Genentech/decima-data

---
license: mit
task_categories:
- tabular-regression
tags:
- biology
- genomics
- single-cell
pretty_name: "Decima Dataset"
size_categories:
- 1M<n<10M
---

# decima-data

## Dataset Summary
This dataset contains gene expression predictions and associated genomic features formatted as an `AnnData` object. It is designed for use with the **Decima** framework to support tasks such as gene expression prediction and genomic sequence modeling. The data provides a comprehensive view of expression across various tissues, organs, and disease states, primarily centered on human brain atlas data.

For more details, please refer to the original paper: https://www.biorxiv.org/content/10.1101/2024.10.09.617507v3.

## Dataset Structure
The dataset is an `AnnData` object with dimensions: **8,856 observations (pseudobulks) × 18,457 variables (genes)**.

### Data Fields

**In `.obs` (Observation metadata):**

| Column | Description |
| :--- | :--- |
| `cell_type` | Specific cell type label |
| `tissue` | Tissue of origin |
| `organ` | Organ of origin |
| `disease` | Clinical status or condition (e.g., healthy) |
| `study` | Source study identifier |
| `dataset` | Source dataset identifier |
| `region` | Anatomical region |
| `subregion` | Specific anatomical subregion |
| `celltype_coarse` | Broad cell type classification |
| `n_cells` | Number of cells aggregated into the pseudobulk |
| `total_counts` | Total read count |
| `n_genes` | Number of genes detected |
| `size_factor` | Sum after normalization |
| `train_pearson` | Pearson correlation on training set |
| `val_pearson` | Pearson correlation on validation set |
| `test_pearson` | Pearson correlation on test set |

**In `.var` (Metadata for variables/genes):**

| Column | Description |
| :--- | :--- |
| `chrom` | Chromosome |
| `start` | Genomic start coordinate (hg38) |
| `end` | Genomic end coordinate (hg38) |
| `strand` | Genomic strand (+/-) |
| `gene_type` | Gene biotype (e.g., protein coding) |
| `frac_nan` | Fraction of missing values |
| `mean_counts` | Average expression counts |
| `n_tracks` | Number of pseudobulks expressing the gene |
| `gene_start` | Gene start position |
| `gene_end` | Gene end position |
| `gene_length` | Total length of the gene |
| `gene_mask_start` | Start of the gene mask in the input sequence |
| `gene_mask_end` | End of the gene mask in the input sequence |
| `frac_N` | Fraction of ambiguous bases (N) in the input |
| `fold` | Borzoi fold assignment |
| `dataset` | Split assignment (e.g., train, test) |
| `gene_id` | Ensembl gene identifier |
| `pearson` | Overall Pearson correlation |
| `size_factor_pearson` | Pearson correlation using size factor |
| `ensembl_canonical_tss` | Canonical Transcription Start Site |

### Data Layers
* **`.layers['preds']`**: Predicted values from the Decima model.
* **`.layers['v1_rep0']` through `.layers['v1_rep3']`**: Predictions from four model replicates.

## Usage

```python
import anndata
from huggingface_hub import hf_hub_download

file_path = hf_hub_download(
repo_id="Genentech/decima-data",
repo_type="dataset",
filename="metadata.h5ad"
)

adata = anndata.read_h5ad(file_path)
```