Skip to content

Releases

v0.2.1

Choose a tag to compare

@timtreis timtreis released this 14 Nov 21:49
f8fa70c

This release fixes the shape test (test="parallel") with tau > 0 when the cells cover only part of the anchor range (#17). In 0.2.0 it returned p = 1 for every gene, so pooled shape tests, where tau defaults to 0.1, called nothing.

The cause was an edge basis function with almost no weight on the anchor support, which made the tau bound explode. Shape directions under 0.1% of the largest eigenvalue are now dropped. On the organoid data the pooled shape test finds 44 genes again, against 0 in 0.2.0, and per-cell-type random-split fits with a false hit drop from 15% to 4%.

v0.2.0

Choose a tag to compare

@timtreis timtreis released this 13 Nov 20:20
283a7e5

The condition tests now use only the part of the anchor range where the cells are, and data with slides gets an optional two-stage test. Together they fix false positives in per-cell-type analyses with replicates (#16).

  • Level, slope, shape and Wald tests are computed over the central 96% of the cells, limited to where both conditions have cells. On SEA-AD MERFISH random donor splits, per-cell-type fits with a false hit drop from 32% to 8%.
  • two_stage=True computes the test from one curve per slide (t-test for level and slope, Hotelling's T^2 for shape). It stays calibrated when the cells cover only part of the anchor range: 0.9% of random-split fits with a hit. fit_model warns and points to it in that case.
  • The joint model stays the default. On human liver fibrosis (7 donors) it is calibrated on random splits and ranks the hepatocyte zonation genes at the top.

v0.1.0

Choose a tag to compare

@timtreis timtreis released this 12 Nov 17:20
d38eadf

First release of diffsvg, which compares how a gene's expression changes with the distance to an anchor (a region, cell type or set of coordinates) between two conditions.

  • Statistical tests:
    • per-gene p-values with FDR correction;
    • test="parallel" for shape differences, "level" for the overall gap, "slope" for steeper or flatter gradients, and "wald" for any difference;
    • slides, sections or donors modelled as replicates with bin_obs(..., sample=...).
  • Exact posterior inference for the Gaussian model, the default, so a panel fits in seconds without a sampler. PyMC is an optional extra (pip install 'diffsvg[pymc]'), only needed for non-normal priors.
  • Plotting: volcano plots, curves with credible bands, and p-value histograms.

Install with pip install diffsvg.