ReconEval

Benchmark for gene-expression reconstruction from single-cell latent representations. The package contains:

• a metrics API that scores a (true, reconstructed) anndata.AnnData pair (statistical, biological, perturbational);

• four tutorial notebooks covering the three benchmark settings;

• training and eval pipelines under experiments/ with Hydra configs and sbatch wrappers;

• analysis notebooks under analysis/ that reproduce the paper figures from cached metric CSVs.

Install

pip install -e . for the metrics API; conda env for the full benchmark.

Installation

Tutorials

Metrics, end-to-end, foundation-model and latent-shift reconstruction notebooks.

Tutorials

API reference

Public functions and classes in sc_reconstruction.

API reference

Experiments

Training drivers, eval scripts, sbatch wrappers and Hydra configs per task.

Experiments

Reproducibility

Analysis notebooks that reproduce the paper figures.

Reproducibility

Manuscript

Preprint: TBD.

https://example.org/preprint