bixverse-rs

Description

A Rust crate for computational biology: statistics, gene set enrichment, graph algorithms, matrix factorisation, ontologies and a full single cell suite. It is the numerical core underneath the bixverse R package, but it has no R dependency of its own and can be used as a plain Rust library.

The code originally lived in src/Rust/ of that R package. Pulling it out into its own crate means the numerics get tested, benchmarked and versioned independently of the wrapper that happens to call them. Design constraint throughout: atlas-scale analysis on a laptop, not on a cluster.

What's in it

Well over a hundred named methods, plus the shared numerical scaffolding they sit on.

Domain	Methods
`core`	Correlations and pairwise similarity, PCA/SVD, randomised SVD, LOESS, RBF kernels, mutual information and entropy, scalar optimisation (R's `optimize`), linear mixed models (REML with Satterthwaite), E-distance and perturbation distances, sparse structures, synthetic data generators
`enrichment`	GSEA (fgsea multi-level), blitzGSEA (gamma-approximated null), GSVA, ssGSEA, singscore, mitch, over-representation
`graph`	Louvain, WalkTrap, spectral clustering, label propagation, personalised and constrained PageRank, connected components, Dijkstra (single and multi-source), Kruskal spanning forests, PAGA-style graph abstraction, graph metrics
`methods`	NMF (bulk, dense HALS, sparse HALS, consensus, refit), ICA, LDA via variational Bayes, sparse multiple CCA, differential correlation, graph diffusion, SNF, RBH, CoReMo, dgRDL, cis-target; bulk DGE via edgeR's quasi-likelihood chain and limma-voom (feature `dge`)
`ml`	k-means, clustering metrics, landmark Gaussian process regression, Matern kernels
`ontology`	GO Elim, Wang and Resnik-style semantic similarity
`single_cell`	I/O for h5ad, 10x h5, mtx, R/Seurat counts and a versioned binary sparse format (single and multi-file, mmap-backed, mergeable); QC, CellSweep ambient removal, HVG, scTransform v2 (per-sample fits, covariate regression, analytic Pearson residuals, corrected counts), PCA, kNN, SNN, MAGIC; doublet detection (Scrublet, scDblFinder, cxds); batch correction (Harmony v1 and v2, BBKNN, fastMNN, Seurat CCA/rPCA anchors); integration metrics (kBET, iLISI/cLISI, silhouette, PC regression); annotation (scType, Symphony); DGE (Mann-Whitney with AUROC, pseudobulk edgeR, NEBULA); analysis (SCENIC with regulon binarisation, AUCell, Hotspot, VISION, DIALOGUE, MELD, miloR, NicheNet, NMF, module scoring, fast k-means/Louvain clustering); meta cells (SEACells, MetaCells2, SuperCell, hdWGCNA, plus density and compactness metrics); trajectories (Palantir, PAGA, diffusion maps, Markov chains, gene trends); Bonsai trees from Sanity posteriors (feature `bonsai`); multi-modal (WNN, DSB)
`gpu`	Sparse randomised SVD, SpMM and sparse GEMM, skinny GEMM, Gram, CholeskyQR2, correlation, PCA, kNN, NMF and consensus NMF, Harmony, BBKNN, SCENIC, Scrublet, SEACells, NEBULA, Sanity for Bonsai, fast clustering

Heavy lifting goes through faer for dense linear algebra (GEMM through Apple Accelerate on macOS), rayon for CPU fan-out, wide for SIMD, and cubecl for GPU kernels. Vector search, distance metrics and k-means come from the sister crate ann-search-rs; the GPU primitives layer from cubecl-utils-rs; the edgeR, limma and NEBULA numerics from edge-rs.

Feature flags

Everything past the base statistics and graph code is gated, so a consumer only pays for what it uses.

Flag	What it turns on
(always on)	Bulk statistics, enrichment, graph, ontology, matrix factorisation
`accelerate`	On by default. Dense CPU GEMM through Apple Accelerate on macOS, no effect elsewhere. See below
`single-cell`	The `single_cell` module, HDF5 and mmap I/O, the binary sparse format
`multi-modal`	`single_cell::multi_modal` (WNN, ADT), implies `single-cell`
`dge`	Negative binomial DGE via `edge-rs`: bulk edgeR and limma-voom. Implied by `single-cell`, usable on its own
`hdf5-static`	Builds the bundled HDF5 from source instead of linking an external `libhdf5`
`hdf5-filters`	Compiles the LZF and Blosc filters into HDF5, so h5ad files written with those codecs load
`bonsai`	Bonsai trees over Sanity posteriors via `bonsai-rs`, implies `single-cell`
`gpu`	The `gpu` module via `cubecl` (wgpu and CPU backends) and `cubek`
`large-test`	Slow but asserting tests: GPU parity gates, large-scale numerical checks
`large_scale_diagnostics`	Unasserted diagnostic sweeps that print tables for a human

macOS and Apple Accelerate

On macOS the default accelerate feature sends every dense CPU GEMM (correlations, covariance, perturbation distances, NMF HALS, Hotspot, Markov chains, the landmark GP) through Apple's Accelerate framework. It ships with every macOS, so there is nothing to install. Turn it off with default-features = false. End to end on an M1 Max:

function	f32	f64
`column_pairwise_cov`, 50k x 2k	2.3x	1.8x
`column_pairwise_cor`, 50k x 2k	1.9x	1.7x
dense NMF HALS, 20k x 2k, k = 10	1.3x	1.6x
dense NMF HALS, 20k x 2k, k = 50	1.8x	1.8x

Using it from Rust

[dependencies]
bixverse-rs = { version = "0.5", features = ["single-cell"] }

use bixverse_rs::prelude::*;

The prelude re-exports the errors, sparse structures, SparseGraph, matrix and vector helpers, the SIMD trait and the assertion macros.

How it plugs into R

The R side is a thin wrapper, and deliberately so. bixverse depends on this crate through extendr, and does the things R is good at: argument checking, S7 classes, documentation, plotting handoff. The Rust side owns the maths.

Each top-level module carries a *_r_wrapper.rs sibling holding the R conversions: from_r_list deserialisers for the parameter structs and a handful of serialisers back to R lists. The #[extendr] entry points themselves live downstream in the R packages, so R types never touch the numerical surface.

Two R packages consume the crate:

bixverse with single-cell and multi-modal. The main package.
bixverse.gpu with gpu and single-cell, adding wgpu-backed versions of the expensive methods.

Anything on the Rust side that never crosses into R works standalone, so the crate is equally usable from a pure Rust binary or behind a PyO3 layer.

Roadmap

Methods

NMF for dense and sparse data
NicheNet for single cell, see Browaeys et al..
Palantir for single-cell trajectories, see Setty, et al..
Slingshot for single-cell trajectories, see Street, et al..

Differential expression

Bulk edgeR quasi-likelihood and limma-voom via edge-rs
Pseudobulk and NEBULA for single cell
limma treat/topTreat, decideTests, and block/correlation in voom

GPU accelerations

GPU-accelerated sparse, randomised SVD
GPU-accelerated Harmony
GPU-accelerated correlations (Metal does not have the absolute greatest performance here unfortunately...)
SEACells with GPU acceleration.
GPU-accelerated BBKNN version
GPU-accelerated NEBULA

Python interface

Data loader to pull the counts out of the binary files and expose them to Python-based deep learning frameworks (JAX, PyTorch).

Updates

Updates on what's happening in this crate can be found here

Licence

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

bixverse-rs 0.6.3