Changes in version 0.99.3                        

Bug fixes

  - tweedspot() failed with mgcv 1.9-4 and later ("link "link" not
    available for tw family") because the Tweedie link was passed to
    mgcv::tw() as a variable. The log link is now passed directly, so
    the default Tweedie model, the examples, and the vignette run on
    current mgcv. Results are unchanged.

                       Changes in version 0.99.2                        

New features

  - Cell-type-specific testing with tweedspot(group = "cell_type"): each
    feature is tested for spatial variation within each group, so pure
    composition effects are not called. Group p-values are adjusted
    stage-wise by default (group_adjust); results are available through
    get_tweedspot_results(group = ) and top_spatial_genes(group = ).
  - tweedie_p = "auto" chooses the variance power per feature by BIC or
    AIC (p_criterion) among Gaussian, Tweedie, Gamma, and inverse
    Gaussian models.
  - size_factor_method = "spanorm" adds optional SpaNorm normalization,
    as size factors or as normalized counts (spanorm_output).
  - New seed argument to tweedspot() for reproducible results.
  - New plotting functions plot_feature() (observed, fitted, and spatial
    effect) and plot_top_features() (top n_top features), which can save
    publication-ready files to output_dir.

Changes

  - filter_genes() is renamed filter_features() with modality-agnostic
    arguments (min_total, min_pct_locations, min_locations, min_mean,
    min_var, drop_empty_locations). It adds named exclusions, a
    model-aware minimum, location QC, and action = "flag"; exclude_mito
    and exclude_ribo remain for transcriptomics.
  - plot_spatial_gene() and plot_spatial_fit() are replaced by
    plot_feature() and plot_top_features().
  - get_tweedspot_results() and top_spatial_genes() return a plain table
    with columns feature, feature_name, stat, pval, padj, edf, dev_expl,
    p, and family; rowData() keeps the tweedspot_ prefix.
  - The two-part model always combines its p-values with the Cauchy
    combination test; the combine argument is removed.
  - All families use a log link, and spatial coordinates are scaled by
    one common factor so the tissue's shape is kept.
  - Sparse and DelayedArray assays are supported without dense copies.

Bug fixes

  - tweedspot() reused one Tweedie family object across features, so
    each fit started from the previous feature's estimate and could get
    stuck at the lower bound, inflating false positives. Results from
    earlier versions with family = "tw" should be recomputed.
  - Stage-wise group adjustment now controls the false discovery rate at
    its target level.
  - Results from an earlier tweedspot() run are cleared before each new
    run, so ordinary and group-specific results cannot be mixed.
  - Mitochondrial and ribosomal gene matching no longer removes genes
    such as MTOR or the RPS6K kinases.
  - Clearer handling of unusual inputs (non-syntactic covariate names,
    missing values, constant or extra coordinates, duplicated row
    names).

Dependencies

  - MatrixGenerics and withr added to Imports; SpaNorm, patchwork,
    svglite, Matrix, and DelayedArray added to Suggests.

                       Changes in version 0.99.1                        

  - Added SCRAN-based size factor estimation through size_factor_method
    = "scran", while preserving total-count library-size scaling as
    size_factor_method = "library_size".
  - Added tweedie_p support for mgcv-compatible variance power settings:
    estimated Tweedie, Gaussian, Poisson, fixed Tweedie powers in
    (1, 2], and inverse Gaussian for strictly positive responses.
  - Removed Stouffer p-value combination from the two-part model
    interface.

                       Changes in version 0.99.0                        

  - Initial Bioconductor-oriented development release.
  - Added agnostic spatially variable gene detection with Tweedie GAMs.
  - Added gene filtering, result extraction, ranking helpers, and
    spatial plotters.