polars-mas

A fast Python library for multiple association studies, built on Polars.

polars-mas is designed as a drop-in replacement for the R PheWAS package, providing 3-14x speedup while producing numerically identical results (within 1e-8 tolerance).

Features

• Fast: Leverages Polars for efficient data processing with multi-threaded computation
• Accurate: Produces results identical to the R PheWAS package
• Flexible: Supports multiple regression models and input/output formats
• PheCode-aware: Built-in PheCode definitions and sex-specific code handling

Supported Models

Supported Formats

• Input: Parquet, CSV, TSV, TXT (tab-delimited)
• Output: Parquet, CSV, TSV, TXT

Installation

pip install polars-mas

Or with uv:

uv add polars-mas

Quick Start

Basic Usage

polars-mas \
  -i data.csv \
  -o results \
  -p exposure \
  -d i:10- \
  -c age,sex,bmi

This runs Firth logistic regression with:

• exposure as the predictor
• Columns 10 onwards as dependent variables (phecodes)
• age, sex, bmi as covariates

PheWAS Analysis

polars-mas \
  -i phewas_data.parquet \
  -o phewas_results \
  -p genetic_variant \
  -d i:20-1850 \
  -c age,sex,pc1,pc2,pc3 \
  --phewas \
  -n 4 \
  -t 8

The --phewas flag enables automatic PheCode annotation in results.

CLI Reference

polars-mas [OPTIONS]

Input Options:
  -i, --input PATH              Input file (parquet, csv, tsv, txt)
  -o, --output PATH             Output file prefix
  -p, --predictors COLS         Predictor columns (comma-separated)
  -d, --dependents COLS         Dependent columns (comma-separated)
  -c, --covariates COLS         Covariate columns (comma-separated)
  -cc, --categorical-covariates Categorical covariate columns
  -nv, --null-values VALUES     Values to treat as null (comma-separated)
  -ot, --output-type TYPE       Output format: parquet, csv, tsv, txt

Association Parameters:
  -m, --model MODEL             firth (default), logistic, or linear
  -mc, --min-case-count N       Minimum cases/controls required (default: 20)
  -mcv, --missing-covariate-values
                                How to handle missing covariates:
                                fail, drop, forward, backward, min, max, mean, zero, one
  -qt, --quantitative           Dependent variables are quantitative

Performance:
  -n, --num-workers N           Number of worker processes (default: 1)
  -t, --threads N               Threads per worker (default: 2)

PheCode Options:
  --phewas                      PheCodes are dependent variables
  --flipwas                     PheCodes are predictor variables
  --sex-col COL                 Sex column name (default: sex)
  --female-code N               Code for female (default: 1)
  --male-only                   Include only male samples
  --female-only                 Include only female samples

Other:
  --dry-run                     Show configuration without running
  -v, --verbose                 Enable verbose logging
  -q, --quiet                   Suppress most output

Column Selection

Columns can be specified by name or index:

# By name in comma-separated list
-p age,sex,bmi

# By index (0-based)
-d i:10          # Column 10
-d i:10-20       # Columns 10-19
-d i:10-         # Column 10 to end

# Can be used in conjuction as well!
-c age,sex,i:8-12

Output

Results include:

Performance

See the benchmarks for detailed comparisons against the R PheWAS package.

Summary: polars-mas achieves 3-14x speedup depending on the number of covariates, with identical numerical results.

Current Limitations

The following features from the R PheWAS package are not yet implemented:

• Multiple testing correction: Only Bonferroni correction is available (FDR/BH planned)
• Variable transformations: RINT and log transformations are not yet functional
• Covariate scaling: Standard and min-max scaling not yet implemented
• Parallel workers: Currently works best with one polars worker; -n flag reserved for future use. More threads with the -t option is fine.
• Python API: Only CLI interface is currently documented; programmatic API in development

Roadmap

Planned features for future releases:

• FDR (Benjamini-Hochberg) multiple testing correction
• Rank-based inverse normal transformation (RINT)
• Log transformation for dependent variables
• Covariate standardization options
• True parallel processing with multiple workers
• Python API with DataFrame input/output
• Manhattan and QQ plot generation
• ICD-to-PheCode mapping utilities

Requirements

• Python >= 3.11
• polars >= 1.9.0
• firthmodels >= 0.4.0
• statsmodels >= 0.14.4

License

MIT

Citation

If you use polars-mas in your research, please cite:

polars-mas: A fast Python library for multiple association studies
https://github.com/PheWAS/polars-mas

Related Projects

• PheWAS R Package - The original R implementation
• Polars - The underlying DataFrame library
• firthmodels - Firth logistic regression implementation