PROTAC-Bench: A Cold-Target Benchmark for PROTAC Degradation Prediction

Dataset Description

PROTAC-Bench is a merged PROTAC degradation dataset containing 10,748 entries across 173 protein targets (9,359 unique SMILES). It combines data from PROTAC-DB 3.0, Ribes et al. (2024), and DegradeMaster with deduplication and canonical SMILES standardization. Each entry has a binary activity label (active: DC50 < 1 μM OR Dmax > 50%) along with target UniProt ID and E3 ligase type (VHL/CRBN/Other). The dataset includes pre-computed 7-property ADMET cascade scores for all entries.

Evaluation Protocol

All models are evaluated under the Leave-One-Target-Out (LOTO) protocol across 65 eligible targets (≥10 entries, activity rate 10–90%). For each fold, all entries for one target are held out as the test set while the remaining entries are used for training. Performance is reported as mean AUROC across the 65 folds. Statistical significance against the RF+Morgan baseline is assessed via paired Wilcoxon signed-rank test.

Dataset Structure

Data Fields

• smiles (string): Canonical SMILES of the PROTAC molecule
• target_uniprot (string): UniProt accession of the target protein
• e3_type (string): E3 ligase type (VHL, CRBN, or Other)
• label (int): Binary activity label (1 = active, 0 = inactive)
• dc50_nm (float): DC50 in nanomolar (when available)
• dmax_pct (float): Dmax percentage (when available)

Data Splits

The dataset ships as a single train split; that name is a packaging convention, not an evaluation split. Evaluation uses the Leave-One-Target-Out (LOTO) protocol with 65 pre-defined folds (see data/loto_folds.json).

Additional Files

• data/admet_scores.csv: 7-property ADMET cascade scores for all 10,748 entries
• evaluation/evaluate.py: Standardized LOTO evaluation script
• evaluation/baselines.py: RF+Morgan baseline reproduction
• examples/example_submission.py: Template for formatting predictions

Quick Start

pip install -r evaluation/requirements.txt

# Run the RF+Morgan baseline (~2 min)
python evaluation/baselines.py

# Evaluate your own predictions
python evaluation/evaluate.py --predictions my_predictions.csv --output results.json

Baseline Results

Model	Mean AUROC	Δ vs RF+Morgan	p-value
RF + Morgan (2048-bit)	0.666	—	—
RF + Morgan + ADMET	0.687	+0.021	<0.05
RF + Morgan + ADMET + k=5	0.700	+0.034	<0.01
EGNN-27	0.801	+0.135	<0.001

Dataset Creation

Source Data

Merged from three public databases:

• PROTAC-DB 3.0: Curated PROTAC degradation data
• Ribes et al. (2024): Published PROTAC activity data
• DegradeMaster: Comprehensive degradation database

Curation

• Canonical SMILES standardization via RDKit
• Deduplication by canonical SMILES + target pair
• Binary labeling: active if DC50 < 1 μM OR Dmax > 50%

Personal and Sensitive Information

This dataset contains no personal or sensitive information. All entries are chemical structures (SMILES) and protein identifiers.

Considerations for Using the Data

Known Biases

• Kinase-dominated: 24 of 65 evaluation targets are kinases
• E3 ligase imbalance: Only VHL and CRBN ligases are well-represented
• Publication bias: Positive results may be over-represented in source databases

Citation

@misc{klamt2026protacbench,
    title={Decomposing the Generalization Gap in PROTAC Activity Prediction:
           Variance Attribution and the Inter-Laboratory Ceiling},
    author={Klamt, Thor and Nejdl, Wolfgang and Tang, Ming},
    year={2026},
    eprint={2605.11764},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    note={Preprint; under revision}
}

License

• Data: CC-BY-4.0
• Code: MIT