Application of Machine Learning algorithms to essential oils to develop classification models of quantitative composition-activity relationship (QCAR)

In this study, the relationship between the chemical composition of essential oils and antibacterial activity against carbapenemase-resistant Acinetobacter baumannii (CRAB) was investigated, exploiting the capabilities of machine learning (ML) algorithms to identify the chemical components most responsible for the experimentally observed biological profile. Four ML models were generated and used to predict the antibacterial activity of 11 experimentally extracted essential oils whose antibacterial activities were evaluated. Two different training sets were compiled, and two separate calculations were performed for each dataset, one for Minimum Inhibitory Concentration (MIC) and the other for Minimum Bactericidal Concentration (MBC) values. The data were extracted from the PyEO database (eo.3d-qsar.com) and then processed on the Jupyter Notebook platform. The datasets were subjected to optimization using a Monte Carlo approach. The final models were analyzed through the Skater library, first determining the Features Importance (FI) for each component and then the Partial Dependence (PD). The PD emphasizes the positive or negative influence of each component within the essential oil against the associated biological activity. The four generated models were experimentally validated using an external test set of 11 essential oils. For both datasets, models with higher accuracy were obtained by separating essential oils into active and inactive with a cutoff value of 0.03% v/v (MIC or MBC). The FI analysis defined that carvacrol, eugenol, thymol, limonene, and eucalyptol were found to be the most influential on antibacterial activity, while for PD analysis carvacrol, eugenol, and thymol revealed positive influence according to their concentration within the essential oils. Limonene and eucalyptol, on the other hand, were assessed to negatively influence antibacterial activity. A mixed result was observed in the case of α-pinene, probably related to a synergistic or anti-synergistic effect depending on the percentage present. External validation of these four optimized models on the 11 essential oils with unknown activity toward CRAB predicted that only the essential oil extracted from Thymus vulgaris would have an activity less than 0.03% v/v. Microbiological tests determined MIC values in a range from 1.25% v/v to >5% v/v. By PD analysis a series of chemical components were selected, and their antibacterial activity was experimentally verified. In agreement with the PD data, the data confirmed an excellent MIC and MBC value for thymol (0.3% v/v) and carvacrol (0.3% v/v) and higher values for compounds such as eucalyptol and α-pinene, 2.5% v/v and 5% v/v, respectively. Of the eleven essential oils experimentally tested, the models correctly predicted that ten samples (91%) had no MIC values below 0.03% v/v. In addition, the PD analysis was confirmed by experimentally obtained data on some chemical components. In light of these results further studies are underway to develop increasingly robust and predictive models that can be used in the future to design "ad hoc" essential oil blends containing mainly the major components indicated by the ML models. The application of machine learning to essential oils thus represents an innovative perspective to address the threat of A. baumannii, paving the way for new natural and sustainable therapies in the fight against multidrug-resistant bacteria.