None means override.

[AlbertoFabbri93]

I think overwriting should require an explicit decision. Layers in AnnData usually have semantic meanings (e.g. counts, log1p, scaled_counts) and often represent different stages of an analysis. Accidentally overwriting one because out=None was omitted or misunderstood could silently destroy important intermediate results. Making overwriting an explicit opt-in would make the API safer.

[flying-sheep]

Interesting idea! The until now is “take X, write X, overwrite unless copy=True”.

How would you suggest this would look like?

[AlbertoFabbri93]

Maybe something like this:

sc.pp.scale(
    adata,
    in=A.layers["counts"],
    out=A.layers["scaled_counts"],  # write to a new layer
)

sc.pp.scale(
    adata,
    in=A.layers["counts"],
    out="inplace",                  # explicit overwrite
)

What do you think?

[selmanozleyen]

Maybe out=None means return a copy (this would get rid of copy) and out=A.layers["counts"] would already mean overwrite no? because out="inplace" and copy=False would conflict

[AlbertoFabbri93]

I like in=A.layers["counts"] and out=A.layers["counts"] when the goal is to overwrite the layer because it is very explicit and obvious but some people might find it too "verbose".

Btw I find overwriting a layer called counts with log transformed or scaled data very confusing... what is the usecase for supporting overwriting the input layer? Is it to save memory?