Machine Learning-Based Prediction of Rule Violations for Drug-Likeness Assessment in Peptide Molecules Using Random Forest Models

Peptide therapeutics often fall outside classical small-molecule heuristics, such as Lipinski’s Rule of Five (Ro5), motivating the development of adapted filters and data-driven approaches to early drug-likeness assessment. We curated >300 k drug (small and peptide) and non-drug molecules from PubChem, extracted key molecular descriptors with RDKit, and generated three rule-violation counters for Ro5, the peptide-oriented beyond-Ro5 (bRo5) extension, and Muegge’s criteria. Random Forest (RF) classifier and regressor models (with 10, 20, and 30 trees) were trained and evaluated. Predictions for 26 peptide test molecules were compared with those from SwissADME, Molinspiration, and manual calculations. Model metrics were uniformly high (Ro5 accuracy/precision/recall = 1.0; Muegge ≈ 0.99), indicating effective learning. Ro5 violation counts matched reference values for 23/26 peptides; the remaining cases differed by +1 violation, reflecting larger structures and platform limits. bRo5 predictions showed near-complete agreement with manual values; minor discrepancies occurred in isolated peptides. Muegge’s predictions were internally consistent but tended to underestimate SwissADME by ~1 violation in several molecules. Four peptides (ML13–16) satisfied bRo5 boundaries; three also fully met Ro5. RF models thus provide fast and reliable in silico filters for peptide drug-likeness and can support the prioritisation of orally developable candidates.