add alternative ssGSEA

[GWMcElfresh]

Hi all,

This is an alternative way to score pathways in transcriptomics data.

It's based on the Wasserstein statistic/earth mover's distance between ranks. This statistic computes how "perturbed" the pathway-ranks for a specific pathway is compared to the whole transcriptome. It is very straightforward to visualize and compute for whole transcriptome data (pseudobulk, spatial, bulk).

The distance is the integral between these two curves:

It is more difficult to visualize the distance in single cell RNASeq, as the distance is pretty dependent on the number of genes expressed within the cell (for us, nFeature_RNA).

I think this can be improved by subsetting to just the expressed genes:

and then normalized (perhaps to the max number of genes across all cells? TBD, I need to experiment more with more single cell data.)

I'll add more details later!

-GW

[GWMcElfresh]

Just for benchmarking, scoring RIRA's T cells with MsigDB's hallmark gene sets takes ~17 hours on 4 cores. BB01 (timecourse) takes about 14 hours with 4 cores.

[bbimber]

@GWMcElfresh: i didnt review this super carefully, but if you feel like this is good you can merge it

[GWMcElfresh]

@GWMcElfresh: i didnt review this super carefully, but if you feel like this is good you can merge it

Thanks, I think the wasserstein based measures are probably well tested (in theory, not code), but the dts based stuff (tested in code, not theory) still passes with a magnitude like 60k higher than I think it should.

still cooking a bit, but I might pull back on the scope of this function and only support lightly support dts as an experimental capability, with wasserstein defaults. The speedup from a6fa79c is really hard to turn down.