PANICLE: Python Algorithms for Nucleotide-phenotype Inference and Chromosome-wide Locus Evaluation

PANICLE finds the DNA variants driving traits — crop yield, disease risk, plant height, anything you can measure — by scanning millions of genetic markers across hundreds or thousands of individuals. It's a fast, modern Python implementation of the four most-used GWAS algorithms (GLM, MLM, FarmCPU, BLINK), built for researchers who want to run GWAS inside their Python pipelines instead of bouncing through R or standalone binaries.

Manhattan-style multi-method comparison on demo-scale data. The 30-second quick start below writes per-method Manhattan and QQ plots for your trait.

Try It in 30 Seconds

The repo ships with a small demo dataset — install, clone for the examples, and run:

pip install panicle
git clone https://github.com/jschnable/PANICLE.git && cd PANICLE
panicle-gwas \
  --phenotype examples/example_phenotypes.csv \
  --genotype  examples/example_genotypes.vcf.gz \
  --traits PlantHeight \
  --methods GLM \
  --outputdir ./results

Equivalent forms: python -m panicle ... or python scripts/run_GWAS.py ... (compatibility wrapper).

Open results/ and you'll find a Manhattan plot, a Q-Q plot, and a CSV of significant markers. That's a full GWAS, from raw VCF to publishable figure, with no setup beyond pip install.

Why PANICLE

• Fast. 5.75M markers × 862 samples in ~9 seconds for GLM, ~28s for MLM on a laptop (benchmarks below).
• Four algorithms, one interface. GLM for speed, MLM for population structure, FarmCPU and BLINK for resolving independent loci — pick any combination in one call.
• Reads what you have. VCF/BCF, PLINK, HapMap, or numeric CSV/TSV. First-run parsing is cached to disk so re-analyses load in seconds.
• Quality-of-life features other implementations lack. Effective marker number for less-conservative Bonferroni thresholds, leave-one-chromosome-out MLM by default, FarmCPU resampling with model inclusion probabilities, optional standard errors.
• Native Python. No R bridge, no shelling out to binaries — drop GWAS into pandas/Jupyter/Snakemake pipelines directly.

Installation

Requires Python 3.9+.

pip install panicle

With optional dependencies for PLINK format support:

pip install panicle[plink]

Or install all optional dependencies:

pip install panicle[all]

Development Installation

To install from source for development:

git clone https://github.com/jschnable/PANICLE.git
cd PANICLE
pip install -e .[all]

Python API

The same workflow as the 30-second quick start, in Python:

from panicle import PANICLE

results = PANICLE(
    phe="examples/example_phenotypes.csv",
    geno="examples/example_genotypes.vcf.gz",
    map_data=None,                # VCF carries its own map; pass a path for numeric CSV/TSV
    n_pcs=3,
    method=["GLM", "MLM", "FarmCPU"],
)
# Results are also saved to CSV files automatically.

For finer control (custom alignment, looping over traits, mlm_mode), see the API Reference and the GWASPipeline class below.

What You Get

Each run writes to --outputdir (default ./GWAS_results/):

• GWAS_<trait>_all_results.csv — every marker, with p-values and effect sizes per method:

  MARKER	CHROM	POS	REF	ALT	MAF	GLM_P	GLM_Effect
  Chr09_4810793	Chr09	4810793	G	A	0.21	3.1e-19	0.412
  Chr06_2110438	Chr06	2110438	C	T	0.14	8.4e-08	-0.047
  …	…	…	…	…	…	…	…

  (With multiple --methods, each adds its own <METHOD>_P and <METHOD>_Effect columns.)

• GWAS_<trait>_significant.csv — only markers passing the significance threshold

• GWAS_<trait>_<METHOD>_manhattan.png — Manhattan plot per method

• GWAS_<trait>_<METHOD>_qq.png — Q-Q plot for diagnostic checking

• GWAS_summary_by_traits_methods.csv — one-row-per-(trait, method) summary across the whole run

CLI Reference

After pip install panicle, the primary entry point is panicle-gwas (also python -m panicle). A complete invocation:

panicle-gwas \
  --phenotype examples/example_phenotypes.csv \
  --genotype  examples/example_genotypes.vcf.gz \
  --traits PlantHeight \
  --methods GLM,MLM,FarmCPU,BLINK \
  --n-pcs 5 \
  --compute-effective-tests \
  --outputs manhattan qq significant_marker_pvalues \
  --outputdir ./results

panicle-gwas --version   # e.g. panicle-gwas 0.4.0
panicle-gwas --help

Related tool: panicle-cache-genotype pre-converts large genotype files for faster reloads.

Parameters

Other useful filters:

• --max-missing (default 1.0), --min-maf (default 0.0)
• --drop-monomorphic / --keep-monomorphic (drop monomorphic is on by default)
• --snps-only, --no-split-multiallelic

Pipeline API (step-by-step control)

For multi-step workflows in scripts or notebooks, use the GWASPipeline class — it exposes loading, alignment, structure, and analysis as separate steps you can inspect between.

from panicle.pipelines.gwas import GWASPipeline

pipeline = GWASPipeline(output_dir="./results")

pipeline.load_data(
    phenotype_file="examples/example_phenotypes.csv",
    genotype_file="examples/example_genotypes.vcf.gz",
    trait_columns=["PlantHeight"],
    loader_kwargs={"compute_effective_tests": True},
)
pipeline.align_samples()
pipeline.compute_population_structure(n_pcs=5)
pipeline.run_analysis(methods=["GLM", "MLM", "FARMCPU", "BLINK"], alpha=0.05)

Input Formats

Phenotype & Covariates

CSV or TSV files with an ID column and numeric columns for traits/covariates. PANICLE auto-detects ID columns named ID, id, IID, sample, Sample, Taxa, taxa, Genotype, genotype, Accession, accession (if multiple, it uses the leftmost). If none match, it uses the first column. Use --phenotype-id-column (or --covariate-id-column) to specify a custom ID column name.

Genotype

• VCF/BCF: .vcf, .vcf.gz, .bcf (Preferred for performance).
• CSV/TSV: Numeric matrix (rows=samples, cols=markers) + genetic map file with MARKER, CHROM, and POS columns (legacy SNP and aliases like Chr, Pos are accepted).
• PLINK: .bed + .bim + .fam.
• HapMap: .hmp.txt.

Performance notes: VCF is typically the slowest format on the first run, but PANICLE caches parsed marker data so subsequent loads are competitive with other formats. BCF is roughly ~2x faster than VCF on the first run, and PLINK/bed is roughly ~4x faster than VCF on the first run (exact speedups depend on marker count, sample size, and hardware).

Tips

1. Effective Tests: Use --compute-effective-tests to calculate a less stringent, more accurate Bonferroni threshold based on marker linkage (Me).
2. Genotype Subsetting: If you align or filter samples manually, use GenotypeMatrix.subset_individuals(...) to preserve pre-imputed fast paths.

Documentation & Examples

Documentation

Detailed documentation is available in the docs/ directory:

• Quick Start Guide - Get up and running in 5 minutes
• API Reference - Complete API documentation for all functions and classes
• Output Files - Understanding result file formats and columns

Interactive Tutorials

• GWAS Tutorial notebook — End-to-end GWASPipeline on the bundled demo VCF/phenotypes (LOCO MLM, FarmCPU, BLINK).
• eQTL Multi-Trait LOCO MLM Acceleration — Synthetic multi-trait LOCO: per-trait vs PANICLE_MLM_LOCO_MULTI speedup and hit table.

Example Scripts

The examples/ directory contains runnable example scripts with included test data:

Run any example:

cd examples
python 01_basic_gwas.py

Algorithms

GLM

General Linear Model for fast single-marker association testing. Uses the Frisch-Waugh-Lovell (FWL) theorem combined with QR decomposition for computational efficiency. The algorithm residualizes the phenotype and genotypes against the covariate matrix (PCs + intercept), then computes per-marker regression statistics in vectorized batches. GLM is the fastest GWAS method but may generate overly optimistic significance values.

MLM

Mixed Linear Model accounting for population structure and cryptic relatedness via a kinship matrix.

Key design decisions:

• LOCO by default: Leave-One-Chromosome-Out kinship avoids proximal contamination (testing a marker against a kinship matrix that includes that marker), increasing power to detect true associations.
• Eigenspace transformation: Data is transformed via eigendecomposition of the kinship matrix, converting the correlated mixed model into an equivalent weighted least squares problem.
• REML variance components: Heritability (h²) is estimated using Brent's method optimization of the REML likelihood.

When map data is available, PANICLE's pipeline MLM path uses LOCO kinship and applies exact LRT refinement to top hits by default. LRT re-estimates variance components per marker, with a GEMMA-inspired derivative solver available for faster exact refinement versus the legacy bounded-Brent optimizer.

FarmCPU

Fixed and random model Circulating Probability Unification. FarmCPU iteratively alternates between a fixed-effect model (GLM) and random-effect model to identify associated markers while controlling for polygenic background. FarmCPU can often detect more independent loci linked to variation in the same trait since it controls for the impact of each significant signal when determining the significance of other signals.

This means FarmCPU will NOT give the "towers" most of us expect from classical manhattan plots which are the result of many different markers in LD with the same causal variant. Instead it will identify only one marker since once the effect of this marker is controlled for the significance of any markers in LD with that marker decline to baseline levels.

FarmCPU Citation: Liu, X., Huang, M., Fan, B., Buckler, E. S., & Zhang, Z. (2016). Iterative usage of fixed and random effect models for powerful and efficient genome-wide association studies. PLoS genetics, 12(2), e1005767.

BLINK

Bayesian-information and Linkage-disequilibrium Iteratively Nested Keyway. BLINK builds on FarmCPU's iterative framework but uses BIC-based model selection to optimize the pseudo-QTN set. Like FarmCPU, BLINK can often identify larger numbers of independent causal variants from the same phenotype/genotype set than GLM or MLM. Like FarmCPU, it will typically identify only one significant marker per causal variant and lacks the expected "towers" in manhattan plots caused by groups of markers that are all in LD.

Blink Citation: Huang, M., Liu, X., Zhou, Y., Summers, R. M., & Zhang, Z. (2019). BLINK: a package for the next level of genome-wide association studies with both individuals and markers in the millions. Gigascience, 8(2), giy154.

Effective Marker Number Estimates

PANICLE includes a python-based based implementation of the effective marker number estimation method implemented in GEC. Accounts for linkage disequilibrium between markers to provide a less conservative multiple testing correction than standard Bonferroni.

GEC citation: Li MX, Yeung JM, Cherny SS, Sham PC. Evaluating the effective numbers of independent tests and significant p-value thresholds in commercial genotyping arrays and public imputation reference datasets. Hum Genet. 2012 May;131(5):747-56.

Benchmarks

Benchmarks based on traits measured from 862 samples, each scored for 5,751,024 markers and run on an Apple M4 CPU (cached VCF).

Data Loading

Note: First run with a given genetic marker file requires substantial time for parsing (≈9 minutes for 5M markers scored for 1000 individuals); subsequent runs use binary cache and load in seconds.

Analysis Times (5.75M markers, 862 samples; excludes data loading/result writing)

Scaling by Marker Count (862 samples; includes cached load, alignment, PCA, kinship where relevant)

License

Distributed under the MIT license. See LICENSE.

Disclaimer: This is an independent Python implementation of algorithms developed by others. Any errors are mine alone. -James