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

Overview Proteolysis-targeting chimeras (PROTACs) are bifunctional molecules that recruit an E3 ubiquitin ligase to a target protein of interest (POI), directing it for proteasomal degradation. Predicting the extent and potency of PROTAC-induced degradation, quantified as Dmax (maximal degradation) and DC50 (concentration at half-maximal degradation) is a key challenge in targeted protein degradation drug design. To address this, we developed TACK (TArgeting Chimeras Knowledge), a statistical machine learning framework trained and evaluated on the largest publicly available curated dataset of PROTAC degradation measurements. TACK integrates data from three sources, PROTAC-DB, PROTACpedia, and TPDdb, and trains MLP and XGBoost models under a rigorous repeated 5×5 cross-validation scheme. Ensemble models are constructed via Caruana's greedy forward selection method. This repository contains the pre-trained ensemble models and ensemble weight specifications required to reproduce the results reported in: "TACK: A statistical evaluation of degradation activity on a novel TArgeting Chimeras Knowledge dataset" by Stefano Ribes, Nils Dunlop, and Rocío Mercado The curated dataset is available on Hugging Face at: https: //huggingface. co/datasets/ailab-bio/TACK The accompanying code is available at: https: //github. com/ribesstefano/TACK Data Available Ensemble Models: The trained models as checkpoint files, organized in directories for ensemble prediction of degradation activity (bin), Dmax, and DC50. Prediction Values: The evaluation results from running the trained models. Cache: Additional data to reproduce training and run the ensemble models. File Descriptions Checkpoints of ensemble models: binbestₐrchₑnsemble: Ensemble of 25 models of the best architecture for predicting binary degradation activity. bincaruanaₑnsemble: Ensemble of 18 models selected via Caruana selection for predicting binary degradation activity. dmaxbestₐrchₑnsemble: Ensemble of 25 models of the best architecture for predicting maximum degradation activity (Dmax). dmaxcaruanaₑnsemble: Ensemble of 33 models selected via Caruana selection for predicting maximum degradation activity (Dmax). dc50bestₐrchₑnsemble: Ensemble of 25 models of the best architecture for predicting DC50. dc50caruanaₑnsemble: Ensemble of 22 models selected via Caruana selection for predicting DC50. Evaluation results: predictions/: Contains the prediction files of all 5×5 trained models. The predictions are on the models’ respective validation fold and the common hold-out set. The predictions are per sample and stored as CSV files. predictionsₚrotacₛtan/: Contains the checkpoints and binary predictions of the PROTAC-STAN model trained on the TACK dataset. Cached data generated during training: cell2cellᵢd. json: Mapping from cell line names to unique cell IDs in CelloSaurus. cell2description. json: Mapping from cell line names to their aggregated textual descriptions (e. g. , tissue of origin, disease state). cell2data. json: Mapping from cell line names to information from CelloSaurus, stored in JSON format. cellₑmbeddingsₘodel=sentence-transformerₚooling=sum. npz: Precomputed cell line embeddings using a sentence transformer model. The file contains a mapping from cell line IDs to their corresponding embedding vectors. morganfpᵣadius16ₛize512. npz: Mapping from SMILES strings to their corresponding Morgan fingerprints with radius 16 and size 512. rdkitdescriptors. npz: Mapping from SMILES strings to their corresponding RDKit descriptors (of size 217). Original PROTAC data used for data curation can be obtained from: PROTAC-DB-v3: http: //cadd. zju. edu. cn/protacdb/. PROTACpedia: https: //protacpedia. weizmann. ac. il/ptcb/main. TPDdb: https: //tpddb. idrblab. net. License This repository is open-source and available under the MIT License. Contact Information For questions or feedback, please contact Rocío Mercado: rocio. mercado@chalmers. se.