viprs 0.1.6

v0.1.6

on Python PyPI

[0.1.6] - 2026-10-28

Added

• Added a browser application that runs the real VIPRS and VIPRSMix inference code locally in a Pyodide
  Web Worker, supports local and hosted summary statistics and LD references, and exports posterior tables
  and detailed JSON run reports without uploading user data to a server.
• Added browser input support for files, folders, URL and glob patterns, hosted UK Biobank LD panels,
  GWAS Catalog accessions, and multi-chromosome analyses, with chromosome-aware ordering, secure ZIP
  extraction, input validation, progress reporting, cancellation, and a bundled chromosome 22 example.
• Added a reproducible WebAssembly build and packaging toolchain for VIPRS and magenpy, including SIMD-enabled
  Cython/C++ kernels, pinned runtime dependencies, wheel and source checksums, build provenance, release-asset
  verification and reuse, and source-build fallback when a compatible magenpy wheel is unavailable.
• Added continuous WebAssembly builds, browser integration tests, release archives, and documentation deployment
  that publishes only a tested application bundle from the same source commit.
• Added comprehensive tests for browser input resolution and streaming, WebAssembly artifacts and release
  selection, archive safety, multi-chromosome inference, user-interface behavior, and parity with native VIPRS.

Changed

• Optimized the VIPRSMix E-step by accumulating component responsibilities, posterior moments, and ELBO terms
  inside the native kernel, using thread-local double-precision reductions and cached sufficient statistics to
  avoid repeated genome-wide scans during M-step and ELBO calculations.
• Added WebAssembly SIMD implementations of the inference kernel's dot-product and update operations while
  preserving the existing native and OpenMP execution paths.
• Updated viprs_fit to obtain all supported summary-statistics formats and aliases from magenpy's parser
  registry instead of maintaining a fixed list, and raised the minimum magenpy version to 0.2.3.
• Centralized GRCh37 GWAS Catalog accession resolution in data_utils and made accessions valid
  summary-statistics inputs in both viprs_fit and the browser application.
• Improved the browser application's performance and memory use by range-streaming hosted Hugging Face LD,
  downloading large range-capable GWAS Catalog files in parallel, parsing summary statistics once, projecting
  large inputs to model-relevant columns with compact numeric storage and categorical alleles, releasing staged
  summary files after parsing, and releasing chromosome data and local LD archives incrementally.
• Normalized browser GWAS-SSF inputs to magenpy's canonical schema before harmonization, including support
  for alternate rsID and odds-ratio headers and LD-based SNP inference when optional rsIDs are absent.
• Improved the browser interface with clearer source selection, model configuration and prior visualization,
  runtime and convergence status, timing summaries, result filtering, download controls, responsive styling,
  and more actionable validation and network errors.
• Refreshed continuous benchmark results for the release changes.

Fixed

• Fixed summary-statistics train/validation splitting to use the sample sizes returned by magenpy, avoid
  compounding sample-size reductions across repeated splits, preserve mixture-model dimensions, and keep the
  model-wide sample size synchronized when restoring or replacing inputs.
• Fixed viprs_fit to use --validation-sumstats-format when reading validation summary statistics rather
  than reusing the training-data format.
• Fixed VIPRSMix parameter updates to preserve the configured floating-point precision and skipped negligible
  LD updates in mixture and grid E-step kernels.
• Fixed browser and WebAssembly build behavior across Windows and POSIX path conventions, along with runtime,
  URL-resolution, and browser-test failures discovered during application development.