Encountering problems with selectFeatures

[sagnikbanerjee15]

Hello,

I am trying to use scTransformed counts instead of log-transformed counts for training and predicting cell types. I created a Seurat object, performed the sctranform and then created a SingleCellExperiment object from the sctransformed counts. The execution fails at this line:
sce_with_sctransformed_counts <- selectFeatures(sce_with_sctransformed_counts, suppress_plot = TRUE), producing the following error:

Error in lm.fit(x, y, offset = offset, singular.ok = singular.ok, ...) : 
  0 (non-NA) cases
In addition: Warning message:
In linearModel(object, n_features) :
  Your object does not contain counts() slot. Dropouts were calculated using logcounts() slot...

It does not seem like there are any NA values in the logcounts

any(is.na(logcounts(sce_with_sctransformed_counts)))
[1] FALSE

I suspected that since the sctransformed values are negative, it could be a potential source of error. Hence, I executed the following:

logcounts(sce_with_sctransformed_counts) <- log2(normcounts(sce_with_sctransformed_counts) + 5)

But I still keep getting the same error.

Could you please help me with this?

This is a portion of the code I am using.

sce_seurat <- CreateSeuratObject(counts = Data[,Train_Idx[[i]]])
sce_seurat <- SCTransform(object = sce_seurat)
sce_with_sctransformed_counts = SingleCellExperiment(assays = list(normcounts = GetAssayData(object = sce_seurat, slot = "scale.data")), 
colData = data.frame(cell_type1 = Labels[Train_Idx[[i]]]))
rowData(sce_with_sctransformed_counts)$feature_symbol <- rownames(sce_with_sctransformed_counts)
logcounts(sce_with_sctransformed_counts) <- log2(normcounts(sce_with_sctransformed_counts) + 5)			
sce_with_sctransformed_counts <- selectFeatures(sce_with_sctransformed_counts, suppress_plot = TRUE)

Thank you.

[mhemberg]

Did you try to create the counts slot instead of the logcounts? Also, I assume that negative values would have a small magnitude and be relatively few, but did you check how many of them you have after doing the normalization with sctransform?

As you point out, we do make the assumption that counts are non-negative and as sctransform was published after scmap we were not able to test this combination (which does seem quite useful so I am curious to see how well it works).