issues/comments re: rewriting dim. red. w/ `scrapper`

[HelenaLC]

Sorry, another one... Just going thru scran/scater deps and trying to figure out what to use instead. I can see PCA, t-SNE, UMAP, but not other's previously supported. I guess I can do without diffusion map, but MDS would be nice -- is that in the plans, maybe :)? (*I can use stats of course, would just be nice to have it to switch-and-call from scrapper depending on input arg.)

[HelenaLC]

I think also UMAP/TSNE are missing a features/subset.row or assay.type argument; I know they are intended to run on an existing DR here (e.g., PCA), but there are data types where people don't do this usually (e.g., flow cytometry/CyTOF might only have ~10 features to being with). As a hack I guess I could duplicate the assay into a mock DR internally.

[HelenaLC]

Also can't seem to find an argument to specify how many PCs to use for UMAP/TSNE; this was something like pca=20 previously, but now the only choice would be to subset PCs? (None of this is detrimental, just leaving notes as I go...)

[HelenaLC]

unrelated... computeRnaQcMetrics does not allow sub=NULL (e.g., specifying mito features or similar is a must at the moment).

[LTLA]

Sorry, another one... Just going thru scran/scater deps and trying to figure out what to use instead. I can see PCA, t-SNE, UMAP, but not other's previously supported. I guess I can do without diffusion map, but MDS would be nice -- is that in the plans, maybe :)? (*I can use stats of course, would just be nice to have it to switch-and-call from scrapper depending on input arg.)

I wasn't planning on doing so. I don't think I can do much better than stats::cmdscale().

I think also UMAP/TSNE are missing a features/subset.row or assay.type argument; I know they are intended to run on an existing DR here (e.g., PCA), but there are data types where people don't do this usually (e.g., flow cytometry/CyTOF might only have ~10 features to being with). As a hack I guess I could duplicate the assay into a mock DR internally.

You could just pass in the assay matrix directly to runUmap() (no .se suffix). I believe it should just work, even with something weird like a DelayedMatrix.

The *.se() functions are just very light-weight wrappers around their non-.se() counterparts. If you want to do something non-standard, calling the non-.se() function will give you more control.

Also can't seem to find an argument to specify how many PCs to use for UMAP/TSNE; this was something like pca=20 previously, but now the only choice would be to subset PCs? (None of this is detrimental, just leaving notes as I go...)

I didn't implement this in scrapper's *.se() functions, to keep those functions simple. If someone wants to do something like that, they can just call the non-se() function:

# Note: the non-SE functions expect the columns to be cells in the PC matrix,
# as the column-major layout is more performant for the various calculations.
out <- rumUmap(t(reducedDim(sce, "PCA")[,1:20]))

... which isn't that much more work but is more explicit/intuitive than having to figure out what a hypothetical num.components= argument actually means.

In principle, if many folks were frequently changing the number of PCs, then there could be a case for a convenient num.components= argument. But I don't think it's all that common, and when people want to change the number of PCs, they tend to do it upstream - either by changing number= in runPca(), or by subsetting and storing it in a new reducedDim entry. This ensures that the effect of the number of PCs can be evaluated on all downstream steps, rather than having to individually set num.components= in each step.

unrelated... computeRnaQcMetrics does not allow sub=NULL (e.g., specifying mito features or similar is a must at the moment).

I was frequently forgetting to specify mitochondrial genes, so I made subset= a required argument. I believe you can set subset=list() if you have no features.