TACK: A Statistical Evaluation of Degradation Activity on a Novel TArgeting Chimeras Knowledge Dataset

If a new PROTAC model beats the last one by two accuracy points, is that a real advance or noise? TACK was built to make that question answerable: a standardised dataset plus a statistically honest benchmarking protocol.

Abstract

Proteolysis-targeting chimeras (PROTACs) represent a promising therapeutic modality that induces targeted protein degradation by hijacking the ubiquitin-proteasome system. We introduce TACK, a novel TArgeting Chimeras Knowledge dataset of 3,514 PROTACs and 6,561 degradation endpoints aggregated from three major repositories with standardized molecular representations, protein annotations, and experimental conditions. Using TACK, we perform a rigorous statistical comparison of three machine learning methods across three tasks: \(DC_{50}\) and \(D_{max}\) regression, and binary activity classification. Models trained on the best performing features show that potency (\(pDC_{50}\), \(R^2 = 0.66\)) is substantially more predictable than maximum degradation (\(D_{max}\), \(R^2 = 0.36\)). Statistical tests support that classical methods (XGBoost and MLP) significantly outperform PROTAC-STAN, a domain-specific graph neural network model (ROC-AUC: 0.85 vs. 0.74, \(p < 0.001\)). Feature ablation demonstrates that cellular context features and simple protein representations rival complex ESM protein embeddings, highlighting the importance of feature engineering over architectural sophistication.

Keywords: PROTAC, targeted protein degradation, benchmark dataset, statistical evaluation, XGBoost, protein embeddings, scaffold split

The Dataset

TACK aggregates and harmonises three major public repositories — TPDdb, PROTAC-DB and PROTACpedia — into what is, to our knowledge, the largest publicly available dataset for training and evaluating PROTAC degradation activity models:

The harmonisation step is the unglamorous part and also the load-bearing one: without consistent cell-line naming, unit conventions and protein identifiers, any cross-source comparison silently measures curation artefacts rather than chemistry.

Three Findings Worth Highlighting

1. Potency is more learnable than efficacy. \(pDC_{50}\) reaches \(R^2 = 0.66\), while \(D_{max}\) stalls at \(R^2 = 0.36\). The two endpoints are not interchangeable targets, and papers that report only one are reporting the easier one.
2. Classical models still win. XGBoost and a plain MLP significantly outperform PROTAC-STAN, a domain-specific graph neural network (ROC-AUC 0.85 vs. 0.74, \(p < 0.001\)). The comparison is made with scaffold-based cross-validation and proper statistical testing rather than a single split.
3. Features beat architectures. Cellular context features and simple protein representations rival much heavier ESM embeddings — a useful result if you are deciding where to spend your compute budget.

Resources

• Dataset: ailab-bio/TACK on Hugging Face
• Code: github.com/ribesstefano/TACK
• Trained models: Zenodo, DOI 10.5281/zenodo.15691822
• Preprint: arXiv:2605.19579

Collaborators

Joint work with Nils Dunlop and Rocío Mercado at the AI Laboratory for Molecular Engineering (AIME), Chalmers University of Technology and University of Gothenburg.

Citation

@misc{ribes2026tackstatisticalevaluationdegradation,
  title={{TACK: A statistical evaluation of degradation activity on a novel TArgeting Chimeras Knowledge dataset}},
  author={Stefano Ribes and Nils Dunlop and Roc{\'i}o Mercado},
  year={2026},
  eprint={2605.19579},
  archivePrefix={arXiv},
  primaryClass={q-bio.QM},
  url={https://arxiv.org/abs/2605.19579}
}

Recommended citation: Ribes, S., Dunlop, N., Mercado, R., "TACK: A statistical evaluation of degradation activity on a novel TArgeting Chimeras Knowledge dataset", 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), AI for Science Track, 2026. arXiv:2605.19579.

Link to paper.