Interpretable Machine Learning for Species-Specific Prediction of Helicobacter pylori Urease Inhibitors
Year: 2026
Type: preprint
Abstract: Helicobacter pylori depends on urease for survival in the acidic stomach, yet inhibitor-prediction studies frequently rely on measurements from different urease sources, limiting their relevance to the pathogen-specific target. To address this problem, an interpretable quantitative structureā€“activity relationship (QSAR) model was developed from 351 com-pounds with experimentally measured H. pylori urease inhibitory activity. PubChem molecular fingerprints were used to predict continuous pIC50 values, with LightGBM selected through cross-validation. The model achieved a five-fold cross-validated \(\:{Q}_{\text{CV}}^{2}\) of 0.525 and an independent-test š¯‘…2 of 0.782, while response randomization substantially diminished test-set predictivity. SHAP analysis revealed contrasting contributions from chemically related carbonyl-containing patterns, alongside associations involving nitrogen, chlorine, and oxygenated molecular fragments. These findings suggest that the model distinguishes local structural environments rather than assigning uniform contributions to broad functional-group classes. Applicability-domain assessment further characterized the chemical space supporting its predictions. The trained model was made accessible through PanURA for individual and batch potency estimation. By integrating pathogen-specific bioactivity data with interpretable molecular predictions, this study establishes a prac-tical framework for prioritizing candidate urease inhibitors and investigating structureā€“activity relationships through subsequent experimental studies.
Source: Research Square
Author: Anthony Peter
Institution: University of Dar es Salaam
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
Funder: University of Dar es Salaam
OpenAlex ID: W7224106657