magenpy 0.2.2

v0.2.2

on Python PyPI

[0.2.2] - 2026-09-25

Added

• Added compute_ld_sumstats_similarity for comparing GWAS summary statistics
  against candidate LD reference panels. It supports regional and genome-wide
  inputs, allele-frequency and regularized LD-likelihood scoring, automatic
  method selection, block-wise/submatrix LD loading, candidate priors, and
  temperature-scaled relative probabilities.
• Added HTTP(S) and Google Cloud Storage support for Zarr-backed LD matrices,
  including consolidated-metadata fallback, optional in-memory caching, and
  authenticated/custom storage options. Existing S3 and Hugging Face support
  now reports actionable installation guidance when dependencies are missing.
• Added a unified SNP harmonization plan that aligns genotype, summary
  statistics, LD, and annotation data in one pass while preserving canonical LD
  ordering and tracking allele flips with positional indexers.
• Added tabular, string, and notebook HTML summaries for GWADataLoader,
  covering sample and variant counts, chromosomes, genome build, available data
  sources, and LD in-memory state without loading matrix entries.
• Added backend-specific optional dependency groups for http, s3, gcs,
  and hf, an aggregate cloud extra, and a profiling extra for psutil.
• Added fits_in_memory for checking an allocation against currently available
  system memory.
• Added documentation for LD/reference similarity and optional cloud/profiling
  dependencies, plus a link to the magenpy LD browser.

Changed

• Reworked GWADataLoader.harmonize_data to materialize source metadata once
  per chromosome, calculate common variants once, and apply a single final LD
  mask. The new path handles differently ordered inputs, missing or duplicate
  identifiers, allele-incompatible variants, and summary-statistic effect/frequency
  flips consistently with the previous implementation.
• Improved summary-statistics parsing by assigning safe string dtypes to variant
  identifiers and alleles, preserving identifiers with leading zeroes, using
  32-bit or nullable 32-bit positions, respecting caller dtype overrides, and
  consolidating format-specific post-processing into the shared parse pipeline.
• Avoided unnecessary groupby and copying when parsed summary statistics are
  already chromosome-specific, both during parsing and when splitting
  SumstatsTable objects.
• Reduced package startup overhead by deferring Zarr, SciPy linear algebra,
  SciPy statistics, tqdm, and psutil until the corresponding functionality
  is used. Frequency-based LD similarity no longer imports SciPy solely for a
  softmax calculation.
• Made process profiling optional: general system utilities now work without
  psutil, and CPU availability uses the standard library.
• Changed LD-store validation to use direct key lookup instead of directory
  listing, allowing stores served by HTTP endpoints that do not support listing.
• Updated dependency metadata and lockfiles, including aiohttp, anyio,
  mkdocs-material, and pymdown-extensions, and synchronized package,
  container, and AI-declaration metadata for version 0.2.2.

Fixed

• Fixed harmonization when source SNPs use different orders, including repeated
  harmonization of already-masked LD matrices and correct propagation of allele
  flips to signed statistics and allele frequencies.
• Fixed PLINK 2 reference-allele inference and SAIGE sample-size inference so
  they occur before missing-value filtering in the common parser pipeline.
• Fixed parsing of missing, nonessential base-pair positions by using pandas'
  nullable Int32 dtype instead of failing during integer conversion.
• Fixed AnnotationMatrix.filter_snps on highly fragmented dataframes by
  consolidating the filtered table and preventing the old index from becoming
  an unintended column.
• Improved type validation for LD masks and corrected minor LD utility
  documentation.