Pathway-Derived Features for Machine Learning-Based Prediction of Multidrug Resistance in Pseudomonas Aeruginosa

Abstract

Antimicrobial resistance (AMR) in Pseudomonas aeruginosa poses a critical threat to healthcare, particularly due to surgical site infections. Current phenotypic testing methods are time-consuming, while existing antimicrobial resistance gene-based machine learning approaches suffer from limited data availability, poor multi-antibiotic prediction performance, and lack of interpretability. This study presents a novel framework using KEGG pathway-derived features to predict multidrug resistance in P. aeruginosa. We annotated whole genome sequences to identify AMR-related metabolic and regulatory pathways, then evaluated these pathway-based features across multiple algorithms, including traditional machine learning methods like logistic regression, random forest, and KNN; deep learning models like MLP, CNN, LSTM, and GRU; and the tabular foundation model TabPFN. All algorithms achieved an accuracy above 0.85, demonstrating the robustness of using pathway-derived features for multi-drug AMR prediction. TabPFN showed superior performance with an accuracy of 0.926, successfully detecting resistance phenotypes across multiple antibiotics simultaneously while providing interpretable insights. In conclusion, the pathway-based features capture complex biological mechanisms underlying resistance beyond simple gene presence/absence, offering a meaningful framework for rapid resistance profiling. We hope this method enables accurate antibiotic selection in clinical settings, potentially reducing inappropriate antibiotic use and combating the spread of the AMR phenomenon.