rsfgsea

High-performance Rust implementation of preranked Gene Set Enrichment Analysis (GSEA), designed as a drop-in, optimized alternative to the R fgsea package.

What it focuses on:

• fgsea-compatible simple and multilevel workflows
• library and CLI use from one Rust crate
• deterministic parity-focused CPU path
• optional hybrid GPU acceleration for large simple-stage screening

Headline results from the main project benchmarks:

• local representative Criterion benchmark, simple: 2.282 s for 10k genes, 1k pathways, 10k permutations
• local representative Criterion benchmark, multilevel: 3.438 s for 10k genes, 1k pathways, nPermSimple=1000
• multilevel, small workload, 1 worker: 2 ms vs R 42 ms (21.0x faster)
• multilevel, large workload, 16 workers: 105 ms vs R 977 ms (9.3x faster)
• simple, small workload, 1 worker: 720 ms vs R 2597 ms (3.6x faster)
• simple, large workload, 16 workers: 674 ms vs R 798 ms (1.18x faster)
• real muscle-comparison validation workload: 81 MB peak RSS vs R 329 MB peak RSS (4.1x lower)

Run the local optimization benchmark with:

cargo bench -p rsfgsea --bench gsea_bench

Current CPU multilevel parity vs R is near floating-point noise in this repo's validation setup:

• max |ES| diff: 4.988e-09
• max |NES| diff: 4.983e-09
• max |pval| diff: 4.975e-09
• max |padj| diff: 4.965e-09

Install

Library:

[dependencies]
rsfgsea = "0.3.4"

CLI:

cargo install rsfgsea

Minimal Rust Example

use rsfgsea::prelude::*;

let ranks = RankedList::new(
    vec!["g1".into(), "g2".into(), "g3".into(), "g4".into()],
    vec![2.0, 1.0, -1.0, -2.0],
);

let pathways = vec![
    Pathway {
        name: "PW_A".into(),
        description: None,
        genes: vec!["g1".into(), "g2".into()],
    },
];

let results = fgsea(
    &ranks,
    &pathways,
    None,
    1000,
    42,
    1,
    ranks.len() - 1,
    1e-50,
    ScoreType::Std,
    1.0,
);

println!("{}", results[0].pathway_name);

CLI

rsfgsea \
  --ranks data/example.rnk \
  --gmt data/pathways.gmt \
  --output results.tsv

For deeper usage, benchmarks, parity notes, and reproducibility details, see the main repository: https://github.com/deminden/rsfgsea