In this study, the predictive and classificatory capabilities of different ML models—CART, RF, and XGB—were evaluated to explain the behavior of various target variables including olive processing type, olive cultivar, country of origin, and isolation matrix of the samples, working with olive bacterial metataxonomic data. RF and XGB algorithms were more suitable for classifying instances of bacterial metataxonomic profiles of table olives. Increasing iterations generally improved results for RF and XGB compared to CART, which uses a single decision tree (CART). No significant differences were noted between RF and XGB, although RF showed slightly higher precision. Similar studies have also reported better results with the RF model for metataxonomic samples obtained from gut microbiota or river studies14,15. Moreover, RF and XGB models showed strong performance in classifying underrepresented categories in olive samples, such as olives darkened by oxidation, with only 7 observations, and the Cypriot cultivar, with only 6 observations, suggesting these categories have distinct bacterial metataxonomic profiles. For this purpose, the characterization of less studied geographical areas, as well as of little-researched cultivars, is essential. Expanding the current dataset could enhance the accuracy of these models; however, it is critical to ensure that the data are incorporated in appropriate proportions and with reliable information. If included in insufficient quantities, the data risk will be excluded during preprocessing. Moreover, the intrinsic environmental factors involved in olive fermentation (pH, salt levels, etc.) could also significantly influence the microbial composition of the samples. Consequently, future work could focus on incorporating a new classification category, such as identifying cases of spoiled fermentation, characterized by unusual fermentation processes as a consequence of deviations of pH, salt, etc.
ML involves techniques that can offer significant insights when applied correctly. Although these techniques have been applied in various fields of microbiology, there are very few studies in which they have been applied to food microbiology16,17. Using supervised ML models it is possible to develop models capable of predicting and understanding complex systems such as table olive fermentation. As demonstrated in this study, robust ensemble models can predict the type of olive processing, cultivar, and country of origin of samples with an accuracy of over 80%. Additionally, these models can predict the sample isolation matrix, whether brine or fruit, with more than 75% accuracy. A significant advantage of supervised ML techniques is that they typically improve their predictions with larger datasets. Therefore, the algorithms developed here could enhance their predictive accuracy with the inclusion of new samples from table olive fermentations. Furthermore, the application of these ML models could support industries and regulatory authorities in ensuring the authenticity and traceability of samples, as well as verifying the accuracy of product preparation. Additionally, the script development could be expanded to include automated analyses for detecting the presence of pathogenic and/or undesirable bacterial genera.
As mentioned previously, decision trees and classification algorithms are advantageous for their interpretability, helping to identify key bacterial genera for classification. The RF models, which performed best overall, were used to assess the importance of predictor variables. For olive processing types, Alkalibacterium had the highest score to discriminate, making it the most important genus for classifying samples into the purest nodes. Spanish-style green olives have the highest abundance of Alkalibacterium, suggesting that this genus is a key indicator for classifying a large number of samples of this olive processing type. This genus includes alkaliphilic and halophilic bacteria, commonly found in olives processed in Spanish style4. This prevalence is likely due to the use of an alkaline solution during one of the processing steps, as well as the use of marine salt as mentioned by others authors18. The second most important predictor variable, Unassigned taxa, also played a significant role in classifying Spanish-style olives, where it was most abundant. These unassigned sequences can arise for several reasons, including a lack of matches in the database used or due to low sequence quality. Consequently, it appears that in the Spanish-style processing, one or both of these factors may have occurred more frequently. Similarly, Marinilactibacillus, the third most important variable, was also key in classifying Spanish-style olives, where it was most abundant. Marinilactibacillus is frequently reported in the fermentation of table olives, particularly in the Spanish style4,19. On the other hand, Enterococcus was the fourth most important discriminant genus and was also more abundant in the production of Spanish-style green olives. Enterococcus is a LAB frequently reported in table olive fermentations1,18,19. Furthermore, Enterococcus has been used as a starter culture in table olive fermentations20. However, in recent years, it has been largely replaced by the genus Lactiplantibacillus due to its significant technological and probiotic potential21. Furthermore, Vibrio was the fifth most important genus and also exhibited higher abundance in the production of Spanish-style green olives. Although Vibrio is less commonly associated with this type of fermentation, it can appear during the early stages, whereas it tends to disappear towards the end of fermentation when the pH becomes more acidic4,19. Vibrio also showed notable abundance in the production of oxidized black olives, similar to Acetobacter (ranked 16th in importance), which was almost exclusively present in samples of oxidized black olives. This is likely due to the specific nature of this fermentation process, where brine is acidified with acetic acid, creating conditions conducive to the proliferation of acetic acid bacteria such as Acetobacter22. Upon reviewing the abundances, Celerinatantimonas and Pediococcus were identified as distinguishing markers for natural black olives, while Halomonas and Pediococcus were more prominent in natural green olives. In natural-style olive fermentation brine, Pediococcus and Lactiplantibacillus are dominant genera, so it is logical to find a certain abundance of Pediococcus in both natural fermentations23,24,25. Furthermore, the abundance data for each fermentation suggests that Pediococcus genus is more closely associated with natural fermentations than with Spanish-style green olive production.
For the olive cultivar category, the variable with the highest discrimination power was Unassigned taxa, which was most abundant in the Gordal, Aloreña, and Manzanilla cultivars. This likely helped in creating purer nodes for these three cultivars. The second most predictive variable was the Lactiplantibacillus genus, which served to differentiate samples from the Nyons cultivar, which showed an almost complete absence of this genus. The absence of Lactiplantibacillus can be attributed to the unique processing of these French PDO olives, which, although a natural black fermentation, has distinct characteristics. As described by Penland et al.26, this fermentation is primarily dominated by the genus Celerinatantimonas. The next genus with a high discrimination power was Alkalibacterium, which was more abundant in the Manzanilla and Hojiblanca cultivars. This might be due to these cultivars frequent use in Spanish-style olive production, though other cultivars like Gordal, Nocellara Etnea, and Halkidiki, also produced as lye-treated olives, did not show notable abundances. Therefore, it can be hypothesized that these two cultivars, particularly Manzanilla, may have a predisposition to Alkalibacterium colonization, possibly due to the use of specific salts containing this genus. Further studies are needed to understand this phenomenon. Additionally, the fourth most significant genera, Loigolactobacillus, was primarily used to differentiate the Hojiblanca cultivar. Loigolactobacillus is another LAB genera capable of fermenting sugars in brine, although it is not commonly found in olive fermentations. The information on this genus is limited, but previous studies suggest that species belonging to this genus participate in cheese and various vegetable fermentations and have been reported in spoilage-related fermentations27,28. The fifth most significant genus was Halomonas, with a greater abundance in the Manzanilla and especially Nocellara Etnea cultivars. A potential explanation offered by other studies is that its presence may result from the use of sea salt in brine29. Thus, the use of sea salt in certain cultivars could result in a distinctive taxonomic profile, making it easier to classify using algorithms like RF. Notably, the genera Cellulosimicrobium and Celerinatantimonas—eighth and ninth in discrimination importance, respectively—were more abundant in the Nyons cultivar. Cellulosimicrobium has previously demonstrated xylanase activity capable of degrading plant cell walls30. However, it is important to mention that the detection of Cellulosimicrobium in amplicon-based metagenomic studies has been subject to controversy. Previous reports have indicated that certain commercial preparations of lyticase may introduce this genus as a contaminant31. Therefore, the presence of Cellulosimicrobium in table olive ecosystems should be interpreted with caution and warrants further investigation using validated lyticase marked as free of DNA contaminants and appropriate negative controls. In the case of Celerinatantimonas, this microorganism has been reported as a bacterium that damages the fruit by creating gas bubbles inside32. The abundance of these two genera and the use of the RF model for classifying the Nyons cultivar make sense, as this PDO is characterized by olives with a melting texture26.
For the target variable country of origin, the most important predictor variable was Unassigned taxa, which was predominantly abundant in olives originating from Spain. The country of origin, along with the previous metadata, likely relates to the results for Unassigned taxa, since samples from Spain were mainly of the Gordal, Manzanilla, and Hojiblanca cultivars, which are often processed as Spanish-style green olives. The second most important variable for the country of origin was Alkalibacterium, which, as previously mentioned, is associated with olive fermentations involving alkaline solutions. In this case, the countries most associated with this genus, with notable abundances, were Spain, France, Algeria, and Greece. The third genus Halomonas was predominantly found in samples from Italy and Spain, correlating with the high abundance in the Nocellara Etnea and Manzanilla cultivars, typical of Italy and Spain, respectively. As mentioned earlier, the presence of this genus may be linked to the use of sea salt in brine preparation. The fourth most important genus was Cellulosimicrobium, which was exclusively abundant in samples from France, particularly due to the PDO Nyons olives discussed earlier, and to a very low extent in Greece, possibly useful for classifying samples from this latter country. The genus Marinilactibacillus ranked fifth in terms of discrimination power, with relative abundances present across all countries but varying in degree. Samples from Spain showed the highest abundance, followed by Italy, Algeria, and, to a lesser extent, France, Cyprus, and Greece. This genus of halophilic bacteria has also been associated with the presence of marine salts during brine preparation29.
Finally, the isolation matrix had Celerinatantimonas as the most significant bacterial genus. This genus has been associated with gas production beneath the fruit’s skin, causing damage, and is thus generally considered a spoilage microorganism32. Celerinatantimonas was predominantly associated with samples isolated from the fruit, where its abundance was greater than in brine, supporting its association with fruit damage. Both Enterococcus, the second most significant genus, and Halomonas, the fifth, were associated exclusively with brine samples, with negligible abundance in fruit-isolated samples. Heinemann et al.33 shown as surface proteins of Lactiplantibacillus plantarum inhibit the adhesion of Enterococcus, which may explain the low abundance of Enterococcus on the fruit’s skin. Conversely, the genera Staphylococcus and Lactiplantibacillus, the third and fourth most significant genera, were more abundant in the fruit. Both genera are well-known for their biofilm-forming capabilities5,34. However, Staphylococcus is not typically found in table olive fermentations; when present, it has always been in low proportions35,36. In contrast, Lactiplantibacillus is the predominant genus in table olive fermentation5. On the other hand, Vibrio and Marinilactibacillus, ranked ninth and sixteenth in importance, respectively, were more abundant in brine samples. These genera, along with Halomonas, are bacteria found in halophilic environments. The data suggest that these halophilic bacteria prefer a free-living existence rather than being part of the fruit’s biofilm.
In conclusion, food metataxonomic database can be very useful when is analyzed using ML models. Results obtained in this work have demonstrated that tree ensemble algorithms, such as RF and XGB, are more accurate in classifying bacterial metataxonomic profiles of table olives samples compared to decision tree algorithms. Furthermore, these models provide valuable insights into the attribute importance, in this case, bacterial genera, for data classification. This allows for a quick identification of which bacterial genera are most associated with each category of the studied metadata. These findings highlight the practical utility of these ML models in the table olive industry and food policy for purposes such as authenticity verification, traceability, quality and safety control, but they can also be applied to other foods including also other microbial groups (fungi). However, we must to take in consideration that this study is limited by the dataset used and their self metadata, which does not include samples from all olive-producing countries, other olive varieties utilized in table olive processing or environmental factors which usually govern table olive fermentations (pH, salt, etc.). Furthermore, most of the data comes from sequencing performed using Illumina technology, and it remains to be determined whether metataxonomic studies conducted with third-generation sequencing platforms, such as PacBio and Oxford Nanopore, can be integrated into the database already generated. Additionally, the methodology used, such as sampling, reagents, DNA extraction, and sequencing technology, varied across the different studies evaluated, which may have influenced the classification efficiency. However, this is considered intrinsic variability in experimental procedures, which is common in microbiome studies. Despite these limitations, we consider that results obtained in this study are highly encouraging and with application in food policy.
