I'm trying to annotate a single cell object with 306k cells with SingleR with a reference from Azimuth website, but it segfaults:

library("SingleR")
library("Seurat")
pbmc_azimuth <- readRDS(url("https://zenodo.org/records/4546839/files/ref.Rds?download=1"))
pbmc_ref <- as.SingleCellExperiment(pbmc_azimuth, assay = 'refAssay') 
predictions4 <- SingleR(test = gf_fast, 
                       ref = pbmc_ref, 
                       labels = pbmc_ref$celltype.l2, 
                       assay.type.test = "reconstructed", 
                       BPPARAM = multi_core)
  |======================================================================| 100%


 *** caught segfault ***
address 0xc, cause 'memory not mapped'
> packageVersion("SingleR")
[1] '2.6.0'
> packageVersion("Seurat")
[1] '5.1.0'
> dim(gf_fast)
[1]  33047 306459
> dim(pbmc_ref)
[1]  5000 36433

I'm sorry I cannot share the data on gf_fast (multi_core is just a MulticoreParam with bpnworkers: 10). The rownames are both gene symbols I'm using R 4.4.1, and Bioc 3.19 both up to date.

In Rstudio I see errors similar to these: https://github.com/rstudio/rstudio/issues/9456#issue-911913250 I could use SingleR with other reference datasets with the same object (gf_fast), so I think it is related to the reference used. I also tried with different DE method but all of them failed.

Any help would be appreciated. Thanks

The training step is usually faster than the classification step, so you might try running trainSingleR on the reference to see if that is where it blows up, and then debug from there.