v0.4.1

Added

• mm.pp.calculate_missingness. Records, per feature, the number and ratio of samples
  without a value (var["n_missing"], var["missing_ratio"]) and, with group_obs, a
  missing_ratio_<level> per group, counting absent and stored-NaN cells of a sparse .X alike.
  Decoys receive 0 / 0.0, so a "missing at most" filter keeps them and the q-value counts stay
  intact. This replaces the hand-written to_df().isna().sum() and dropna(how="all") steps, which
  provenance could not record; the processing tutorial now drops empty features through it.
• mm.dt.save_layer and mm.dt.load_layer. save_layer copies a modality's .X into a
  named layer and load_layer copies a layer back into .X. Both are recorded in provenance,
  keep dense and sparse matrices as they are, and leave the source and destination independent.
  An existing layer is not overwritten unless overwrite=True, and a backed MuData is rejected
  with a clear error. The PTM tutorial takes its unadjusted snapshot with save_layer.
• mm.dt.drop deletes table columns and mapping entries. Besides uns entries, drop now
  removes obs/var columns, DataFrame columns inside obsm/varm, and entries of global or
  modality obsm/varm mappings, recording only the target and key for replay. _log and the
  global modality masks are protected; removing samples or features remains the job of
  pp.apply_filter, which keeps the aligned axes consistent.
• mm.dt edits the decoy table. dt.assign accepts on="decoy", and the table path
  <modality>.decoy addresses uns["decoy"] in dt.map, dt.replace and dt.drop, so the
  columns a variable filter needs on both targets and decoys are recorded in provenance instead
  of being written by hand. pp.add_filter now refuses a variable filter whose column is missing
  from the decoy table with a message that names these calls, instead of a bare KeyError. The
  processing tutorial carries the cell annotation onto the PSM var and decoy table with dt.map.

Changed

• import msmu takes about a third of the time. umap-learn and inmoose are imported when
  tl.umap, pp.correct_batch_effect(method="combat") or the limma engine run, not when the package
  is imported; on a laptop import msmu drops from about 5.7 s to 1.7 s. polars, a required
  dependency, is imported at module level in the readers instead of inside each function, and SDRF
  validation no longer describes sdrf-pipelines as an optional package -- it is a required
  dependency, so the ImportError branch is gone.
• Median polish returns NaN where samples cannot be compared. Breaking for values:
  to_protein(agg_method="median_polish") and to_ptm (its default rollup) now leave a sample
  NaN when none of the group's features observed in that sample is observed in at least one
  other sample and, when the observed samples split into groups that share no feature, keep the
  largest group and leave the others NaN (every sample when groups tie for largest). Any two
  non-NaN values of a feature are therefore on one common scale, and their differences are
  those of medpolish on the features observed in more than one sample; previously such samples
  were placed on the group's overall level, which reads as "unchanged". Groups whose every
  feature is observed in at least two samples are unaffected.
• to_protein uses every unique peptide by default. Breaking for callers that relied on
  the implicit top_n=3: pass top_n=3 to keep the cut. The docstring already stated None;
  the default now matches it. A top-N cut ranks on a NaN-skipping median and so favours
  peptides observed in few samples, which are the ones median polish cannot compare.

Fixed

• Median polish no longer flattens the difference between samples when most peptides are
  observed in one sample. A peptide observed in a single sample has a zero residual by
  construction and used to vote "no difference" in that sample's column median; with such
  peptides in the majority (e.g. 1 vs 1 LFQ without match-between-runs) both samples collapsed
  onto the overall level and the protein reported a log2 fold change of exactly 0. Those
  peptides now set the protein level only; the between-sample difference comes from the peptides
  observed in more than one sample.

Merged pull requests

What's Changed

• BID-321 Extend dt.drop to tables and axis mappings by @jlthepi in #47
• BID-329 Add recorded layer save and load operations by @jlthepi in #48
• BID-330 Leave samples that share no peptide NaN in median polish and use every peptide by default by @hyungwookc07 in #49
• BID-331 Import umap and inmoose on use and settle the lazy-import rule by @hyungwookc07 in #50
• BID-332 Keep the largest comparable group in median polish and fit single-sample peptides after the sweeps by @hyungwookc07 in #51
• BID-333 Release MSMU 0.4.1 by @hyungwookc07 in #52
• BID-335 Record decoy-table column edits through dt and explain missing decoy columns in add_filter by @hyungwookc07 in #54
• BID-336 Add pp.calculate_missingness to record per-feature missing counts in var and on the decoys by @hyungwookc07 in #55
• BID-333 Release MSMU 0.4.1 by @hyungwookc07 in #56

Full Changelog: v0.4.0...v0.4.1