viprs 0.1.5

v0.1.5

on Python PyPI

Added

• Added a summary-statistics evaluation mode to viprs_evaluate, with support for independent GWAS
  summary statistics, LD reference panels, fitted VIPRS effect sizes, custom summary-statistic formats,
  and the Pseudo_Pearson_R and Pseudo_R2 metrics.
• Added a continuous benchmarking workflow that fits VIPRS and VIPRSMix across five folds, records
  prediction accuracy, runtime, and peak memory usage, and publishes CSV and Markdown benchmark histories.
• Added a reusable CLI benchmark utility and regression tests for benchmark result recording and aggregation.
• Added PGS Catalog scoring-file format 2.0 import and export, including metadata validation, compressed
  output, selectable effect columns, and viprs_fit --output-format pgs_catalog support.
• Added LDPredInfGrid, which tunes the LDpred-inf penalty using PUMAS summary-statistic splitting and
  pseudo-validation, then optionally refits the selected model on the full GWAS.
• Added allele-frequency-dependent priors to LDpred-inf through the alpha parameter, with optional joint
  search over alpha values and penalty factors in LDPredInfGrid.
• Added CLI documentation for viprs_fit, viprs_evaluate, and viprs_score, plus an FAQ covering required
  summary-statistic fields, allele conventions, genome builds, variant matching, and VIPRSMix selection.
• Added broader unit and integration coverage for VIPRSMix, E-step implementations, CLI argument handling,
  summary-statistics evaluation, PGS Catalog files, and LDpred-inf model fitting and selection.

Changed

• Reparameterized effect-size quantities during VIPRS optimization using a sample-size-based scale to improve
  numerical stability, while restoring parameters to their original scale after fitting and in public outputs.
• Simplified model selection in viprs_fit: --n-components 1 selects VIPRS, while values greater than one
  select VIPRSMix, removing the need for a separate model flag.
• Refactored viprs_evaluate into explicit individual-level and summary-statistics workflows with stricter
  input validation and support for multiple fitted models in one effect-size table.
• Updated VIPRSMix initialization, fixed-parameter handling, component precision updates, prior summaries,
  posterior moment calculations, and theta-table reporting to consistently support scalar, vector-valued,
  and chromosome-indexed parameters.
• Optimized mixed-precision E-step operations to avoid temporary converted arrays and skip negligible
  posterior-mean updates.
• Updated continuous benchmarks to download their input archives directly instead of depending on
  Hugging Face, and refreshed the recorded benchmark results.
• Reimplemented LDPredInf using standardized marginal effects, chromosome-wise MINRES solves, current
  magenpy LD interfaces, solver convergence checks, and scalar or chromosome-specific heritability.
• Extended Hugging Face LD-path handling to CLI preflight validation and validation LD panels.
• Included chromosome and position columns in inferred parameter tables to support portable scoring-file
  formats and position-based harmonization.
• Improved error logging when writing inferred parameters and hyperparameter tables.

Fixed

• Fixed VIPRSMix prior initialization and heritability accounting, including component precision ratios,
  mixture probabilities, null probabilities, average prior variance, and chromosome-specific parameters.
• Fixed VIPRSMix E-step calculations for the null component and regularization terms, including vector-valued
  lambda_min inputs and cached variational precisions.
• Fixed numerical precision loss in VIPRSMix posterior second-moment calculations by accumulating in
  double precision.
• Fixed LDpred-inf compatibility with current magenpy, use of unstandardized GWAS effects, unchecked solver
  termination, unsupported chromosome-specific heritability, and unnecessary genome-wide block matrices.
• Fixed Windows compatibility in CLI evaluation tests.