Skip to contents

Changelog

Version 0.1.1

  • ✨ find_markers(): against argument for sibling-aware extraction ("nearest" = most-correlated type), balanced top = c(pos, neg) budgets (default c(10, 10)), and max_pct_out pan-lineage exclusion guard.
  • ✨ Add model-based scoring: train_weights(return_models = TRUE) persists per-type linear models (glmnet/lr) and hitype_score_models() scores cells with the model equations (linear predictors) and a margin-based Unknown threshold.
  • feat: enhance train_weights functions to support class-specific markers and update documentation

Version 0.1.0

  • 🐛 Fix RunHitype() with a metadata-column ident (cluster-level assignments) assigning every cluster the same arbitrary cell type: the column was taken as an unnamed vector, so hitype_assign() could not map the score columns back to cells, every cluster scored all-zero, and the ties collapsed to one cell type. The cluster vector is now named from the cell names (as the Idents() branch already was).

Version 0.0.8

  • ✨ train_weights() (and compile_weights()) gain a pos_only argument (default FALSE): when TRUE, markers whose trained weight is not positive are dropped from the output, leaving only the markers overexpressed in each cell type (no direction = "negative" rows).
  • 🐛 Fix train_weights() silently returning the rescale midpoint as all weights when a marker gene has no variance across the cells (e.g. an unexpressed gene that still exists in the expression matrix): after scaling, its column is all-NA, every supervised fit fails with “x has missing values”, and each tryCatch-wrapped fit is silently swallowed (all-zero coefficients -> uniform midpoint weights, e.g. 3 for range = c(1, 5)). prepare_data_for_training() now drops zero-variance markers (with a warning) from both the training matrix and the gene sets, so they never reach the model or the compiled weights.
  • 🐛 Fix hitype_score() collapsing to a zero-row score matrix when a marker gene has no variance across the given (e.g. small held-out test) cells: its scaled row is all-NA, NA propagates to every cell type sharing the marker, all-NA rows are removed, and downstream hitype_assign() crashes with “arguments imply differing number of rows”. Zero-variance genes now score as 0.
  • 🐛 Fix the on-test assignment printed by train_weights() (with run_weights_on_test) being all-zero when every cell type shares the same marker pool: the scores were computed with the default use_sensitivity = TRUE, and markers present in every gene set get a sensitivity of 0, zeroing the whole score matrix. On-test scoring now uses use_sensitivity = FALSE (as recommended for scoring with learned weights, per hitype_score()).
  • ✨ compile_weights() drops exactly-zero trained weights from the output by default (drop_zero = TRUE, new argument on train_weights()), so zero means “no direction”: markers the penalty zeroed out are absent from the cell type’s list. Pass drop_zero = FALSE to keep every candidate marker (with format = "db" they are then encoded as *).
  • ✨ The universal output of train_weights()/compile_weights() (the default format) now keeps the raw trained weights as-is instead of range-rescaling them. The rescale was pure distortion — it erased the sign of genuinely negative markers under a positive-only range like c(1, 5) and injected a per-cell-type offset proportional to marker count — and hitype_score() consumes only signed raw weights anyway. range now applies to format = "db" only: it must be the 4-element sign-split form c(-low, -high, low, high) (default c(-5, -1, 1, 5)), which scales positive weights into [low, high] and negative ones into [-high, -low] before rounding, so all-positive training data cannot produce negative db markers. The db weights are integer-encoded so they round-trip exactly through gs_prepare(): a weight of w markers as gene + "+" × (w − 1), a weight of −w as gene + "-" × w (the decoder counts positive repeats from 1).

Version 0.0.7

  • ✨ train_weights() now derives the training cell types from the data instead of the marker file. With a Seurat exprs, clusters can be NULL (default, using Seurat::Idents()) or a column name in the meta.data; with a matrix, it stays a named per-cell vector. A cell type in the marker file with no cells of that type in the data is ignored (with a warning), and its markers are pooled for the cell types in the data that are not covered by the marker file, so those cell types can still be trained. No cells are dropped from the training data, and the returned weights cover only the cell types present in the data.

  • 🐛 Fix train_weights() failing on Seurat input (exprs as a Seurat object): clusters were converted to integer codes while data$clusters stayed a factor, so the per-method clusters == ct comparisons (against unique(data$clusters)) were all-FALSE, every supervised fit errored inside a tryCatch and was silently swallowed, and all trained weights came out as the rescale midpoint (e.g. uniform 3 for the default range = c(1, 5)). Clusters now stay as character cell-type labels throughout, which compile_weights() also needs to match output_node against the gene-set cell-type names.

Version 0.0.6

  • ✨ gs_prepare() now auto-detects the universal marker format (a long table with cell_type and gene columns, shared with biopipen’s CellTypeAnnotation) in addition to the native wide ScType-style format. Column aliases are matched case-insensitively (celltype/type → cell_type; marker/gene_symbol → gene; sign → direction; tissueType → tissue). Optional columns: direction (positive/negative, aliases pos/neg/+/-), weight (numeric; positive markers get abs(weight), negative markers get -abs(weight), signed weights are used as-is when direction is missing), tissue (filtered by tissue_type), and level. Marker files can be .txt/.tsv, .csv, .xlsx, .rds, or .qs/.qs2 (the latter only when the qs/qs2 package is installed).
  • ✨ train_weights() and find_markers() gain a format argument ("universal" by default, "db" for the legacy wide format), and so does the internal compile_weights(). The default universal output carries exact numeric weights — the legacy db format’s weight-suffix magnitude shift no longer applies to the default flow, including run_weights_on_test_data.

Version 0.0.5

  • ✨ Add find_markers() to discover marker genes from your data with dependency-light fc (default), Seurat, or presto backends, directly producing a db data.frame consumable by gs_prepare().
  • ✨ Add method argument to train_weights() with 7 weight-learning backends: uniform, correlation, lr, glmnet (default, recommended), rf, xgb, lrp. epochs and batch_size now apply to the lrp method only.
  • ✨ Add cv_folds argument for cross-validated weight estimation: stratified folds, per-fold weights averaged for stability. Used by lr, glmnet, and lrp.
  • ✨ Add norm and use_sensitivity arguments to hitype_score() and hitype_score_level():
    • norm = "weight" normalizes scores by sum(abs(weights)) (fairer across cell types with different marker numbers);
    • use_sensitivity = FALSE disables marker-sensitivity weighting — recommended when scoring with learned weights to avoid double-penalizing shared markers.
  • 🔧 Statistical soundness: removed input masking in prepare_data_for_training() (masked zeros let the model learn the mask pattern instead of expression).
  • 🔧 Seurat v5 compatibility: use Seurat::GetAssayData(x, layer = "data") instead of direct @assays$RNA@data / @assays$RNA@scale.data access.
  • 🧩 New suggested dependencies: glmnet, ranger, xgboost (required per method, optional overall).

Version 0.0.4

  • ✨ Add weight_encoding argument to gs_prepare

Version 0.0.3

  • Move keras and innsight as suggested

Version 0.0.2

  • Add logo
  • Reduce keras requirement to 2.11

Version 0.0.1

  • Add make_unique to RunHitype

Version 0.0.0

  • Initial release