Rosemeis / halkoSVD

Fast Python/Cython implementation of the PCAone Halko algorithm

Python implementation of Halko algorithm (PCAone)

This is a fast implementation of the PCAone (H+Y) Halko algorithm in Python/Cython for genetic data. It takes binary PLINK format (*.bed, *.bim, *.fam) as input. For simplicity, mean imputation is performed for missing data.

It is inspired by the lovely PCAone software! Have a look here.

Installation

# Build and install via PyPI
pip install halkoSVD

# Download source and install via pip
git clone https://github.com/Rosemeis/halkoSVD.git
cd halkoSVD
pip install .

# Download source and install in new Conda environment
git clone https://github.com/Rosemeis/halkoSVD.git
conda env create -f environment.yml
conda activate halkoSVD


# You can now run the program with the `halkoSVD` command

Quick usage

Provide halkoSVD with the file prefix of the PLINK files.

# Check help message of the program
halkoSVD -h

# Extract the top 10 PCs
halkoSVD --bfile input --threads 32 --pca 10 --out halko

Options

• --power, specify the number of power iterations (12)
• --extra, number of extra vectors for oversampling (16)
• --batch, specify the batch size to process SNPs (4096)
• --full, load the entire genotype matrix into matrix
• --loadings, save the SNP loadings
• --raw, only output eigenvectors without FID/IID