.github/pkgdown/extra.css

Skip to contents

Changelog

Source: NEWS.md

HDAnalyzeR 1.1.0

A maintenance release focused on removing dependencies that no longer exist, getting continuous integration green again, and fixing the bugs that a rewritten test suite uncovered.

Breaking changes

  • multiROC and vip have been removed. Both were archived from CRAN (multiROC on 2026-05-04, vip on 2026-07-08), which made the package uninstallable from a clean library. Their functionality is now implemented in-package:

    • Multiclass ROC AUC is computed with yardstick. The per-class one-vs-rest AUCs and the micro average are numerically identical to what multiROC returned. The macro value now uses the standard definition, the unweighted mean of the per-class AUCs, rather than multiROC’s curve-averaged variant, so it may differ slightly (by up to ~0.02 on small datasets).
    • Variable importance for the random forest and logistic regression engines is read directly from the fitted model and matches vip::vi() exactly.

  • Many packages moved from Imports to Suggests. Together with the removals below the package now installs with 28 hard dependencies instead of

    1. arrow, cluster, clusterProfiler, easyPubMed, embed, enrichplot, fpc, missForest, ppsr, readxl, WGCNA and writexl are only needed by the specific functions that use them, and those functions now raise a clear error naming the package and the exact install command when it is missing. Install everything with:
    # CRAN
install.packages(c("arrow", "cluster", "easyPubMed", "embed", "fpc",
                   "missForest", "ppsr", "readxl", "WGCNA", "writexl"))
# Bioconductor
BiocManager::install(c("clusterProfiler", "enrichplot", "org.Hs.eg.db", "ReactomePA"))

  • purrr, tidyselect and umap have been dropped entirely. umap was never actually used: hd_umap() runs through embed::step_umap(), which calls uwot. The check for umap in hd_umap() has been replaced with a check for embed. knitr, patchwork and viridis moved from Imports to Suggests; they are only used by the vignettes.

  • broom, forcats, readr, scales, stringr and tidytext have been dropped. Each was pulled in for one or two small functions, which are now implemented in R/compat.R or taken from a package that was already a hard dependency. Together with their own dependencies this removes 15 packages from a clean install (backports, bit, bit64, broom, clipr, crayon, forcats, hms, janeaustenr, progress, readr, SnowballC, tidytext, tokenizers, vroom), taking the install closure from 147 to 132 packages. There is no user-visible change in behaviour:

  • rlang and withr stay in Imports even though only rlang::sym()/syms() and withr::local_seed() are used. Both remain in the install closure via dplyr and ggplot2 regardless, so replacing them would churn hundreds of call sites without removing a single package.

  • glmnet and ranger stay in Imports. They are the engines hd_model_rreg() and hd_model_rf() fit through, so the package cannot work without them even though it never calls them directly.

  • extract_protein_list() (internal) no longer takes an UpSetR membership matrix; it derives the combinations from the feature lists directly.

Bug fixes

Differential expression

  • hd_de_limma() no longer fails with length of 'dimnames' [2] not equal to array extent when correct contains a categorical variable with more than two levels. Only the two group columns of the design matrix are renamed now.
  • hd_de_ttest() returned CI.L, CI.R and t as character strings, because each row was assembled with c() across mixed types. All statistics are numeric again.
  • hd_plot_de_summary() labelled every feature "up" in both proteins_df_up and proteins_df_down. The direction is now passed explicitly.
  • hd_plot_de_summary() and hd_plot_model_summary() no longer error when summarising a single analysis. UpSetR::fromList() degenerates to a named vector for one set; the UpSet plot is now skipped with a message and returned as NULL when fewer than two groups have features.

Dimensionality reduction

  • hd_plot_dim(plot_loadings = ...) drew every loading arrow on the diagonal, because the tip used the same loading value for both xend and yend. Arrows now run from the origin to the feature’s (x, y) loadings, and carry an arrowhead.
  • hd_pca() returned loadings for every component prcomp() produced rather than the components that were requested, so pca_loadings disagreed with pca_res and pca_variance.
  • The point layer used an invalid Color label, which made ggplot2 emit Ignoring unknown labels. hd_plot_model_summary() had the same problem: its metrics bar plot labelled color while mapping fill, so the legend title was dropped.

Clustering

  • hd_cluster(normalize = TRUE) left the first feature completely unscaled and discarded the sample identifiers. It moved the sample ID into row names before calling hd_normalize(), which then treated the first feature as the ID column. Normalisation now happens before the row names are set.
  • hd_cluster() returned the sample ID column as a factor; it is a character vector again.
  • hd_assess_clusters() produced duplicated rows in cluster_assessment whenever two clusters had the same size and stability, because the old and new cluster labels were matched on (n, Mean_ji). They are matched on the original cluster label now.

Enrichment

  • hd_gsea(ranked_by = "both") computed the combined logFC / significance score and then ranked by adj.P.Val anyway. It now ranks by the score it computed.
  • hd_gsea() renamed the ranked gene vector to ENTREZ identifiers by position. Because symbol mapping drops unmappable genes, every score after the first failure was attached to the wrong gene. Scores are now matched by symbol.
  • hd_gsea() is reproducible. GSEA p-values come from a permutation test, so every call returned a slightly different set of enriched terms. A new seed argument (123 by default, NULL to opt out) fixes the random number generator for the duration of the call.
  • hd_ora() and hd_gsea() no longer stop when nothing is enriched. They warn and return the empty enrichment object instead. A null result is a legitimate outcome, and aborting on it broke the vignettes and the examples in CI whenever the permutation test or a Bioconductor annotation update happened to push everything above the threshold. hd_plot_ora() and hd_plot_gsea() warn and return their input unchanged in that case, rather than failing inside clusterProfiler.
  • hd_ora() tested for missing adjusted p-values with is.na(any(p.adjust)), which coerces the numbers to logicals and never detects anything. An all-NA column now reports that nothing was enriched instead of failing with missing value where TRUE/FALSE needed.
  • The returned hd_enrichment object carries the pval_lim it was built with, so the plotting functions know which terms were considered significant.

Co-expression networks

  • hd_wgcna() passed NA straight to WGCNA::blockwiseModules() when pickSoftThreshold() found no power meeting the scale-free topology cut-off. It now falls back to the WGCNA default power of 6 and says so.
  • hd_plot_wgcna() errored when clinical_vars contained only continuous variables.
  • hd_plot_wgcna() bound the module eigengenes and the metadata side by side, silently pairing modules with the wrong samples whenever the metadata rows were in a different order. They are joined on the sample ID now.

Models

  • hd_model_rf() failed with The 'range' lower bound (2) must not exceed upper bound (1) on datasets with fewer than about nine predictors, because mtry was tuned over c(floor(sqrt(p)), floor(p / 3)). The range is clamped now.

Utilities and reporting

  • hd_import_data() returned one of the function’s own local variables instead of the stored object when reading .rda files, because it used ls()[1]. It uses the names returned by load() now.
  • hd_save_data() corrupted values containing a double-quote when writing TSV. utils::write.table() defaults to escaping an embedded quote as \", which no reader parses back, so he said "hi" reimported as he said \hi\. Quotes are doubled now, the way utils::write.csv() already did for CSV.
  • hd_qc_summary(), hd_impute_median(), hd_impute_knn() and hd_impute_missForest() printed tables as deparsed R code (c("f2", "f1")c(1, 1)). Tables are rendered the way they print at the console, truncated to the first ten rows.
  • hd_filter() and hd_save_path() produced messages with no spaces between the words and the values (Rows remaining:3, Directoryoutalready exists.).
  • hd_normalize() left scaled:center and scaled:scale attributes on every column of the returned tibble.

Performance

hd_qc_summary() ran out of memory on large datasets (10,000 features, 5,000 samples). The results are unchanged on every input; only the way they are computed differs.

  • The correlation heatmap is capped. hd_plot_cor_heatmap() and hd_qc_summary() gained max_heatmap_features (default 1000). Above the limit the correlation matrix and the reported pairs are still returned, but cor_heatmap is NULL and a warning explains why. Clustering 10,000 features needs a 400 MB distance matrix and hours of hclust, for a plot with 100 million unreadable cells.
  • The high-correlation pairs are read straight off the matrix. The pairs used to be found by reshaping the whole correlation matrix to long format first, which is one row per entry: 100 million rows before any filtering, several GB. The new cor_pairs_above() scans blocks of columns and keeps only what exceeds the threshold, in the same order as before.
  • calc_na_percentage_row() no longer uses rowwise(), which evaluated once per sample. Counting with rowSums() over blocks of columns is roughly 2000 times faster on a 1,000 × 3,000 dataset (21 s to 0.01 s) and uses a bounded amount of memory.
  • calc_na_percentage_col() counts a column at a time instead of building a one-row, 10,000-column summary and pivoting it.
  • hd_correlate() substitutes use = "everything" for "pairwise.complete.obs" when the input has no missing values. The two are equivalent in that case, and the pairwise code path is about three times slower because it compares every pair of columns separately.
  • check_numeric_columns() no longer coerces already-numeric columns with as.numeric(), which allocated a copy of every column of the dataset.
  • hd_qc_summary() now says up front when a correlation is going to take a while.

Continuous integration

  • Fixed .Rbuildignore: the patterns for inst/extdata, inst/cheatsheet and inst/hdanalyzer_app used \e, \c and \h instead of /, so they never matched. The source tarball was 49 MB; it is now 604 KB.
  • R-CMD-check also runs on dev/** and ka/** branches and on workflow_dispatch, not only on main. Pushes to working branches were never building.
  • Added bioc-version to the R setup so the Bioconductor dependencies resolve reliably, and oldrel-1 to the test matrix.
  • The R-devel job is now continue-on-error. Bioconductor routinely lags R-devel for weeks after a release, which used to fail the whole matrix for reasons unrelated to this package.
  • Added a job timeout and concurrency cancellation so superseded runs stop instead of queueing.
  • Fixed a malformed Suggests field in DESCRIPTION (a trailing comma) and a redirecting URL.
  • The hd_literature_search() example queried PubMed live during R CMD check, so a throttled request could stall the examples step on every platform. It is wrapped in \donttest{} now.
  • The hd_gsea() examples and the post_analysis vignette failed the examples and pkgdown steps with No significant terms found. GSEA is a permutation test, so whether anything cleared the threshold varied between machines and between runs. hd_gsea() is seeded and no longer treats an empty result as an error (see Enrichment above).
  • Every vignette declared a \VignetteIndexEntry{} that did not match its YAML title, so each build printed a warning about it. The index entries now carry the real titles.
  • Examples that need a Suggests package are now conditional. Twelve help topics ran examples that call into clusterProfiler, org.Hs.eg.db, enrichplot, WGCNA, ppsr, missForest, cluster, fpc, embed or easyPubMed unconditionally. Suggested packages are best-effort on CI — the Windows runner had no usable clusterProfiler, and R CMD check then failed with The 'clusterProfiler' package is required ... but is not installed while macOS and Ubuntu passed. Each of those topics now carries an @examplesIf requireNamespace(...) guard, so the examples run where the package is available and are skipped where it is not.
  • The pkgdown workflow survives, and reports, a crash in the article subprocess. pkgdown renders each article with callr::r_safe() and polls the subprocess every 200 ms until it exits. The site build died twice while polling the render of classification.Rmd, the longest article, inside processx’s own internals — chain_call() on one run, assert_that() on the next. pkgdown cannot format that condition because it carries no $stderr, so it reported subscript out of bounds from wrap_rmarkdown_error() and the real cause never reached the log. This is not in the package: every article renders cleanly in-process on the same runner, which has around 14 GB of memory and 79 GB of disk free. The workflow now:
    • installs the newest callr and processx rather than whichever version is cached, since that is where the crash happens;
    • builds with quiet = FALSE, so the subprocess output is streamed into the log and a genuine article error stays readable;
    • retries once with clean = FALSE, lazy = TRUE on failure, which resumes from the articles that already rendered instead of starting over;
    • reports the runner’s memory and disk, and re-renders every article in-process, both only when the build fails.
  • Added build and check artefacts (*.Rcheck/, *.tar.gz, Rplots.pdf) to .gitignore.

Testing

  • The test suite has been rewritten. It now checks behaviour and numeric results rather than only object shapes: differential expression is verified against stats::t.test, imputation and normalisation against hand-computed values, multiclass AUC against a one-vs-rest computation, and variable importance against the model coefficients.
  • Added end-to-end tests covering the full workflows on the shipped example data: QC to imputation to normalisation to PCA to differential expression to classification, plus the multi-class and multi-model summary paths.
  • Tests for optional packages are skipped rather than failed when the package is not installed.
  • The PubMed tests are opt-in. NCBI throttles unauthenticated clients and a throttled request can stall for minutes, which is exactly the kind of hang that makes CI look broken. Run them with HDANALYZER_TEST_PUBMED=true.
  • The end-to-end tests run on a five-disease, forty-assay subset of the example data. They cover the same code paths as the full dataset but the whole suite finishes in about a minute, so it is cheap enough to run on every push.

The suite is 15 files, 229 test_that() blocks and 608 assertions, passing with no failures.

R CMD check --as-cran now completes with 0 errors and 0 warnings. Two NOTEs remain and neither is a defect: “New submission” (the package is not on CRAN) and a handful of examples running longer than five seconds, which is inherent to fitting real models on the example data.

Internal

HDAnalyzeR 1.0.1 (development)

Changes

  • In multi-classification models the feature importance is now returned for each class separately.

Bugs

  • Fix bug in hd_filter().

HDAnalyzeR 1.0.0 (2025-11-17)

Release of HDAnalyzeR.