---
license: cc-by-4.0
tags:
  - biology
  - transcriptomics
  - single-cell
  - bulk-rna-seq
---

# OmicsFM transcriptomics datasets

Bulk and single-cell expression matrices used to train and evaluate
[OmicsFM](https://huggingface.co/rednaSander/omicsfm), prepared as model-ready
AnnData from public databases.

The proteomics modality lives separately in
[omicsfm-data-proteomics](https://huggingface.co/datasets/rednaSander/omicsfm-data-proteomics).

## Naming

Every dataset is `<modality>_<identifier authority>`, so the name states both
what was measured and which identifiers index the features:

| suffix | authority | example |
|---|---|---|
| `_uniprot` | UniProt accessions | `A0A024RBG1` |
| `_hgnc` | HGNC gene symbols | `TSPAN6` |
| `_ensembl` | Ensembl gene IDs | `ENSG00000237491` |

The `_uniprot` datasets share one 20,272-accession vocabulary with the
proteomics data, so a feature index means the same protein across every
modality. That shared space is what the models consume.

The gene-space datasets are the native measurement spaces, kept for comparison
against methods that operate on genes. Bulk and single-cell use **different**
gene identifiers, so those two are not comparable feature by feature; only the
`_uniprot` versions are.

## Contents

| dataset | split | samples | features |
|---|---|---:|---:|
| `bulk_transcriptomics_uniprot` | train | 614,169 | 20,272 |
| `bulk_transcriptomics_uniprot` | valid | 34,945 | 20,272 |
| `bulk_transcriptomics_uniprot` | test | 31,102 | 20,272 |
| `bulk_transcriptomics_hgnc` | train | 614,169 | 67,186 |
| `bulk_transcriptomics_hgnc` | valid | 34,945 | 67,186 |
| `bulk_transcriptomics_hgnc` | test | 31,102 | 67,186 |
| `sc_transcriptomics_uniprot` | train | 3,546,382 | 20,272 |
| `sc_transcriptomics_uniprot` | valid | 361,037 | 20,272 |
| `sc_transcriptomics_uniprot` | test | 642,687 | 20,272 |
| `sc_transcriptomics_ensembl` | train | 3,546,382 | 61,497 |
| `sc_transcriptomics_ensembl` | valid | 361,037 | 61,497 |
| `sc_transcriptomics_ensembl` | test | 642,687 | 61,497 |

Splits are grouped so that no study spans a boundary: bulk by ARCHS4 series,
single-cell by Census dataset. Assignment is deterministic given seed 42.

## Sources

| modality | source |
|---|---|
| bulk | [ARCHS4](https://archs4.org) human gene-level counts (kallisto, raw estimated counts) |
| single-cell | [CELLxGENE Census](https://chanzuckerberg.github.io/cellxgene-census/) |

These are model-ready derivatives, not the primary data: quality filtered,
harmonised to a controlled vocabulary, and projected onto the shared UniProt
space. The construction pipeline is in the OmicsFM repository under
`transcriptomics/`.

## Use

```python
from huggingface_hub import hf_hub_download
import anndata as ad

path = hf_hub_download("rednaSander/omicsfm-data",
                       "sc_transcriptomics_uniprot/test.h5ad",
                       repo_type="dataset")
adata = ad.read_h5ad(path)
```

Tokenisation caches are not included: `omicsfm.data.ExpressionDataset` rebuilds
them on first use.

## Citation

Publication in preparation. Please also cite ARCHS4 and CELLxGENE Census.