README.md

---
pretty_name: Embpy Static Embeddings
license: other
tags:
- embpy
- biology
- genomics
- gene-embeddings
- protein-embeddings
- zarr
---

# Embpy Static Embeddings

This dataset contains static gene and protein embeddings packaged with embpy. Embedding values are stored in Zarr arrays, while row identifiers, provenance, species metadata, and AnnData-like `.uns` metadata are stored in sidecar metadata files.

## Summary

- Repository: `theislab/Embpy_Data`
- Schema version: `embpy.static_embedding_package.v1`
- Generated at: `2026-06-01T15:07:46+00:00`
- Number of embeddings: `11`
- Total indexed entities: `201,717`
- Species keys: `human_9606`
- NCBI taxonomy IDs: `9606`
- Default gene identifier policy: `ensembl_id`
- Unresolved gene identifier policy: `drop`

## Available Embeddings

| key | entity | species | id type | rows | dims | description |
| --- | --- | --- | --- | --- | --- | --- |
| crispr_gene_effect | gene | human_9606 | ensembl_id | 17,087 | 1,178 | DepMap CRISPR gene effect matrix, genes as rows after transposition. |
| crispr_gene_effect_1178 | gene | human_9606 | ensembl_id | 17,087 | 1,178 | DepMap CRISPR gene effect embedding, scaled, 1178d. |
| crispr_gene_effect_205 | gene | human_9606 | ensembl_id | 17,916 | 205 | DepMap CRISPR gene effect embedding, scaled, 205d. |
| gene2vec | gene | human_9606 | ensembl_id | 18,795 | 200 | Gene2Vec co-expression embedding, 200d. |
| genept | gene | human_9606 | ensembl_id | 18,807 | 3,072 | GenePT GPT-3.5 text embedding, 3072d, Ensembl-keyed. |
| genept_scaled | gene | human_9606 | ensembl_id | 17,728 | 3,072 | GenePT GPT-3.5 text embedding, z-scored, 3072d. |
| omics | gene | human_9606 | ensembl_id | 19,385 | 256 | Omics 256d static gene embedding, Ensembl-keyed. |
| pops | gene | human_9606 | ensembl_id | 18,383 | 256 | PoPS 256d gene features, Ensembl-keyed. |
| string_functional_9606 | protein | human_9606 | string_protein_id | 19,699 | 512 | STRING/SPACE functional PPI embedding for human proteins, species 9606, 512d. |
| string_node2vec_9606 | protein | human_9606 | string_protein_id | 19,622 | 128 | STRING node2vec PPI embedding for human proteins, species 9606, 128d. |
| wikicrow | gene | human_9606 | ensembl_id | 17,208 | 4,096 | WikiCrow text embedding, scaled, 4096d. |

## File Layout

```text
manifest.json
embeddings/<model_key>/
  values.zarr/
  metadata/
    index.parquet
    index.csv
    metadata.json
    uns.json
```

The dense matrix is stored under `values.zarr`. The `metadata/index.parquet` file maps row positions to `entity_id` values and any preserved aliases such as source IDs or gene symbols.

## Loading With embpy

```python
from embpy.pp import HFHandler

hf = HFHandler("theislab/Embpy_Data")
embedding = hf.download_embedding("crispr_gene_effect")
matrix = embedding["embeddings"]
ids = embedding["ids"]
```

For a local checkout or downloaded snapshot:

```python
from embpy import load_static_embedding_package

store = load_static_embedding_package("/path/to/package", key="crispr_gene_effect")
tp53 = store.get("ENSG00000141510")
tp53_by_symbol = store.get("TP53", id_type="symbol")
```

Missing identifiers raise by default. Use `missing="drop"` or `missing="nan"` when a partial result is acceptable.

## Metadata And Species

This card reports species using package-level defaults. Regenerate the package with a current embpy build to also write per-embedding `species`, `taxonomy_id`, and `species_key` fields into `metadata/metadata.json`.

## Validation

The package was designed to be validated locally before upload:

```bash
python -m embpy.scripts.package_static_embeddings validate --package /path/to/package
```

This package intentionally skips 2 source collection(s) containing 2642 file(s). Those collections usually contain per-species artifacts and should be packaged only when a species/taxonomy ID is selected explicitly.

## License And Attribution

This repository aggregates embeddings derived from multiple upstream resources. Please check the per-embedding metadata and upstream sources for the applicable licenses and citation terms.