scCCVGBen

A 200-dataset single-cell representation-learning benchmark and a Python package for the centroid-coupled variational graph attention autoencoder (100 scRNA-seq + 100 scATAC-seq, balanced).

The repository carries:

• the dataset catalogue with GEO-verified metadata,
• the method and metric registries,
• the Python package used to regenerate the benchmark,
• the Hugo source for the public site,
• the Next.js companion explorer (webapp/).

Install

pip install -e ".[dev]"

Reproduce

See REPRODUCE.md. Sweep scripts are idempotent and resume per (dataset, method) row.

Layout

• scccvgben/ — Python package
• scripts/ — data curation and sweep runners
• data/ — dataset metadata CSVs
• figures/ — generated figures (raw screenshots in figures/site_shots/ stay local)
• site/ — Hugo source for the public site
• webapp/ — Next.js benchmark explorer
• tests/ — smoke tests

Figures

• Figure 1 — site and benchmark overview. Render with scripts/make_figure1_site.py.
• Figure 2 — model architecture and benchmark axes. Render with scripts/make_figure2_model_architecture.py.
• Generated figure binaries are not committed by the figure-pipeline PR.

Sites

• Hugo public site: https://peterponyu.github.io/scCCVGBen/
• Next.js companion explorer: https://peterponyu.github.io/scccvgben-next/

License

MIT — see LICENSE.