dawnn 2.2.0 (7 September 2026)

This release contains two major bugfixes, both of which can change results.

Bugfix in p-value computation. Previously, we were accidentally computing one-tailed p-values but reporting them as two-tailed. Now, we actually compute two-tailed p-values, as claimed. This will likely alter results.
Bugfix in Benjamini-Yekutieli procedure. It now steps up correctly, rather than stopping at the first p-value above its cutoff. This could alter results, but in our testing it does not seem to affect which cells Dawnn calls as in regions of DA whatsoever.

dawnn 2.1.1 (21 August 2026)

Documentation changes requested by CRAN: removed examples for unexported functions.

Mainly changes with a view to submitting to CRAN.

Model now hosted on Zenodo.
download_model() verifies the size and MD5 checksum of a model downloaded from the default URL, deleting it and failing if either does not match.
download_model() now saves the model to the user cache directory (tools::R_user_dir("dawnn", "cache")) by default. Other functions updated accordingly.
download_model() reports an unreachable URL with an informative message, and no longer leaves the timeout option modified if a download fails.
run_dawnn() no longer alters the global random number generator state.
The default verbosity of run_dawnn() is now 1 rather than 2, so the progress output of the underlying predict() calls is suppressed by default.
Added more sanity checks.
Vectorised p-value calculation.
Model now downloaded in binary mode, fixing corrupted downloads on Windows.
Added a vignette, a CITATION file, and package URLs.
R (>= 4.0.0) is now required, up from R (>= 3.5.0).

Simultaneously test for local and global differential abundance.
Only take single label from user (since two labels are assumed, the other need not be passed).

Fixed a bug where the alpha parameter was not being respected (the default value of 0.1 was always being used).