[parashardhapola]

HVG detection bounds were absolute (min_cells=20, max_cells=n-20), so subsets lost markers, for example Iapp, Pyy, Chgb, and Chga on 150 beta cells. This also adds batch-aware selection and the seurat_v3 and Pearson residual flavors.

Changes

• Default min_cells is 1% of the selected cells (between 3 and 20), and genes detected in more than 99.5% of the cells are excluded; the resolved bounds are recorded.
• Filters run before ranking, and the default blacklist covers current histone symbols and keeps CCN1 to CCN6.
• flavor="seurat_v3" (local regression matching scikit-misc to 1e-12) and flavor="pearson_residuals" (Lause et al. 2021), batch= (Seurat's integration-feature ranking), and PearsonResidualNormalizer.
• Against scanpy 1.12.4 on PBMC data: seurat_v3 shares 0.998 of its selection, Pearson residuals 1.0.
• features.hvgs without matplotlib warns and returns the selection.
• Revision 2 of select_hvgs for selections that record min_cells 20, a finite variance bound, or the earlier blacklist.

Behaviour changes

• Default selections on fewer than 2,000 cells change; stored selections with the earlier defaults are recomputed.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!