Feature list - beware of Broad's blocklist

[dlogan]

Hi @apahl , in a JUMP cell painting meeting someone pointed to this file in this repo here (https://github.com/mpimp-comas/2021_grigalunas_burhop_zinken/blob/master/cpa.py). In it, I noticed that there were some CellProfiler features that Broad/CellProfiler folks have noted to be suspicious. Namely this is their 'blocklist' of features:
https://figshare.com/articles/dataset/Blacklist_Features_-_Cell_Profiler/10255811

So this is not necessarily a problem, but I wanted to warn you that in the future you may want to exclude these feature categories as uninformative or not trustworthy.

Note this discussion as well: cytomining/pycytominer#176

[apahl]

Hi @dlogan, thanks a lot for the heads up, I will look into this. We selected our set of features from their reproducibility between two biological repeats of the same plate.

[dlogan]

No problem. In our experience, some of these Correlation (in particular, RWC and Manders) and Granularity features look like noise.

[apahl]

I just compared our feature list (our channels are named differently) with the feature block list and it seems that only two of our 579 features overlap with the 55 features from the block list:

Nuclei_Correlation_RWC_Mito_Syto
Nuclei_Correlation_RWC_Syto_Mito

We are using no Nuclei_Correlation_Manders_* features and only 5 from Nuclei_Granularity_* (which are not part of the block list):

Nuclei_Granularity_1_Mito
Nuclei_Granularity_2_Hoechst
Nuclei_Granularity_3_Mito
Nuclei_Granularity_3_Ph_golgi
Nuclei_Granularity_4_Mito

I think this could actually be a nice confirmation of the feature selection by reproducibility procedure.

Again, thanks a lot for pointing this out.

KR Axel