"Computational Discovery of Novel NUDT5 Inhibitors for ER+ Breast Cancer Therapy"

"The identification of novel chemical scaffolds for underexplored therapeutic targets is severely constrained by limited bioactivity data, rendering conventional machine learning approaches unreliable. Here, we address this challenge for Nudix hydrolase 5 (NUDT5), an ADP-ribose pyrophosphatase that mediates nuclear ATP synthesis in ER+ breast cancer, by developing a Transferability-Weighted Consensus Scoring (TWCS) framework. TWCS integrates structure-based docking (Glide XP), ligand-based similarity (ECFP4 Tanimoto), and three supervised classifiers (random forest, gradient-boosted trees, SVM) trained on a deliberately constructed dataset of 20 NUDT5 actives, 347 MTH1 proxy actives with transfer weights, and 520 property-matched decoys.Critically, we benchmark TWCS against each individual method in a retrospective validation. TWCS achieves consensus AUC = 1.000, EF1% = 27.0, and BEDROC20 = 0.999, matching the best individual classifiers while providing robustness against single-method failure modes. Y-randomization (30 permutations, all p < 0.001) and leave-scaffold-out cross-validation confirm that model performance reflects genuine structure\u2013activity learning rather than dataset artifacts.From a virtual library of 18,412 compounds, the five-tiered TWCS cascade identified 10 structurally diverse candidates with zero Lipinski violations, mean molecular weight of 348 Da, mean Tanimoto similarity to TH5427 of 0.18, and consensus scores ranging from 0.31 to 0.82. All candidates pass PAINS filters and exhibit meaningful scaffold diversity. We provide complete SMILES, computed properties, and multi-method scores to enable immediate experimental follow-up, while critically assessing the translational limitations of NUDT5 inhibition for ER+ breast cancer therapy."