outlier detection in scanpy?

[maximilianh]

Hi there,

instead of a fixed mito-percent cutoff that I have to set each time, does scanpy have some automated way of finding outliers of some heuristic to find a reasonable cutoff?

If not, do you think 5% is a good cutoff for most datasets?

I'm looking for a scanpy function similar to detect_outliers() in the R package scPipe.

Sorry if this has been asked before, I did some searching in the issues but couldn't find anything.

best
Max

[ivirshup]

Personally, I spend a bit of time looking at distributions of various qc metrics using sns.distplot and sns.jointplot. It would be useful if we could suggest cutoffs, but I'm not sure what best practice for this would be. Just for reference, here's what scPipe does.

I think it's hard to suggest a good all-round cutoff, since the biology can vary so much. I'd like to think any filtering should be based off some model of what a "poor quality cell" looks like for the dataset at hand.

[LuckyMD]

Based on my experience setting a single cutoff for all datasets will not work, as I've used a lot of different cutoffs depending on the distributions. I would echo @ivirshup's suggestion of looking at distributions. Joint distributions being a lot more important than individual histograms. There's a small discussion about it in our best practices paper

[maximilianh]

Thanks everyone! I wonder how this affects one-pipeline-for-everything portals, like the EBI single cell expression atlas... and standarized pipelines like cellranger

…

On Mon, Jul 1, 2019 at 3:29 PM MalteDLuecken ***@***.***> wrote: Based on my experience setting a single cutoff for all datasets will not work, as I've used a lot of different cutoffs depending on the distributions. I would echo @ivirshup <https://github.com/ivirshup>'s suggestion of looking at distributions. Joint distributions being a lot more important than individual histograms. There's a small discussion about it in our best practices paper <https://www.embopress.org/lookup/doi/10.15252/msb.20188746> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#718>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AACL4TKKTTZ4IHBJJDFAPKLP5IBDPANCNFSM4H4DUZEA> .

[LuckyMD]

As far as I'm aware Cell Ranger works with global cutoffs. I recall seeing 4k umis being used. That seems quite excessive to me. I've seen small cells fall out of the analysis with that threshold. I guess if you just want to work with sth, then a conservative cutoff might be fine, but you can always do better with your data.

[maximilianh]

Ah. I see. Thanks! Does it requires a minimum of 4000 UMIs or a maximum of 4000 UMIs ? And do we know the other thresholds, like %-mito, number of cells, number genes ?

…

On Mon, Jul 1, 2019 at 3:40 PM MalteDLuecken ***@***.***> wrote: As far as I'm aware Cell Ranger works with global cutoffs. I recall seeing 4k umis being used. That seems quite excessive to me. I've seen small cells fall out of the analysis with that threshold. I guess if you just want to work with sth, then a conservative cutoff might be fine, but you can always do better with your data. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[LuckyMD]

I recall a minimum of 4000 UMIs, although I'm not sure if that's still true in newer versions. I don't know about other thresholds. I think you should be able to find this on their website though (somewhere hidden).

[maximilianh]

Google can't find 4000 and UMI on their website: https://www.google.com/search?safe=off&ei=PC4aXYaLBcWAad_4vdAG&q=%224000%22++inurl%3A10xgenomics+umi&oq=%224000%22++inurl%3A10xgenomics+umi&gs_l=psy-ab.3...3637.3826..3984...0.0..0.76.218.3......0....1..gws-wiz.......0i71.oYAb2OqUy7I

…

[LuckyMD]

I remember it being quite difficult to find out how CellRanger does things... I found it in some pdf at the time. The 4k UMI threshold might be old by now as well though.