SE and RangeSE (and SCE) differences in accessing rowData

[lgatto]

I recently observed a substantial difference in using rowData() when using SE and RangedSE instances:

> example(RangedSummarizedExperiment)
*** output flushed ***
> se <- as(rse, "SummarizedExperiment")
> microbenchmark::microbenchmark(rowData(se), rowData(rse))
Unit: microseconds
         expr     min       lq     mean   median       uq      max neval
  rowData(se) 129.264 134.4920 138.2764 136.8520 141.9285  154.541   100
 rowData(rse) 598.731 621.2365 638.1464 633.3445 646.5985 1074.851   100

This is because rowData(), or more precisely mcols on an RSE first gets the rowRanges, and then that object's mcols.

I discovered this because I noticed significant differences when manipulating SE and SCE object. One reason is that SCE inherits from RSE, and hence suffer from the above overhead.

> sce <- as(se, "SingleCellExperiment")
> microbenchmark::microbenchmark(mcols(se), mcols(rse), mcols(sce))
Unit: microseconds
       expr      min        lq      mean   median        uq      max neval
  mcols(se)  123.715  137.6745  163.4424  150.011  177.8125  412.293   100
 mcols(rse)  589.028  635.6750  683.9597  667.770  716.7420 1086.609   100
 mcols(sce) 2664.261 2835.4035 3029.9824 2935.945 3115.0290 4567.160   100

I understand that there are historic reasons why SCE derives from RSE, but would it be conceivable to have the main SCE derive from SE, and have a RangedSCE as there's a RangedSE? I can imagine this would be a lot of work with risks of breaking things?

I am not really requesting any change here, just asking out of curiosity, as we are considering dropping using SCE for single-cell proteomics in favour of SEs because of that, and only switch the SCE when needed (for example when the reducedDim slot is populated in the downstream analyses)

Also adding @LTLA and @drisso to the discussion.