Continuous Inhibition-Zone Modeling and Binary Classification for Pseudomonas aeruginosa Hit Prioritization: A Retrospective QSAR Evaluation

Background/Objectives: Antibiotic-resistant Pseudomonas aeruginosa and limited experimental validation capacity motivate efficient prioritization of antibacterial candidates from large chemical libraries. Quantitative structure–activity relationship (QSAR) benchmarks often binarize disk diffusion inhibition-zone (IZ) measurements, obscuring activity gradients and imposing threshold dependence. We examined whether continuous-IZ modeling provides complementary retrospective prioritization relative to calibrated binary classification in a highly imbalanced dataset. Methods: In this retrospective matched-data evaluation, we revisited a published ChEMBL-derived P. aeruginosa disk diffusion dataset using preserved training and locked external validation partitions. A calibrated support vector classifier using Molecular ACCess System keys (SVC/MACCS) provided conservative binary active calls. RegressionStack combined source-descriptor extreme gradient boosting (XGBoost) and ElasticNet regressors, Morgan-fingerprint random forest and gradient-boosting regressors, and a MACCS-key XGBoost regressor through an XGBoost meta-regressor to predict continuous IZ. Results: On the locked external set (n = 1130; 87 actives), SVC/MACCS achieved a positive predictive value (PPV) = 0.619, receiver operating characteristic area under the curve (ROC-AUC) = 0.857, precision–recall area under the curve (PR-AUC) = 0.479, and enrichment factor at 1% (EF@1%) = 7.58. RegressionStack achieved a mean absolute error (MAE) = 3.20 mm, ROC-AUC = 0.896, PR-AUC = 0.545, and EF@1% = 9.74. Neither paired permutation tests (ROC-AUC, p = 0.501; PR-AUC, p = 0.442) nor paired bootstrap confidence intervals resolved these differences. The y-randomization analyses supported non-random signals; scaffold-grouped validation retained early enrichment but showed reduced broader performance. The consensus-positive tier contained 27 actives among 35 nominations (PPV = 0.771). At measured IZ ≥ 30 mm, MAE increased to 9.46 mm, and all 30 compounds were underpredicted. Conclusions: Continuous-target modeling retained the IZ scale during training and generated a threshold-flexible predicted-IZ prioritization coordinate complementary to, but not statistically superior to, calibrated binary classification. The methodological contribution is a matched-data evaluation of complete workflows and their nomination behavior under the same partitions, yielding a retrospective compound-tiering scheme. Because the pipelines differed in architecture and molecular representation, their differences cannot be attributed solely to endpoint formulation. The workflows were not prospectively evaluated on compounds lacking pre-existing IZ measurements, and whether retrospective enrichment improves experimental hit discovery or reduces screening workload remains to be established.