bgen

C library and Python bindings for reading BGEN files and encoding layout-2 genotype payloads (format specifications 1.2 and 1.3).

This repository is a maintained fork of the original limix/bgen and limix/cbgen projects, merged into a single codebase with a modern build system, nanobind-based Python bindings, and S3 support.

Features

• Fast C library for parsing BGEN genotype files
• Python bindings via nanobind
• Supports BGEN 1.2 and 1.3 format specifications
• zlib and zstd decompression support
• Metafile indexing for efficient random access
• Native bounded-batch layout-2 probability and hard-call encoding

Quick start (pixi)

pixi install
pixi run test

Building the C library

cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build
ctest --test-dir build --output-on-failure

System dependencies: zlib, zstd (installed via your system package manager or conda).

Installing the Python package

pip install .

Or in development mode:

pip install -e ".[test]"
pytest python/tests

For a self-contained local CPython 3.12 wheel, including the private nanobind backend and bundled native libraries:

pixi run -e wheel python -m pip wheel --no-deps --no-build-isolation -w dist \
  -Cbuild-dir=build-wheel \
  -Ccmake.define.CMAKE_DISABLE_FIND_PACKAGE_bgen=ON \
  -Ccmake.define.BGEN_NANOBIND_BACKEND=private -Cwheel.py-api= .

The empty wheel.py-api override is intentional: a private backend is specific to its build interpreter, so its wheel must be tagged cp312-cp312, not abi3. Bundled dependencies' license notices are installed with the wheel.

Python usage

from cbgen import BgenFile, BgenMetafile

# Open a BGEN file
with BgenFile("example.bgen") as bgen:
    print(f"Variants: {bgen.nvariants}, Samples: {bgen.nsamples}")

    if bgen.contain_samples:
        samples = bgen.read_samples()

    # Create a metafile index for efficient access
    bgen.create_metafile("example.bgen.metafile")

# Read variant data via metafile
with BgenFile("example.bgen") as bgen:
    with BgenMetafile("example.bgen.metafile") as mf:
        partition = mf.read_partition(0)
        for variant in partition.variants:
            gt = bgen.read_genotype(variant.offset)
            print(f"{variant.rsid}: {gt.probability.shape}")

bgen.layout reports the native header's layout version (1 or 2), also available on the legacy bgen_file API. Sample and variant counts are unsigned 32-bit values exposed as Python integers. Unsupported header flags and truncated headers are rejected by the native reader.

Layout 1 supports raw and zlib-compressed, biallelic, diploid, unphased records, including metafile indexing. Its three independently rounded probabilities are decoded as q / 32768 without renormalization; all-zero triplets are marked missing and returned as NaNs. Layout 2 supports raw, zlib and zstd blocks with 1--32-bit probabilities, variable ploidy, and multiple alleles. Phased records whose samples all have zero ploidy return shape (nsamples, 0), retaining their ploidy and missingness arrays. Both layouts support precision=32 and precision=64; returned Python arrays remain valid after subsequent reads and after closing the file.

Structured batch reads

The modern BgenFile supports native batches for both layouts and every supported compression method, including mixed phase, allele counts and ploidy. One batch can supply multiple conversion fields without rereading each variant:

with BgenFile("example.bgen") as bgen:
    with BgenMetafile("example.bgen.metafile") as mf:
        offsets = [v.offset for v in mf.read_partition(0).variants]
    # Use bounded chunks, not the entire file, for large datasets.
    metadata = bgen.read_genotype_metadata(offsets)
    width = int(metadata.ncombs.max()) if len(offsets) else 0
    batch = bgen.read_genotypes_batch(offsets, max_ncomb=width, precision=64)
    probabilities = batch.probability  # (variants, samples, padded width)
    ploidy, missing = batch.ploidy, batch.missing
    for v in range(len(offsets)):
        row = probabilities[v, :, :batch.ncombs[v]]
        if batch.phased[v]:
            haplotypes = row.reshape(bgen.nsamples, -1, int(batch.nalleles[v]))

GenotypeBatchResult exposes columnar, C-contiguous, NumPy-owned arrays:

ncombs is each variant's single-reader width, based on its declared maximum ploidy, not the caller's padded width. probability_count is each sample's structural width even when missing: ploidy * nalleles if phased, otherwise choose(ploidy + nalleles - 1, nalleles - 1). Zero-ploidy unphased samples have one state; zero-ploidy phased samples have no probabilities. All-missing variants retain their structural widths. Missing samples and unused columns are NaN; distinguish them using missing and probability_count, not NaNs. Phased probabilities are haplotype-major; unphased probabilities retain BGEN's colex order. Layout-1 nbits is 16 and its exact independent rounding is preserved.

Offsets preserve order and duplicates. max_ncomb is capacity, never truncation: too-small or invalid widths raise an exception. An explicit width reads/decompresses and decodes each selected block once, returning metadata from that same read. None first scans widths, then rereads blocks to decode; compressed layout-2 blocks therefore decompress twice, but probability values are decoded only once. read_genotype_metadata returns GenotypeBatchMetadata without allocating or decoding probabilities. It still decompresses compressed layout-2 blocks and reads layout-1 triplets for missingness, and validates framing/header/payload lengths, not numerical probability sums. get_ncombs provides a narrower width-only scan.

Empty batches return (0, S, W), with W=0 for automatic sizing. Explicit W=0 is valid for wholly zero-ploidy phased batches. All arrays survive later reads and reader closure. Compute dosage from float64 probability views before casting output to float32 if rounding accuracy matters.

Output storage is V*S*W*itemsize + 6*V*S + 8*V bytes, including probability padding and metadata, plus offsets and native reusable one-variant encoded/ decoded scratch. There is no second full probability tensor or Python object per sample/variant. Choose V from a byte budget when S or W is large; sample slicing is not supported. File handles are not thread-safe. Legacy bgen_file.read_genotypes_batch remains probability-only float64 and does not allocate the new metadata arrays.

Lightweight dimensions and compact capacity (4.14.0)

with BgenFile("example.bgen") as bgen:
    dimensions = bgen.read_dimensions(
        offsets, ploidy_source="actual", probability_model="auto"
    )
    batch = bgen.read_genotypes_batch(
        offsets, max_actual_ncomb=dimensions.decode_width, precision=64
    )
    # All output fields can reuse this one batch.

GenotypeDimensions contains Python integer maxima gp_width, hp_width and decode_width = max(gp_width, hp_width); native counts must fit uint32. It does not return per-sample arrays. GP width zero means no unphased model was selected; HP width zero means no phased model was selected. HP width is at least one if phased is selected, even if all haplotypes have ploidy zero. This preserves model presence for output schema creation; actual decoded phased rows may still have zero probability entries. Empty selections return all zeros.

ploidy_source="actual" bounds dimensions using all sample ploidies, including missing samples. "declared" uses the encoded maximum ploidy and may reserve more space than any sample needs. probability_model="auto" reads original phase. "unphased" or "phased" assumes that model, without validating it during prefix discovery; "mixed" reserves both GP and HP bounds for every record. The phase byte occurs after the sample-ploidy bytes, so it is not available from the first eight bytes. Layout 1 always has intrinsic K=2/P=2, so no genotype payload is needed; explicit phased/mixed assumptions still reserve HP=4 (mixed also GP=3), and the full reader reports original unphasedness.

Optional max_ploidy=bound (0--63) substitutes a caller-provided capacity assumption instead of either source of ploidy. Raw/zlib then skip scanning sample ploidies: auto still reads phase, while explicit models only need the header. This avoids overflow from an unused, loose declared maximum. The bound is not validated against samples in discovery; consumers must check decoded ploidy and model against their assumptions. With known allele counts, a supplied ploidy bound and explicit model, consumers can calculate capacity and call compact batch directly without invoking discovery at all.

Discovery is not an integrity-validation pass. Raw/zlib prefix strategies may accept a truncated tail, skipped invalid ploidies/phase or a missing checksum that full reads reject. Zstd deliberately uses the full-block native header path; actual/override strategies validate compact actual counts, while declared discovery validates declared counts too. Layout-1 fixed dimensions do not validate payloads or authenticate index offsets. Numerical probability sums are not validated by discovery or native decoding; consumers requiring numerical validation must check decoded probabilities.

max_actual_ncomb=W is mutually exclusive with max_ncomb. It reads each block once, validates the full native header/payload lengths and bounds actual structural sample widths before writing that variant. Missing samples count towards capacity. Too-small capacity raises, never truncates. Even overflowing unused declared combination counts do not force a huge intermediate allocation. In this mode only, batch.ncombs[v] is the effective maximum actual sample width, not the declared maximum-ploidy width; it is zero for zero-sample variants. Probabilities, precision, metadata ownership, ordering, duplicate offsets and NaN padding otherwise retain their usual semantics. There is no width pre-scan and no declared-width probability scratch; output and optional one-variant decoded scratch are bounded by W. All default APIs keep their original behavior.

Performance checks

scripts/benchmark_decode.py compares two Release shared-library builds through their exported native public APIs. It generates deterministic, compressed, biallelic diploid fixtures and measures single reads, batch reads, combination scans and indexing, including float32 single reads and wider padded batches. Timings exclude fixture generation, initial native output-buffer allocation and warm-up; old/new process order alternates across repetitions. For example, on macOS:

pixi run python scripts/benchmark_decode.py --generate build-benchmark/fixtures \
  --zstd .pixi/envs/default/lib/libzstd.dylib
pixi run python scripts/benchmark_decode.py \
  --manifest build-benchmark/fixtures/fixtures.json \
  --old /path/to/before/libbgen.dylib --new /path/to/after/libbgen.dylib \
  --repeats 11 --loops 10 --output build-benchmark/results.json

Use matching compilers, optimization flags and compression libraries, and run without competing builds/tests. Results are local, warm-cache measurements, not a guarantee for other hardware, remote storage or arbitrary datasets. To compare actual installed Python wheels with the same interpreter and dependencies, use --python-old /path/to/old-install --python-new /path/to/new-install --modes single single32 batch padded scan instead of --old/--new. These measurements include NumPy output allocation and bypass editable-install hooks.

Encoding layout-2 genotype blocks (4.15 API)

The encoder returns uncompressed genotype payloads, not complete BGEN files. File headers, variant identifiers, length prefixes and compression remain the caller's responsibility. It does not depend on VCZ or interpret format-specific NaN sentinel bit patterns.

import numpy as np
from cbgen import encode_layout2, encode_layout2_hardcalls

payload, offsets = encode_layout2(
    allele_counts=np.array([2], dtype=np.uint16),
    ploidy=np.array([[2, 1]], dtype=np.uint8),
    missing=np.array([[False, True]], dtype=bool),
    phased=np.array([False], dtype=bool),
    gp=np.array([[[0.5, 0.25, 0.25], [np.nan, np.nan, np.nan]]]),
    bits=16,
)
first_block = payload[int(offsets[0]):int(offsets[1])]

hard_payload, hard_offsets = encode_layout2_hardcalls(
    genotypes=np.array([[[0, 1], [-1, -2]]], dtype=np.int8),
    allele_counts=np.array([2], dtype=np.uint16),
    phased=np.array([False], dtype=bool),
)

Both interfaces are keyword-only:

encode_layout2(*, allele_counts, ploidy, missing, phased,
               gp=None, hp=None, bits=8, max_output_bytes=268435456)
encode_layout2_hardcalls(*, genotypes, allele_counts, phased,
                         ploidy=None, missing=None, bits=8,
                         max_output_bytes=268435456)
    -> (bytes, uint64[nvariants + 1] offsets)

All inputs must be aligned, native-endian, C-contiguous NumPy arrays with matching variant/sample dimensions. No implicit dtype or layout conversions are made. Read-only inputs are accepted; do not mutate inputs concurrently. Offsets start at zero and end at the payload length; every variant starts on a byte boundary. Empty batches, zero samples and zero ploidy are supported. The native measurement and encoding passes release the GIL.

For each unphased sample, the first C(P+K-1,P) GP entries are active. For phased samples, the first P*K HP entries are active, arranged as P simplexes of K alleles. Width must cover all samples, including missing ones. Padding beyond active width is ignored, not validated. The caller must validate any external format's fill/missing conventions. Only the selected GP or HP field is read for a variant; an unused field may be omitted.

Active nonmissing simplexes must be finite, in [0,1], and sum to one within absolute tolerance 1e-6 (no relative tolerance). They are normalized and quantized with stable largest-remainder rounding; exact ties favor earlier state indexes. The last state of each simplex is omitted. All bit depths 1..32 use little-endian least-significant-bit-first packing. Explicitly missing probability samples skip numeric validation and store only zeros. At P=0, unphased GP still has one state (which must be one if nonmissing); phased HP has no states. Missingness remains explicit even at P=0.

For hard calls, ploidy is derived from trailing -2 fill. Optional explicit ploidy must agree. Without an explicit missing array, any -1 allele marks the whole sample missing. With one, a nonmissing call containing -1 is an error; missing calls may contain valid alleles or -1, but all allele indexes and trailing fill are still checked. Multiallelic unphased hard calls map to the BGEN colex genotype rank. Hard calls are packed directly without materializing a dense probability array.

The default allocation limit is 256 MiB of payload per batch, overridable by max_output_bytes. Each individual block must also fit a BGEN uint32 length field. Offsets and one variant descriptor per input variant are allocated separately; quantization scratch is reused and scales with the largest active simplex (approximately 24 bytes per state on 64-bit platforms). Chunk inputs to bound both input and scratch memory. Invalid values, insufficient widths, overflow and allocation limits raise errors; per-call data errors identify the zero-based variant and sample. No partial payload is returned.

C API

Link against the bgen library:

find_package(bgen REQUIRED)
target_link_libraries(mytarget PRIVATE BGEN::bgen)

See include/bgen/bgen.h for the full C API. The public bgen/encode.h declares bgen_layout2_size and bgen_layout2_encode, supporting the same numeric and hard-call models. Each bgen_layout2_variant contains explicit input buffer byte lengths and padding width. Measure first, allocate the payload and nvariants+1 uint64 offsets, then encode. Returns are 0 on success and -1 on error; an optional per-call bgen_encode_error gives the location and message (SIZE_MAX means not applicable). The size pass validates structure and hard calls; numeric probabilities are validated by the encode pass. On encode failure, C output buffers may be partially written and must be discarded. Output/input aliasing and concurrent input mutation are unsupported. No global encoder state is used.

Specialized biallelic compatibility and fixed-stride encoding

The following public APIs keep the specialized legacy biallelic hard-call paths in the native library, rather than in downstream format adapters:

encode_layout2_hardcalls_legacy(genotypes, phased, *, max_output_bytes=268435456)
    -> (uint8[variants, 10 + 3*samples] padded_blocks, uint32[variants] lengths)
layout2_fixed_size(num_samples, uniform_ploidy, total_string_length) -> int
encode_layout2_fixed(varid, rsid, chrom, allele1, allele2, position,
                     genotypes, phased, out, uniform_ploidy, total_string_length)
    -> None

genotypes must be contiguous int8[variants,samples,2], phased is bool[variants]. Haploid calls use -2 in slot 1; -1 in either allele makes the whole call missing. These compatibility APIs reject -2 in slot 0, including all-fill calls. Use the general hard-call API for P=0. Empty sample axes retain the legacy declared ploidy range 2/2. Only padded_blocks[v, :lengths[v]] is encoded; trailing bytes are unspecified. No Python payload copy is required.

encode_layout2_fixed fuses genotype encoding, identifying-field framing and zlib level-0 stored-block serialization into caller-owned output, with one reusable native genotype scratch buffer. Every sample must have the declared uniform_ploidy (1 or 2); mixed ploidy raises NotImplementedError. String fields are contiguous uint8[variants,width] padded byte rows, ending at the first NUL or row width. The sum of their actual lengths must equal total_string_length for every variant; callers can pad an identifier field to achieve a fixed total. Identifier lengths must fit uint16 and allele lengths uint32. position is contiguous non-negative int32[variants]. Invalid alleles, lengths, metadata and string budgets raise ValueError with variant/sample context where applicable.

out must be a writable contiguous uint8 array, either 1D with at least variants * layout2_fixed_size(...) capacity, or exactly (variants, layout2_fixed_size(...)). Input/output overlap is rejected. Insufficient 1D capacity raises cbgen.BufferTooSmall, a ValueError subclass; malformed shapes/dtypes and read-only buffers raise ValueError. On failure, caller-owned output may be partially written and must be discarded. On success, any excess 1D capacity remains untouched. The output contains complete identifying fields and C/D length framing, but no file header; it requires layout-2/zlib flags in the caller's BGEN header. Stored DEFLATE chunks are deterministically at most 65535 bytes, with the RFC1950 Adler-32 supplied by the optimized zlib-ng dependency.

Both specialized calls release the GIL and validate and encode hard calls in one native sample pass, without general combinatorial calculations or dense probability arrays. The general encode_layout2_hardcalls also selects the direct biallelic kernel for int8 inputs with K=2, bits=8 and width 1/2 when explicit ploidy/missing arrays are absent; all other inputs use the general checked encoder. Its P=0 and missingness semantics do not change. The public C counterparts and bounded buffer descriptors are declared in bgen/encode.h.

Reproduce installed-kernel timings with pixi run -e wheel python scripts/benchmark_encode_hardcalls.py. The optional --backend vcz runs an installed older VCZ implementation in its own environment for byte-hash and throughput comparisons; it is not a library dependency. The specialized implementation is independently written against the BGEN and RFC1950/1951 formats, preserving legacy observable behavior without incorporating Apache-licensed VCZ implementation source.

Acknowledgments

• Marc Jan Bonder for bug-reporting and improvement suggestions.
• Yan Wong for bug-reporting and improvement suggestions.
• Carl Kadie for bug-reporting and improvement suggestions.

Authors

• Danilo Horta
• Julianus Pfeuffer

License

This project is licensed under the MIT License.