---
license: mit
task_categories:
- tabular-regression
tags:
- perturbation-prediction
- Perturb-seq
- benchmark
- zero-shot
- protein-embeddings
size_categories:
- 1M<n<10M
configs:
- config_name: cohort
  data_files:
  - split: train
    path: viewer/cohort/*.parquet
- config_name: annotations
  data_files:
  - split: train
    path: viewer/annotations/*.parquet
- config_name: responses
  data_files:
  - split: train
    path: viewer/responses/*.parquet
- config_name: embeddings_esm2_650m
  data_files:
  - split: train
    path: viewer/embeddings/embedding_esm2_650m.parquet
- config_name: embeddings_esm2_3b
  data_files:
  - split: train
    path: viewer/embeddings/embedding_esm2_3b.parquet
- config_name: embeddings_protrek
  data_files:
  - split: train
    path: viewer/embeddings/embedding_protrek.parquet
- config_name: splits
  data_files:
  - split: train
    path: viewer/splits/splits.parquet
- config_name: rpe1_splits
  data_files:
  - split: train
    path: viewer/splits/rpe1_splits.parquet
---

# AMBIMOD benchmark data (K562)

Derived evaluation data for the AMBIMOD zero-shot perturbation-prediction
benchmark. The data is organized into Dataset-Viewer configs (browse each
tab above) plus a ready-to-use archive for the benchmark code.

## Dataset Access

Two equivalent access paths:

**A. Load individual tables with `datasets`** (what the Dataset Viewer shows):

```python
from datasets import load_dataset

cohort = load_dataset("wozhen/ambimod", "cohort", split="train")
splits = load_dataset("wozhen/ambimod", "splits", split="train")
responses = load_dataset("wozhen/ambimod", "responses", split="train")
```

**B. Download the benchmark archive** (exact layout the code expects —
`cache/`, `data/splits/`, checksums):

```bash
huggingface-cli download wozhen/ambimod --repo-type dataset --local-dir .
tar xzf ambimod-benchmark-data.tar.gz
```

## Configs (Dataset Viewer tabs)

| Config | Rows | Columns | Content |
|---|---|---|---|
| `cohort` | 490 | tf, n_cells, sequence_length, sequence, uniprot_header | quality-controlled K562 TF cohort (UniProt-reviewed human sequences) |
| `annotations` | 490+490 | tf ↔ DBD family (InterPro); tf ↔ paralog list | leakage-control annotations |
| `responses` | 2,450,000 | tf, replicate (1–5), gene_idx, gene_id, log1p_delta | held-out evaluation responses: log1p-delta per TF × replicate × 1,000 target genes |
| `embeddings_esm2_650m` | 627,200 | tf, dim, value | ESM2-650M embeddings (490 × 1280, L2-normalized) |
| `embeddings_esm2_3b` | 1,254,400 | tf, dim, value | ESM2-3B embeddings (490 × 2560) |
| `embeddings_protrek` | 501,760 | tf, dim, value | ProTrek-650M embeddings (490 × 1024) |
| `splits` | 7,350 | design, seed, role, tf | the 15 frozen K562 splits (random / dbd_family / embed_ball × 5 seeds), train/test membership |
| `rpe1_splits` | 6,534 | seed, role, perturbation | frozen RPE1 split manifests (reference for the raw-data pipeline) |

Embeddings are stored in long format (`tf, dim, value`) for Viewer
compatibility; the benchmark archive (`ambimod-benchmark-data.tar.gz`)
contains the same values as dense `(490, d)` float32 arrays.

## Raw Data Sources

- K562 GWPS + RPE1 essentialome (Perturb-seq): https://gwps.wi.mit.edu/
- K562/RPE1 figshare mirror (RPE1 h5ad, direct download): https://figshare.com/ndownloader/files/42362275
- Norman 2019 CRISPRa (GEARS-official file, Harvard Dataverse): https://dataverse.harvard.edu/api/access/datafile/6154020

## Embedding provenance

All embedding caches were **computed by us** from public model weights applied
to the cohort's UniProt-reviewed sequences — no third-party embedding files
are redistributed. Checkpoints used (all public):

| Embedding | Checkpoint | Pooling |
|---|---|---|
| `esm2_650m` | `facebook/esm2_t33_650M_UR50D` | residue mean (1022-aa length-weighted chunks), L2-normalized |
| `esm2_3b` | `facebook/esm2_t36_3B_UR50D` | same |
| `protrek` | `westlake-repl/ProTrek_650M` | official `get_protein_repr` |

`scripts/recompute_embeddings.py` in the code repository rebuilds any cache
from these checkpoints and fails on checksum mismatch. The RPE1 patched
ProTrek cache and its gated sanity receipt are documented in
`EMBEDDINGS.md` in the code repository.

## Usage with the benchmark code

```bash
pip install -e git+https://github.com/ambimod/ambimod#egg=ambimod
python -c "from mechid.benchmark import load_benchmark; s = load_benchmark(); print(len(s))"
# -> 15
```

## License

MIT (derived arrays and manifests). Embeddings were computed from public
protein sequences with public models (ESM2, ProTrek). The raw datasets
remain under their original licenses and are not redistributed.