Interest in python implementation

[LinearParadox]

Hi all,

I've been interested in working on implementing this tool in python and putting it out there, so people can more easily use it in the python ecosystem. I just wanted to seek permission from the authors first as I don't want to step on any toes.

Thanks!

[plger]

Hi,
thanks for asking, but I've got no intention of doing that ATM, I suppose neither does @LTLA (assuming you want to also port the other, more classical methods included in the package) ?
So as far as I'm concerned, feel free to go ahead.
Do you intend to wrap around the R package, or to reimplement it? In the latter case, I can probably help you avoid some of the clutter relative to the main scDblFinder method : there's a lot of code, parameters and options that reflect the history of the package's development, when I didn't know what actually worked best, and that one wouldn't necessarily need to port.
Best,
Pierre-Luc

[jkanche]

If you are interested in developing these methods on Bioconductor representations like single cell experiment, delayed arrays, genomic ranges etc implemented in Python, consider using BiocPy

[LTLA]

I don't have any particular intentions on this matter; I've already used up my Python tolerance for Q1 in updating scranpy. Some of that stuff might be useful for porting over scDblFinder, along with the knncolle Python package (same as BiocNeighbors).

My Python packages also use BiocPy classes, so I don't have to think when mapping between languages.

[LinearParadox]

So I have a little bit implemented, and at least what I've had mostly works with the anndata format? I'm
happy to write other classes if needed and I have the ability, but the stuff I've had so far seems to soon with the Python framework of anndata. It would be great to know about methods or params that are no longer needed, it would definitely save some time.

The initial plan for me was to not use the R methods such as BioC neighbors, but to use Python equivalents when/if available, like scikit learn. I don't think results will be identical, but I'm hoping to have the same general algorithm, with similar performance for benchmarks.