Changes for the Bioconductor review (BiocContributions #171).
citation("prolfqua") works again: inst/CITATION contained a non-ASCII dash that utils::readCitationFile() could not read.vignette("prolfqua"): introduction, installation with BiocManager, a quick start, and links to all tutorials on the package website.Comparing2Groups describes the simulated data it actually uses (the text still referred to MaxQuant Ionstar data) and no longer looks up protein IDs that do not exist in it. Both vignettes use BiocStyle.?prolfqua opens a package overview page.script_copy_helper_vec() gains an overwrite argument (default TRUE, as before).summary_ROPECA_median_p.scaled() is renamed to summary_ROPECA_median_p_scaled() (Bioconductor function names must not contain .). There is no alias.data_spectronautDIA250_A, data_skylinePRMSample_A and data_skylineSRM_HL_A; their examples use simulated data now. data_IonstarProtein_subsetNorm and data_SAINTe_output move to prolfquasaint, data_benchmarkExample and data_test_confusion_matrix_scores to prolfquabenchmark (prolfqua_data(<name>, package = <package>)). The unused ExperimentHub manifest inst/extdata/metadata.csv is gone. inst/script/README.md documents the four datasets that remain.lm_missing method (ContrastsLMMissingFacade) and ContrastsMissing (group-mean substitution with a pooled-variance t-test). MissingHelpers keeps get_stats() and get_lod(); its imputation and contrast methods and the weighted argument are gone. All of it lives on in the new package prolfquatakeoutmissing: loading it registers method = "lm_missing" again. In prolfqua, use lm_impute, which refits the proteins plain lm cannot estimate with LOD-imputed values.This release removes about 3,600 lines (20%) of dead or duplicated code from R/. Everything removed had no callers in prolfqua or in the downstream packages (prolfquapp, prophosqua, prolfquappPTMreaders, prolfquasaint, prolfquabenchmark, ptm-pipeline).
filter_proteins_by_peptide_count() and nr_B_in_A() are back, internal: the filtered() method of the saved data_ionstar dataset calls the filter, so its removal broke every caller of that dataset, prolfquapp's QC vignette among them. nr_B_in_A(pdata, level_a, level_b) only counts the children of each parent; the filter, LFQData$filter_proteins_by_peptide_count() and nr_features_experiment() count through it instead of each repeating the count.sigma() for MASS::rlm fits took the square root twice. The rlm method's contrast sigma and all moderated statistics derived from it (std.error, statistic, df, p.value, FDR, confidence intervals) change; fold changes do not. On simulated data the moderated df drops from about 8,000 to about 25.moderated_p_limma(variance_floor = ) with robust = FALSE returned NA moderated p-values for rows above the floor, because the scalar prior df was padded with NA when the floored rows were set to Inf. This is the default path of the binomial_nested facade (binomial_bound = TRUE). Rows above the floor now keep the estimated prior df.merge_contrasts_results(): when both inputs share a model name, modelName is now <name>_prefer / <name>_add instead of NA for every row.setup_analysis() no longer fails when sample_name and file_name are the same column.options(prolfqua.vectorize) is removed; contrasts always use the per-row path, which was already the default. Setting the option has no effect.ContrastsFirth: proteins whose contrast fails are reported with the contrasts_linfct() warning listing the dropped protein IDs, instead of a message.ContrastsInterface provides default get_Plotter() and to_wide() methods, so subclasses no longer need to implement them.estimate_intensity() takes method = "medpolish" or "rlm" instead of a function in .func.config field is filled on the first get_config() call; use get_config() rather than reading $config.center_to_reference_cfg() always returns a median-centred copy (summary and copy removed). scale_with_subset() gains colname and loses the unused get_scales.plot_stat_violin_median() checks stat with match.arg(), so the default stat now works (it used "CV" only when given explicitly).get_p_values_pbeta() returns NA for an NA n.obs instead of stopping.script_copy_helper_vec() reports the source path it actually copies from and warns once for a missing file; find_package_file() returns NULL invisibly when the file is not found.contrasts_fisher_exact(), contrasts_linfct_firth(), effective_contaminant_pattern(), effective_decoy_pattern(), filter_difference(), hierarchy_counts_sample(), is_singular_lm(), lfq_power_t_test_quantiles(), linfct_all_possible_contrasts(), linfct_factors_contrasts(), medpolish_estimate(), medpolish_estimate_df(), medpolish_estimate_dfconfig(), moderated_p_limma_long(), names_to_matrix(), nr_B_in_A(), plot_hierarchies_add_quantline(), plot_hierarchies_line_df(), plot_lmer_peptide_predictions(), remove_na_rows(), rlm_estimate(), rlm_estimate_dfconfig(), scatter_plotly(), summarize_stats_all(), tidy_to_wide(), and the function forms of filter_proteins_by_peptide_count() and remove_small_intensities() (the LFQData methods of the same names are unchanged). unregister_facade() is no longer exported.LFQDataStats$power_t_test_quantiles() or lfq_power_t_test_quantiles_V2(); summarize_stats(x, factor_key = NULL) for summarize_stats_all(x); the Summariser method $hierarchy_counts_sample().data_checksummarizerobust and data_correlatedPeptideList.LFQData$filter_difference(), $omit_na(), $contaminant_proportion(); LFQDataPlotter$write(), $write_pltly(); Model$write_anova_figures(), $write_coef_figures() and the same on ModelFirth; ContrastsPlotter$score_plotly(), $histogram_estimate() (use $histogram_diff()); AnalysisConfiguration$id_required().report_columns (argument and field of all strategy_*() and Strategy* classes), is_mixed, $model_fun(get_formula = ) (use $formula), strategy_logistf(test = )ModelLimma$design, $formula and the matching $new() arguments; Model$anova_df, ModelFirth$anova_df; AnovaExtractor$col_fdr; Contrasts$protein_annot; ContrastsMissing$global, $present, $minsdContrastsPlotter prefix and protein_annot fields and $new(protein_annot = ); LFQDataPlotter file_paths_pdf, file_paths_htmlAnalysisConfiguration$hierarchy_keys(rev = ); setup_analysis(cc = , from_factors = ); plot_hierarchies_line(separate = ); transform_work_intensity(intensity_new_name = ); compute_pooled(method = ) and poolvar(method = ) with the "V2" pooled variance; sim_make_model_lmer(singular = ); internal facade fields ContrastsLMMissingFacade$missing_contrast, $merged, ContrastsLimpaNestedFacade$.lfqdata_nestedbuild_contrast_analysis() no longer honours an undocumented builder field on facade registry entries.ContrastsFacadeBase$.setup(), StrategyBase, AggregatorBase, plotting methods on ModelInterface); the four build_model_limma*() builders share one core. Class vectors of strategies and aggregators gain StrategyBase / AggregatorBase.MissingHelpers$get_lod() no longer returns NA for data without groups seen once. It falls back to the partly observed groups with the fewest observations, and to the lowest group mean when every group is complete or empty. build_model_impute() therefore refits a protein missing a whole group in nearly complete data, where the NA LOD made the refit fail silently and left the protein unfitted.
New impute_from_model() fills the missing response cells of an LFQData with predictions from its per-protein lm models (from build_model() or build_model_impute()), leaving observed values unchanged, and reports per protein how many cells were observed and imputed and by which route (complete, fitted, lod_refit or none). Additive designs such as ~ group_ + pair are supported; a model that is not lm-based is an error.
A mixed-model analysis (lmer_nested) no longer aborts partway through when a
single protein has no estimable contrast. Such a protein is now dropped with a
warning naming it, like a rank-deficient one, instead of failing the whole run
with subscript out of bounds from lmerTest::contest().
A ContrastConfiguration built with pvalue_col = NA_character_ now stores
"" instead. HDF5 has no character NA, so a configuration serialized into an
AnnData artifact came back with pvalue_col set to the literal string "NA"
and has_pvalue() reported a p-value column that no backend had produced.
Backends without a p-value, such as SAINTexpress, are now read back
correctly.
The peptide count in simulated data is now called nrPeptides, matching the
name every reader in the ecosystem produces. It was the only place the column
was called nr_peptides, which meant code written against simulated data did
not work on real data -- and vice versa -- and downstream packages carried
aliases between the two spellings. The nr_peptides option (minimum
peptides per protein) is unchanged; only the data column is renamed.
New list_to_ContrastConfiguration() rebuilds a ContrastConfiguration from the plain list produced by
R6_extract_values(), so a serialized column-role mapping (e.g. from SummarizedExperiment metadata or an AnnData
uns entry) can be restored with its methods, such as has_pvalue(), intact.
Moderated t/Wald contrast tables now report the posterior contrast standard error in std.error, while
std.error.unmoderated and df.unmoderated preserve the pre-moderation test inputs for downstream comparisons.
Moderated confidence intervals now use the contrast scale rather than the residual-variance scale; this corrects
intervals that were too wide when the raw contrast scale factor was below one and too narrow when it was above one.
The separate infinite-prior-df fallback is unchanged and remains a distinct follow-up decision.
main_effect_contrasts() now emits averaging contrasts with a top-level LHS - RHS, making its output directly valid for the stricter contrast parser.
Added the binomial_nested facade for nested LFQ data. It reuses the Firth backend's completed peptide-detection
representation, fits protein-by-sample detected/undetected peptide counts with a quasibinomial GLM, and reports
moderated protein-level log odds-ratio contrasts. A symmetric pseudo-count stabilizes separated fits, and posterior
dispersion is bounded below by one by default (binomial_bound = FALSE disables the bound).
Variance moderation now uses limma::squeezeVar() directly. moderated_p_limma() — and therefore every
moderated Wald-test facade (lm, rlm, rfit, lm_impute, lm_missing, rfit_impute, lmer_nested) and
the DEqMS small-group fallback — now calls limma::squeezeVar() instead of the vendored squeezeVarRob() fork
(limma is already a dependency). Modern limma estimates the empirical-Bayes prior robustly on the fractional / low
residual degrees of freedom these estimators produce, without the min_df workaround prolfqua inherited from MSqRob.
This slightly changes moderated p-values, FDR and confidence bounds; on the IonStar spike-in benchmark AUC is
unchanged, FDR calibration is maintained, and ≥99.7% of significance calls are identical. See the
SqueezeVar_comparison vignette in prolfquabenchmark. (ropeca_nested is unaffected — it aggregates raw
peptide-level p-values and never used variance moderation. A contrast with fewer than three features now returns
unmoderated results, matching limma, where the empirical-Bayes prior is undefined.)
Fixed the scale used to moderate rlm contrasts. StrategyRLM$sigma() now returns the robust scale
MASS::rlm builds its standard errors from (fit$s) instead of stats::sigma() (the ordinary-residual scale).
Because the two differ, ContrastsModerated's sigma / sqrt(var.post) rescaling was leaving a spurious
per-protein factor in moderated rlm statistics, p-values, FDR and confidence bounds. Only the rlm facade's
moderated output is affected — lm and rfit were already coherent, and rlm's unmoderated statistics
(diff, std.error) were always correct.
Removed the exported squeezeVarRob() and its internal helpers (fitFDist_LG, fitFDistRobustly_LG); the whole
vendored squeezeVarRob.R file is gone. Use limma::squeezeVar() instead. DEqMS moderation now calls
limma::trigammaInverse() (identical to the copy that was removed). This is a breaking change only for code that
called prolfqua::squeezeVarRob() directly — the ecosystem does not. Also dropped the now-unused statmod from
Imports.
User-facing errors at prolfqua's public entry points now carry typed condition classes (all inheriting from
prolfqua_error), so callers and tests can catch failures by class instead of matching message text. Covered:
setup_analysis(), build_contrast_analysis(), the LFQData$set_data() / set_config_value() mutators,
LFQData$get_Aggregator(), the aggregators' "must aggregate a multi-level hierarchy" / "aggregate first" checks,
LFQDataTransformer$center_to_reference(), and ContrastsSimpleImpute$get_contrasts(). LFQData$set_data() and
set_config_value() additionally now reject data that no longer contains the columns required by the current
configuration, instead of accepting invalid state silently.
Abundance-density plots now carry per-sample Plotly highlight keys, allowing interactive reports to fade non-hovered sample curves while preserving the existing ggplot output.
Breaking — contrast schema. The modelName column of get_contrasts() output is now the selected facade key (lm, rlm, rfit, lm_impute, lm_missing, limma, limma_impute, limma_voom, limma_voom_impute, limpa, deqms, deqms_voom, firth, lmer_nested, ropeca_nested, firth_nested, limpa_nested) instead of the testing-schema label (WaldTest_moderated, *_DEqMS, *_imputed, …). Rescue/imputation state moved to a new estimate_type column with values observed, lod_imputed, or missing_fallback. The redundant facade column was removed (its role is now played by modelName). The moderated-Wald-test wording belongs in methods text, not per-row labels. Downstream code that filtered on modelName == "WaldTest_moderated" (or the _imputed/_DEqMS variants) or read the facade column must migrate to the facade key and estimate_type.
build_contrast_analysis() and the exported FACADE_REGISTRY now derive method dispatch and the available-method list from a single seeded registry (list_facades()), removing three hand-maintained copies that could drift. New exported base class ContrastsFacadeBase holds the shared facade plumbing; the 19 built-in facades are now thin subclasses.
ContrastsPlotter colours volcano/MA/score plots by estimate_type when present (and the colour column was left at its default), using black for observed estimates, green for LOD-imputed estimates, and blue for group-mean fallback estimates.
Correctness fixes: AggregateTopN now validates func via match.arg (an invalid value errors instead of silently meaning mean); Model$get_anova() drops degenerate rows by is.na(factor) rather than a never-matching "NULL" string; sim_lfq_data() now honours its mean_prot argument; LFQDataTransformer/LFQDataSummariser deep-clone their input like the other decorators (no more mutating the caller's object); ContrastsFirth$get_linfct() returns the linfct-annotated model copies instead of mutating the shared model; contrasts_fisher_exact() pre-allocates nrow(x) results; ContrastsMissing now actually validates its output schema.
Fixed the vectorized contrast path (options(prolfqua.vectorize = TRUE)) to match the default loop path on rank-deficient (aliased) model fits: a contrast whose weights fall on non-estimable coefficients now returns NA instead of a spurious fold-change of 0 with a NaN p-value. The invalid-row guard now counts nonzero weights rather than summing signed weights, so contrasts with canceling +1/-1 weights on missing coefficients are correctly flagged. The default (non-vectorized) path was never affected.
The vectorized contrast path now assigns stable contrast_1, contrast_2, ... labels to unnamed contrasts, matching
the documented naming behavior instead of leaking an expression into the result row names.
get_contrast() now derives group_1/group_2 from the contrast's left/right side expressions, fixing mislabeled per-group columns for averaging contrasts such as (group_A + group_B)/2 - group_Ctrl (previously it used the first two group tokens). Simple A - B contrasts are unaffected. A contrast that is not a difference LHS - RHS now errors with a clear message instead of silently extracting the first two group tokens. Nested contrasts that reference an earlier contrast by name (e.g. Interaction = "AvsB_gv_X - AvsB_gv_Z") remain supported in the ContrastsMissing / lm_missing path.
setup_analysis() now stops with an informative error (listing the offending keys) when a hierarchy-key/sample combination has more than one observation, instead of silently returning a different-schema count table that crashed downstream. Pass debug = TRUE to recover the old behaviour and return the count table for inspection.
Removed the unused impute_with_zcomp(), estimate_lod_global(), and function_lod_quantile() exports (and the zCompositions dependency). For missing-value imputation use AggregateLimpa$new(lfqdata, impute_only = TRUE)$aggregate().
Hardened plot_pca(): errors early on duplicated sample names, an all-missing matrix, or too few samples instead of returning NULL (which broke pca_plotly()); joins scores to annotation with an explicit by; makes prcomp(center = TRUE, scale. = FALSE) explicit.
Hardened abundance heatmaps for sparse significant-feature subsets: when row or column distances are non-finite because of missing values, plot_heatmap() now falls back to the input order instead of returning a ComplexHeatmap object that fails during drawing.
Abundance heatmaps now retain every partially observed feature that can be meaningfully row z-scored by default, while automatically removing constant, single-observation, and all-missing rows. Significant-feature reports no longer hide bait-only results or require a dataset-specific missing-value threshold, and sparse matrices render safely even when clustering distances are unavailable.
LFQDataPlotter$heatmap() now shows only the top_n most variable features (default 1000), ranked by the prolfqua per-feature statistic (CV for untransformed data, sd for transformed, via LFQDataStats). Row clustering uses stats::hclust, which errors above 65536 features, so peptide-list / entrapment searches with tens of thousands of degenerate protein groups no longer crash the QC heatmap. Pass top_n = NULL (or Inf) to keep every feature.
StrategyLogistf now uses Wald confidence intervals instead of profiling every coefficient, preventing firth_nested
analyses from stalling for days on proteins with hundreds or thousands of peptides.
check-bioc Makefile target and a Bioconductor Docker check image.ContrastsPlotter$volcano() and volcano_plotly() no longer cap positive FDR/p-value scores at 1e-4 by default.LFQData$to_wide() to data_wide() and removed deprecated compatibility wrappers.nr_obs_experiment() into nr_children_experiment() and nr_features_experiment().LFQData data/config internals private and moved callers to accessor methods.LFQData APIs.