Skip to main content

figwig

PyPI Downloads Unit Tests Documentation Status

[docs][release notes]

A fast, multithreaded reader and writer of bigWig and figwig files, into and out of numpy. Also contains a formal specification of the figwig file format, which holds any number of per-base tracks over one genome with a JSON record of attributes for each, and reads windows from any subset of them in one call, into (windows, tracks, width). figwig bw2fw, fw2bw and bam2fw convert bigWigs and reads to and from a figwig file.

Built on work from Nezar Abdennur and Jack Huey.

Loading training data for a genomics model means reading the signal under tens or hundreds of thousands of windows: 1,000 bp around every peak and background region of an experiment, say. figwig reads all of them in one call, straight into a float32 array, from one bigWig or from several at once, and writes bigWigs from the same windows, such as a model's predictions, or from intervals. The common readers make one call per window, and each call holds Python's GIL, so adding threads does not help them. figwig decompresses and compresses with zlib or libdeflate and decodes with numba, all without the GIL, so its threads run in parallel. It depends only on numpy and numba.

Its command, figwig bam2bw, turns SAM/BAM files of reads, or BED/tsv files of fragments, into bigWigs of per-base counts. It is bam2bw with faster readers, the same outputs, and -c added.

figwig fastest other tool
Read 167,750 windows of 1,000 bp from an ATAC-seq bigWig 0.103 s on 8 threads 1.25 s, pybigtools
Write 15.9 million per-base counts 0.14 s on 8 threads 4.20 s, pyBigWig
Write 21.2 million intervals, with zoom levels 0.45 s on 8 threads 5.47 s, pybigtools
Convert a 2.4 GB ATAC-seq BAM to two stranded bigWigs 4.65 s, 910 MB peak, at -p 2 58.48 s, 3,039 MB peak, bam2bw 0.5.1
The same BAM's per-base coverage, like deepTools 1.6 s at -p 8 14.9 s on 8 processes, bamCoverage --binSize 1, the same value at every base
Read 8,192 windows of 2,048 bp from 48 count tracks on chr1 0.071 s on 8 threads, 0.040 s on 32, from one figwig file 0.19 s and 0.083 s from 48 bigWigs, figwig; 5.9 s pybigtools; 4.0 s zarr
A DataLoader batch of 64 windows x 48 tracks as a CUDA tensor 4 ms from one figwig file 30 ms from the bigWigs, figwig; 120 ms pybigtools
10 pseudobulks of 10,246 cells over 8,192 windows 0.23 s on 8 threads 4.9 s from an in-memory scipy CSR; 95 s from one bigWig per cell
Convert 257.6 million 10x fragments to 10,246 per-cell tracks 8 s at -p 8, 3.9 s at -p 32, 19 GB peak 140 s, 24 GB, a pandas recipe; 217 s, 8 GB, SnapATAC2 (fragments, not counts)

Measured on a 2x AMD EPYC 9575F with Python 3.13 and numba 0.68. The benchmarks page has every setup, the other tables (figwig files against bigWigs, zarr and HDF5 at 6, 48 and 512 tracks; writers; threads) and the plots.

Read time, write time and file size by assay, against what most people use today

Each cell is figwig against the baseline, figwig first, in bold where it is smaller. figwig ran on 8 threads unless a note says otherwise; each baseline is given its fastest setting (the other tools mostly do not gain from threads). Reads are windows of 2,048 bp unless noted. The Hi-C reads run on one thread, figwig and cooler alike.

Setting Baseline Read Write Size
1 ATAC-seq track, 167,750 windows of 1,000 bp on random signal bases and negatives one bigWig; read by pybigtools, written by pyBigWig 98.4 ms vs 1.86 s (pybigtools, 1 thread, sorted windows) 0.245 s vs 8.13 s (pyBigWig, 1 thread) 64.4 MB vs 64.4 MB (pyBigWig)
48 count tracks over chr1, 8,192 windows of 2,048 bp one bigWig per track; read by pybigtools, written by pyBigWig 72.7 ms vs 4.68 s (pybigtools, 1 thread) 2.04 s vs 69.4 s (pyBigWig, 1 thread) 97.2 MB vs 1.08 GB (pyBigWig)
1,024 count tracks over chr1, 1,024 windows of 2,048 bp one bigWig per track; read by pybigtools, written by figwig 0.172 s vs 22.4 s (pybigtools, 1 thread) 33.2 s vs 497 s (bigWigs written by figwig from a file of groups of 8; pyBigWig not run) 2.04 GB vs 21.2 GB (the bigWigs written by figwig; pyBigWig not run)
10x PBMC scATAC-seq, 10,246 cells, 492 million cut sites; reads: 10 pseudobulks over 8,192 windows of 2,048 bp on K562 ATAC peaks SnapATAC2 (write, size) and a scipy CSR matrix in memory (read) 0.289 s vs 0.445 s (scipy CSR, 8 threads) 8.27 s vs 218 s (SnapATAC2, 8 threads) 1.11 GB vs 1.02 GB (SnapATAC2, which stores fragments rather than per-base counts)
K562 Fiber-seq BAM, 99,685 alignments, 886 MB; reads: m6A summed over all reads in 8,192 windows of 2,048 bp a pysam loop with modified_bases (write), the BAM (size) and a scipy CSR matrix in memory (read) 86.4 ms vs 0.263 s (scipy CSR, 1 thread) 9.52 s vs 245 s (pysam, 1 thread, m6A and 5mC placed, nothing written) 605 MB vs 886 MB (the BAM)
K562 Hi-C, 5 kb, 907M contacts; 2 Mb diagonal window as a dense 400 x 400 block cooler (Python API) window: 0.878 ms vs 18.9 ms 47.6 s vs 42.3 s 979 MB vs 748 MB
Single-cell Hi-C, mouse ES cells (Nagano 2017), 3,882 cells, 10 kb, 626M contacts; 2 Mb diagonal window summed over the cells of 4 labels, (4, 200, 200), and one whole cell scool (cooler) pseudobulk window: 14.8 ms vs 23.6 s; one cell: 8.09 s vs 14 ms 60 s vs 314 s 1.45 GB vs 1.95 GB
  • The 1,024-track baseline write and size are the 1,024 bigWigs figwig wrote with 10 zoom levels (497 s, 21.2 GB); pyBigWig was not run at that size.
  • scATAC-seq: figwig and SnapATAC2 hold the same 491,845,610 fragment ends of the 10,246 called cells (every cell's figwig total is twice its SnapATAC2 fragment count), and figwig and the scipy matrix gave identical sums over the windows.
  • The scipy matrix is 14.8 GB in memory (the read times leave out building it and opening the figwig file, which takes 5 ms), where the figwig file is 1.11 GB.
  • SnapATAC2 stores each cell's fragments, so counts over windows or peaks are still to be computed from its output.
  • Fiber-seq: figwig and the pysam pass agree on the m6A total of each of the 99,685 alignments over its span, and on the sums over the windows (the file and the matrix both hold the 0 to 255 probabilities).
  • The pysam loop places 204,449,545 m6A and 15,488,000 5mC calls and keeps only the m6A calls, in a scipy matrix of 13.4 GB that the read time leaves out building; the figwig file also holds the 5mC, nucleosome, MSP and FIRE channels.
  • Reading times leave out opening the figwig file, which takes 1.39 s here, and loading the scipy matrix.
  • With both triangles (1.79 GB at 5 kb) the figwig file also reads a bin's whole row in 36.2 ms against cooler's 8.97 s.
  • With one track per cell in track order (level 6, 2.47 GB) the figwig file reads one whole cell in 12.2 ms.

Measured by benchmarks/readme_table/run.py on 2026-11-11 on AMD EPYC 9575F 64-Core Processor (256 cores), figwig 0.3.0 at 2858597; the numbers are in benchmarks/readme_table/results.json.

Installation

pip install figwig

Or, with uv:

uv add figwig

It needs Python 3.10 or later, numpy 1.23 or later, and numba 0.58 or later. Optional extras:

  • figwig[fast] adds the deflate package, which writes bigWigs several times faster and reads figwig files about twice as fast.
  • figwig[isal] adds the isal package, whose ISA-L reads bigWigs 1.3 to 1.6 times faster, with the same values.
  • figwig[bam2bw] adds pysam, pyfaidx, biopython, tqdm, isal and deflate, which figwig bam2bw needs. figwig bam2bw runs on Linux and macOS, since pysam does not support Windows.
  • figwig[cuda] adds cupy (for CUDA 12; install cupy-cuda13x or cupy-cuda11x otherwise), with which FigWigReader.read(..., device="cuda") lands windows in GPU memory, for torch.from_dlpack and its JAX and TensorFlow equivalents.
pip install "figwig[fast]"
pip install "figwig[isal]"
pip install "figwig[bam2bw]"
pip install "figwig[cuda]"

Development install

git clone https://github.com/jmschrei/figwig.git
cd figwig
uv sync --extra dev
uv run pytest -n 8    # the tests on 8 cores; the first run compiles the kernels

Claude Code Skill

figwig ships a Claude Code skill that teaches a coding agent to use figwig in any project: reading windows, writing predictions, intervals and counts, converting reads with figwig bam2bw, figwig files and their converters, the rules a write follows, what each error means, and how much threads gain. Install it into ~/.claude/skills/figwig with:

figwig install-skill

After upgrading figwig, run figwig install-skill --force to replace the installed copy, which it otherwise leaves, with a message. -d DIRECTORY installs into another skills directory, and --symlink links to the copy inside the installed package instead of copying it.

Python API

The examples read two public ENCODE bigWigs: the ATAC-seq signal of HG02943 (ENCFF830RWF, 238 MB) and the DNase-seq signal of dorsolateral prefrontal cortex (ENCFF989SAK, 645 MB).

wget https://www.encodeproject.org/files/ENCFF830RWF/@@download/ENCFF830RWF.bigWig
wget https://www.encodeproject.org/files/ENCFF989SAK/@@download/ENCFF989SAK.bigWig

Reading windows

import numpy
from figwig import BigWigReader

bw = BigWigReader("ENCFF830RWF.bigWig")
chroms = numpy.array(["chr1", "chr1", "chr2"])
starts = numpy.array([1_000_000, 2_500_000, 300_000])
y = bw.read(chroms, starts, width=1000, n_jobs=8)
print(y.shape, y.dtype)
# (3, 1000) float32
print(y.sum(axis=1))
# [116.  43.  10.]

Window j covers [starts[j], starts[j] + width) on chroms[j], 0-based as in BED files, and is written into row j, whatever order the windows are given in. A base that no interval covers is missing, 0 unless given, and a base past the end of its chromosome is NaN, or oob if given. A BigWigReader reads the file's index once and keeps it, and can be pickled to DataLoader workers.

Reading several bigWigs into one array

import numpy
from figwig import read_bigwig

chroms = numpy.array(["chr1", "chr1", "chr2"])
starts = numpy.array([1_000_000, 2_500_000, 300_000])
y = read_bigwig(["ENCFF830RWF.bigWig", "ENCFF989SAK.bigWig"], chroms, starts,
	width=1000)
print(y.shape)
# (3, 2, 1000)

Channel i holds the values from the i-th file, in the (batch, channels, length) layout that sequence models take, and every file's work shares one pool of threads.

Writing windows

import numpy
from figwig import read_bigwig
from figwig import write_bigwig

chrom_sizes = {"chr1": 248_956_422, "chr2": 242_193_529}
chroms = numpy.array(["chr2", "chr1", "chr1"])
starts = numpy.array([300_000, 2_500_000, 1_000_000])
y_hat = numpy.random.default_rng(0).random((3, 2, 1000), dtype=numpy.float32)

write_bigwig(["plus.bw", "minus.bw"], chrom_sizes, chroms, starts, y_hat)
y = read_bigwig(["plus.bw", "minus.bw"], chroms, starts, width=1000)
print(numpy.array_equal(y, y_hat))
# True

write_bigwig takes what read_bigwig gives, such as a model's predictions for the same windows. A base whose value is missing, 0.0 unless given, or NaN is left out of the file, so that reading with the same missing gives the values back.

Writing a batch at a time

from figwig import BigWigReader
from figwig import BigWigWriter

chrom_sizes = {"chr1": 248_956_422, "chr2": 242_193_529}
with BigWigWriter("example.bw", chrom_sizes) as writer:
	writer.write("chr1", [1000, 2000], [0.5, 2.0], ends=[1500, 2100])
	writer.write("chr1", [5000, 5003, 5004], [[3], [1], [1]])
	writer.write("chr2", [10_000], [[0.0, 1.5, 1.5, 0.0, 2.5]])

y = BigWigReader("example.bw").read(["chr1", "chr2"], [5000, 10_000], width=6)
print(y)
# [[3.  0.  0.  1.  1.  0. ]
#  [0.  1.5 1.5 0.  2.5 0. ]]

BigWigWriter writes intervals (with ends), single bases (windows of width

  1. and windows as they come, so a track larger than memory can be written a batch at a time. Each call's items come after the last call's, in the order of chrom_sizes. The header, index and zoom levels are written when the writer is closed.

Reading and writing figwig files

import numpy
from figwig import FigWigReader
from figwig import FigWigWriter

chrom_sizes = {"chr1": 248_956_422, "chr2": 242_193_529}
tracks = [
	{"name": "atac.+", "assay": "ATAC-seq", "strand": "+"},
	{"name": "atac.-", "assay": "ATAC-seq", "strand": "-"},
]
counts = numpy.random.default_rng(0).poisson(0.1, (3, 2, 2114))
chroms = numpy.array(["chr2", "chr1", "chr1"])
starts = numpy.array([300_000, 2_500_000, 1_000_000])

with FigWigWriter("example.fw", chrom_sizes, tracks, genome="hg38") as fw:
	fw.write(chroms, starts, counts)

fw = FigWigReader("example.fw")
y = fw.read(chroms, starts, width=2114, tracks=fw.select(strand="+"))
print(y.shape, y.dtype, numpy.array_equal(y[:, 0], counts[:, 0]))
# (3, 1, 2114) uint8 True

A figwig file holds many tracks over one genome, each with attributes that select matches, and is read like a bigWig: many windows of one width in one call, from any subset of the tracks, into (windows, tracks, width) in the tracks' own dtype. The same read gives the nonzero bases as a sparse list with sparse=True, the tracks summed per label with labels=, or each window summed per bin with bins=; sums gives each track's total over intervals of any lengths; tile_sums gives each track's sum per tile, or per bin of a zoom level, without decoding a chunk; and device="cuda" lands the output of read and sums in GPU memory, where torch.from_dlpack takes it without a copy. A base no track wrote is 0, as is a bigWig's no-data unless bw2fw --missing nan keeps them apart.

A file is written in position order, for reading windows across many tracks, or in track order, for reading a few tracks over the genome, and figwig reorder rewrites it in the other; write_sparse takes nonzero bases as entries, such as the cut sites of single cells. fw.tracks, fw.stats and fw.chrom_sizes say what a file holds, as figwig info does from the shell. The first read of each kind in a new environment compiles its kernels, with nothing printed meanwhile; numba caches them beside the package, or in the user's cache directory when the package directory is read-only, and NUMBA_CACHE_DIR chooses another.

The figwig files guide covers the layout, the API and the converters, and the format page specifies the format byte by byte, with a Kaitai Struct definition and a reference reader. The reading and writing guides cover values and coordinates, chromosomes a file does not have, PyTorch data loaders, binned values, the files figwig refuses and falling back to another reader, compression, zoom levels, and files identical to pyBigWig's.

Command line

figwig bam2bw

figwig bam2bw my.bam -s hg38.chrom.sizes -n test-run -p 8                  # test-run.+.bw, test-run.-.bw
figwig bam2bw fragments.tsv.gz -s hg38.chrom.sizes -n test-run -f -u -p 8  # test-run.bw

figwig bam2bw is bam2bw 0.5.1 with the same arguments, output files and messages, except that -p is a number of cores, 1 unless given, -c is new, and a BAM whose header disagrees with the sizes is reported on stderr first. By default it counts the 5' end of every mapped read at each base and writes the two strands to two bigWigs; -u writes one, -f counts both ends of each fragment, -3p the 3' ends, -c every aligned base of each read (the coverage deepTools' bamCoverage gives), -ps, -ns, -sf and -r shift and scale the counts, and -z writes zoom levels. SAM files, --mate_pairs, remote files and malformed records go through bam2bw's own pysam or line loop, so that their result, or their error, is bam2bw's. The bigWigs hold bam2bw's entries but not its bytes. figwig bam2bw -h lists the arguments, and the documentation says how files are read and shared across cores.

figwig bw2fw, fw2bw, bam2fw, fiberbam2fw, reorder, merge and info

figwig bw2fw -o k562.fw -m tracks.tsv -g hg38 -p 8 atac.+.bw atac.-.bw dnase.bw   # bigWigs -> k562.fw
figwig fw2bw k562.fw -o bigwigs -p 8 --select assay=ATAC-seq                      # tracks -> bigwigs/<name>.bw
figwig bam2fw a.bam b.bam -s hg38.chrom.sizes -n k562 -m files.tsv -p 8          # reads -> k562.fw
figwig bam2fw fragments.tsv.gz -s hg38.chrom.sizes -n pbmc --barcode-column 4 --cells cells.tsv -f -p 8   # a track per cell
figwig fiberbam2fw fibers.bam -n fibers -g hg38 -p 8                               # Fiber-seq reads -> a track per read and channel
figwig reorder k562.fw by_track.fw --order tracks                                # the other order
figwig merge -o all.fw k562.fw hepg2.fw                                          # files -> one file
figwig info k562.fw                                                              # what the file holds

figwig bw2fw writes bigWigs as the tracks of one figwig file, named after their files or by a metadata table whose other columns become the tracks' attributes, or adds them to a file with --append; figwig fw2bw writes a figwig file's tracks back as bigWigs with such a table; figwig bam2fw takes figwig bam2bw's arguments and writes each input file as its own pair of stranded tracks, or a fragments file as one track per cell with --barcode-column; figwig fiberbam2fw writes a Fiber-seq BAM as five tracks per read (m6A and 5mC probabilities, nucleosomes, MSPs, FIRE quality) with the read's alignment and tags as attributes; figwig reorder rewrites a file in the other order; figwig merge copies several files' tracks into one; and figwig info prints a file's tracks, attributes and statistics. The converters page has the options.

figwig install-skill

Installs the Claude Code skill; see above.

Origin

figwig began as the bigWig reader inside tangermeme's extract_loci (PR #107). It was checked there against pybigtools, bit for bit, on the training and validation sets of 92 Cherimoya models and on 3,131 synthetic cases.

Metadata

Release files for figwig 0.3.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for figwig 0.3.1
File Size Uploaded Provenance Actions
figwig-0.3.1.tar.gz 523.1 kB Without provenance Details for figwig-0.3.1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for figwig 0.3.1
File Interpreter ABI Platform Actions
figwig-0.3.1-py3-none-any.whl Python 3 none any Details for figwig-0.3.1-py3-none-any.whl

Total release size: 839.0 kB

Release files / figwig-0.3.1.tar.gz

Download URL figwig-0.3.1.tar.gz
Size 523.1 kB
Tags Source
SHA-256 checksum
How to use checksums
dd829905bccc63c13c90882ee1605fa6b2fc823da40a9b93370a50106914098a
BLAKE2b-256 checksum
How to use checksums
3608aeefc0127b2a15a9b940de0a4d4e04df392b00089e3c3feb5ec94168bc1a
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.5

Release files / figwig-0.3.1-py3-none-any.whl

Download URL figwig-0.3.1-py3-none-any.whl
Size 315.9 kB
Tags Python 3
SHA-256 checksum
How to use checksums
5927df3744cf3fc386885012adbae19df66aa7b663f8d444600b4d9c9aa66d8b
BLAKE2b-256 checksum
How to use checksums
0f68d898ff7905d389803933e81fdbc3d4f9da5f29d7a3ff20ba21bd97a69bbe
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.5

Security for figwig 0.3.1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page