TACK

A statistical evaluation of degradation activity on a novel TArgeting Chimeras Knowledge dataset

TACK predicts PROTAC-induced protein degradation — Dmax, DC50 and binary activity — from a molecule plus its experimental context: target protein, E3 ligase, cell line, assay and treatment time. The repository ships the curated dataset, its curation pipeline, and two sets of pre-trained models.

Quick start

The fast ensembles are the default way to run TACK: gradient-boosted ensembles over a fused molecular + context design matrix. CPU only, no torch, milliseconds per molecule, and molecules are featurised on the fly so unseen compounds can be scored.

git clone https://github.com/ribesstefano/TACK.git && cd TACK
uv venv --python 3.13 && source .venv/bin/activate
uv pip install -e .

from tackai.fusion import FusionData, XGBStackedEnsemble

context_encoder = FusionData.from_pretrained("ailab-bio/TACK-Context-Encoding")
model = XGBStackedEnsemble.from_pretrained("ailab-bio/TACK-Fast-Ensemble", subfolder="dmax")

X = context_encoder.transform([{
    "smiles": "N#Cc1cccc(-c2ccc(=O)n(Cc3cccc(-c4ncc(OCC5CCN(CCCCC(=O)Nc6cccc7c6C(=O)N(C6CCC(=O)NC6=O)C7=O)CC5)cn4)c3)n2)c1",
    "poi_seq": "MKAPAVLAPGILVLLFTLVQRSNG...",  # target protein, full sequence
    "e3_seq": "MAGEGDQQDAAHNMGNHLPLLPAE...",   # E3 ligase, full sequence
    "cell_id": "CVCL_0333",                    # Cellosaurus accession
    "assay": "western blot",
    "assay_time": 72.0,                        # hours
}])

model.predict(X)  # {'mean': ..., 'std': ..., 'std_noise': ..., 'std_disagreement': ...}

Every prediction carries its own error bar, split into model noise and ensemble disagreement. FusionData.help() and XGBStackedEnsemble.help(show_extras=True) describe the accepted record fields, the cached vocabularies and every output key.

Tutorial: notebooks/fast_TACK_models_tutorial.ipynb

Install options

Pre-computed context_encoder embeddings land in TACKAI_CACHE (copy .env.example to .env to set it; default ~/.cache/tackai/). On macOS, export OMP_NUM_THREADS=1 before importing — torch and xgboost bundle separate libomp copies and segfault when both run multi-threaded in one process.

Limitations of the fast ensembles

Please be aware of the following caveats when using the fast ensembles:

• The context_encoder vocabularies are closed. POI (protein of interest) and E3 blocks are keyed by the exact full amino-acid sequence, and an unseen sequence raises an error. A cell line must be a known Cellosaurus accession (blank falls back to unknown, an unrecognised accession raises). The assay vocabulary is closed; anything outside it falls back to unknown with a warning.
• Predicitng outside the training domain is risky. The model was trained on a limited set of POIs, E3 ligases, cell lines and assays, and the training set is heavily biased towards AR, BRD4, VHL and CRBN. Predictions for other targets or contexts might be unreliable. To register a new sequence embedding, i.e., 640-d ESM-2 embedding, use: context_encoder.encoder.register_sequence().
• No check of PROTAC SMILES. The model accepts any new PROTAC molecule, but does not check that the SMILES is a valid PROTAC-like structure. The training set is mostly bivalent molecules with a linker, but the model will happily "score" any small molecule resulting in a meaningless prediction.
• Thresholds are fixed. activity means Dmax > 80% and pDC50 > 6, hard-coded in tackai.fusion.data.
• One timepoint, no kinetics. assay_time is the mean of the reported DC50/Dmax hours and defaults to 24 h when absent.
• Accuracy ceiling. In 5×5 scaffold-grouped cross-validation the surrogate benchmark reaches R² ≈ 0.42 (Dmax), R² ≈ 0.55 (pDC50) and ROC-AUC ≈ 0.87 (activity). That is enough to rank PROTACs within a series; it does not replace a dose–response curve, and accuracy on unseen chemotypes is markedly lower than those numbers suggest.
• Uncertainty is a hint, not a gate. std widens outside the training domain, but nothing checks the applicability domain for you, and the conformal lower/upper bounds exist only for an ensemble that has been through calibrate(). The provided models are calibrated.
• These are not the paper's models. They are trained on an extended curation (~8.4k measurements) that is not yet published and not peer-reviewed. The results in the paper come from the legacy pipeline below.
• Censored measurements are taken at face value. During training, DC50 > 10 µM is read as 10 µM: the operator columns are ignored, and repeated measurements are not de-duplicated.

Dataset

The curated dataset is on the Hub at ailab-bio/TACK:

from datasets import load_dataset

dmax = load_dataset("ailab-bio/TACK", "Dmax", split="train")
dc50 = load_dataset("ailab-bio/TACK", "DC50", split="train")
multi = load_dataset("ailab-bio/TACK", "multitask", split="train")

It integrates TPDdb, PROTAC-DB and PROTACpedia with deduplication, POI/E3 normalisation and scaffold clustering, and ships SMILES_Held_Out, SMILES_Scaffold_Cluster and SMILES_Butina_Cluster columns so the held-out scaffolds never leak into training. To re-run the curation, see tack_dataset/README.md.

Reproducing the paper

The legacy pipeline is kept for reproducibility: Lightning/XGBoost models trained with repeated 5×5 cross-validation and combined by Caruana's greedy forward selection.

uv pip install -e ".[dev,training]"

Inference with the published ensembles. Checkpoints and the shared embedding cache are on Zenodo (doi:10.5281/zenodo.15691822) as ensembles.zip (per-task checkpoints, _hparams.yaml/_state.pt data-processing state and ensemble_weights_*.json) and cache.zip (ESM protein, cell-line and molecular embeddings, which must match the versions used during training). Unzip cache.zip and point TACKAI_CACHE at it, or let the Hub do both:

from tackai.ensemble_predictor import EnsemblePredictor

predictor = EnsemblePredictor.from_pretrained("ailab-bio/TACK-ensembles", subfolder="dmax_caruana")
result = predictor.predict({"SMILES": "CCO", "POI_Name": "BRD4", ...})

from_pretrained also takes a local directory, for compute nodes without outbound internet. See notebooks/ensemble_predictor_tutorial.ipynb.

Training and evaluation. Runs are composed with Hydra from configs/train.yaml by picking a feature set (data=, a file stem under configs/data/) and a model (model=, a file stem under configs/model/); any leaf can be overridden inline.

tack train model=xgboost data=fp task=dmax group=scaffold    # one 5×5 CV experiment
tack collect --predictions-dir ./predictions                 # merge per-fold prediction CSVs
tack evaluate --predictions-dir ./predictions --output-dir ./evaluation
python scripts/ensemble_comparison.py --task dmax --prediction_dir ./predictions --output_dir ./ensemble_results

tack evaluate ranks methods per task with autorank and writes report.md, ranking.csv, metrics.csv and critical-difference diagrams. There is no sweep subcommand — loop over configurations from the shell:

for enc in minmax onehot embedding; do
    tack train model=xgboost data=fp task=dmax data.categorical_encoding=$enc
done

Run identity is encoded in prediction filenames (preds-model=…-task=…-group=…-fold=…-split=….csv) and in per-row method/task/group/ fold/split columns. Batch inference over a CSV of PROTACs:

tack predict --checkpoints-dir ensembles/dmax \
    --weights ensemble_weights_dmax_caruana_ensemble.json \
    --input-csv compounds.csv --output-csv predictions.csv

For the PROTAC-STAN baseline, see protac_stan/README.md.

Repository layout

TACK/
├── tackai/
│   ├── fusion/          # fast surrogates: design matrix, context encoder, ensembles
│   ├── data/            # legacy datamodule, tasks, embeddings
│   ├── models/          # MLP / BERT / multi-embedding backbones
│   ├── training/        # cross-validation, splitting, persistence
│   ├── evaluation/      # discovery, ranking, reports
│   └── cli.py           # tack train / predict / collect / evaluate
├── tack_dataset/        # dataset curation (scraping, dedup, clustering)
├── configs/             # Hydra configs for the legacy pipeline
├── notebooks/           # tutorials and the fusion experiment record
├── protac_stan/         # PROTAC-STAN baseline
├── scripts/             # ensemble selection and legacy helpers
└── test/                # pytest suite

To run the full test suite, install the dev dependencies and run:

OMP_NUM_THREADS=1 pytest    # the full suite needs ".[dev,gp,training]"

License

MIT — see LICENSE.

Citation

@inproceedings{ribes_tack_2026,
	address = {New York, NY, USA},
	series = {{KDD} '26},
	title = {{TACK}: {A} {Statistical} {Evaluation} of {Degradation} {Activity} on a {Novel} {TArgeting} {Chimeras} {Knowledge} {Dataset}},
	isbn = {979-8-4007-2259-2},
	shorttitle = {{TACK}},
	url = {https://dl.acm.org/doi/10.1145/3770855.3819052},
	doi = {10.1145/3770855.3819052},
	urldate = {2026-10-12},
	booktitle = {Proceedings of the 32nd {ACM} {SIGKDD} {Conference} on {Knowledge} {Discovery} and {Data} {Mining} {V}.2},
	publisher = {Association for Computing Machinery},
	author = {Ribes, Stefano and Dunlop, Nils and Mercado, Rocío},
	month = aug,
	year = {2026},
	pages = {11912--11923},
	file = {Full Text PDF:/Users/ribes/Zotero/storage/QN359RNF/Ribes et al. - 2026 - TACK A Statistical Evaluation of Degradation Activity on a Novel TArgeting Chimeras Knowledge Datas.pdf:application/pdf},
}

Acknowledgements

The authors acknowledge funding provided by the Chalmers Gender Initiative for Excellence (Genie), and by the Wallenberg AI, Autonomous Systems, and Software Program (WASP), supported by the Knut and Alice Wallenberg Foundation. The authors thank Yossra Gharbi, Alexander Persson, and Felix Erngård for helpful discussions. The computations and data storage were enabled by resources provided by Chalmers e-Commons and by the National Academic Infrastructure for Supercomputing in Sweden (NAISS), partially funded by the Swedish Research Council through grant agreement no. 2022-06725.