Electronic Nose for Monitoring Odor Changes of Lactobacillus Species during Milk Fermentation and Rapid Selection of Probiotic Candidates

Probiotic bacteria have been associated with a unique production of aroma compounds in fermented foods but rapid methods for discriminating between foods containing probiotic, moderately probiotic, or non-probiotic bacteria remain aloof. An electronic nose (e-nose) is a high-sensitivity instrument capable of non-invasive volatile measurements of foods. In our study, we applied the e-nose to differentiate probiotic, moderately probiotic, and non-probiotic Lactobacillus bacteria strains at different fermentation time points (0th, 4th, and 11th) of milk fermentation. The pH of the changing milk medium was monitored with their corresponding increase in microbial cell counts. An e-nose with two gas chromatographic columns was used to develop classification models for the different bacteria groups and time points and to monitor the formation of the aromatic compounds during the fermentation process. Results of the e-nose showed good classification accuracy of the different bacteria groups at the 0th (74.44% for column 1 and 82.78% for column 2), the 4th (89.44% for column 1 and 92.22% for column 2), and the 11th (81.67% for column 1 and 81.67% for column 2) hour of fermentation. The loading vectors of the classification models showed the importance of some specific aroma compounds formed during the fermentation. Results show that aroma monitoring of the fermentation process with the e-nose is a promising and reliable analytical method for the rapid classification of bacteria strains according to their probiotic activity and for the monitoring of aroma changes during the fermentation process.


Introduction
Milk is an important dairy product from mammalian sources such as cows, goats, buffalo, or sheep. Its fermented form has yielded many different foods in the dairy industry such as ice cream, cheese, yogurt, etc. Fermentation of milk into food products is often associated with starter cultures such as Streptococcus thermophilus and Lactobacillus delbrueckii subsp. bulgaricus [1]. Lactobacilli are gram-positive bacteria consisting of both anaerobic and aerobic species. The main product of their fermentation is lactic acid but secondary products are also generated as diacetyl, acetoin, and acetone [2]. Lactobacillus strains have significant industrial importance due to their fermentation activity and health benefits

Analyzed Bacteria Strains
Fifteen different types of Lactobacillus bacteria strains were acquired in a freeze-dried state from reputable sources and separately used to produce fermented milk products based on their probiotic grouping. Depending on the growth rate, optical density, biomass production, minimal inhibitory concentration of bile, and best recovery after 3 h at low pH and pepsin, the bacteria strains could be grouped into non-probiotic (S2, S3, S4, S29, S30), moderately probiotic (S1, S7, S8, S9, Y12), and probiotic Foods 2020, 9,1539 3 of 16 (R1, S6, S10, S11, S22) [9]. All the species used in this study were Lactobacilli bulgaricus. They ferment the glucose to lactic acid as a major end product by homolactic fermentation (Embden-Meyerhof-Parnas pathway and subsequent pyruvate reduction).

Preparation of Reconstituted Milk
Skimmed milk powder for microbiological use was acquired from Sigma Aldrich (St. Louis, Missouri, USA) and used in this study. According to the producer, the skimmed milk powder contained 4.7-6.0% total nitrogen (N), ≤1.5% lipid, ≥50.0% reducing sugars (as lactose monohydrate), and ≤5% loss on drying. On each experimental day, the skimmed milk powder was freshly diluted to 10% w/v in sterilized distilled water until it reached 3.5% protein content as recommended by the producer. For purposes of this study, the resulting mixture was referred to as reconstituted milk and could be said to contain <1% ash, <0.15% lipid, 0.47-0.6% total nitrogen, and about 5% reducing sugar content per the calculations after the 10% w/v dilution.

Preparation of Strain Suspension (Activation of Freeze-Dried Bacteria)
The freeze-dried bacteria strain (10 mg) was weighed into a 10 mL flask then filled up to volume with the reconstituted skimmed milk. This was referred to as the strain suspension and was cultured for 24 h at 37 • C to obtain the activated bacteria before initial cell number counting was performed. Cell numbers were determined using the Breed staining method [17] with a light microscope at 100-times magnitude with immersion oil in three parallel repeats. Each of the 15 Lactobacillus bacteria strains suspension was prepared on different days using a random design.

Preparation of the Milk Suspension
The activated strain suspension was further diluted with reconstituted milk to set the initial cell count of the milk suspension at 10 6 CFU/mL using the average cell number determined by the Breed staining method. This sample preparation was applied to provide similar conditions for the milk fermentation for all 15 bacteria strains having different growing characteristics. A total of 400 mL of the milk suspension was prepared for the entire study for each of the 15 strains on different days.

Determination of the Cell Count at the Different Fermentation Time Points
The fermented milk suspension was incubated for eleven hours and colony counts were determined at the beginning (0 h), after 4 h (4 h), and after eleven hours (11 h) of fermentation in three replicates using the layered plating method on MRS agar (Biolab Co. Ltd., Budapest, Hungary). The plating was performed by a 10-times dilution in six steps, depending on the fermentation time points that had the most influence on the colony count. These fermentation time points were determined to be the time zero: 10 1 , 10 2 , 10 3 ; fourth hour: 10 2 , 10 3 , 10 4 ; eleventh hour: 10 4 , 10 5 , 10 6 , respectively. The plates were incubated for 72 h at 37 • C before counting the colonies.

Analysis of pH
The pH was measured in a controlled environment for 20 h at 37 • C from the fermented milk suspension. The pH change during the milk fermentation was monitored with a Mettler Toledo Seven Multi pH meter every four minutes and resulted in 300 pH values for each fermented milk suspension.

Analysis of the Aroma Composition of the Milk Suspension during the Fermentation Process Using the E-Nose
During the fermentation process, 100 mL of each milk suspension was collected from the fermenting batch for the electronic nose measurements at the beginning, at 4 h, and at 11 h of fermentation and was stored at −18 • C until e-nose measurements. E-nose measurements were performed with a Heracles Neo 300 ultra-fast GC analyzer (Alpha MOS, Toulouse, France). The Heracles e-nose is composed Foods 2020, 9, 1539 4 of 16 of a rapid, selective, and highly sensitive gas chromatography system that is specifically designed for the analysis of volatile compounds. The system is equipped with an odor concentrator that is called a trap. Its principle of operation is completed when the samples are injected and concentrated in the cold trap, then the trap is heated, and the concentrated odor is injected and divided into two columns-Restek MXT-5: length 10 m; ID 0.18 mm; thickness: 0.40 µm; low-polarity stationary phase; and Restek MXT-1701: length 10 m; ID 0.18 mm; thickness: 0.40 µm; mid-polarity stationary phase (Restek, Co., Bellefonte, PA, USA). Both columns are metal capillary columns, MXT-5 is composed of Crossbond 5% diphenyl/95% dimethyl polysiloxane, while MXT-1701 is composed of Crossbond 14%cyanopropylphenyl/86% dimethyl polysiloxane. The volatile compound is separated by both columns and detected with two flame ionization detectors (FID). The autosampler and the analyzer were operated with the software AlphaSoft ver. 16 (Alpha MOS, Toulouse, France), and the same software was used for data acquisition and basis data transformations. During data acquisition, the retention time of the volatiles was recorded, where retention time is characterized by the elution time of the molecules. Retention times were converted to retention indices. The Kovats retention index compared the retention time of n-alkanes with the retention time of the investigated volatile molecules of a sample under the same conditions [18]. The Kovats index (KI) characterized the volatile compounds on the specific columns and could be assigned to a specific aroma, which was recorded in the AroChemBase v7 [19]. Throughout the manuscript, as an identifier after the KI, the "1A" appears for the column MXT-5 and "2A" for the column MXT-1701. The autosampler's tray of the e-nose was equipped with a self-developed thermostat which allowed the samples to be stored at 8 • C during the sampling period. Samples were measured on three different days: on the first day, the samples of the 0th hour; on the second day, the samples of the 4th hour; and on the third day the samples of the 11th hour. The frozen milk suspensions were thawed right before the respective e-nose tests and kept in a water bath at 40 • C for 10 min. Each sample was prepared and injected into the e-nose in three repeats.
Before the analysis, a method was created with the following parameters of the PAL-RSI Autosampler and the Heracles GC analyzer. Autosampler: 1 g of sample in 20 mL headspace vials with PTFE cap, incubation at 70 • C for 20 min with 500 rpm agitation to generate headspace, 5 mL of headspace injected into the Heracles analyzer, flushing time between injections of 90 s. Analyzer: hydrogen carrier gas, flow of carrier gas at 30 mL/min, trapping temperature at 50 • C, initial oven temperature at 50 • C, endpoint of oven temperature at 250 • C, heating rate at 2 • C/s, acquisition duration of 110 s, acquisition period of 0.01 s, injection speed of 125 µL/s, cleaning phase of 8 min.

Microbial Assessment
The change of the pH value during the fermentation was approximated with the model described by Torrestiana et al. [20] for each of the tested bacteria strains using the following equation: where: A is the initial culture pH; B relates to the slope of the linear decreasing region from the pH-time curve; C represents the time at which the initial pH decreased to half of its fixed value; D is the final culture pH. The fitted curves were used to determine the pH and fermentation time at the inflection point for each of the strains. A logarithm with base ten (log 10) of each colony count was calculated and the averaged value of the three parallel samples was recorded for each strain to determine the cell number at the 0th, 4th, and 11th hours of incubation. The initial cell count in the milk suspension was corrected to 10 6 CFU/mL by dilution based on the cell number determined by the Breed staining method but the layered plating method resulted in slight differences. This was, however, expected due to the relatively high uncertainty of the Breed staining method. Thus, relative colony counts were calculated to reduce Foods 2020, 9, 1539 5 of 16 the differences that occurred in the initial cell count: the average value of colony counts in 1 mL fermented milk suspension counted at the 4th and 11th hours of fermentation was divided by the average value of the initial colony count for each of the tested strains, separately.
The pH inflection points and relative cell numbers were calculated and illustrated using Microsoft Excel 2013. ANOVA was used to identify any significant differences among the groups of non-probiotic, moderately probiotic, and probiotic strains in the case of relative numbers and inflection points of pH curves. Where ANOVA indicated, a Tukey HSD test (p < 0.05) was used for detecting the significant differences between the groups [21]. The ANOVA tests were performed in R-project (ver. 3.6.3) software (R Core Team, Vienna, Austria).

E-Nose Data Construction and Analysis
The results of the e-nose tests obtained during the three days experiment were combined into one file using the AlphaSoft (ver. 16) software (Alpha MOS, Toulouse, France). The chromatograms were transformed into individual variables, in the AlphaSoft (ver. 16) software called sensors, based on the area and respective Kovats indexes of the identified peaks. Classification models were built using linear discriminant analysis (LDA) based on the results of the most selective sensors. Models were built for the discrimination of the strains according to their probiotic activity. This was done separately for the 0th, 4th, and 11th hour. Using the results of the non-probiotic, moderately probiotic, and probiotic strains, LDA models were also built for the differentiation of the different fermentation times (0, 4, and 11 h). The LDA calculations based on the most distinctive sensors were done in the AlphaSoft (ver. 16) software. The models were characterized by the distance and pattern discrimination index (%) between the classified groups.
The results of the e-nose were also analyzed using continuous chromatograms for the two columns separately. LDA models were separately built to classify the groups of the tested strains with regards to their probiotic activity and the three fermentation times. The models were validated with fivefold-cross-validation. The data of the continuous chromatograms were analyzed using R-project (ver. 3.6.3) software.

Growth Characterization of Bacteria Strains in Reconstituted Milk
Results of the average pH change for the probiotic, moderately probiotic, and non-probiotic bacteria groups during the fermentation of milk are shown in Figure 1. A decrease in pH was observed with the increase in the fermentation time. A decrease in pH (from neutral to acidic) was observed from the first hour to the fifteenth. The pH values of the probiotic group decreased at the highest rate especially between the 7th and 9th hours. The general decrease in the pH of all the different bacteria groups can be attributed to the production of lactic acid but the variation in the curves could be due to the different lactic acid-producing activities of the different strains. The pH of all groups, however, became relatively stable after the 15th hour of fermentation when it dropped below pH 4.
Based on the pH curves in Figure 1, the fourth (4th) and eleventh (11th) hours were selected for the comparison of the pH, relative cell count, and aroma composition of the tested bacteria strains. The 4th hour was identified as the lag phase which could be the earliest time to discriminate between the probiotic, moderately probiotic, and non-probiotic groups. On the other hand, the 11th hour coincided with the end of the log phase and was characterized by the highest pH difference between the non-probiotic and the probiotic groups. In addition, this time point was found to be the best for the discrimination of non-probiotic and probiotic candidates in similar studies [9]. Based on the pH curves in Figure 1, the fourth (4th) and eleventh (11th) hours were selected for the comparison of the pH, relative cell count, and aroma composition of the tested bacteria strains. The 4th hour was identified as the lag phase which could be the earliest time to discriminate between the probiotic, moderately probiotic, and non-probiotic groups. On the other hand, the 11th hour coincided with the end of the log phase and was characterized by the highest pH difference between the non-probiotic and the probiotic groups. In addition, this time point was found to be the best for the discrimination of non-probiotic and probiotic candidates in similar studies [9]. Figure 2a, shows the inflection points of the pH curve and the time of the inflection points for each analyzed bacteria strain, while Figure 2b, presents the average and standard deviation of these values for the non-probiotic, moderately probiotic, and probiotic groups. There were no significant differences found between the inflection points and times of inflection for the analyzed three groups, but the variation in inflection points revealed a lower standard deviation for the probiotic and moderate groups compared to the non-probiotic groups.    Figure 2b, presents the average and standard deviation of these values for the non-probiotic, moderately probiotic, and probiotic groups. There were no significant differences found between the inflection points and times of inflection for the analyzed three groups, but the variation in inflection points revealed a lower standard deviation for the probiotic and moderate groups compared to the non-probiotic groups. Based on the pH curves in Figure 1, the fourth (4th) and eleventh (11th) hours were selected for the comparison of the pH, relative cell count, and aroma composition of the tested bacteria strains. The 4th hour was identified as the lag phase which could be the earliest time to discriminate between the probiotic, moderately probiotic, and non-probiotic groups. On the other hand, the 11th hour coincided with the end of the log phase and was characterized by the highest pH difference between the non-probiotic and the probiotic groups. In addition, this time point was found to be the best for the discrimination of non-probiotic and probiotic candidates in similar studies [9]. Figure 2a, shows the inflection points of the pH curve and the time of the inflection points for each analyzed bacteria strain, while Figure 2b, presents the average and standard deviation of these values for the non-probiotic, moderately probiotic, and probiotic groups. There were no significant differences found between the inflection points and times of inflection for the analyzed three groups, but the variation in inflection points revealed a lower standard deviation for the probiotic and moderate groups compared to the non-probiotic groups.   Colony counts for the 4th and 11th hours of fermentation relative to the initial count of each tested bacteria strain are separately shown in Figure 3. The a, c, and e diagrams show the relative values at 4 h while the b, d, and f show them at 11 h of fermentation. Colony counts for the 4th and 11th hours of fermentation relative to the initial count of each tested bacteria strain are separately shown in Figure 3. The a, c, and e diagrams show the relative values at 4 h while the b, d, and f show them at 11 hours of fermentation. Generally, there was only a slight increase in colony counts for all the 15 tested bacteria strains in the first four hours. The increase in colony counts reached 0.5 order of magnitude in only a few cases for moderately probiotic and non-probiotic groups (Figure 3c,e). A further increase in colony counts was observed after 11 h of fermentation for all the strains resulting in 1.5-2 order of magnitude higher colony counts from the 10 6 initial number. This resulted in an average of 5 × 10 7 to 10 8 CFU/mL of the cells after 11 h of fermentation in the milk. Strain Y12 in the moderate group had the highest relative value compared to all the other strains after the 11th hour. This was followed by strain S11 belonging to the probiotic group.
Some differences were observed when the averages of the relative value per colony were calculated (Figure 3g,h) for the different bacteria groups but the differences were not significant. The average relative value of the probiotic group was the smallest at the 4th and 11th hour, but the growth of the moderate group was the most intense at the 11th hour. This agreed with the results of the averaged inflection point for the pH curves. It can be also noted that there was a relatively small increase in colony counts after four hours of fermentation in the probiotic group. This was similar to those of the moderately probiotic and non-probiotic groups after 11 h of fermentation. Figure 4 shows the change in relative values for the three groups from the 4th hour to the 11th. Between the defined time points, the growth of the strains in the probiotic group was the highest and the non-probiotic strains were the slowest. This proves that among all the three different bacteria groups investigated, the probiotic strains provided the highest growth rate when pH started to decrease i.e., in the logarithmic phase of the growth curve. Generally, there was only a slight increase in colony counts for all the 15 tested bacteria strains in the first four hours. The increase in colony counts reached 0.5 order of magnitude in only a few cases for moderately probiotic and non-probiotic groups (Figure 3c,e). A further increase in colony counts was observed after 11 h of fermentation for all the strains resulting in 1.5-2 order of magnitude higher colony counts from the 10 6 initial number. This resulted in an average of 5 × 10 7 to 10 8 CFU/mL of the cells after 11 h of fermentation in the milk. Strain Y12 in the moderate group had the highest relative value compared to all the other strains after the 11th hour. This was followed by strain S11 belonging to the probiotic group.
Some differences were observed when the averages of the relative value per colony were calculated (Figure 3g,h) for the different bacteria groups but the differences were not significant. The average relative value of the probiotic group was the smallest at the 4th and 11th hour, but the growth of the moderate group was the most intense at the 11th hour. This agreed with the results of the averaged inflection point for the pH curves. It can be also noted that there was a relatively small increase in colony counts after four hours of fermentation in the probiotic group. This was similar to those of the moderately probiotic and non-probiotic groups after 11 h of fermentation. Figure 4 shows the change in relative values for the three groups from the 4th hour to the 11th. Between the defined time points, the growth of the strains in the probiotic group was the highest and the non-probiotic strains were the slowest. This proves that among all the three different bacteria groups investigated, the probiotic strains provided the highest growth rate when pH started to decrease i.e., in the logarithmic phase of the growth curve.

Discrimination of Probiotic Strains Based on Their Aroma Composition Using E-Nose
LDA models built for the discrimination of the non-probiotic, moderately probiotic, and probiotic groups inoculated in milk have been presented in Figure 5. Figure 5a presents the model for the onset of the milk fermentation (0th hour), Figure 5b, after four hours of fermentation, and Figure 5c, after 11 h of fermentation. Improving classification tendency was observed with the longer fermentation time. The average cross-validation score of the samples at the 0th hour was 38% using the 27 selected sensors but increased to 54% in the 11th hour. At the 0th and 4th hour, there was some overlapping between the three groups, but a clear separation was observed at the 11th hour.

Discrimination of Probiotic Strains Based on Their Aroma Composition Using E-Nose
LDA models built for the discrimination of the non-probiotic, moderately probiotic, and probiotic groups inoculated in milk have been presented in Figure 5. Figure 5a presents the model for the onset of the milk fermentation (0th hour), Figure 5b, after four hours of fermentation, and Figure 5c, after 11 h of fermentation. Improving classification tendency was observed with the longer fermentation time. The average cross-validation score of the samples at the 0th hour was 38% using the 27 selected sensors but increased to 54% in the 11th hour. At the 0th and 4th hour, there was some overlapping between the three groups, but a clear separation was observed at the 11th hour. Figure 5c shows a complete separation of the three tested groups according to their probiotic activity mainly through the first discriminant factor (DF 1), which describes 81.979% of the total variance between groups. From the left to the right, the tendency according to the increasing probiotic activity was observed. The loading vectors show that sensors 985.4-1A, 1000.33-1A, 922.4-2A, and 1202.47-1A contributed the most to differentiate among the groups. The identified compounds based on the retention indices obtained by the electronic nose are shown in Table 1. Sensor 1000.33-1A can be assigned to the aroma compound of 2-octanol (expressing fatty or oily aroma), decane (sweet aroma), and 2,4-heptadienal (fatty aroma), whereas, 1202.47-1A can be assigned to the compound of pyridine, 2-pentyl (fatty and tallowy aroma). This shows that non-probiotic samples can be characterized more by the fatty aroma. As per literature data, 2-octanol was found during cheese fermentation [22], while 2,4-heptadienal was associated with milk during fermentation [23,24].
In The different aromas have also been identified in probiotic yogurts with the GC-MS method [23,26]. Models built based on the continuous data of the chromatograms of the two columns separately showed similar results for the two columns. Classification models built for the discrimination of the non-probiotic, moderately probiotic, and probiotic groups can be seen in Table 2 separately for the three sampled time points. Slightly better classifications were obtained using column 2 (MXT-1701). The average recognition and prediction abilities were 74.44% and 26.66 % for column 1, and 82.78% and 48.99% for column 2 at the 0th hour, respectively. Misclassifications were found in all of the cases, but generally probiotic and non-probiotic bacteria were misclassified as moderately probiotic and vice versa. Probiotic and non-probiotic groups also had 20% and 13.33% interclass misclassification in the case of column 1 and column 2, respectively. At the 4th hour, the recognition and prediction abilities improved compared to the 0th hour for all the bacteria groups. Recognition and prediction were 89.44% and 60% for column 1, and 92.22% and 66.97% for column 2, respectively. At the 11th hour, there was only slight improvement observed in the average recognition and prediction accuracies, however, misclassification between non-probiotic and probiotic groups was further reduced. The average recognition and prediction abilities were 81.67% and 55.55% for column 1, and 81.67% and 59.99% for column 2, respectively.   Table 1. Sensor 1000.33-1A can be assigned to the aroma compound of 2-octanol (expressing fatty or oily aroma), decane (sweet   and 1-hexen-3-one (metallic). The metallic aroma was also found to be associated with the probiotic samples at the 11th hour (1434.32-2A).

Conclusions
Measurements of pH showed a sharp decrease from the 1st hour to the 15 th , which could be attributed to the production of lactic acid by the different probiotic bacteria strains. Evaluation of relative values for the different bacteria groups at the 4th and 11th hour showed an increase in the average relative colony counts of the probiotic group in the fermented milk. Bacteria strains S11 and Y13 had the highest relative values after 11 h of fermentation suggesting that those may be the ideal choices for yogurt fortification. The change in relative values was the highest in the probiotic group implying that those were the most sensitive to the conditions of yogurt fermentation.

Conclusions
Measurements of pH showed a sharp decrease from the 1st hour to the 15th, which could be attributed to the production of lactic acid by the different probiotic bacteria strains. Evaluation of relative values for the different bacteria groups at the 4th and 11th hour showed an increase in the average relative colony counts of the probiotic group in the fermented milk. Bacteria strains S11 and Y13 had the highest relative values after 11 h of fermentation suggesting that those may be the ideal choices for yogurt fortification. The change in relative values was the highest in the probiotic group implying that those were the most sensitive to the conditions of yogurt fermentation.
Electronic nose analysis showed an increasing average recognition and prediction accuracy for the different bacteria groups from the 0th hour to the 4th and 11th. The highest accuracy was observed after 11 h of fermentation, confirming an improved classification tendency with longer fermentation time. Better classifications were obtained using column 2 (MXT-1701) of the electronic nose. The loadings vector of the different classification plots revealed that some of the sensors of the electronic nose could be associated with specific aromas in yogurt. Separate models developed for the acquisition times of the different bacteria groups showed a complete separation of the time points for the probiotic, moderate, and non-probiotic groups. Generally, distances between the time groups were higher between the 0th and 11th hour sample points than between the 0th and 4th or 4th and 11th hours. The electronic nose presents a non-invasive and reliable method for rapid classification and prediction of these parameters.