Secondary analysis was performed on data obtained from a clinical trial conducted on 42 healthy adults which evaluated the safety and immune response to daily consumption of L. johnsonii N6.2 (13, 14). The clinical trial consisted of three periods, a 1-week baseline, 8-week intervention, and 4-week washout. During the intervention period, subjects received either 1 capsule per day of 10⁸ colony-forming units (CFU) of L. johnsonii N6.2 or placebo (skim milk) for 8 weeks. The study was approved by the Institutional Review Board 1 (# 201400370) at the University of Florida. Total LAB counts were analyzed on stool samples collected at baseline, week 2, 4, and 8 of treatment, and a final collection after washout (week 12). Details of the stool collection methods can be found in Marcial et al. (10). LAB were enumerated by plating fresh stool samples (1 g) diluted in phosphate buffer solution (pH 7.4) on acidified MRS agar media (pH 5.5 ± 0.1) and incubating 48 h at 37°C under microaerobic conditions. Enumeration values were referred to as CFU per wet gram stool (CFU/g).

Study participants recorded their dietary intake for 2 days during each period using the online software Automated Self-Administered 24-Hour (ASA24®) Dietary Assessment Tool (version 2018) for a total of up to six recalls (13). Participants had 24 h to complete their dietary recall from the date of the intake and were not alerted beforehand of which dates would be scheduled. Participants who completed less than 3 days of intake were excluded from statistical analysis (n = 5). See Table 1 for the adjusted demographics of the data analyzed.

For initial analyses, the data were separated by intervention, participants receiving L. johnsonii N6.2 (n = 20), and the placebo group (Pla; n = 18; See Supplementary Figure S1). The data were then stratified based upon the participants’ stool counts of LAB. Three groups were previously defined in Marcial et al. (10) as subjects with high counts of LAB throughout the study (>10⁵ CFU/g) termed High LAB (H-LAB, n = 18), subjects that increased in counts of LAB over the course of the study (10⁴ to 10⁸ CFU/g) termed Low-to-High LAB (LH-LAB, n = 10), and subjects with low counts of LAB throughout the study (<10⁵ CFU/g) termed Low LAB (L-LAB, n = 10). Associations between dietary intake, CFU/g wet stool LAB, and serum untargeted metabolomic data of these subjects were then explored based on this stratification. Each stratification and its purpose are described for each separate set of data and analysis in the methods below.

Individuals were grouped based on their intervention group (L. johnsonii N6.2 versus placebo) and by LAB counts in separate analyses (Supplementary Figure S1). Individual average nutrient intakes were adjusted to estimate the habitual diet, defined as the average food consumption over a period of time over which a dietary pattern is maintained (18). To remove within-person variance, the equation put forward by the US National Academy of Science Subcommittee on Criteria for Dietary Evaluation was used,

where the average is the mean value for the group, x_i is the value observed for each individual, S_b/S_obs is the inverse of the within-person variance (19). Finally, the residual method was used to normalize for energy intake and gender (20). Using the adjusted average intake values and the stratifications described, several multivariate analyses were performed to evaluate and describe dietary patterns and differences among the groups of individuals. The nutrition data were also stratified by groups as defined by ASA24: macronutrients, vitamins, minerals, other, grains, meats (referred to as proteins in this text), vegetables, fruits, dairy, and “extra” which includes discretionary oils and fats (21). Total fats, within macronutrients, are subdivided into saturated fatty acids, monounsaturated fatty acids, and polyunsaturated fatty acids (Supplementary Table S1).

One-way analysis of variance (ANOVA) was used to compare the adjusted average intake of all nutrients and food groups for individuals in the H-and L-LAB groups. Linear regressions were completed with the adjusted intake value of each nutrient and food group and the log (CFU/g wet stool) lactic acid bacteria. After multiple test correction, significance was considered as a value of p of less than 0.05. Principal component analysis (PCA) plots were assembled using the R package FactoMineR (22). Visualization of eigenvalues and PCA plots were done using corrplot package (21). Linear regression was used in the R stats package to complete principal component regression with log (CFU/g wet stool) LAB used as the dependent variable and PC’s from previous PCA used as independent variables (23). Beta coefficients and R² values were calculated and recorded for each regression. Significance for all analyses was considered as a value of p of less than 0.05.

Extreme gradient boosting was utilized for the categorical prediction of LAB grouping using the R package xgboost (24). Each model was trained separately with groups of variables (dietary intake as listed in Supplementary Table S1). The depth of the trees was set to three for each model for maximum accuracy. The minimum testing root-mean-square deviation (RSME) was determined for each model and used to set the optimal number of rounds of boosting for reduction of overfitting. Percent accuracy of model predictions, value of ps, and variable importance were determined for each model. Next, classification type random forests were generated using the R package randomForest to predict individual classification of H-and L-LAB groups based on nutrition data (22). Each model was trained separately with groups of variables (dietary intake as listed in Supplementary Table S1). Groups included macronutrients, vitamins, minerals, saturated fats, polyunsaturated fats, monounsaturated fats, grains, vegetables, fruits, proteins, dairy, and food additives. Each model was improved by optimizing the number of variables randomly sampled as candidates at each split, referring to each random subset of variables, within the random forest tree, known as the mtry value, and removing variables to reduce error rate. Each model was repeated 500 times with distribution of predictive rates, variable importance, and frequency at which a variable is predictive within the models calculated for each model. Variable importance is presented by mean decrease Gini value which is a measure of how each variable contributes to the homogeneity of the random forest. A higher the value of mean decrease Gini score equates to a higher level of importance within the model, thus allowing the ranking of usefulness of variables. For both modeling techniques, the data were separated into two groups, 70% used to train the data set (n = 20), and 30% used to test the dataset (n = 8). A total of 17 models were then refined, one per dietary group, and run 500 times to determine their accuracy and reproducibility of predictions. Potential biomarkers were chosen based upon their decrease gini value, highest frequency of prediction, and importance within a model of >85% accuracy of prediction.

Global metabolomics data was obtained as described in Marcial et al. (10) For identification of the features, the mass-to-charge ratios and ionization mode were used to search on-line databases such as the Human Metabolome Database (25).

The bacterial strains used are found in Supplementary Table S4. All strains were first activated in MRS-T (10 g/L peptone, 10 g/L beef extract (Bacto), 5 g/L yeast extract (Bacto), 20 g/L glucose, 2 g/L di-potassium hydrogen phosphate, 5 g/L sodium acetate, 2 g/L ammonium citrate tribasic, 0.2 g/L magnesium sulfate, 0.05 g/L manganese sulfate, 1 ml/L Tween 80). To evaluate the growth requirement of fatty acids, growth experiments were performed in four variations of MRS-T. First, MRS-NT was formulated by removing Tween 80 from MRS-T. Then, oleic or erucic acid was added at 0.1% (v/v) to make MRS-O or MRS-E, respectively. MRS-T with 0.1% DMSO was used as a vehicle control (MRS-TD). For growth curves, cells were subcultured twice in MRS-T for 16 h at 37°C. From the second activation culture, cells were pelleted by centrifugation at 3,000× rpm for 8 min and washed with MRS-NT. Cells were then suspended in MRS-NT and used to inoculate the media with the different fatty acids at an initial optical density (OD₆₀₀) of 0.05 in biological triplicates. Cells were incubated at 37°C and OD₆₀₀ recorded at 2 h intervals for a total of 18 h. Growth kinetics μmax and duplication time (DT) were estimated using a linear model of exponential phase. Statistics were performed in R as described previously by ANOVA and t-test, with p < 0.05 being considered statistically significant.

Principal component analysis (PCA) was performed as an exploratory method to visualize differences in the diet of L-LAB and H-LAB individuals. Models were trained with the entire nutrition dataset and with subsets of nutrients defined in Supplementary Table S1. First, the nutrition intake of subjects in the two intervention groups was evaluated. PCA of nutrition intake data where individuals were labeled by LAB counts, H-LAB, and L-LAB were performed next. PCA models obtained from groups for proteins, vegetables, dairy, and monounsaturated fats showed visible separation of H-LAB and L-LAB groups (Figure 1). Within proteins, the intake of fish, soy, nuts, franks, and poultry explained the variance between the two LAB groups (Figure 1A). Similarly, the nutrients that significantly contributed to the separation of the H-LAB and L-LAB individuals in the vegetables PCA were starchy vegetables, other vegetables, and tomatoes (Figure 1B), cheese in the dairy food group PCA (Figure 1C), and erucic acid (C22:1) in the monounsaturated fats group (Figure 1D). These results suggest that intake of a defined set of nutrients could be associated to the LAB stool count of individuals.

Given the results obtained with PCA, the nutrition data were then further investigated to identify correlations between LAB CFUs and dietary intake. To first determine differences in habitual intake between the two groups, single factor ANOVA was performed. After multiple test corrections for the 97 different variables included in the analyses (Supplementary Tables S1–S3), it was found that the food groups consumed in significantly higher amounts by the H-LAB group were soy (p = 6e-05), cheese (p = 1e-05), nuts and seeds (p = 8.5e-05), franks (p = 0.04), and fish high in N-3 fatty acids (p = 1.9e-07). On the contrary, dietary variables consumed in higher amounts by the L-LAB group are starchy vegetables (p = 0.003) and tomatoes (p = 3.6e-07; Table 2). In agreement with this observation, the L-LAB group had a higher intake of added vitamin E (p = 0.0002), and lycopene (p = 9.2e-05; Supplementary Tables S2, S3). While the total intake of fats was not significantly different between the groups, it was found that the H-LAB group consumed significantly higher amounts of the saturated fatty acids caprylic acid (C8:0; p = 2.4e-06), capric acid (C10:0; p = 0.004), lauric acid (C12:0; p = 0.001), and myristic acid (C14:0; p = 0.009), as well as the monounsaturated acid erucic acid (C22:1; p = 0.0001) and the polyunsaturated acid stearidonic acid (C18:4; p = 2.7e-06). To confirm the relationship between these nutrient variables and LAB CFUs, linear regression was performed. A positive correlation was observed between the intake of cheese, soy, nuts and seeds, fish high in N-3 fatty acids, and other vegetables, while a negative correlation was observed for tomatoes and lycopene (Table 3).

Finally, a principal component regression was performed using the previously described PCs and log (CFU/g wet stool) LAB. Regression of the principal components confirmed the results obtained, with the most significant result being that of dimension 2 (Dim2), the second principal component, of the vegetable model which is negatively correlated to CFU LAB (R² = 0.52, p = 1.35e-05; Supplementary Figure S2A). Additionally, proteins dimension 1 (Dim1) showed significant positive correlation to log CFU’s LAB (R² = 0.59, p = 1.73e-06; Supplementary Figure S2B). Taken together, the variables of the second dimension of the vegetable model including starchy vegetables, and tomatoes are negatively correlated to the abundancy of LAB. Additionally, the variables of dimension one of the protein model including fish high in N3 fatty acids, soy, nuts and seeds, fermented meats, and poultry have a positive correlation.

To determine potential biomarkers of LAB in the gut of healthy adults, the machine learning techniques extreme gradient boosting and random forest were employed. Extreme gradient boosting was used to build several highly accurate models for the prediction of H-LAB and L-LAB categorization. The most accurate models include those trained with all food groups, proteins, fatty acids (sfats, mfats, and pfats), and vegetables (Table 4). Each separate model had a 100% accuracy and singular variables with 100% importance in the model. The important features of each model were tomato, fish high in M3 fatty acids (M_fish_hi), S080, M221, P184, and tomato, respectively. To validate the results of the extreme gradient boosting decision tree modeling, random forest models were trained with the dietary intake variables and refined over 500 iterations using an 80/20 cross-validation scheme (Table 5). Models with >85% prediction accuracy were considered significant, and variables with both high importance and high frequency of prediction were chosen as potential biomarkers of LAB. First, all nutrition variables were analyzed together to create a predictive model. This model, along with all others, was repeated over 500 iterations to determine its reproducibility. The model with all 97 variables was not accurate in its predictions with highly variable accuracy of 93.8 ± 26% (Table 4). Next, dietary intake variables were then separated into categories (given in Supplementary Table S1). Of these, only two accurate models were found: proteins and monounsaturated fatty acids. The model of protein intake was both accurate and predictive of LAB grouping (accuracy = 98.6 ± 11.8%), with predictor variables being fish high in n-3 fatty acids (“M_FISH_HI”) and nuts and seeds (“M_NUTSD”; Figure 2A). All other food groups were not predictive of LAB group (Table 4). Interestingly, the monounsaturated fatty acid model trained with all monounsaturated fatty acids has a high error rate (accuracy = 87.2 ± 33.4%; Figure 2B). However, the model trained with only erucic acid (“M221,” C22:1) yielded an accuracy of 100%, meaning intake of M221 alone can be used to accurately predict LAB grouping (Figure 2B, Table 4). From these models nuts and seeds, fish high in N-3 fatty acids, and erucic acid (C22:1) were identified as potential biomarkers of LAB.

Untargeted metabolomic profiling was performed to evaluate whether nutrient intake maybe reflected in metabolite changes. The fluctuations in metabolites were analyzed by PCA and expressed as fold-change. Similar to the nutrition analyses, it was found that comparing data by intervention group (Ljo versus placebo) did not result in any significant differences between metabolite levels (Fig. S3A-B). Next, the data were stratified by H-and L-LAB groups. While PCA of the metabolite profiles showed only a slight separation between H-LAB and L-LAB groups (Fig. S3C), volcano plot analysis shows four significant features (Supplementary Figure S3D, Table S4). Based on their mass and the Human Metabolome Database (25), these features were putatively identified as chondroitin sulfate E (mass = 134.072, p = 0.01) and pyrazine (mass = 81.0448) which were higher in the H-LAB group, while polypropylene glycol (mass = 152.1283 p = 0.09) and caproic acid (mass = 117.0912, p = 0.08) were higher in the L-LAB group (Supplementary Table S5). When analyzed by timepoint (T), several more significant features were identified with 6 significant features at T1 (Initial sample, no intervention), 16 at T2 (2 weeks into the intervention phase), 15 at T3 (4 weeks into the intervention phase), 53 at T4 (8 weeks into the intervention phase), and 17 at T5 (12 weeks, end point; Supplementary Figures S3E–N). These metabolites were also identified by their mass and are listed in Supplementary Table S4. The relationship between metabolites, dietary intake, and their effect on the presence of LAB is overtly complex and will require further analyses.

The analysis of habitual diet indicated a relationship between the intake of monounsaturated fatty acids erucic acid (C22:1) and oleic acid (C18:1) and LAB counts. To further investigate these results, in vitro growth kinetic experiments were performed with several species formerly classified as of Lactobacillus to determine their ability to utilize the alternate fatty acid erucic acid opposed to Tween 80 as its fatty acid source. The strains (listed in Supplementary Table S5) were each grown in triplicate in MRS-E, MRS-O, MRS-T, MRS-NT, and MRS-TD (MRS-T with DMSO 0.1%, vehicle control) for a total of 18 h. Growth kinetics μmax and doubling time (DT) were calculated for each culture. Strains were chosen based on their modes of fermentation to compare strains of various metabolic capabilities as follows: homofermentative (L. johnsonii N6.2, L. johnsonii ATCC 33200, Lactobacillus amylovorus ATCC 33620, Ligilactobacillus murinus ATCC 35020; heterofermentative (Limosilactobacillus fermentum ATCC14391, Limosilactobacillus Reuteri TD1), and facultatively heterofermentative (Lactipantibacillus plantarum ATCC25302, and Lacticaseibacillus casei ATCC 334; Supplementary Table S5). While most of the Lactobacillus strains are auxotrophic for fatty acids and require the addition of Tween 80 to the media, three of the selected strains, L. plantarum ATCC25302, L. casei ATCC 334, and L. fermentum ATCC14391, have the genes that encode for the necessary enzymes of the FASII pathway required for de novo biosynthesis of fatty acids (Supplementary Table S5). As expected, these three strains were able to grow at comparable speed in all culture media (Figure 3).

It was found that all the homofermentative Lactobacillus species required fatty acids for growth as it is evidenced by the absence of growth in MRS-NT. Interestingly, all the strains tested were able to grow in presence of erucic acid albeit with significant differences. Lactobacillus johnsonii N6.2 was able to utilize Tween 80 (MRS-T), erucic acid (MRS-E), and oleic acid (MRS-O) as well as the DMSO vehicle control (MRS-TD). The extended lag phase observed in MRS-E seems to be due to the DMSO vehicle as the growth patterns are visually similar (Figure 3A). However, there is a significant decrease in μ-max and increase in DT in cells grown in MRS-E as compared to MRS-O, MRS-T, and MRS-TD (Supplementary Figure S5). Cells grown in MRS-O have an 8 h lag as well as significantly lower μ-max compared to MRS-T, but not MRS-TD. Cells in MRS-O also had lower DT than any other condition (Supplementary Figure S5). Lactobacillus johnsonii ATCC 33200 also utilized both oleic and erucic acid, though in MRS-E, there was a 10 h lag phase (Figure 3B). Cells grown in MRS-O had significantly higher μ-max compared to each condition and lower DT compared to the vehicle control MRS-TD (Supplementary Figure S5). Lactobacillus amylovorus ATCC 33620 had no detectible growth in MRS-O until 14 h incubation time (Figure 3C). In MRS-E, there were no significant differences in μ-max compared to MRS-TD but had significantly longer DT compared to MRS-DT (Supplementary Figure S5). Ligilactobacillus murinus ATCC 35020 had measurable growth in all conditions, with higher μ-max in MRS-O and lower DT compared to MRS-E (Figure 3D, Supplementary Figure S5).

Within the heterofermentative L. reuteri TD1 had no significant differences in μmax between MRS-E and MRS-O, however in MRS-E the DT was significantly lower (Figure 3E). Ligilactobacillus fermentum 14391 which encodes the necessary enzymes of the FASII pathway, had growth in MRS-NT as expected. It also had no difference in μ-max comparing MRS-O and MRS-E but had significantly lower DT in MRS-O compared to MRS-E (Figure 3F). The facultative heterofermentative strains L. plantarum ATCC 81 and L. casei ATCC 334 both encode the necessary enzymes of the FASII pathway and thus, as expected, both were able to grow in MRS-NT (Figures 3G,H). Lactipantibacillus plantarum ATCC 25302 had lower μ-max in MRS-E compared to MRS-O, and lower DT in MRS-O compared to MRS-E. Lacticaseibacillus casei ATCC 334 had lower μ-max and DT in MRS-O compared to MRS-E (Supplementary Figure S5). Together, these results indicate that the use of erucic acid by some strains of lactic acid bacteria may allow for a competitive advantage in the human gut and may be used as a natural dietary prebiotic to consider during probiotic interventions.