Sort columns without splitting lower levels

[lingyixu]

First, thanks for making and maintaining this package for Python users.

In the context of sorting multilevel columns, i.e. passing a dataframe to col_split, would it be possible to split columns on the first level only and keep lower level ones unsplit? An example could be an extension of this Clustermap in Advanced Usage. When there're multiple AB values under a group, or multiple Groups values under one AB, passing both AB and Groups to col_split and specifying a non-zero col_split_gap will result in gaps between every adjacent AB and Groups. This is sometimes not ideal for identifying high-level groups. On the other hand, passing one single level to col_split doesn't necessarily guarantee a consistent order of lower levels across the high level. I'd appreciate it if an option could be provided for specifying whether to split a sorted level. Also, please kindly point me to any reference if this function is already incorporated within the package.

Thanks again and happy to chat more.

[DingWB]

Hi @lingyixu
Thanks for your feedback.
To generate the heatmap you described, you can sort the dataframe by lower level first and then set col_split =upper level and col_cluster=False.
Do you have any suggestions for the parameter? For example, if col_split=df_col.loc[:,['A','B']] and you don't need the gap between groups in B (lower level), what parameter should I add? Please give me your idea.

[lingyixu]

Hi @DingWB , thanks for getting back.

Passing a pre-sorted dataframe as you described is a good workaround. To my mind, it's worth documenting this with a line or two.

Speaking of the extra parameter, my thought was to bring it into plot_matrix() function and to use it along with col_order. It would be great to allow col_order to support multiple lists, such as taking in a dict in the form of {col_name:col_order}. The new param would then specify which column(s) to split, such as taking in a list in the form of [col_name1, col_name2, ...], with column names consistent with the keys in col_order. This way, ncols = len(col_order) could be updated with ncols = len(col_order[col_name]), and self.wspace will be updated accordingly. Does this make sense?

So far this is what I have in mind. I didn't get a chance to verify whether this is valid or not, so I could be biased. Please feel free to let me know what you think. Thanks.

[DingWB]

Hi @lingyixu ,

Thanks for your suggestion. I will think about it.
For the documentation, I do have a sentence here, maybe it is not very obvious.
Please don't hesitate to update the documentation (if something is missing) under the notebooks directory and make a pull request.

[lingyixu]

Hi @DingWB ,

Thanks for pointing that out. I see your notes there, appreciate it.

May I suggest adding a few words in the docstring, for instance, "setting it to False will preserve the column order of the input data", given that col_cluster it's set to True by default. Same for row_cluster. If you don't mind, I can proceed with making the edits and creating a pull request.

PyComplexHeatmap/PyComplexHeatmap/clustermap.py

Lines 1016 to 1017 in f5dd95c

	col_cluster :bool
	whether to perform cluster on rows/columns.

Thanks again for all your comments and feedback. For now, I'm going to mark this discussion as answered. Please feel free to add any further comments in this thread, and I’ll check back to see if any additional input is needed from my side.

[DingWB]

Hi @lingyixu ,

Thanks for your suggestion, please go ahead and add whatever you want and make a pull request.