Skip to contents

Changelog

Source: NEWS.md

multideconv (development version)

New features

  • New aggregate_cell_groups(): sums cell types into user-defined groups (e.g. Myeloid cells = macrophages + monocytes + dendritic cells) within each method-signature combination, adding a feature <method>_<signature>_<group>. The original features are kept, combinations that already estimate the group are left as they are and combinations with fewer than min_types (default 2) cell types of the group are skipped. It prints which cell types were summed in each combination.
  • aggregate_cell_groups() is an optional step after compute.deconvolution(), independent of compute.deconvolution.analysis(). Its output is a deconvolution matrix as any other; group names that are not in the nomenclature (e.g. Lymphocytes) are given in cells_extra to compute.deconvolution.analysis() and compute.benchmark().

multideconv 0.2.0

Breaking changes

  • compute.deconvolution.analysis() now builds cell subgroups with complete-linkage hierarchical clustering of the feature correlations: features form a subgroup only if every pair of them correlates at least corr. Results no longer depend on the order of the columns.
  • Subgroups are named <CellType>_Subgroup.<i>: the .Iteration.<k> suffix was removed.
  • The pruning step was removed together with removeCorrelatedFeatures(): compute.deconvolution.analysis() and prepare_multideconv_folds() no longer have the prune_thr and seed arguments.
  • The output of compute.deconvolution.analysis() no longer contains “Discarded groups with equal method” (same-method subgroups are kept) nor “High correlated deconvolution groups (>0.9) per cell type”.
  • The low-variance filter now removes features whose coefficient of variation (sd / mean) is below cv_thr (default 0.1) instead of the 25% least variable features: var_quantile was replaced by cv_thr in compute.deconvolution.analysis() and prepare_multideconv_folds(). Rare cell types are no longer removed only because their values are small.
  • computeCBSX(), computeDWLS() and computeMOMF() are no longer exported. Second-generation methods are run with compute.deconvolution(sc_deconv = TRUE).

New features and improvements

  • New cell types: Dendritic.plasmacytoid.cells, Myeloid.cells, Basophils, Epithelial, Pericytes, Mural.cells and T.cells.proliferative. get_cell_type_nomenclature() is the single source of the cell type vocabulary.
  • A factor/character batch is adjusted with one indicator per batch, which is correct for 3 or more batches.
  • create_metacells() works with Seurat 4 and Seurat 5 objects and with a default assay other than “RNA”.
  • create_sc_pseudobulk() now sums the counts of each sample’s cells (previously the mean; identical after TPM normalisation), keeps sample names unchanged, saves a comma-separated CSV (previously tab-separated) and has a return argument to skip saving. cells_labels is no longer needed.
  • create_metacells() metacells are now the sum of the counts of their cells (integer counts; previously the average), and max_shared defaults to 10 (previously 15, which with k = 15 allowed metacells to overlap completely).
  • CIBERSORTx is retried with fixed input/output folders when the container fails with error code 139.
  • compute.deconvolution() skips CIBERSORTx with a warning when no credentials are given.
  • compute.subgroup.pathways() returns the correlations and p-values per cell type (in addition to the PDFs), has a corr_type argument (“pearson” or “spearman”), uses width/height when given, keeps pathway and feature names unchanged and stops with a clear error when sample names do not match.
  • create_sc_signatures() accepts method names in any case (and “CBSX” for “CIBERSORTx”), warns about unknown ones, replaces _ in name_signature and reports overwritten signature files.
  • compute.benchmark(): the scatter plots report one correlation per cell type instead of a single pooled correlation, and the heatmap labels show how many cell types each average is based on. For subgrouped input no “average” row is computed (a column such as Subgroup.1 holds unrelated features).

Bug fixes

multideconv 0.0.1

This is the first release version of multideconv! 🎉