scenic-rs 0.3.0

Rust implementation of the SCENIC single-cell GRN pipeline (GRN + cisTarget + AUCell)

scenic-rs

A memory-efficient Rust implementation of the pySCENIC single-cell gene regulatory network (GRN) pipeline. Not associated with the lab, just built this as a quick project!

Benefits

• Implements the same algorithms as pySCENIC (GRNBoost2, GENIE3, AUCell, CTX trimming)
• Can use modern numpy/pandas in your environment
• Memory usage is constant with increased parallelism, allowing for faster execution without OOM

Requirements

Runtime (using scenic-rs)

• Python ≥ 3.9
• numpy, pandas
• scanpy (optional) — only if you use it to load your data (Cell Ranger / .h5ad); scenic-rs itself just takes a numpy matrix + gene/TF names
• For the ctx step only: the cisTarget ranking databases (*.genes_vs_motifs.rankings.feather) and the motif2TF annotations (.tbl), from resources.aertslab.org/cistarget

From source (optional — only for development; pip install uses prebuilt wheels)

• A recent stable Rust toolchain (cargo/rustc) — ≥ 1.70 (tested on 1.86); install via rustup.rs.
• maturin ≥ 1.0 (pip install maturin), then maturin develop --release

Benchmarks / validation only (optional)

• A separate pySCENIC environment to compare against: python3.10 -m venv ~/venvs/pyscenic_clean && ~/venvs/pyscenic_clean/bin/pip install "numpy<1.24" "pandas<2" pyscenic
• matplotlib + scipy for the plotting and validation scripts

Install

pip install scenic-rs   # prebuilt wheels: Linux / macOS / Windows, Python >=3.9

scenic-rs depends only on (unpinned) numpy and pandas, so it drops into an existing environment — including a scanpy env — without forcing a downgrade.

conda — a ready-made environment that includes scanpy:

conda env create -f environment.yml   # then: conda activate scenic-rs

Docker / Nextflow / Singularity — a version-pinned image is published per release for reproducible pipelines:

docker pull ghcr.io/nglaszik/scenic-rs:latest      # or a pinned tag, e.g. :0.1.1

// In a Nextflow process, just point at the image:
process scenic {
    container 'ghcr.io/nglaszik/scenic-rs:0.1.1'
    // ... your script that `import scenic_rs`
}

Usage

scenic-rs takes a cells × genes float32 matrix plus the gene names and a TF list — load these however you like (e.g. with scanpy). Raw counts; do your own cell/gene QC first. From Cell Ranger output:

import numpy as np, scanpy as sc

adata = sc.read_10x_mtx("sample/outs/filtered_feature_bc_matrix")   # or sc.read_10x_h5(".../filtered_feature_bc_matrix.h5")
adata.var_names_make_unique()
sc.pp.filter_genes(adata, min_cells=3)                              # basic QC (filter cells/genes as you see fit)
X = np.asarray(adata.X.todense() if hasattr(adata.X, "todense") else adata.X, dtype="float32")
gene_names = adata.var_names.tolist()

# ...or from an existing AnnData .h5ad:
# adata = sc.read_h5ad("counts.h5ad")
# X = np.asarray(adata.layers["counts"].todense(), dtype="float32"); gene_names = adata.var_names.tolist()

# TFs = a TF list (e.g. allTFs_hg38.txt from aertslab) intersected with the genes present
tf_names = [g for g in open("allTFs_hg38.txt").read().split() if g in set(gene_names)]

Then run the pipeline:

from scenic_rs import grnboost2, genie3, aucell, RankingDb, ctx

# 1. GRN: TF -> target importances  (pandas DataFrame [TF, target, importance])
adj = grnboost2(X, gene_names, tf_names)        # default — fast, OOB early stopping
# adj = genie3(X, gene_names, tf_names)         # alternative — random forest, slower, ~same result (use one)

# 2. ctx: prune co-expression modules to motif-supported regulons
dbs = [RankingDb("hg38_...genes_vs_motifs.rankings.feather", "hg38")]
regulons = ctx(adj, X, gene_names, dbs, "motifs-...hgnc-...tbl")   # -> [Regulon(...)]

# 3. AUCell: per-cell regulon activity
auc = aucell(X, gene_names, {r.name: r.genes for r in regulons})

adj matches pySCENIC's adjacencies format and ctx matches its regulon output, so any step can also be swapped in individually alongside pySCENIC.

Validation & benchmarks (vs real pySCENIC)

• Running each step (grnboost2/genie3/aucell/ctx) in both scenic-rs and pySCENIC 0.12.1 / ctxcore 0.2.0
• Uses the same inputs (pbmc3k 2700 cells × 758 genes, 300 TFs), (hg38 cisTarget DB 5876 motifs × 27015 genes + motif2tf annotations)
• Equal cores — rayon threads = pySCENIC's Dask-worker count
• Backend startup counted on both sides

Correctness — scenic-rs reproduces pySCENIC's numbers

step	concordance vs pySCENIC
GRN / GRNBoost2	Spearman 0.74, top-1000 edge Jaccard 0.56 — at the stochastic ceiling
GRN / GENIE3	Spearman 0.99, per-target median 0.99
ctx / cisTarget	per-(module,motif) NES Spearman 1.00 (max diff 1.4e-13)
AUCell	Spearman 0.98, max abs diff 0.06

GRNBoost2 is stochastic — running the algorithm twice (different seed) only agrees ~0.73 per target.

Memory & speed

• Peak memory (PSS) is constant in scenic-rs as cores increase, but grows ~linearly in pySCENIC (per-worker copies). Ratios = pySCENIC / scenic-rs:

step	memory: 16c → 96c	speed (equal cores)
GRN / GRNBoost2	13× → 26× less	2.8× faster @96c
GRN / GENIE3	30× → 123× less	equivalent
ctx / cisTarget	14× → 26× less	~10× faster
AUCell	18× → 51× less	~170× faster

At 96 cores pySCENIC peaks at ~11.7 GB (GRNBoost2), ~20.5 GB (GENIE3) and ~17 GB (ctx); scenic-rs stays at ~0.45 GB, ~0.17 GB and ~0.66 GB respectively.

Speedup on GRNBoost2 is likely due to Dask setup, relatively negligible on larger datasets.

Reproduce

Needs a pySCENIC env, e.g. python3.10 -m venv ~/venvs/pyscenic_clean && ~/venvs/pyscenic_clean/bin/pip install "numpy<1.24" "pandas<2" pyscenic:

# correctness + cached adjacencies (GRN + AUCell)
python bench/benchmark_pyscenic.py --workers 16 --genie3
# memory/time scaling sweep across cores (GRNBoost2, GENIE3, AUCell)
python bench/mem_benchmark.py --sweep 16,48,96 --genie3
# ctx scaling sweep (real hg38 DB)
python bench/benchmark_ctx.py --sweep 16,48,96
# ctx per-step parity checks (math, DB, modules, regulons, NES concordance)
PYTHONPATH=python ~/venvs/pyscenic_clean/bin/python bench/validate_ctx_regulons.py
# render figures
python bench/plot_benchmark.py        # concordance.png, grn_per_target.png
python bench/plot_mem_benchmark.py    # mem_scaling.png, time_scaling.png (all steps)

License & attribution

GPL-3.0-or-later. scenic-rs is a Rust reimplementation of, and a derivative work of, the GPL-3.0 projects pySCENIC and ctxcore (aertslab, VIB-KU Leuven) — their GRN/cisTarget/AUCell workflow and the recovery/NES/module/pruning logic. It is an independent project, not affiliated with or endorsed by aertslab. See NOTICE for details.

If you use scenic-rs, please cite the original SCENIC work:

• Aibar et al. (2017) SCENIC: single-cell regulatory network inference and clustering. Nature Methods.
• Van de Sande et al. (2020) A scalable SCENIC workflow for single-cell gene regulatory network analysis. Nature Protocols.
• Moerman et al. (2019) GRNBoost2 and Arboreto… Bioinformatics.