Scanpy not working correctly with scikit-learn 0.21.1

[danielStrobl]

Hey!

Scanpy does not seem to work correctly together with scikit-learn 0.21.1.
When running the PBMC clustering tutorial (https://github.com/theislab/scanpy-tutorials/blob/master/pbmc3k.ipynb), the produced UMAP plots look very different to the reference.

By downgrading scikit-learn to 0.20.0, everything works fine.
The problem seems to arise already at the computation of the neighborhood graph, as the clustering is also different.

[LuckyMD]

Just to add to this, PCA plots look fine with the newer scikit-learn I believe. Maybe it's the umap neighbourhood graph function depending on sklearn for something?

[gokceneraslan]

Oh that's reaaaally bad. I did a quick git bisect on sklearn:

Here is the commit that broke our umaps: scikit-learn/scikit-learn#13554

[LuckyMD]

Should this be relayed to scikit-learn then? If so, that should probably be done by someone who knows where in the sc.pp.neighbors() function this is breaking...

[LuckyMD]

@flying-sheep mentioned this was known and already fixed though?

[gokceneraslan]

OK, seems to be fixed in sklearn master branch (probably scikit-learn/scikit-learn#13910), but this is such a huge bug and it has been going on since May 9th :( We could have blacklisted sklearn versions 0.21.0 and 0.21.1 if it was known, no? Some colleagues mentioned weird UMAP results with scanpy actually, it turns out they upgraded their sklearn...

[flying-sheep]

@flying-sheep mentioned this was known and already fixed though?

I meant the other breakage due to the scipy update, sorry.

We could have blacklisted sklearn versions 0.21.0 and 0.21.1 if it was known, no?

We should do that now. We can do sklearn >= 0.19.1, != 0.21.0, != 0.21.1 I think.

[gokceneraslan]

Thanks!