Applying machine learning to classify table olives using bacterial metataxonomic data

Abstract

In recent years, metataxonomic analysis has been increasingly used to characterize microbial communities in fermented foods. Moreover, advances in bioinformatics and machine learning (ML) have expanded resources for analyzing these metataxonomic data. Particularly tree-based algorithms are valuable for their interpretability. This work compares the use of three tree-based ML algorithms-Classification and Regression Tree, Random Forest (RF), and Extreme Gradient Boosting- for the analysis of a database composed of 442 samples of 16S rRNA bacterial profiles obtained from table olives. Our findings show that ML techniques can effectively classify bacterial profiles based on olive processing type, cultivar, country of origin, and isolation matrix. The RF model achieved the highest accuracy, reaching 97% in the best cases, with a kappa coefficient above 0.8 for most categories. This approach holds potential applications in the table olive sector and in other food products, where the industrial application of ML techniques could enhance traceability, authenticity, and quality control.

Fig. 1. Application of AI to olive bacterial metataxonomic data.

Results of the 3 ML models CART, RF, and XGB for data classification according to olive processing type [A Prediction accuracy results and B Kappa coefficient results], olive cultivar [C Prediction accuracy results and D Kappa coefficient results], country of origin [E Prediction accuracy results and F Kappa coefficient results], and isolation matrix [G Prediction accuracy results and H Kappa coefficient results].

Fig. 2. Ranking of discriminant bacterial genera.

Top 20 most important bacterial genera deduced from the RF model for data classification according to their A type of olive processing, B olive cultivar, C country of origin, D isolation matrix.

Fig. 3

Workflow diagram used in the present study including data acquisition, preprocessing, building supervised ML models, and finally validation with test data.