rank_genes_groups refactoring

[Koncopd]

Hi, @falexwolf
Do you have any specific things in mind for rank_genes_groups refactoring? What should be done?

[ivirshup]

I think there was some discussion of this (among other topics) here: #562

[falexwolf]

Some notes from a brief discussion with Sergei.

1. make helper functions for each method so that level of indentation and length is decreased
2. replace lists rankings_gene_... by DataFrame
3. think about simplifying the wilcoxon implementation, compare with scipy stats implementation and potentially update the test
4. investigate how the logreg implementation behaves for different choices of reference groups

[fidelram]

Can rank_genes_groups be linked to use diffxpy on top of the available methods?

I am using the following code to convert the output of rank_genes_groups to a data frame, in case is useful:

def rank_genes_groups_df(adata, key='rank_genes_groups'):
    # create a data frame with columns from .uns['rank_genes_groups'] (eg. names, 
    # logfoldchanges, pvals). 
    # Ideally, the list of columns should be consistent between methods
    # but 'logreg' does not return logfoldchanges for example

    dd = []
    groupby = adata.uns['rank_genes_groups']['params']['groupby']
    for group in adata.obs[groupby].cat.categories:
        cols = []
        # inner loop to make data frame by concatenating the columns per group
        for col in adata.uns[key].keys():
            if col != 'params':
                   cols.append(pd.DataFrame(adata.uns[key][col][group], columns=[col]))
        
        df = pd.concat(cols,axis=1)
        df['group'] = group
        dd.append(df)

    # concatenate the individual group data frames into one long data frame
    rgg = pd.concat(dd)
    rgg['group'] = rgg['group'].astype('category')
    return rgg.set_index('group')

This results on a table like this: