Relevance of running `compute_for_mcview` prior to importing data set

[tzeitim]

What are the gains of running compute_for_mcview on a data set before importing it into an MCView project? Is it just a matter of computing outliers independently from MCView::import_dataset ?

I am encountering issues where a massive number of threads are triggered by BLAS (issue 24) when running such function (compute_for_mcview) so I would like to understand the difference between loading a data set with and without this step.

This is an example of the python command that clogs BLAS

outliers = mc.pl.compute_for_mcview(adata=clean, gdata=metacells, random_seed=123456, compute_var_var_similarity=dict(top=50, bottom=50))

[aviezerl]

It is listed in https://metacells.readthedocs.io/en/latest/pipeline.html?highlight=mcview#metacells.pipeline.mcview.compute_for_mcview

1. UMAP projections (umap_x, umap_y per obs; also umap_u, umap_v, umap_w for chameleon)
2. Inner fold factors (inner_fold layer).
3. Most similar MC for each outlier (most_similar per obs).
4. Deviant fold factors (deviant_fold layer).
5. Gene-gene similarity (var_var_similarity in varp, only top and bottom 50 by default).
6. Significant genes (significant_gene mask per var).
7. Outliers fold factors per cell (most_similar_fold layer in the slice of the per-cells data containing only the outliers).

The app should run without it though, with some features missing (inner-fold, deviant-fold, projected-fold, significant genes, outliers), and some features would take more time at the import_dataset stage (UMAP and gene/gene correlations).

[tzeitim]

Thanks for the reference! I should have been more precise, I was aware of this part of the problem (python documentation) but I am missing one for the pre-processing done on the R front.

Is there an equivalent documentation for the steps done by MCView::import? Or how to control some of its more technical parametes (e.g. number of cores used when computing stats)