Lab Adoption¶

This page is for collaborators deciding whether RustScenic is worth testing on their own data.

When To Try It¶

RustScenic is worth testing if one of these problems is blocking your analysis:

• pySCENIC, arboreto, pycisTopic or pycistarget will not install cleanly on a modern Python stack.
• AUCell, cisTarget, enhancer linking or GRN inference is becoming a runtime or memory bottleneck.
• You need deterministic CPU execution under a fixed seed.
• You want a single Python package covering GRN, AUCell, motif support, topics, enhancer links and eRegulons.

Public Proof Points¶

Use these when introducing the package to a lab or collaborator:

Proof point	Evidence
Released	v0.4.7 on PyPI, with Python 3.10 to 3.13 release wheels.
Faster tested path	11x to 52x faster than SCENIC+ in controlled real-data core E2E rows.
Lower atlas-scale memory pressure	6.34 GB peak RSS on a 100k-cell four-stage scale check; legacy pySCENIC reports exceed 40 GB on similar workloads.
Lab artefact	Collaborator human brain GEM-X full monolith run recovered 16 of 17 expected brain TFs.

Minimal Adoption Test¶

Run one small dataset first:

pip install rustscenic
python -m rustscenic.quickstart

Then run your own AnnData object through GRN plus AUCell:

import anndata as ad
import rustscenic.data
import rustscenic.grn
import rustscenic.aucell

adata = ad.read_h5ad("your_data.h5ad")
tfs = rustscenic.data.tfs("hs")
grn = rustscenic.grn.infer(adata, tf_names=tfs, n_estimators=500, seed=777)

Evidence To Record¶

For a useful adoption report, record:

• Dataset name and source.
• Number of cells, genes, peaks and transcription factors.
• RustScenic version.
• Python version and operating system.
• Command or script.
• Wall time and peak memory if available.
• Output checks: number of GRN edges, regulons, AUCell matrix shape, non-empty motif results.
• Biological checks: expected marker TFs and their expected cell types.

What Counts As A Strong Adoption Result¶

A strong report does not need to prove RustScenic is better than every reference tool. It should prove:

• The tool installs.
• It runs on real data without manual source edits.
• Outputs are non-empty and shape-correct.
• Known biological controls are recovered or failures are explained.
• The command is reproducible by another person.

Best First Dataset¶

Use a small real dataset before a full atlas. The most useful first adoption run has enough cells to show biological signal, but not so many that debugging becomes a compute problem. A good first target is 1,000 to 5,000 cells with a known set of marker TFs.