magenpy 0.2.5

v0.2.5

on Python PyPI

[0.2.5] - 2026-11-13

Critical correctness warning: recompute affected results. Earlier releases
that used the magenpy backend could silently produce incorrect results after
sample or ancestry subsetting.
Allele-frequency QC could use pooled samples or stale MAF/sample-count caches.
Separately, native BED-backed chromosome/block splits could use an incorrect
sample count to calculate BED offsets, read the wrong genotypes, and corrupt
LD matrices and LD scores.

If your workflow used sample subsets, review its outputs and rerun affected
frequency estimation, variant QC, LD matrices/LD scores, and dependent analyses
from the original unfiltered BED/FAM/BIM files with the intended sample lists.
Installing this update does not repair previously saved results.

Fixed

• Replace deprecated pandas inplace arguments in table manipulation with
  explicit assignments to avoid pandas 3.1 warnings and prepare for pandas 4.

• Store mgp_compute_ld --zip Zarr keys at the archive root so local and remote
  LD readers can open the same archive without a chromosome-folder mismatch.

• Apply sample selection, including phenotype/covariate-driven sample removal,
  before MAF/MAC filtering. Invalidate cached MAF and per-variant sample counts
  whenever the selected samples change.

• Preserve the full source FAM sample count through chromosome and variant/block
  splits, keeping native BED byte offsets independent of retained sample indices.

• Keep xarray genotype columns aligned with variant metadata during QC and
  harmonization; prevent repeated extraction IDs from duplicating variants and
  isolate sample-table changes across splits.

• Correct MAC cutoff rounding and exclude variants without observed genotypes
  when applying frequency/count QC.

• Implement cleanup for all genotype backends and ensure PLINK temporary
  directories are removed even when an operation fails.

• Compute native GWAS residual variance using each variant's observed phenotype
  subset, correcting standard errors when genotype calls are missing; use the
  finite-sample t distribution for native linear-GWAS p-values.

• Preserve PLINK scoring rows with opposing nonzero weights, and make missing
  calls contribute zero after genotype standardization.

• Allow PLINK block/sample LD estimation without genetic-map distances.

Changed

• Drop Python 3.10 support. Supported versions, CI test matrices, and wheel
  builds are limited to Python 3.11, 3.12, and 3.13. Exclude Python 3.14 from
  package metadata until the Zarr 2 / numcodecs dependency stack can be
  installed reliably across platforms.

• Standardize chromosome labels before splitting and matching across genotypes,
  summary statistics, annotations, and LD metadata. Encodings such as chr1,
  chr:1, chr_1, and 1 share a canonical ID. Compact categorical storage uses
  PLINK chromosome definitions, preserves additional contigs, and excludes unused
  categories from chromosome discovery and grouping.

• Consolidate chromosome and position-dtype normalization in summary-statistics
  post-processing, and simplify harmonization with shared planning and allele
  correction utilities, fewer table copies, and fewer LD-mask updates.

• Rename GenotypeMatrix.identifier_cols to variant_id_cols; callers using the
  former property should update their code.

Added

• Add continuous native speed/memory benchmarks for windowed/block LD, scoring,
  GWAS, and variant statistics, with base/current comparisons on the same runner
  and persistent CSV/Markdown history.
• Support optional mean imputation in native GWAS with impute_missing=True,
  retaining per-SNP exclusion of missing calls as the default. Regression sample
  counts, standardization, residual variance, and p-values follow the chosen mode.
• Add automatic safe Python fixes/formatting, C/C++ formatting, Cython quote
  normalization, and independent lint checks in GitHub Actions, with pinned tools
  and matching local make fix / make lint commands.
• Expand regression coverage with independently encoded BED fixtures for subset
  frequencies, QC, split offsets, scoring, GWAS, LD/LD scores, chromosome
  normalization, memory reuse, and cleanup.
• Require native-versus-PLINK parity tests on pushes, pull requests, and before
  PyPI uploads, with pinned PLINK downloads, fresh native builds, and zero skips.