vcfv

Fast VCF parser to Polars/Pandas DataFrame with INFO and FORMAT field expansion.

from vcfv import read_vcf

df = read_vcf("variants.vcf", parse_info="wide", parse_format="wide")

Why vcfv?

Existing VCF parsers either load the entire file into memory at once, keep INFO and FORMAT as raw strings, or are built on aging dependencies. vcfv is built on Polars which gives you:

• Lazy evaluation — filters and projections are pushed down before reading
• Native INFO and FORMAT expansion — split into typed DataFrame columns in a single call
• Simple API — one function, sensible defaults
• Gzip support — reads .vcf.gz transparently

Installation

pip install vcfv

With Pandas support:

pip install vcfv[pandas]

Usage

Basic

from vcfv import read_vcf

# Returns a Polars DataFrame
df = read_vcf("variants.vcf")

INFO and FORMAT expansion

# Expand INFO fields into typed columns (INFO_AC, INFO_AF, INFO_DP, ...)
df = read_vcf("variants.vcf", parse_info="wide")

# Expand FORMAT fields per sample (Sample1_GT, Sample1_DP, ...)
df = read_vcf("variants.vcf", parse_format="wide")

# Both at once
df = read_vcf("variants.vcf", parse_info="wide", parse_format="wide")

Lazy evaluation

import polars as pl
from vcfv import read_vcf

# Returns a LazyFrame — nothing is read until .collect()
lf = read_vcf("variants.vcf", lazy=True)

# Combine with any Polars lazy operations before collecting
df = lf.filter(pl.col("CHROM") == "chr1").collect()

Filtering

# Filter by minimum QUAL score — rows with QUAL=. are excluded
df = read_vcf("variants.vcf", min_qual=30.0)

# Filter by genomic region — format: "chrom:start-end"
df = read_vcf("variants.vcf", region="chr1:1000000-2000000")

# Select specific output columns
df = read_vcf("variants.vcf", fields=["CHROM", "POS", "REF", "ALT"])

# Combine filters freely
df = read_vcf(
    "variants.vcf",
    parse_info="wide",
    parse_format="wide",
    min_qual=30.0,
    region="chr1:1000000-2000000",
    fields=["CHROM", "POS", "REF", "ALT", "INFO_AF"],
)

Gzip support

df = read_vcf("variants.vcf.gz")

Chunked iteration for large files

from vcfv import iter_vcf

for chunk in iter_vcf("large.vcf", chunk_size=10_000):
    process(chunk)  # each chunk is a Polars DataFrame

Pandas output

from vcfv import read_vcf_pandas

df = read_vcf_pandas("variants.vcf")  # returns a Pandas DataFrame

Incomplete INFO/FORMAT headers

Some VCF files have INFO or FORMAT fields in the data that are not declared in the header. By default vcfv reads keys only from the header (infer_info_keys_from="header") which is fastest. If you suspect your file has undeclared fields you can scan a sample of rows or the entire file:

# Scan first 50,000 rows to detect undeclared INFO keys
# Emits a UserWarning if any undeclared keys are found
df = read_vcf("variants.vcf", infer_info_keys_from=50_000)

# Scan the entire file — safe but slower (streams in chunks, no full RAM load)
df = read_vcf("variants.vcf", infer_info_keys_from="all")

# Same options available for FORMAT keys
df = read_vcf("variants.vcf", infer_format_keys_from="all")

API reference

read_vcf

read_vcf(
    path: str | Path,
    *,
    lazy: bool = False,
    parse_info: False | "wide" | "struct" | "auto" = "auto",
    infer_info_keys_from: int | "header" | "all" = "header",
    parse_format: False | "wide" | "struct" | "auto" = "auto",
    infer_format_keys_from: int | "header" | "all" = "header",
    min_qual: float | None = None,
    region: str | None = None,
    fields: list[str] | None = None,
) -> pl.DataFrame | pl.LazyFrame

Parameter	Type	Default	Description
path	str | Path	—	Path to .vcf or .vcf.gz file
lazy	bool	False	Return a LazyFrame instead of DataFrame
parse_info	see below	"auto"	How to handle the INFO column
infer_info_keys_from	int | "header" | "all"	"header"	Where to discover INFO field keys
parse_format	see below	"auto"	How to handle FORMAT and sample columns
infer_format_keys_from	int | "header" | "all"	"header"	Where to discover FORMAT field keys
min_qual	float | None	None	Keep only rows where QUAL >= min_qual. Rows with QUAL=. are excluded
region	str | None	None	Genomic region filter in format "chr1:1000000-2000000"
fields	list[str] | None	None	Select specific output columns by name

parse_info and parse_format options

Value	Behaviour
"auto"	Picks the best strategy automatically (see below)
"wide"	Each field becomes a separate typed column (INFO_AF, Sample1_GT, ...)
"struct"	Fields stored as a list of strings per row — memory efficient for many samples
False	No parsing — raw strings kept as-is

parse_format="auto" selects:

• "wide" for ≤ 20 samples
• "struct" for ≤ 200 samples
• False for > 200 samples

parse_info="auto" always selects "wide".

infer_info_keys_from and infer_format_keys_from options

Value	Behaviour
"header"	Read keys from ##INFO/##FORMAT header lines — fastest, default
int (e.g. 10_000)	Scan first N rows and merge with header keys — emits UserWarning for undeclared keys
"all"	Stream entire file in chunks to find all keys — safe for incomplete headers, slowest

iter_vcf

iter_vcf(
    path: str | Path,
    chunk_size: int = 10_000,
    *,
    parse_info: False | "wide" | "struct" | "auto" = "auto",
    infer_info_keys_from: int | "header" | "all" = "header",
    parse_format: False | "wide" | "struct" | "auto" = "auto",
    infer_format_keys_from: int | "header" | "all" = "header",
    min_qual: float | None = None,
    region: str | None = None,
    fields: list[str] | None = None,
) -> Iterator[pl.DataFrame]

Yields chunk_size rows at a time as Polars DataFrames. Accepts all the same parameters as read_vcf.

Benchmarks

Tested on 500,000 variants, 10 runs, trimmed mean (top/bottom 10% dropped). RAM measured via psutil RSS delta per subprocess (includes Rust/C heap). Tested on: Windows 11, AMD Ryzen 7 7730U, 16 GB RAM.

vcfv

Mode	Mean	Median	Stdev	RAM
eager (INFO+FORMAT expanded, header)	3.39s	3.38s	±0.31s	570 MB
lazy (INFO+FORMAT expanded, header)	3.55s	3.57s	±0.11s	571 MB
eager (INFO+FORMAT expanded, all)	9.71s	9.70s	±0.32s	646 MB
eager (no expansion)	0.15s	0.14s	±0.01s	273 MB
lazy (no expansion)	0.15s	0.15s	±0.01s	277 MB

Baselines (no INFO/FORMAT expansion — apples-to-apples)

Tool	Mean	Median	Stdev	RAM
scikit-allel	2.10s	2.10s	±0.05s	115 MB
pandas raw TSV	3.23s	3.21s	±0.08s	358 MB

vcfv with no expansion is ~14x faster than scikit-allel (0.15s vs 2.10s).

Neither scikit-allel nor pandas natively expand INFO or FORMAT fields — vcfv is the only tool in this comparison that does so in a single call.

To reproduce: python benchmarks/benchmark.py your_file.vcf --runs 10

Supported formats

Format	Support
.vcf	✅ Full lazy streaming
.vcf.gz	✅ Decompressed into memory before parsing
.bcf	❌ Not supported yet

Requirements

• Python ≥ 3.11
• Polars ≥ 1.0
• PyArrow ≥ 14.0

Contributing

Issues and pull requests are welcome. Please open an issue before submitting a PR for larger changes.

License

MIT