README.md

---
license: mit
task_categories:
  - tabular-classification
  - tabular-regression
tags:
  - chemistry
  - drug-discovery
  - molecules
  - ADMET
  - SMILES
  - molecular-properties
size_categories:
  - 100K<n<1M
---

# MoleculeBench

Pre-computed molecular property scores for **417,500 unique drug-like molecules** from ZINC250K and dockstring, scored with three computational chemistry tools.

## Quick Start

```python
import pandas as pd

df = pd.read_parquet("hf://datasets/yoonholee/MoleculeBench/all_results.parquet")
print(df.shape)  # (417500, 126)

# Filter drug-like molecules
druglike = df[
    (df["admet_ai.QED"] > 0.5) &
    (df["rdkit_properties.LipinskiViolations"] == 0) &
    (df["admet_ai.BBB_Martins"] > 0.8)
]
```

## Dataset Details

| | |
|---|---|
| **Molecules** | 417,500 unique canonical SMILES |
| **Sources** | ZINC250K (249K) + dockstring (260K), deduplicated |
| **Columns** | 126 (1 SMILES + 98 ADMET-AI + 24 RDKit + 3 similarity) |
| **Format** | Parquet (291 MB) |

## Column Reference

### ADMET-AI (98 columns)

Predicted with [ADMET-AI](https://github.com/swansonk14/admet_ai) (Chemprop-RDKit models trained on TDC datasets). Each property has a raw prediction and a DrugBank-approved percentile.

**Physicochemical:** `molecular_weight`, `logP`, `hydrogen_bond_acceptors`, `hydrogen_bond_donors`, `Lipinski`, `QED`, `stereo_centers`, `tpsa`

**Absorption:** `HIA_Hou`, `Bioavailability_Ma`, `Caco2_Wang`, `PAMPA_NCATS`, `Pgp_Broccatelli`, `Solubility_AqSolDB`, `HydrationFreeEnergy_FreeSolv`, `Lipophilicity_AstraZeneca`

**Distribution:** `BBB_Martins`, `PPBR_AZ`, `VDss_Lombardo`

**Metabolism:** `CYP1A2_Veith`, `CYP2C9_Veith`, `CYP2C9_Substrate_CarbonMangels`, `CYP2C19_Veith`, `CYP2D6_Veith`, `CYP2D6_Substrate_CarbonMangels`, `CYP3A4_Veith`, `CYP3A4_Substrate_CarbonMangels`

**Excretion:** `Clearance_Hepatocyte_AZ`, `Clearance_Microsome_AZ`, `Half_Life_Obach`

**Toxicity:** `AMES`, `Carcinogens_Lagunin`, `ClinTox`, `DILI`, `hERG`, `LD50_Zhu`, `Skin_Reaction`, `NR-AR`, `NR-AR-LBD`, `NR-AhR`, `NR-Aromatase`, `NR-ER`, `NR-ER-LBD`, `NR-PPAR-gamma`, `SR-ARE`, `SR-ATAD5`, `SR-HSE`, `SR-MMP`, `SR-p53`

Each property also has a `*_drugbank_approved_percentile` column (49 total).

### RDKit Properties (24 columns)

Computed with RDKit. Prefixed with `rdkit_properties.`:

`MolWt`, `LogP`, `QED`, `TPSA`, `HBondDonorCount`, `HBondAcceptorCount`, `RotatableBondCount`, `RingCount`, `AromaticRingCount`, `HeavyAtomCount`, `FractionCSP3`, `MolarRefractivity`, `FormalCharge`, `BasicAmineCount`, `AccessibleSurfaceArea`, `SlogP_VSA5`, `PEOE_VSA6`, `Kappa1`, `BalabanJ`, `BertzCT`, `SyntheticAccessibility`, `PAINS`, `LipinskiViolations`, `MurckoScaffold`

### Similarity Search (3 columns)

Nearest neighbors from a FAISS index over ChEMBL/PubChem/BindingDB (~2M compounds):

- `similarity.similar_compounds` — JSON list of top-k neighbors with SMILES, Tanimoto similarity, source database, and metadata
- `similarity.count` — number of neighbors returned
- `similarity.error` — error message if search failed (rare)

## Source Datasets

- **ZINC250K**: 249,455 drug-like molecules from [ZINC](https://zinc.docking.org/) via the [chemical_vae](https://github.com/aspuru-guzik-group/chemical_vae) paper
- **dockstring**: 260,155 molecules from the [dockstring](https://github.com/dockstring/dockstring) benchmark (ExCAPE-DB subset with docking scores for 58 protein targets)

After canonicalization and deduplication, 1,041 overlapping molecules were removed.

## Generation

Scored using [MoleculeBench](https://github.com/yoonholee/MoleculeBench), which runs each tool as an isolated HTTP worker service:

```bash
uv run moleculebench --start-services
uv run python benchmark_all.py
```