aggregateAcrossCells Error: "New 'sample_id's must map uniquely" only in SpatialExperiment

[lahuuki]

Hi scuttle team,

I have found an issue with scuttle::aggregateAcrossCells() . Trying to pseudobulk a SpatialExperiment object on a single variable in colData results in the error Error in .local(x, ..., value) : New 'sample_id's must map uniquely.

The traceback() on this error is unhelpful, and the error message leaves me unsure of what to change.

However if I convert the object to a SingleCellExperiment the function runs as expected.

library("spatialLIBD")
library("scuttle")

spe <- fetch_data("spe")
class(spe)
# [1] "SpatialExperiment"

table(spe$spatialLIBD)
# L1    L2    L3    L4    L5    L6    WM 
# 5322  2858 17587  3547  7300  6201  4514

spe_pseudo <- scuttle::aggregateAcrossCells(spe, id = spe$spatialLIBD)
# Error in .local(x, ..., value) : New 'sample_id's must map uniquely

## conver to single cell 
spatialCoords(spe) <- NULL
imgData(spe) <- NULL

spe <- as(spe, "SingleCellExperiment")

## runs successfully
spe_pseudo <- scuttle::aggregateAcrossCells(spe, id = spe$spatialLIBD)

[LTLA]

This is an unfortunate interaction involving several factors.

First we have aggregateAcrossCells()'s attempt to collapse column data for the pseudo-bulk samples. The idea is to generate sensible pseudo-bulk colData from the per-cell colData, typically by seeing if there is a single unique value. So, for example, if all cells in a pseudo-bulk come from the same treatment condition/patient/whatever, that pseudo-bulk sample will be assigned the relevant treatment/patient/whatever in its colData. This is useful for generating sane colData that can be immediately fed into design matrices, e.g., for edgeR.

The next factor is that aggregateAcrossCells() will try to preserve the type of the input object. So if you give it a SingleCellExperiment, it will return a SCE. If you give it a RangedSummarizedExperiment, it will return an RSE. And yes, if you give it a SpatialExperiment, it will return a SpatialExperiment.

Finally we have SpatialExperiment's insistence that the sample IDs must map uniquely. This is because the SpatialExperiment uses a sample_id column in the colData to assign each column (i.e., spot) to a sample (which may have one or more images). Obviously each spot must belong to exactly one sample. But if your pseudo-bulks are not nested within each sample, the aggregated colData will not have one unique sample_id per pseudo-bulk, in which case it is no longer consistent with the expectations for a SpatialExperiment object.

There are several solutions, but I would say that the most sensible approach is to convert the SpatialExperiment to an SCE before calling aggregateAcrossCells(). Even if we were to put aside the sample_id error, the SpatialExperiment concept might not make sense for pseudo-bulks. For example, what are the "coordinates" of a pseudo-bulk sample? That concept would only make sense if your pseudo-bulks were formed from spatially-adjacent spots in the same sample. (By comparison, aggregateAcrossCells() for an SCE will compute the average of the reduced dimension coordinates for each pseudo-bulk, under the assumption that the pseudo-bulks refer to similar clusters or whatever.)

If you must keep it in a SpatialExperiment, I'd make sure that each pseudo-bulk is nested within a sample. This would at least get rid of the sample_id error as each pseudo-bulk will now be associated with exactly one sample_id. Someone could even consider writing an aggregateAcrossCells() method for SpatialExperiment to average the coordinates within each sample, though I don't know how useful that really is.

[lahuuki]

Hi Aaron,
That makes sense, will just convert to SCE when needed.
Thanks for the fast response!
-Louise