Abstract
<jats:p>The intrinsic resistance of Gram-negative bacteria to many antibiotics is largely attributed to the low permeability of their outer membrane, making the accurate prediction of compound permeation a major challenge in antibiotic discovery. Many existing approaches rely on high-dimensional representations that limit interpretability and reproducibility. In this study, we present an interpretable machine learning framework for predicting Gram-negative outer membrane permeability using a reduced set of physicochemical descriptors. From a predefined pool of 59 descriptors, we apply a two-step feature selection strategy that combines hierarchical clustering with cluster-wise Random Forest modelling. The resulting reduced feature sets comprise 19 descriptors for Pseudomonas aeruginosa (P. aeruginosa) and 12 for Escherichia coli (E. coli). The framework was evaluated on two public datasets containing 600 compounds for P. aeruginosa and 1,293 compounds for E. coli using 5-fold cross-validation and scaffold-based validation to assess generalisation to unseen scaffolds. Across multiple classifiers, the reduced descriptor models achieved predictive performance comparable to the full descriptor and Extended Connectivity Fingerprint fingerprints (ECFP4), with representative AUC values reaching approximately 0.92 for P. aeruginosa and 0.93 for E. coli under standard cross-validation. Functional group augmentation further improved the performance of the reduced descriptor set, particularly for E. coli. Validation using a benchmark dataset demonstrated that the Reduced descriptor set augmented with functional group counts, generalised more robustly to independent datasets compared with ECFP4 fingerprints. Overall, our results demonstrate that computationally efficient models can achieve competitive predictive performance while providing mechanistic insight into Gram-negative outer membrane permeability to support antibiotic design against such pathogens.</jats:p>