README.md

---
pretty_name: BioAssayAlign Assay-Compound Data
tags:
- biology
- chemistry
- drug-discovery
- bioassay
- screening
- ranking
- parquet
language:
- en
license: other
size_categories:
- 100M<n<1B
---

# BioAssayAlign Assay-Compound Data

<p align="center">
  <img src="./bioassayalign.png" alt="BioAssayAlign logo" width="280">
</p>

## What this dataset is

BioAssayAlign Assay-Compound Data is a **frozen assay-and-molecule dataset for assay-conditioned ranking and retrieval**.

It answers questions like:
- given an assay description, which molecules in a submitted list should rank first?
- which historical assays are closest to this assay?

It is not:
- a chatbot dataset
- a generic pretraining corpus
- a clinical or patient dataset

Companion model:
- [BioAssayAlign Qwen3-Embedding-0.6B Compatibility](https://huggingface.co/lighteternal/BioAssayAlign-Qwen3-Embedding-0.6B-Compatibility)

Companion Space:
- [BioAssayAlign Compatibility Explorer](https://huggingface.co/spaces/lighteternal/BioAssayAlign-Compatibility-Explorer)

## What is included

This public release is focused on the **prepared compatibility-ranking subset** used by the published model.

Directory: `prepared/compatibility-ranking/`

Files:
- `compat_assays.parquet`
- `compat_candidate_pools.parquet`
- `compat_train_groups.parquet`
- `COMPATIBILITY_PREPARED_MANIFEST.json`
- `SOURCE_DATASET_MANIFEST.json`

This prepared subset is the one used to train the published compatibility model linked above.

For lineage and reproducibility, the release also includes:
- `raw/DATASET_MANIFEST.json`

That manifest records the frozen upstream sources and hashes for the full raw corpus derived from:
- PubChem BioAssay snapshot dated `2026-03-01`
- ChEMBL release `chembl_36`

The full raw parquet pair is **not** included in this compact public repo. This repo is intentionally scoped to the prepared subset that reproduces the public model.

## Why there are multiple parquet files

### `prepared/compatibility-ranking/compat_assays.parquet`
Prepared assay rows used for compatibility ranking.

### `prepared/compatibility-ranking/compat_candidate_pools.parquet`
Held-out assay candidate pools used for evaluation.

### `prepared/compatibility-ranking/compat_train_groups.parquet`
Training groups with:
- one assay
- one positive molecule
- explicit same-assay negative molecules

## Dataset scale

### Source frozen corpus referenced by `raw/DATASET_MANIFEST.json`

| Source table | Rows |
|---|---:|
| assays | `3,800,882` |
| measurements | `323,706,180` |

### Prepared ranking subset used by the public model

| File | Rows |
|---|---:|
| `compat_assays.parquet` | `11,195` |
| `compat_candidate_pools.parquet` | `1,432,532` |
| `compat_train_groups.parquet` | `508,216` |

Split counts:

| Split | Assays |
|---|---:|
| train | `8,967` |
| val | `1,117` |
| test | `1,111` |

## Sanitization and privacy

This public dataset does **not** contain patient data or direct personal identifiers.

Before release, I removed internal-only publishing clutter such as:
- shard outputs from HF CPU prep jobs
- precomputed training feature stores
- private training-only intermediate files

This public repo intentionally excludes:
- shard directories from HF CPU prep jobs
- precomputed training feature stores
- internal benchmark artifacts unrelated to the released model
- local build outputs unrelated to the public model

## File schemas

### `prepared/compatibility-ranking/compat_train_groups.parquet`

Important columns:
- `assay_uid`
- `positive_smiles`
- `positive_smiles_hash`
- `negative_smiles`
- `negative_smiles_hashes`
- `example_weight`

This is the core ranking supervision format used by the public model.

## Example row

Conceptually, one training observation looks like:

```json
{
  "assay_uid": "pubchem:720659",
  "positive_smiles": "CC1=CC(=O)N(C)C(=O)N1",
  "positive_smiles_hash": "4d6f0d...abc",
  "negative_smiles": [
    "CCOC1=CC=CC=C1",
    "CCN(CC)CCOC1=CC=CC=C1",
    "COC1=CC=CC=C1O"
  ],
  "negative_smiles_hashes": [
    "a1...",
    "b2...",
    "c3..."
  ],
  "example_weight": 1.34
}
```

## How to load it locally

### Python / pandas

```python
import pandas as pd

train_groups = pd.read_parquet("prepared/compatibility-ranking/compat_train_groups.parquet")
compat_assays = pd.read_parquet("prepared/compatibility-ranking/compat_assays.parquet")
candidate_pools = pd.read_parquet("prepared/compatibility-ranking/compat_candidate_pools.parquet")
```

### Python / pyarrow

```python
import pyarrow.parquet as pq

train_groups = pq.read_table("prepared/compatibility-ranking/compat_train_groups.parquet")
```

## How this relates to the public model

The published model was trained on:
- `prepared/compatibility-ranking/compat_assays.parquet`
- `prepared/compatibility-ranking/compat_candidate_pools.parquet`
- `prepared/compatibility-ranking/compat_train_groups.parquet`

Published model:
- [lighteternal/BioAssayAlign-Qwen3-Embedding-0.6B-Compatibility](https://huggingface.co/lighteternal/BioAssayAlign-Qwen3-Embedding-0.6B-Compatibility)

## Upstream sources

This dataset is derived from public upstream resources including:
- PubChem BioAssay
- ChEMBL

Users are responsible for complying with the attribution and usage terms of the upstream sources.