scregclust takes too long time

[Pang-Ka]

Dear authors,
I used a dataset with 20541 features and 19840 samples to run scregclust with the following codes:
fit <- scregclust( z, genesymbols, is_regulator, penalization = seq(0.1, 0.5, 0.05), n_modules = 10L, n_cycles = 50L, noise_threshold = 0.05)

It took about 18 hours to complete one cycle of one penalization score. Is there any way to speed up this process?
Many thanks!

[cyianor]

Hi,
How many tentative regulators are you considering, i.e. what is sum(is_regulator)?

Also, what kind of computer are you running this on?

[Pang-Ka]

Hi,
I found the sum(is_regulator) is 1, is it right?
My code is as follows:
z <- GetAssayData(sc, layer = "scale.data")
dim(z) #[1] 20527 19840
out <- scregclust_format(z, mode = "TF")
is_regulator <- out$is_regulator

I run the code with my MacBook Pro (Apple M1 Pro, MacOS 14.6.1)

[idacharlottalarsson]

Hi,

Adding to Felix's answer above, I would strongly recommend that you include a feature selection step before running scregclust, e.g. only including the top XX most variable genes in your gene expression matrix. For sc-data generated from a UMI-based platform, e.g. 10X Genomics, I usually normalize my data using sctransform which by default returns the 3000 most variable genes (but this can be of course be changed to return more or less genes). Usually, including all 20000+ genes is unnecessary since most of them have zero variance and are therefore uninformative (as you can see already in your first run more than 16000 genes were placed in the noise cluster). Hopefully this will speed things up.

Let us know how it goes!

[Pang-Ka]

Hi,

Adding to Felix's answer above, I would strongly recommend that you include a feature selection step before running scregclust, e.g. only including the top XX most variable genes in your gene expression matrix. For sc-data generated from a UMI-based platform, e.g. 10X Genomics, I usually normalize my data using sctransform which by default returns the 3000 most variable genes (but this can be of course be changed to return more or less genes). Usually, including all 20000+ genes is unnecessary since most of them have zero variance and are therefore uninformative (as you can see already in your first run more than 16000 genes were placed in the noise cluster). Hopefully this will speed things up.

Let us know how it goes!

Thank you for your suggestions. I tried to use top 2000 variable genes to run scregclust, but I still got some ERROR:
out <- scregclust_format(z, mode = "TF")
genesymbols <- VariableFeatures(sc)
sample_assignment <- out$sample_assignment
is_regulator <- out$is_regulator
mouseTF <- read.table('/Volumes/Data/allTFs_mm.txt')
ix<-which(genesymbols %in% mouseTF$V1)
is_regulator[ix]<-1
fit <- scregclust(z, genesymbols, is_regulator, penalization = seq(0.1, 0.5, 0.05), n_modules = 10L, n_cycles = 50L, noise_threshold = 0.05)