Changelog#

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased#

0.3.0 - 2026-11-16#

Added#

  • tl.gene_set_test and tl.replicate_summaries: gene-set tests with replicates, as a rotation test (default) or with decoupler’s univariate linear model; gene sets can be lists or decoupler networks (#18)

  • A separate replicate scale for the slope direction of each replicate’s curve (sample_slope_sd), and the columns level_eff_df, slope_eff_df and shape_eff_df (#19)

  • datasets.legnini23: the Legnini et al. organoid data, downloaded from the scverse example-data CDN and pinned by sha256

  • Logging through the shared scverse logger (scverse.diffsvg): fit_model reports its choices and run time, and bin_obs and gene_set_test report what they leave out; see the logging page (#21)

  • Docs: “diffsvg in 10 minutes”, and guides for comparing conditions with replicates, choosing settings and clustering curves, and data without replicates; the method page has a validation section

Changed#

  • Warnings point at the caller’s line, and predictive_seed warns with a FutureWarning through scverse_misc.deprecated_arg (#21)

  • The docs are built with scverse-doc (#21)

  • Docs: full_tutorial is now settings_and_clustering, single_slide_test is now without_replicates, and the test-comparison notebook is replaced by the Validation section of the method page

  • With replicates, the shape test at tau = 0 uses F with max(n_replicates - 2, shape_eff_df) denominator degrees of freedom instead of n_replicates - 2 (#19)

  • Exact fits with replicates take about 4 times longer; sample_kwargs={"slope_scale": False} gives the previous model

0.2.1 - 2026-11-14#

Fixed#

  • The shape test with tau > 0 returned p = 1 for every gene when the cells covered only part of the anchor range (#17)

0.2.0 - 2026-11-13#

Added#

  • fit_model(..., two_stage=True): tests from one curve per slide (t-test for level and slope, Hotelling’s T² for shape) (#16)

Changed#

  • The condition tests on contiguous bins use only the part of the anchor range where both conditions have cells; with replicates and two_stage=False, fit_model warns when that part covers less than 75% of the range (#16)

Removed#

  • The inference="advi" alias (#16)

0.1.0 - 2026-11-12#

Added#

  • Binning (pp.bin_obs), model setup and fitting (tl.setup_model, tl.fit_model), profiles and one-vs-rest tests (tl.fit_profiles, tl.one_vs_rest), stability checks (tl.check_stability) and plots (pl.plot_volcano, pl.plot_prediction_curve, pl.plot_pvalue_histogram)

  • tl.cluster_curves and pl.plot_cluster_curves: k-means on the coefficients of the combined curve or the interaction, which fit_model and fit_profiles now store

  • Exact posterior inference (inference="auto", the default); PyMC, nutpie and arviz moved to the optional pymc extra (#8)

  • test="level" and test="slope", and the Wald test with slides (#10)

Changed#

  • The default test is the shape (parallel curves) test on contiguous bins weighted by their cell counts (#7)

  • tau defaults to 0 when slides are modelled (#10)

  • tl.one_vs_rest tests every pair of genes with the stored posterior covariance (#9)

  • pl.plot_prediction_curve and pl.plot_cluster_curves draw with matplotlib and return axes; holoviews is no longer a dependency

Removed#

  • ADVI inference; the exact posterior covers the same models (#9)