NEWS.md

prolfqua 1.7.0

• 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.

prolfqua 1.6.1

• Added a check-bioc Makefile target and a Bioconductor Docker check image.
• Updated BiocCheck-related documentation and vignette metadata.
• ContrastsPlotter$volcano() and volcano_plotly() no longer cap positive FDR/p-value scores at 1e-4 by default.
• Renamed LFQData$to_wide() to data_wide() and removed deprecated compatibility wrappers.
• Split nr_obs_experiment() into nr_children_experiment() and nr_features_experiment().
• Made LFQData data/config internals private and moved callers to accessor methods.
• Refactored transformer and aggregation helpers around current LFQData APIs.