Inconsistency between assay data used and what is advised in SingleR?

[denvercal1234GitHub]

Hi there,

Q1: In the tutorial, logcounts were used, but in SingleR documentation, it was strongly advised against using of any transformed data and prefer raw counts. Would you mind clarifying if there is a reason you guys use logcounts here?

Q2: If my sce object is flow data, I will simply set Matrix::Matrix(sparse = F)?

Thank you.

# Get cell type reference data
blueprint <- celldex::BlueprintEncodeData()

# Infer cell identities
cell_type_df <-

    assays(pbmc_small_UMAP)$logcounts %>%
    Matrix::Matrix(sparse = TRUE) %>%
    SingleR::SingleR(
        ref = blueprint,
        labels = blueprint$label.main,
        method = "single"
    ) %>%
    as.data.frame() %>%
    as_tibble(rownames="cell") %>%
    select(cell, first.labels)

From SingleR (https://bioconductor.org/books/release/SingleRBook/classic-mode.html): "For the test data, the assay data need not be log-transformed or even (scale) normalized. This is because SingleR() computes Spearman correlations within each cell, which is unaffected by monotonic transformations like cell-specific scaling or log-transformation. It is perfectly satisfactory to provide the raw counts for the test dataset to SingleR(), which is the reason for setting assay.type.test=1 in our previous SingleR() call for the Grun dataset."

[denvercal1234GitHub]

Does not need to be but ok if is. "Technically speaking, the test dataset does not need log-expression values but we compute them anyway for convenience."

[stemangiola]

could you point to which tutorial you are referring to?

[stemangiola]

@susansjy22 is this the case for HPCell?

[denvercal1234GitHub]

could you point to which tutorial you are referring to?

This guide: https://stemangiola.github.io/tidySingleCellExperiment/articles/introduction.html

[stemangiola]

@william-hutchison could you please help with this?

[william-hutchison]

Regarding point 1, I think the tutorial material is okay as is. The test data does not have to have to be log-transformed, but it does not have to be raw either:

"For the test data, the assay data need not be log-transformed or even (scale) normalized. This is because SingleR() computes Spearman correlations within each cell, which is unaffected by monotonic transformations like cell-specific scaling or log-transformation." https://bioconductor.org/books/release/SingleRBook/classic-mode.html

I assume this is why you closed the issue @denvercal1234GitHub ? Please let us know if you have any further concerns though.

I could add a note on this topic to the tutorial. Although given any assay data is fine, maybe this information is unnecessary for the user.

[stemangiola]

The thing I would like to understand is, is proving lo-transformed data an error?

Either SingleR has a method to detect if data is logged, or applying a statistics designed to non-logged data (and the other way around), to logged data, is almost never a good idea.

[susansjy22]

@susansjy22 is this the case for HPCell?

For HPCell, logNormCounts() is used to transform the test data prior to annotation with SingleR(). This should be fine, as @william-hutchison has stated. Since SingleR() uses Spearman correlation which is relies on a rank order of value rather than their actual magnitudes, so monotonic transformations like log-transformation or scaling wouldn’t affect the analysis.

[william-hutchison]

Yes I can confirm, I just tested the SingleR with both logcounts and counts and the output was identical.

[stemangiola]

Very well explained! OK, in this case, @susansjy22 let's omit the transformation for SingleR execution, @william-hutchison we can omit it in the tidySCE documentation.

Link the pull requests on the respective repositories in this issue (under development menu on the right on the pull request)