Can densMAP (in development version of Seurat) yield reproducible clustering?

[NicolaasVanRenne]

I"m running densMAP in the development version of Seurat, and it seems to work (hallelujah!) but the clustering is not reproducible. If I run the same pipeline twice over, the clustering is similar yet different. Is there a way to make it reproducible?

setting the seed doesn't help. Is this intrinsic to the densMAP algorithm? Or is there a way to make it reproducible?

thanks,

Nicolaas

ps command I run is:
RunUMAP(MySeuratObject, dims = 1:UMAP.dimensions, densmap=TRUE)

[alanocallaghan]

This'll be because the seed is controlled in python, which has a separate RNG stream to R. Setting the R seed doesn't affect the python one.

https://github.com/satijalab/seurat/blob/13b615c27eeeac85e5c928aa752197ac224339b9/R/dimensional_reduction.R#L1298

If you want reproducible densmap in R might I suggest using densvis? Otherwise flag the issue with the Seurat team, as I'm fairly sure the issue isn't with the densmap algorithm itself.

[alanocallaghan]

eg,

library("densvis")
set.seed(42)
x <- matrix(rnorm(200), ncol=2)
d1 <- densmap(x, random_state = 42L)
d2 <- densmap(x, random_state = 42L)
identical(d1, d2)
# TRUE

vs

library(Seurat)
data("pbmc_small")
set.seed(42)
pbmc_small1 <- RunUMAP(pbmc_small, dims = 1:5, densmap = TRUE)
set.seed(42)
pbmc_small2 <- RunUMAP(pbmc_small, dims = 1:5, densmap = TRUE)
identical(Embeddings(Reductions(pbmc_small1, "umap")), Embeddings(Reductions(pbmc_small2, "umap")))
# FALSE

densvis also uses basilisk which means the umap version is conda-controlled.

[alanocallaghan]

This is fixed now