Tracking issue for ark-analysis interoperability

saeyslab / harpy

Other

2 stars 0 forks source link

Tracking issue for ark-analysis interoperability #5

Open berombau opened 3 months ago

berombau commented 3 months ago

Some interop with ark-analysis would be nice for running analyses and comparing and benchmarking. Seems like they are starting with AnnData interoperability themselves, mostly for the downstream analysis (https://github.com/angelolab/ark-analysis/pull/1112).

[ ] make sure the Harpy sdata.table work with anndata_dir (https://github.com/angelolab/ark-analysis/blob/3bb2f09b6eab5668ec74e27f2b25fed24af02f7f/templates/cell_neighbors_analysis.ipynb)
[ ] most interesting are "2. Pixel clustering with Pixie" and "3. Cell clustering with Pixie", as they contain a Jupyter notebook GUI for manual cluster adjustment and annotation

ArneDefauw commented 2 months ago

This PR https://github.com/saeyslab/harpy/pull/21 implements features in the notebook 2. Pixel clustering with Pixie".

Resulting table is compatible with the ark gui to visualize clusters and meta clusters via a heatmap. See notebook on how to use it.

ArneDefauw commented 2 months ago

This PR https://github.com/saeyslab/harpy/pull/22 implements much of the features of notebook 3. Cell clustering with pixie

See notebook and https://github.com/saeyslab/harpy/blob/34164eb34adc73475e5c9bba1feea9e242a28fe7/src/sparrow/_tests/test_table/test_cell_clustering/test_cell_clustering.py#L12 for illustration on how to use it

ArneDefauw commented 1 month ago

This PR https://github.com/saeyslab/harpy/pull/23 implements the multiple fov scenario as in notebook cell clustering with pixie.

See notebook for an example on how to use it, and the unit test: https://github.com/saeyslab/harpy/blob/49575620b5464a30a3a75a9c97696131072106e3/src/sparrow/_tests/test_table/test_cell_clustering/test_cell_clustering.py#L11

I've compared our implementation of pixel and cell clustering with the implementation of ark:

All preprocessing gives same results as ark,
for flowsom clustering there is some randomness involved, so it is harder to compare, but qualitatively, we obtain similar results.
our approach, working with dask arrays for pixel clustering instead of dataframes is more scalable.