Add 'equal_frequency' option to highly_variable_genes

[davidhbrann]

This fixes #415, by allowing one to find variable genes using the equal_frequency option. It also adds and option to change the number of bins for cell ranger flavor.

I originally tried to copy the implementation in Seurat, which would allow a test similar to what's already present for the equal_width implementation. However the Seurat code has an error:

else if (binning.method=="equal_frequency") {
        data_x_bin <- cut(x = gene.mean, breaks = c(-1,quantile(gene.mean[gene.mean>0],probs=seq(0,1,length.out=num.bin))))
}

The -1 in the code makes it such that there is always only one value in the first bin, which goes from -1 to the minimum value. Not sure why they have this, but then we get different answers since the Scanpy code in highly_variable_genes always makes bins that have only one gene significant (to correct the other error from Seurat that normally excludes these bins/genes, which often contain some highly-expressed genes). Additionally, the cut function in R sometimes returns bin edges with different rounding than the Seurat implementation since Seurat does not modify the default dig.lab = 3. In contrast, I believe pandas uses the actual cutoffs in the data.

[falexwolf]

Great, thank you!

It will be merged very soon; the change in the cell ranger option might be a backwards breaking change, hence, I added this here: #453.

Alex

[gokceneraslan]

Hey @falexwolf, if starting the percentile bins from 10 is not intentional (as mentioned in #624 ) I can resolve the conflicts and merge this along with the fix in #624 .

[falexwolf]

Sounds good, @gokceneraslan!