Parallelizing by subsampling TFs

[jrknoedler]

I'm having a problem with running SCENIC on a large-ish dataset (>70K cells) as my university's HPC only allows 24 hours for highly memory-intensive jobs. Would it be reasonable to parallelize this by running multiple jobs on lists of 100 TFs from the TF list and then concatenating all retained regulons? It seems like this depends on whether the enrichment scores and statistics are in any way dependent on the number of candidate regulons being tested (in which case this is bad) or whether it only looks across cells within regulons (in which case this seems like it would be fine).