bandnorm-rs 0.1.0

Rust port of BandNorm's scGAD + bandnorm normalization for single-cell Hi-C — bit-equivalent to the R reference, ~10-30× faster.

rust-bandnorm

Rust port of BandNorm (Shen, Zheng, Keleș 2022) — single-cell Hi-C scGAD scoring and band-normalisation for cell-type discovery and downstream embedding.

This is a separate package that re-implements the algorithm in Rust; it is not a fork of the upstream R tree. The original R implementation continues to be available at https://github.com/sshen82/BandNorm.

What it computes

Two functions, mirroring BandNorm/R/scGAD.R and BandNorm/R/bandnorm.R:

import bandnorm_rs as bn

# scGAD: gene × cell matrix from contact pairs.
gad = bn.scgad(hic_df, genes, resolution=10_000, depth_norm=True)

# BandNorm: per (chrom, diagonal, cell) depth-normalisation of contact counts.
norm_df = bn.bandnorm(hic_df)

Both are bit-equivalent (within f64 ε) to the R reference and faster by ~10-30× depending on the dataset.

Algorithm parity

tests/test_parity.py compares the Rust output against an R reference generated via pure GenomicRanges (no GenomicInteractions dependency) on the official BandNorm-shipped scGADExample.rda + mm9Annotations data: 100 cells × ~1.3 M contact pairs × 10180 genes.

Tolerance: max |Rust − R| < 1e-10 on raw counts (integer-exact), < 1e-6 on Z-scored GAD output.

Install

git clone https://github.com/omicverse/rust-bandnorm
cd rust-bandnorm
maturin develop --release

PyPI release coming soon (pip install bandnorm-rs).

Layout

rust-bandnorm/
├── pyproject.toml
├── rust/Cargo.toml
├── rust/src/lib.rs              # ~250 LoC: gene index + scGAD + bandnorm
├── python/bandnorm_rs/        # Python wrapper
└── tests/
    ├── reference_scgad.R        # R reference generator
    └── test_parity.py           # pytest parity vs R

License

MIT.