Add key_added option to sc.tl.paga

[gokceneraslan]

It'd be cool to allow multiple abstract representations of the same graph (e.g. w.r.t different groups). What do you think about storing PAGA info under adata.uns['paga'][key_added] instead of directly adata.uns['paga']?

[LuckyMD]

Is this backward compatible for cases where old AnnData object are loaded?

[gokceneraslan]

The default behavior is the same, if key is not given, everything will be stored under adata.uns['paga'] as before and the plotting functions will also look for adata.uns['paga'] unless a key='key' is specified.

[LuckyMD]

I see... but as it's always adata.uns['paga'] what would happen if I have an old object and then run a new sc.tl.paga() with key set to something. I assume the new version has a dict in adata.uns['paga'] with the keys of the individual paga runs, whereas the old case just had the paga run data in there directly.

[gokceneraslan]

I see... but as it's always adata.uns['paga'] what would happen if I have an old object and then run a new sc.tl.paga() with key set to something.

Then you'll have it under both adata.uns['paga'] and adata.uns['paga'][key]. That's not backwards compatibility though. BC is running the old code (so no key='key') with an old object (where paga is under adata.uns['paga']) with new scanpy and getting the old result, which is satisfied here.

There are weird failure modes though, like using key='groups' or key='connectivities' might override some parts of an existing, old-style paga result. We can forbid keys like these.

Actually this is related to the versioning of AnnData specification. We should keep some sort of version like /attrs/LOOM_SPEC_VERSION in loom. Then it would be easier to understand what to expect from an anndata object that is created at any point in time.

[LuckyMD]

There are weird failure modes though, like using key='groups' or key='connectivities' might override some parts of an existing, old-style paga result. We can forbid keys like these.

this is exactly what i was worried about. 'groups' is probably not such a rare case as a key. But good to take care of that. I guess in the old & new case, you just have a bit of a messy adata.uns['paga'] dictionary, but it will still work (in most cases).

[falexwolf]

I'd actually be fine with merging this - @gokceneraslan, you can merge yourself, can't you?

[falexwolf]

We'll have 1.5.0 in a couple of days, and so it's fine to make a few behavior changes. We should put a warning in the release notes, though.