v1.4.5

Patch release focused on pandas 3 and Python 3.12/3.13 compatibility, plus a switch to
numpy.random.Generator.

Fixes

• pandas 3 Copy-on-Write compatibility. Under pandas 3 (mandatory Copy-on-Write),
  DataFrame.values returns a read-only view, so several in-place mutations raised
  ValueError: assignment destination is read-only. run_palantir failed in _differentiation_entropy
  (the terminal-state identity block, now built with np.eye(...)), and select_branch_cells failed on
  its NaN fill of the fate-probability array (now copied before mutation). The package now works under
  both pandas 2 and pandas 3. (#180)
• anndata ≥0.11 / Python 3.12–3.13 compatibility. cluster_gene_trends passed the dtype= keyword
  to AnnData(...), which anndata ≥0.11 (the version that resolves on Python 3.12/3.13) removed, raising
  TypeError: got an unexpected keyword argument 'dtype'. The array is now cast with .astype(np.float32)
  before construction; behaviour is identical and compatible with all anndata versions. (#180)

Changes

• Random number generation now uses numpy.random.Generator in place of the legacy global
  numpy.random.RandomState, and the NumPy lower bound is raised to numpy>=1.17. Because the default
  bit generator is now PCG64 rather than the Mersenne Twister, numerical outputs are expected to differ
  from previous versions. If you need to reproduce results from an earlier release exactly, pin that
  prior version. (#178)

Thanks

Thanks to @Angie92Usa for reporting the pandas 3 incompatibility in #180.

Full changelog: v1.4.4...v1.4.5