Tool: rsfgsea: Rust implementation of preranked GSEA built for large transcriptome workflows where GSEA became the bottleneck

Hi,

I am sharing rsfgsea, a Rust implementation of preranked GSEA built for a large transcriptome analysis workflow where GSEA had become a bottleneck.

The goal was not to replace fgsea as a method, but to provide a faster implementation while staying as close as possible to fgsea results. rsfgsea has Rust crate, CLI, Python and R bindings. Max absolute difference from R fgsea is around 5e-9 for ES, NES, p-value, and padj. Benchmark speedups from about 3x to 21.5x, depending on workload. Also it has experimental hybrid CPU/GPU path.

Install Python binding

pip install rsfgseapy

Minimal Python usage:

import rsfgseapy

res = rsfgseapy.run_gsea_py(
    ranks=ranks,
    gmt_path="pathways.gmt",
)

I would be interested in feedback on:

whether this looks useful beyond a single pipeline
whether the parity and benchmark results are convincing
what would be important to see before a small software paper
whether the hybrid CPU/GPU direction is interesting in practice

Comments from users of GSEA in real workflows would be especially helpful.

rust python gsea tool fgsea

• 723 views • link

written 5 months ago by Denis Demin •

1 answer

I think it would be good to have some awareness that pre-ranked GSEA is very different in statistical behaviour from the original published GSEA method, which was based on sample permutation. There have never been any publications supporting the pre-ranked GSEA statistical method, and it is known in some circles to give spectacularly inflated statistical significance. You say on your github page that rsfgsea and fgsea use a "robust statistical model", but I do not think there is any published literature supporting that remark. Robust to what and according to whom? We have shown in unpublished work that pre-ranked GSEA is spectacularly non-robust to inter-gene correlations https://genome-2026.p.asnevents.com.au/days/2026-02-16/abstract/133160.

See https://support.bioconductor.org/p/9158100/ for more discussion.

Try out this experiment: take any large human expression dataset. Randomly split it into two sets of samples. Use fgsea or rsfgsea to do a pathway analysis between the two random sets.

You discuss accuracy in terms of absolute differences, but relative differences might be more relevant for very small p-values. fgsea often gives very small p-values like 1e-30, especially for larger gene sets. How close an approximation is rsfgsea in cases like that?

• 0 views • link

written 5 months ago by Gordon Smyth

Thanks, these are fair and important points.

My goal with rsfgsea is much narrower than validating preranked GSEA as a statistical framework. It is primarily a high-performance implementation of fgsea-style preranked GSEA, with a focus on numerical agreement, reproducibility, and runtime in larger workflows.

I agree that fgsea is often used in correlation-prone settings such as differential expression workflows, and I have therefore replaced “robust statistical method” with “widely used” in the GitHub description.

On the implementation side, I am already checking relative differences in the low-p tail.

Thanks again for the comment and for pointing to these alternatives.

• 0 views • link

written 5 months ago by Denis Demin •

I rechecked this on synthetic low-p benchmark cases spanning approximately 1e-20 down to 1e-125, with matched seed handling on both sides, and in that setting the p-values were effectively identical to machine precision.

• 0 views • link

written 5 months ago by Denis Demin •