Machine learning decision tree-based models for predicting the antibacterial activity of Lamiaceae essential oils against Staphylococcus aureus

Essential oils (EOs) are complex mixtures of volatile specialised metabolites derived from plants with well-documented antimicrobial properties. Their antibacterial activity is influenced by the composition, concentration, and synergistic interactions of their constituents. Despite their therapeutic potential, predicting the antibacterial efficacy of EOs remains scarce due to their chemical diversity and variability. This study applies computational approaches to develop predictive models for the antibacterial activity of EOs from Lamiaceae plants against Staphylococcus aureus. A curated dataset (NAEOL – natural antimicrobial essential oils Lamiaceae dataset) was constructed from the literature, incorporating chemical composition, Kovats index, and antibacterial activity data using minimal inhibitory concentration (MIC) values. A multi-step feature selection strategy combining ANOVA filtering, Random Forest-based variable importance ranking, and J48-based evaluation was implemented to identify robust molecular predictors. Machine learning approaches, including decision tree models (Naive-Bayes-ClassifiersTree, Random Forest, Random Tree and J48) and artificial neural networks (ANN), were employed to classify EOs as active or inactive and to investigate key patterns of bioactive compounds in their chemical composition. The models were evaluated using statistical validation metrics, and strategies for data refinement were explored to enhance the predictive performance. The J48 model assumed the best results for the predictive antibacterial activity of EOs. This model was boosted and hyper-parametrized to improve prediction performance, integrating it into an automated open KNIME workflow. Chemical compounds estragole, bicyclogermacrene, o-cymene, γ-caryophyllene, and sabinene hydrate were relevant for discriminating the presence of antibacterial activity, demonstrating the potential of computational approaches for accelerating the search for promising bioactive EOs.