Improve readQFeatures

[leopoldguyot]

#264

[leopoldguyot]

For a subset of 500 cells (of 2310) from the data of Wu et al. 2026 (https://zenodo.org/records/22283331/files/peptide_report.tsv)
3 strategies:

1. zero as NA and log transformation before import into QFeatures

        qfeat1 <- readQFeatures(subpep,
            runCol = "cellID",
            quantCols = c("Precursor.Quantity"),
            fnames = "Precursor.Id",
            zeroToNA = TRUE,
            logTransform = TRUE
        )

2. zero as NA and log transformation after import into QFeatures

       qfeat2 <- readQFeatures(subpep,
            runCol = "cellID",
            quantCols = c("Precursor.Quantity"),
            fnames = "Precursor.Id",
            zeroToNA = FALSE,
            logTransform = FALSE
        )
        qfeat2 <- zeroIsNA(qfeat2, i = names(qfeat2))
        qfeat2 <- logTransform(qfeat2,
            i = names(qfeat2),
            name = paste0("log_", names(qfeat2))
        )

3. zero as NA and log transformation before import into QFeatures + directly import the psms joined (some information are lost, ie. quality metrics specific to a run)

        metadata_cols <- subpep %>%
            group_by(Precursor.Id) %>%
            summarise(
                across(-c(cellID, Precursor.Quantity), ~
                    n_distinct(.x)),
                .groups = "drop"
            ) %>%
            select(-Precursor.Id) %>%
            select(where(~ all(.x == 1L))) %>%
            names()
        pep_wide <- subpep %>%
            pivot_wider(
                id_cols = all_of(c("Precursor.Id", metadata_cols)),
                names_from = cellID,
                values_from = Precursor.Quantity
            )
        qfeat3 <- readQFeatures(pep_wide,
            quantCols = unique(subpep$cellID),
            fnames = "Precursor.Id",
            zeroToNA = TRUE,
            logTransform = TRUE
        )

	Elapsed_Time_sec	Peak_RAM_Used_MiB
1	107.389	1902.5
2	1200.995	1903.5
3	22.032	1620.7

[leopoldguyot]

With multiple sets, we could split the data before importing it as a SummarizedExperiment.

[cvanderaa]

Two quick thoughts:

• If log transformation becomes a default, we must be careful with the way CVs are computed. It should be computed on a linear scale, otherwise we need to adapt the CV formula (see here). I can't remember if QFeatures had a CV related function, but this is the case in scp. To be honest, we could also smply drop CVs, as I think there are quite some issues.
• It seems you are doing this for computational efficiency. In fact, filtering would also be much faster if performed before the QFeatures object. I've seen script were part of the filtering is performed upstream readQfeatures, because it offers more flexibility (eg using tidyverse functions). I doesn't require any functionality on our side as it occurs upstream, but maybe usefull to document somewhere?

[leopoldguyot]

Two quick thoughts:

• If log transformation becomes a default, we must be careful with the way CVs are computed. It should be computed on a linear scale, otherwise we need to adapt the CV formula (see here). I can't remember if QFeatures had a CV related function, but this is the case in scp. To be honest, we could also smply drop CVs, as I think there are quite some issues.

Indeed, that is something we should be careful about. I am still not sure whether logTransform should be the default, as this would break existing scripts. Concerning CV, we could adapt the documentation of medianCVperCell or implement a version that is appropriate for log-transformed data.

• It seems you are doing this for computational efficiency. In fact, filtering would also be much faster if performed before the QFeatures object. I've seen script were part of the filtering is performed upstream readQfeatures, because it offers more flexibility (eg using tidyverse functions). I doesn't require any functionality on our side as it occurs upstream, but maybe usefull to document somewhere?

Good point, I think it could be part of the vignette suggested here: UCLouvain-CBIO/scp#97 Still not sure if we should add a new vignette or improve others vignettes (eg. readSCP).