bunbun 0.1.2

Fast, unified genotype parser for PLINK, VCF, PLINK2, and BGEN formats

🐰 bunbun

Fast, unified genotype parser for PLINK, VCF, PLINK2, and BGEN formats.

bunbun reads genotype data from any major format and gives you back the same three things every time:

output	type	what
.samples	polars.DataFrame	sample metadata (ID, sex, family, phenotype, …)
.variants	polars.DataFrame	variant metadata (chrom, pos, ref, alt, …)
.dosages	numpy.ndarray	genotype matrix — (n_samples × n_variants), float32

---

Install

uv pip install bunbun

If you want to build it from the source

git clone https://github.com/nahid18/bunbun.git && cd bunbun

uv venv 
source .venv/bin/activate # OR
source .venv/Scripts/activate

uv pip install maturin numpy pytest polars
uv pip install -e .

Quick start

import bunbun

# Auto-detect: pass any file from the fileset
data = bunbun.read("cohort.bed")

data.samples    # polars DataFrame: sample_id, family_id, sex, …
data.variants   # polars DataFrame: chrom, pos, id, ref, alt, …
data.dosages    # numpy array: shape (500, 10000), dtype float32

data.shape      # (500, 10000)

Format-specific readers

data = bunbun.read_plink("cohort.bed")
data = bunbun.read_vcf("calls.vcf.gz")
data = bunbun.read_plink2("imputed.pgen")
data = bunbun.read_bgen("ukb.bgen", sample_path="ukb.sample")

Operations

# Allele frequencies (parallel, NaN-aware)
freqs = data.allele_frequencies()   # numpy array, length n_variants

# Missing rate per variant
miss = data.missing_rates()

# Mean-impute missing values (in place)
data.mean_impute()

# Subset to a genomic region
chr1_data = data.region("1", 100_000, 500_000)

# Subset to specific samples
sub = data.subset_samples(["SAMPLE_001", "SAMPLE_042"])

Works with your existing stack

import polars as pl
import numpy as np
from sklearn.decomposition import PCA

data = bunbun.read("cohort.bed")

# Filter variants with polars
common = data.variants.filter(pl.col("chrom") == "1")

# Feed dosages straight into sklearn
pca = PCA(n_components=10)
components = pca.fit_transform(np.nan_to_num(data.dosages))

---

Supported formats

Format	Extensions	Compression	Dosage
PLINK 1.x	.bed .bim .fam	—	hard calls
VCF	.vcf .vcf.gz	gzip, bgzf	GT or DS field
PLINK 2	.pgen .pvar .psam	—	hard calls
BGEN	.bgen .sample	zlib, zstd	probabilistic

Output schema

data.samples — polars DataFrame

column	dtype	notes
sample_id	Utf8	IID (PLINK) or sample column (VCF)
family_id	Utf8	FID or "."
paternal_id	Utf8	or "0"
maternal_id	Utf8	or "0"
sex	Utf8	"male" / "female" / "unknown"
phenotype	Float64	or NaN

data.variants — polars DataFrame

column	dtype	notes
chrom	Utf8	string for X/Y/MT compatibility
pos	UInt32	1-based bp position
id	Utf8	rsID or chrom:pos:ref:alt
ref	Utf8	reference allele
alt	Utf8	alternate allele(s), comma-separated
cm	Float64	genetic distance or NaN
qual	Float64	quality score or NaN
filter	Utf8	"PASS" / filter string

data.dosages — numpy ndarray

• Shape: (n_samples, n_variants)
• Dtype: float32
• Encoding: 0.0 = hom-ref, 1.0 = het, 2.0 = hom-alt, NaN = missing
• For imputed data (BGEN, VCF with DS): continuous values in [0, 2]

---

Rust API

use bunbun::{read, PlinkReader, GenotypeReader};

// Auto-detect
let data = bunbun::read("cohort.bed").unwrap();

// Or explicit
let reader = PlinkReader::new("cohort.bed");
let data = reader.read_all().unwrap();

println!("{} × {}", data.n_samples(), data.n_variants());
println!("{}", data.samples.df);   // polars DataFrame
let freqs = data.dosages.allele_frequencies();

---

• One output type. GenotypeData is the same regardless of input format.
• Zero-copy where possible. PLINK .bed is memory-mapped; numpy gets the ndarray directly.
• Parallel by default. Genotype decoding uses rayon across variants.
• Small-string optimization. The Allele type stores ≤7-byte alleles inline (covers >99% of SNPs).
• Transparent compression. gzip, bgzf, zstd, bzip2 detected from magic bytes.

---

License

MIT