Prodrug-ML: prodrug-likeness prediction via machine learning on sampled negative decoys

A prodrug is a pharmacologically inactive (or attenuated) derivative that undergoes bioreversible transformation in vivo to release an active parent drug, enabling temporary optimization of properties such as solubility, permeability, and targeting. Despite expanding catalogs of known prodrugs, in silico screening remains limited by the absence of reliable negative examples: training/evaluation sets often contain only positives or ad-hoc decoys, leading to class imbalance, property-mismatch shortcuts, and irreproducible benchmarks. Unfortunately, the limitation of reliable negatives has resulted in there being no efficient machine learning-based prodrug screening approach. Therefore, we introduce Prodrug-ML, an efficient machine learning-based screen for prodrug-likeness that prioritizes candidates rather than asserting mechanistic truth. Prodrug-ML helps medicinal chemists triage prodrugging ideas during hit-to-lead and lead optimization, filter enumerated libraries of promoiety–attachment variants before ADMET assays, and retrospectively mine internal/ChEMBL-like collections to surface likely prodrug chemotypes. In practice, users (i) generate or collect candidate structures (e.g., parent drug ± pro-moieties), (ii) score them with Prodrug-ML, and (iii) advance only high-scoring candidates to synthesis/assay, thereby reducing wet-lab load while maintaining chemical diversity. In order to achieve such practical usage, the Prodrug-ML framework, containing the default classifier, LightGBM, addresses these issues by (i) constructing three complementary, property-controlled negative cohorts (DUD-E–style near-misses, random ChEMBL, and strictly filtered ChEMBL), (ii) hardness control and label-noise guardrails on decoys, (iii) domain-bias control, and (iv) cross-decoy validation with multimodel feature selection. Produg-ML has been evaluated five times on hold-out data and an unseen test benchmark, after 80% of training data. In the benchmarks, the multimodel ensemble consistently improves early retrieval and overall discrimination, attaining EF@1% ≈ 6--8, EF@5% ≈ 5--6, BEDROC₂₀ ≈ 0.78--0.82, BEDROC₅₀ ≈ 0.90--0.95, and BEDROC₈₀ ≈ 0.95--0.99, alongside ROC AUC ≈ 0.86--0.87, average precision ≈ 0.60--0.65, and F1 ≈ 0.58--0.62. As a result, these results, especially high BEDROC scores, are consistent with concentrating at least a prodrug within the top ∼ 2--3% of ranked candidates, implying ∼ 97--98% reductions in experimental time and cost when using standard wet-lab workflows that assay only the early tranche.