NEWS.md
In presto: Fast Functions for Differential Expression Using Wilcoxon and AUC

presto 1.1.0

First CRAN release. Version 1.0.0 refers to the pre-CRAN package installed from GitHub.

wilcoxauc() accepts disk-backed DelayedMatrix input (e.g. HDF5-backed assays from HDF5Array), processed in feature blocks so the whole matrix never has to be loaded in memory (#26). Other matrix-like classes with an as(., "dgCMatrix") method (e.g. BPCells) are coerced automatically.
wilcoxauc() gains a transposed argument for observations x features input (samples in rows), which runs directly on that layout without materializing a transposed copy (#18).
wilcoxauc() gains an nthreads argument: the per-feature ranking of sparse (dgCMatrix) input can run multithreaded. Results are identical at any thread count; roughly 2.6x faster with 4 threads on a 3,000 x 80,000 matrix. The default (nthreads = 1) is unchanged.
New toy_seurat() and toy_sce() generate small example Seurat / SingleCellExperiment objects on the fly with the installed package versions, so they can never go stale.
Two vignettes: Getting started with presto and Pseudobulk differential expression with DESeq2, both worked through on a real 25,341-cell colon CD8 T-cell dataset (GSE206299, Thomas et al. 2024).

wilcoxauc() now stops with an informative error when X contains NA values, instead of silently returning incorrect results (#25).
The shipped object_seurat and object_sce datasets were removed; use toy_seurat() and toy_sce() instead.
Rcpp and data.table moved from Depends to Imports, so they are no longer attached to the search path when presto is loaded.

The Wilcoxon variance correction now includes every tie group: the final tie group per feature and the implicit-zero group of all-zero features used to be dropped, inflating p-values on heavily tied data (#29). P-values now agree with stats::wilcox.test() to machine precision; on typical count data the change is tiny (< 1e-4), but it is meaningful for binary or heavily tied features. Constant features report pval = 1 instead of an arbitrary value.
wilcoxauc() and rank_matrix() no longer modify their input in place: dense input used to be silently overwritten with ranks via an aliased memory buffer (#7). Dense avgExpr and logFC are computed from the original values (also #7).
The Seurat dispatcher of wilcoxauc() uses the layer argument required by Seurat 5 (#44).
pseudobulk_within() handles character and multi-level contrast variables.
pseudobulk_pairwise() no longer fails on single-column meta_data.
collapse_counts() no longer drops a single-column meta_data to a vector, which had broken the collapse_background = TRUE path of pseudobulk_deseq2().
sumGroups() and nnzeroGroups() accept a character group vector y again. A character y previously coerced to NA via as.integer(), producing an out-of-bounds index that crashed on some platforms; the labels are now factored first, and NA labels raise a clear error.
summarize_dge_pairs() no longer prints debug output, and pseudobulk_deseq2() no longer warns unconditionally.
Replaced deprecated .data$ tidyselect usage and superseded tidyr::spread() with tidyr::pivot_wider() in top_markers() and top_markers_dds() (#20).

The sparse-input statistics of wilcoxauc() (transpose, ranking, and all group-wise reductions) are fused into a single-pass C++ kernel, about 12% faster serially before any threading.