---
license: mit
tags:
- biology
- single_cell
- deep_neural_networks
- benchmark
pretty_name: scREF, all cells
---

# scREF

This dataset contains human single cell RNA-sequencing (scRNA-seq) data collected from 46 studies and standardized 
by Diaz-Mejia JJ et al. (2025) for the paper [Benchmarking and optimizing organism wide single-cell RNA alignment methods](https://arxiv.org/abs/2503.20730) presented at the LMRL Workshop at the International Conference on Learning Representations (2025).

* Folder `Phenomic-AI/scref_ICLR_2025/zarr` contains standardized single-cell RNA data for each study in `zarr` format.
* Sub-folder names show: `{first author, last name}_{journal}_{year}_{Pubmed ID}`.
* `zarr` files can be loaded as AnnData objects in Python with [Dask + Zarr](https://anndata.readthedocs.io/en/latest/tutorials/notebooks/%7Bread%2Cwrite%7D_dispatched.html)
* Cell-metadata includes an `obs` slot with columns:
    - `barcode`: unique cell identifier
    - `authors_celltype`: original author cell type annotations
    - `standard_true_celltype`: cell type annotations standardized across studies
    - `sample_name`: unique sample identifier
    - `tissue_collected`: tissue where the sample was collected from
    - `included_scref_train`: boolean indicating if the cell was included in downsampled training and benchmark analyses.
* Code to compute Batch Adversarially trained single-cell Variational Inference (BA-scVI) is available at https://github.com/PhenomicAI/bascvi