v2.1.0

Removed

• Breaking: li.mt.lric(pair_chunk=...) is removed, after being deprecated and ignored in 2.0.0.
• Breaking: li.mt.lric(transform_fn=...) is removed. Its default, mean-normalising each gene to 1, was redundant: after introducing analytical normalisation for the area of interest, any per-gene rescaling cancels exactly in every LRIC ratio (g, g_expr). To weight by a nonlinear transform of expression, store it as a layer and pass it -- e.g. adata.layers["sqrt"] = np.sqrt(adata.X), then li.mt.lric(..., layer="sqrt"). Unlike transform_fn, a layer goes through the non-negativity check.

Changed

• Permutation-based methods now reject n_perms values that are 0, negative or not integers with a clear ValueError; None skips permutations.
• The permutation statistics behind cellphonedb, cellchat, geometric_mean and rank_aggregate p-values are gathered and reduced a block of permutations at a time instead of as one (2, n_perms, n_interactions) tensor, so their peak memory no longer scales with n_perms (#282, by @pwwang). P-values are bit-identical for a given seed. Peak RSS / wall time, cellphonedb, MacBook Pro M3, 24 GB: 462k interactions, 1000 permutations 14.7 GB / 27 s -> 1.6 GB / 9 s; 4.2M interactions, 300 permutations 15.2 GB (memory-bound) / 510 s -> 3.7 GB / 35 s. Up to ~15% slower when the whole tensor fits in memory. Internally, the pipeline now owns the permutation test: a permutation-based method only declares how it combines ligand and receptor statistics (MethodMeta(combine=...)), plus an optional observed-only adjustment (observed_fn=, CellPhoneDB's zero mask), instead of each method re-implementing proximity weighting and p-value counting.

Fixed

• Explicit zeros stored in a sparse X no longer count as expressing cells in ligand_props/receptor_props (and thus expr_prop). li.pp.zi_minmax now prunes its sub-cutoff zeros, so with it as x_transform/y_transform far fewer interactions pass expr_prop.
• MISTy raises if a view's obs_names differ from the intra-view's; views were paired by row position, so a reordered view gave silently wrong results.
• li.mt.lric/li.mt.cross_pcf: pairs exactly on a radius-tile edge (e.g. gridded spots with radius_step equal to the spacing) no longer drift into the neighbouring tile; the number of annuli is floor(max_radius / radius_step), so float steps no longer add a spurious annulus (e.g. 0.3 / 0.1) and max_radius < radius_step raises.
• li.mt.get_lric_auc normalises by the radial width it actually integrates, so an empty interior bin no longer pulls the score toward 0.
• Docs build with sphinx-autodoc-typehints>=3.17 (the alias_typehints shim is removed); README and decision-tree links point to liana.readthedocs.io (#283).
• The pyCrossTalkeR tutorial runs in CI again; the tutorials extra requires pycrosstalker>=2.1.8 (#271).