`msmu.pp.collapse_obs`

Collapse obs rows by sample_key, aggregating duplicates into one row per group.

Common use cases (with recommended agg_method):

LFQ fractions (sample × LC fraction → sample): agg_method="sum". Standard LFQ rollup since each fraction holds a different portion of the sample's peptides.
Technical replicates (sample × injection/process repeats → sample): agg_method="median" or "mean". Replicates measure the same analyte, so the central tendency is the right estimate ("sum" would inflate).
Other obs-axis duplicates (plate replicates, etc.): pick agg_method that matches the meaning of the duplication.

The effect on a given modality depends on its var structure:

When var rows are file-specific (e.g. DIA-NN psm where the var index is filename.Precursor.Id), each (sample, var) cell still holds at most one non-NaN value, so this function only reduces the obs dimension; the actual cross-fraction precursor aggregation happens later in to_peptide.
When var rows are identity-based (e.g. DDA-LFQ peptide modality where the var index is the peptide identity), a single (sample, var) cell may collect values from multiple obs rows, and agg_method performs the real rollup.

Workflow contexts

DIA-NN fractionated: read → psm/precursor → collapse_obs → to_peptide → to_protein.
DDA-LFQ fractionated: read → psm + peptide → collapse_obs (peptide gets real sum) → to_protein.
Any workflow with technical replicates: collapse_obs(agg_method="median") early in the pipeline (typically before normalisation).

Applied uniformly across all modalities so the obs axis stays consistent.

Parameters:

Name	Type	Description	Default
`mdata`	`MuData`	MuData whose obs rows contain duplicates to be collapsed.	required
`sample_key`	`str`	Column in `obs` identifying the group each row belongs to.	required
`agg_method`	`Literal['sum', 'max', 'median', 'mean']`	Aggregation across rows of the same group for each `(group, feature)` pair. One of `"sum"`, `"max"`, `"median"`, `"mean"`. NaN values are skipped; an all-NaN group yields NaN (not 0) for `"sum"` as well.	`'sum'`
`layer`	`str \| None`	Layer to aggregate. If None, `.X` is used.	`None`
`log_transformed`	`bool`	Whether the input quantification is in log2-space. Defaults to `False` because `collapse_obs` is typically called before `log2_transform` (read → collapse → log2 → normalise → ...). When `log_transformed=True` is combined with `agg_method="sum"`, the function internally converts back to linear space, sums, and re-applies log2 (i.e. `log2(sum(2^x))`) — the correct LFQ rollup for log2 input. Other `agg_method` choices (`max`, `median`, `mean`) operate directly in the given space; `mean` on log2 data corresponds to the geometric mean of linear intensities.	`False`

Returns:

Type	Description
`MuData`	New MuData with collapsed obs. Obs columns uniform within a group retain
`MuData`	their scalar value; non-uniform columns become lists preserving the original
`MuData`	row order, so per-row metadata (e.g. filenames) is not lost.

Notes