figwig: A Multidimensional Data Format for Genomic Machine Learning

On October 28, developer jmschreiber91 released and live-demoed a new genomics data format called figwig: it generalizes the traditional bigWig format to N-dimensional arrays, efficiently storing massive track collections, single-cell experiments, fiber-seq, and even Hi-C and scHi-C data. The core idea is merging many experiments into a single data structure, making files smaller and I/O faster. The format is open-sourced on GitHub under the MIT license.

Confirmed

• figwig does everything bigWig can do, but is designed for modern genomic machine learning: a universal kernel loads data directly onto GPUs, with compatibility for JAX, TensorFlow, PyTorch, and other frameworks
• Benchmarked performance: for 90,543 fiber-seq reads (a 62 MB file, one track per 1,024 reads), single-threaded writing took 6.8 seconds and genome-wide reads were as fast as 0.42 seconds; the author claims loading is an order of magnitude faster than bigwig
• Single-cell use case: whole-genome data for 10,246 PBMC ATAC cells fits in just 1.0 GB with position-sorted figwig files (2.4 GB track-sorted), with block reads significantly faster than per-cell bigwig
• Toolchain: fast command-line tools support bigwig↔figwig conversion, converting fragment files with cell barcodes to figwig, and bigwig-style zoom operations
• Ships with an agent-oriented agentic skill for easy automated tool invocation

Why it matters

• bigWig is the de facto standard for storing genomic tracks but cannot natively represent high-dimensional data such as single-cell, fiber-seq, or Hi-C; figwig brings these scenarios into one unified structure that plugs directly into GPUs and mainstream deep learning frameworks, providing a more efficient data layer for large-scale genomic modeling (especially ML pipelines)
• Moving single-cell data from many per-cell files to a single compact file can substantially cut storage and random-access costs

2026-10-28 ~ 2026-10-28 · 8 related posts

• figwig: new genomics format generalizes bigWig to N-dimensional arrays — jmschreiber91 · 2026-10-28
• figwig stores 10,246 single-cell ATAC profiles in 1 GB, reads chunks up to 100x faster — jmschreiber91 · 2026-10-28
• [source] figwig benchmarks: 90,543 fiber-seq reads read in 0.42s vs 27s for scipy CSR — jmschreiber91 · 2026-10-28
• Benchmark: figwig genomic data format loads GPU batches ~10x faster than bigWigs — jmschreiber91 · 2026-10-28
• More details on the genomic ML file format: sparsity and pseudobulk support — jmschreiber91 · 2026-10-28
• A bigwig-style file format built for genomic ML with direct GPU loading — jmschreiber91 · 2026-10-28
• [source] figwig format ships with CLI tools and an agentic skill for bioinformatics sparse data — jmschreiber91 · 2026-10-28

1 near-duplicate retellings: jmschreiber91