To comprehensively characterize PPGRs, we recruited 800 individuals aged 18–70 not previously diagnosed with TIIDM (Figure 1Figure 1A, Table 1Table 1). The cohort is representative of the adult non-diabetic Israeli population (Israeli Center for Disease Control, 2014xHealth 2013. Israeli Center for Disease Control.
See all ReferencesIsraeli Center for Disease Control, 2014), with 54% overweight (BMI ≥ 25 kg/m²) and 22% obese (BMI ≥ 30 kg/m², Figures 1Figures 1B, 1C, and S1S1). These properties are also characteristic of the Western adult non-diabetic population (World Health Organization, 2008xGlobal Health Observatory Data Repository. World Health Organization.
See all ReferencesWorld Health Organization, 2008).

Figure 1

Profiling of Postprandial Glycemic Responses, Clinical Data, and Gut Microbiome

(A) Illustration of our experimental design.

(B and C) Distribution of BMI and glycated hemoglobin (HbA1c%) in our cohort. Thresholds for overweight (BMI ≥ 25 kg/m²), obese (BMI ≥ 30 kg/m²), prediabetes (HbA1c% ≥ 5.7%) and TIIDM (≥6.5%) are shown.

(D) Example of continuous glucose monitoring (CGM) for one participant during an entire week. Colored area within zoom-in shows the incremental area under the glucose curve (iAUC) which we use to quantify the meal’s PPGR.

(E) Major food components consumed by energy intake.

(F) Distribution of meals (dots) by macronutrient content. Inset shows histogram of meals per macronutrient.

(G) Bray-Curtis based PCoA of metagenome-based bacterial abundances of stool samples in our cohort and in the U.S. HMP and European MetaHIT cohorts. Inset shows PCoA when samples from other HMP body sites are added. See also Figure S2Figure S2.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Figure S1

Histograms Showing Distribution of Major Risk Factors in the Main 800-Person Cohort and the Validation 100-Person Cohort, Related to Table 1Table 1 and Figure 3Figure 3

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Each participant was connected to a continuous glucose monitor (CGM), which measures interstitial fluid glucose every 5 min for 7 full days (the “connection week”), using subcutaneous sensors (Figure 1Figure 1D). CGMs estimate blood glucose levels with high accuracy (Bailey et al., 2014xAccuracy and acceptability of the 6-day Enlite continuous subcutaneous glucose sensor. Bailey, T.S., Ahmann, A., Brazg, R., Christiansen, M., Garg, S., Watkins, E., Welsh, J.B., and Lee, S.W. Diabetes Technol. Ther. 2014; 16: 277–283
Crossref | PubMed | Scopus (26)See all ReferencesBailey et al., 2014) and previous studies found no significant differences between PPGRs extracted from CGMs and those obtained from either venous or capillary blood (Vrolix and Mensink, 2010xVariability of the glycemic response to single food products in healthy subjects. Vrolix, R. and Mensink, R.P. Contemp. Clin. Trials. 2010; 31: 5–11
Abstract | Full Text | Full Text PDF | PubMed | Scopus (13)See all ReferencesVrolix and Mensink, 2010). We used blinded CGMs and thus participants were unaware of their CGM levels during the connection week. Together, we recorded over 1.5 million glucose measurements from 5,435 days.

While connected to the CGM, participants were instructed to log their activities in real-time, including food intake, exercise and sleep, using a smartphone-adjusted website (www.personalnutrition.org) that we developed (Figure S2Figure S2A). Each food item within every meal was logged along with its weight by selecting it from a database of 6,401 foods with full nutritional values based on the Israeli Ministry of Health database that we further improved and expanded with additional items from certified sources. To increase compliance, participants were informed that accurate logging is crucial for them to receive an accurate analysis of their PPGRs to food (ultimately provided to each of them). During the connection week, participants were asked to follow their normal daily routine and dietary habits, except for the first meal of every day, which we provided as one of four different types of standardized meals, each consisting of 50 g of available carbohydrates. This resulted in a total of 46,898 real-life meals with close-to or full nutritional values (median of 54 meals per participant) and 5,107 standardized meals. The PPGR of each meal was calculated by combining reported meal time with CGM data and computing the incremental area under the glucose curve in the 2 hr after the meal (iAUC; Wolever and Jenkins, 1986xThe use of the glycemic index in predicting the blood glucose response to mixed meals. Wolever, T.M. and Jenkins, D.J. Am. J. Clin. Nutr. 1986; 43: 167–172
PubMedSee all ReferencesWolever and Jenkins, 1986; Figure 1Figure 1D).

Figure S2

Logging of Lifestyle and Nutritional Information, Related to Figure 1Figure 1

(A) Key features of the smartphone-adjusted website that we developed for meal logging and daily activity collection. (B–D) Major food components consumed by the cohort by protein (B), carbohydrates (C), and fat (D).

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Prior to CGM connection, a comprehensive profile was collected from each participant, including: food frequency, lifestyle, and medical background questionnaires; anthropometric measures (e.g., height, hip circumference); a panel of blood tests; and a single stool sample, used for microbiota profiling by both 16S rRNA and metagenomic sequencing.

With a total of ∼10,000,000 Calories logged, our data provide a global view into the cohort’s dietary habits, showing the fraction that each food source contributes to the cohort’s overall energy intake (e.g., dairy, 7%; sweets, 6%; Figure 1Figure 1E), and macronutrient intake (Figures S2Figures S2B–S2D). Analysis of the caloric breakdown of every meal by macronutrients revealed that protein intake varies relatively little across meals (80% of meals have 5%–35% protein), while fat and carbohydrates have a wide and bimodal distribution, where one of the modes corresponds to fat-free meals and constitutes 18% of all meals (Figure 1Figure 1F).

Principal coordinates analysis (PCoA) on the Bray-Curtis dissimilarity between metagenome-based relative abundances (RA) revealed a similar degree of variability in the microbiomes of our cohort and stool samples of the US HMP (Human Microbiome Project Consortium, 2012xStructure, function and diversity of the healthy human microbiome. Human Microbiome Project Consortium. Nature. 2012; 486: 207–214
Crossref | PubMed | Scopus (2141)See all ReferencesHuman Microbiome Project Consortium, 2012) and European MetaHIT (Nielsen et al., 2014xIdentification and assembly of genomes and genetic elements in complex metagenomic samples without using reference genomes. Nielsen, H.B., Almeida, M., Juncker, A.S., Rasmussen, S., Li, J., Sunagawa, S., Plichta, D.R., Gautier, L., Pedersen, A.G., Le Chatelier, E...., MetaHIT Consortium, and MetaHIT Consortium. Nat. Biotechnol. 2014; 32: 822–828
Crossref | PubMed | Scopus (167)See all ReferencesNielsen et al., 2014) cohorts (Figure 1Figure 1G). The first two principal coordinates show some distinction between our cohort and the other cohorts, but when HMP samples from other body sites are added to the PCoA, stool samples from all three cohorts cluster together and separate from the rest, indicative of overall similarity in the gut microbiota composition of individuals from these three distinct geographical regions (Figure 1Figure 1G).

Our data replicate known associations of PPGRs with risk factors, as the median standardized meal PPGR was significantly correlated with several known risk factors including BMI (R = 0.24, p < 10^−10), glycated hemoglobin (HbA1c%, R = 0.49, p < 10^−10), wakeup glucose (R = 0.47, p < 10^−10), and age (R = 0.42, p < 10^−10, Figure 2Figure 2A). These associations are not confined to extreme values but persist along the entire range of PPGR values, suggesting that the reduction in levels of risk factors is continuous across all postprandial values, with lower values associated with lower levels of risk factors even within the normal value ranges (Figure 2Figure 2A).

Utilizing the continuous nature of the CGMs, we also examined the association between risk factors and the glucose level of each participant at different percentiles (0–100) with respect to all glucose measurements from the connection week. These levels are affected by the PPGRs while also reflecting the general glycemic control state of the participant. All percentiles significantly associated with risk factors (wakeup glucose, BMI, HbA1c%, and age; Figures S3Figures S3A–S3D). The percentile at which the glucose level correlation was highest varied across risk factors. For example, BMI had the highest correlation with the 40^th glucose value percentile, whereas for HbA1c% percentile 95 had the highest correlation (Figures S3Figures S3A and S3C). These results suggest that the entire range of glucose levels of an individual may have clinical relevance, with different percentiles being more relevant for particular risk factors.

Figure S3

Glycemic Data Are Interpersonally Variable and Correlated with Risk Factors, Related to Figure 2Figure 2

(A–D) For each participant, the CGM measured glucose level at all percentiles (0th, 1st, 2nd, …, 100th percentiles) across the entire connection week (x axis) were correlated with BMI (A), age (B), HbA1c% (C) and wakeup glucose levels (D). For each such percentile, shown is the Pearson R (left y axis, full line) and FDR-corrected q-value (right y axis, dashed line). (E–H) Histograms as in Figure 2Figure 2B, without smoothing, of the postprandial glycemic response (PPGR) to glucose (E), bread (F), bread and butter (G) and fructose (H), the standardized meals provided to participants (each with 50 g of available carbohydrates). (I–K) Histograms as in Figure 2Figure 2B, demonstrating similarly high variability in the PPGR of each participant to bread (I), bread & butter (J) and fructose (K), except here PPGRs were normalized to the PPGR of the participant to 50 g. glucose. Purple vertical line denotes published glycemic index (GI) of bread and fructose (Foster-Powell et al., 2002xInternational table of glycemic index and glycemic load values: 2002. Foster-Powell, K., Holt, S.H.A., and Brand-Miller, J.C. Am. J. Clin. Nutr. 2002; 76: 5–56
PubMedSee all ReferencesFoster-Powell et al., 2002).

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Next, we examined intra- and interpersonal variability in the PPGR to the same food. First, we assessed the extent to which PPGRs to three types of standardized meals that were given twice to every participant (Figure 1Figure 1A), are reproducible within the same person. Indeed, the two replicates showed high agreement (R = 0.77 for glucose, R = 0.77 for bread with butter, R = 0.71 for bread, p < 10^−10 in all cases), demonstrating that the PPGR to identical meals is reproducible within the same person and that our experimental system reliably measures this reproducibility. However, when comparing the PPGRs of different people to the same meal, we found high interpersonal variability, with the PPGRs of every meal type (except fructose) spanning the entire range of PPGRs measured in our cohort (Figures 2Figures 2B, 2C , and S3S3E–S3H). For example, the average PPGR to bread across 795 people was 44 ± 31 mg/dl^∗h (mean ± SD), with the bottom 10% of participants exhibiting an average PPGR below 15 mg/dl^∗h and the top 10% of participants exhibiting an average PPGR above 79 mg/dl^∗h. The large interpersonal differences in PPGRs are also evident in that the type of meal that induced the highest PPGR differs across participants and that different participants might have opposite PPGRs to pairs of different standardized meals (Figures 2Figures 2D and 2E).

Figure 2

High Interpersonal Variability in the Postprandial Glycemic Response to the Same Meal

(A) PPGRs associate with risk factors. Shown are PPGRs, BMI, HbA1c%, age, and wakeup glucose of all participants, sorted by median standardized meal PPGR (top, red dots). Correlation of factors with the median PPGRs to standardized meals is shown along with a moving average line.

(B) Kernel density estimation (KDE) smoothed histogram of the PPGR to four types of standardized meals provided to participants (each with 50 g of available carbohydrates). Dashed lines represent histogram modes (See also Figure S3Figure S3).

(C) Example of high interpersonal variability and low intra-personal variability in the PPGR to bread across four participants (two replicates per participant consumed on two different mornings).

(D) Heatmap of PPGR (average of two replicates) of participants (rows) to three types of standardized meals (columns) consumed in replicates. Clustering is by each participant’s relative rankings of the three meal types.

(E) Example of two replicates of the PPGR to two standardized meals for two participants exhibiting reproducible yet opposite PPGRs.

(F) Box plot (box, IQR; whiskers, 10–90 percentiles) of the PPGR to different real-life meals along with amount of carbohydrates consumed (green; mean ± std).

(G) Same as (E), for a pair of real-life meals, each containing 20 g of carbohydrates.

(H) Heatmap (subset) of statistically significant associations (p < 0.05, FDR corrected) between participants’ standardized meals PPGRs and participants’ clinical and microbiome data (See also Figure S4Figure S4 for the full heatmap).

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Interpersonal variability was not merely a result of participants having high PPGRs to all meals, since high variability was also observed when the PPGR of each participant was normalized to his/her own PPGR to glucose (Figures S3Figures S3I–S3K). For white bread and fructose, for which such normalized PPGRs were previously measured, the mode of the PPGR distribution in our cohort had excellent agreement with published values (Foster-Powell et al., 2002xInternational table of glycemic index and glycemic load values: 2002. Foster-Powell, K., Holt, S.H.A., and Brand-Miller, J.C. Am. J. Clin. Nutr. 2002; 76: 5–56
PubMedSee all ReferencesFoster-Powell et al., 2002), further validating the accuracy of our data (bread: 65 versus 71; fructose: 15 versus 19, Figures S3Figures S3I and S3K).

Next, we examined variability in the PPGRs to the multiple real-life meals reported by our participants. Since real-life meals vary in amounts and may each contain several different food components, we only examined meals that contained 20–40 g of carbohydrates and had a single dominant food component whose carbohydrate content exceeded 50% of the meal’s carbohydrate content. We then ranked the resulting dominant foods that had at least 20 meal instances by their population-average PPGR (Figure 2Figure 2F). For foods with a published glycemic index, our population-average PPGRs agreed with published values (R = 0.69, p < 0.0005), further supporting our data (Table S1xDownload (.31 MB )

Document S1. Supplemental Experimental Procedures and Table S1Table S1). For example, the average PPGR to rice and potatoes was relatively high, whereas that for ice cream, beer, and dark chocolate was relatively low, in agreement with published data (Atkinson et al., 2008xInternational tables of glycemic index and glycemic load values: 2008. Atkinson, F.S., Foster-Powell, K., and Brand-Miller, J.C. Diabetes Care. 2008; 31: 2281–2283
Crossref | PubMed | Scopus (489)See all References, Foster-Powell et al., 2002xInternational table of glycemic index and glycemic load values: 2002. Foster-Powell, K., Holt, S.H.A., and Brand-Miller, J.C. Am. J. Clin. Nutr. 2002; 76: 5–56
PubMedSee all References). Similar to standardized meals, PPGRs to self-reported meals highly varied across individuals, with both low and high responders noted for each type of meal (Figures 2Figures 2F and 2G).

We found multiple significant associations between the standardized meal PPGRs of participants and both their clinical and gut microbiome data (Figures 2Figures 2H and S4S4). Notably, the TIIDM and metabolic syndrome risk factors HbA1c%, BMI, systolic blood pressure, and alanine aminotransferase (ALT) activity are all positively associated with PPGRs to all types of standardized meals, reinforcing the medical relevance of PPGRs. In most standardized meals, PPGRs also exhibit a positive correlation with CRP, whose levels rise in response to inflammation (Figure 2Figure 2H).

Figure S4

Interpersonal Differences in Postprandial Glycemic Responses (PPGRs) to Standardized Meals Associate with Clinical and Microbiome Features, Related to Figure 2Figure 2

Complete set of statistically significant associations from Figure 2Figure 2H. Shown are all statistically significant associations (p < 0.05, FDR corrected) found between participants’ standardized meals PPGRs and participants’ blood parameters, anthropometrics, 16S rRNA and metagenomic derived abundances, and abundances of KEGG pathways and modules. Yellow and blue entries indicate significant positive and negative associations, respectively.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

With respect to microbiome features, the phylogenetically related Proteobacteria and Enterobacteriaceae both exhibit positive associations with a few of the standardized meals PPGR (Figure 2Figure 2H). These taxa have reported associations with poor glycemic control, and with components of the metabolic syndrome including obesity, insulin resistance, and impaired lipid profile (Xiao et al., 2014xA gut microbiota-targeted dietary intervention for amelioration of chronic inflammation underlying metabolic syndrome. Xiao, S., Fei, N., Pang, X., Shen, J., Wang, L., Zhang, B., Zhang, M., Zhang, X., Zhang, C., Li, M. et al. FEMS Microbiol. Ecol. 2014; 87: 357–367
Crossref | PubMed | Scopus (60)See all ReferencesXiao et al., 2014). RAs of Actinobacteria are positively associated with the PPGR to both glucose and bread, which is intriguing since high levels of this phylum were reported to associate with a high-fat, low-fiber diet (Wu et al., 2011xLinking long-term dietary patterns with gut microbial enterotypes. Wu, G.D., Chen, J., Hoffmann, C., Bittinger, K., Chen, Y.-Y., Keilbaugh, S.A., Bewtra, M., Knights, D., Walters, W.A., Knight, R. et al. Science. 2011; 334: 105–108
Crossref | PubMed | Scopus (1413)See all ReferencesWu et al., 2011).

At the functional level, the KEGG pathways of bacterial chemotaxis and of flagellar assembly, reported to increase in mice fed high-fat diets and decrease upon prebiotics administration (Everard et al., 2014xMicrobiome of prebiotic-treated mice reveals novel targets involved in host response during obesity. Everard, A., Lazarevic, V., Gaïa, N., Johansson, M., StÃ¥hlman, M., Backhed, F., Delzenne, N.M., Schrenzel, J., François, P., and Cani, P.D. ISME J. 2014; 8: 2116–2130
Crossref | PubMed | Scopus (86)See all ReferencesEverard et al., 2014), exhibit positive associations with several standardized meal PPGRs (Figure 2Figure 2H). The KEGG pathway of ABC transporters, reported to be positively associated with TIIDM (Karlsson et al., 2013xGut metagenome in European women with normal, impaired and diabetic glucose control. Karlsson, F.H., Tremaroli, V., Nookaew, I., Bergström, G., Behre, C.J., Fagerberg, B., Nielsen, J., and Bäckhed, F. Nature. 2013; 498: 99–103
Crossref | PubMed | Scopus (494)See all ReferencesKarlsson et al., 2013) and with a Western high-fat/high-sugar diet (Turnbaugh et al., 2009xA core gut microbiome in obese and lean twins. Turnbaugh, P.J., Hamady, M., Yatsunenko, T., Cantarel, B.L., Duncan, A., Ley, R.E., Sogin, M.L., Jones, W.J., Roe, B.A., Affourtit, J.P. et al. Nature. 2009; 457: 480–484
Crossref | PubMed | Scopus (2698)See all ReferencesTurnbaugh et al., 2009), also exhibits positive association with several standardized meal PPGRs (Figure 2Figure 2H). Several bacterial secretion systems, including both type II and type III secretion systems that are instrumental in bacterial infection and quorum sensing (Sandkvist, 2001xType II secretion and pathogenesis. Sandkvist, M. Infect. Immun. 2001; 69: 3523–3535
Crossref | PubMed | Scopus (205)See all ReferencesSandkvist, 2001) are positively associated with most standardized meal PPGRs (Figure 2Figure 2H). Finally, KEGG modules for transport of the positively charged amino acids lysine and arginine are associated with high PPGR to standardized foods, while transport of the negatively charged amino acid glutamate is associated with low PPGRs to these foods.

Taken together, these results show that PPGRs vary greatly across different people and associate with multiple person-specific clinical and microbiome factors.

We next asked whether clinical and microbiome factors could be integrated into an algorithm that predicts individualized PPGRs. To this end, we employed a two-phase approach. In the first, discovery phase, the algorithm was developed on the main cohort of 800 participants, and performance was evaluated using a standard leave-one-out cross validation scheme, whereby PPGRs of each participant were predicted using a model trained on the data of all other participants. In the second, validation phase, an independent cohort of 100 participants was recruited and profiled, and their PPGRs were predicted using the model trained only on the main cohort (Figure 3Figure 3A).

Figure 3

Accurate Predictions of Personalized Postprandial Glycemic Responses

(A) Illustration of our machine-learning scheme for predicting PPGRs.

(B–E) PPGR predictions. Dots represent predicted (x axis) and CGM-measured PPGR (y axis) for meals, for a model based: only on the meal’s carbohydrate content (B); only on the meal’s Caloric content (C); our predictor evaluated in leave-one-person-out cross validation on the main 800-person cohort (D); and our predictor evaluated on the independent 100-person validation cohort (E). Pearson correlation of predicted and measured PPGRs is indicated.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Given non-linear relationships between PPGRs and the different factors, we devised a model based on gradient boosting regression (Friedman, 2001xGreedy Function Approximation: A Gradient Boosting Machine. Friedman, J.H. Ann. Stat. 2001; 29: 1189–1232
CrossrefSee all ReferencesFriedman, 2001). This model predicts PPGRs using the sum of thousands of different decision trees. Trees are inferred sequentially, with each tree trained on the residual of all previous trees and making a small contribution to the overall prediction (Figure 3Figure 3A). The features within each tree are selected by an inference procedure from a pool of 137 features representing meal content (e.g., energy, macronutrients, micronutrients); daily activity (e.g., meals, exercises, sleep times); blood parameters (e.g., HbA1c%, HDL cholesterol); CGM-derived features; questionnaires; and microbiome features (16S rRNA and metagenomic RAs, KEGG pathway and module RAs and bacterial growth dynamics - PTRs; Korem et al., 2015xGrowth dynamics of gut microbiota in health and disease inferred from single metagenomic samples. Korem, T., Zeevi, D., Suez, J., Weinberger, A., Avnit-Sagi, T., Pompan-Lotan, M., Matot, E., Jona, G., Harmelin, A., Cohen, N. et al. Science. 2015; 349: 1101–1106
Crossref | PubMed | Scopus (33)See all ReferencesKorem et al., 2015).

As a baseline reference, we used the “carbohydrate counting” model, as it is the current gold standard for predicting PPGRs (American Diabetes Association, 2015bx(4) Foundations of care: education, nutrition, physical activity, smoking cessation, psychosocial care, and immunization. American Diabetes Association. Diabetes Care. 2015; 38: S20–S30
Crossref | PubMed | Scopus (70)See all References, Bao et al., 2011xImproving the estimation of mealtime insulin dose in adults with type 1 diabetes: the Normal Insulin Demand for Dose Adjustment (NIDDA) study. Bao, J., Gilbertson, H.R., Gray, R., Munns, D., Howard, G., Petocz, P., Colagiuri, S., and Brand-Miller, J.C. Diabetes Care. 2011; 34: 2146–2151
Crossref | PubMed | Scopus (17)See all References). On our data, this model that consists of a single explanatory variable representing the meal’s carbohydrate amount achieves a modest yet statistically significant correlation with PPGRs (R = 0.38, p < 10^−10, Figure 3Figure 3B). A model using only meal Caloric content performs worse (R = 0.33, p < 10^−10, Figure 3Figure 3C). Our predictor that integrates the above person-specific factors predicts the held-out PPGRs of individuals with a significantly higher correlation (R = 0.68, p < 10^−10, Figure 3Figure 3D). This correlation approaches the presumed upper bound limit set by the 0.71–0.77 correlation that we observed between the PPGR of the same person to two replicates of the same standardized meal.

We further validated our model on an independent cohort of 100 individuals that we recruited separately. Data from this additional cohort were not available to us while developing the algorithm. Participants in this cohort underwent the same profiling as in the main 800-person cohort. No significant differences were found between the main and validation cohorts in key parameters, including age, BMI, non-fasting total and HDL cholesterol, and HbA1c% (Table 1Table 1, Figure S1Figure S1).

Notably, our algorithm, derived solely using the main 800 participants cohort, achieved similar performance on the 100 participants of the validation cohort (R = 0.68 and R = 0.70 on the main and validation cohorts, respectively, Figures 3Figures 3D and 3E). The reference carbohydrate counting model achieved the same performance as in the main cohort (R = 0.38). This result further supports the ability of our algorithm to provide personalized PPGR predictions.

To gain insight into the contribution of the different features in the algorithm’s predictions, we examined partial dependence plots (PDP), commonly used to study functional relations between features used in predictors such as our gradient boosting regressor and an outcome (PPGRs in our case; Hastie et al., 2008xThe Elements of Statistical Learning: Data Mining, Inference and Prediction. Hastie, T., Tibshirani, R., and Friedman, J.
See all ReferencesHastie et al., 2008). PDPs graphically visualize the marginal effect of a given feature on prediction outcome after accounting for the average effect of all other features. While this effect may be indicative of feature importance, it may also be misleading due to higher-order interactions (Hastie et al., 2008xThe Elements of Statistical Learning: Data Mining, Inference and Prediction. Hastie, T., Tibshirani, R., and Friedman, J.
See all ReferencesHastie et al., 2008). Nonetheless, PDPs are commonly used for knowledge discovery in large datasets such as ours.

As expected, the PDP of carbohydrates (Figure 4Figure 4A) shows that as the meal carbohydrate content increases, our algorithm predicts, on average, a higher PPGR. We term this relation, of higher predicted PPGR with increasing feature value, as non-beneficial (with respect to prediction), and the opposite relation, of lower predicted PPGR with increasing feature value, as beneficial (also with respect to prediction; see PDP legend in Figure 4Figure 4). However, since PDPs display the overall contribution of each feature across the entire cohort, we asked whether the relationship between carbohydrate amount and PPGRs varies across people. To this end, for each participant we computed the slope of the linear regression between the PPGR and carbohydrate amount of all his/her meals. As expected, this slope was positive for nearly all (95.1%) participants, reflective of higher PPGRs in meals richer in carbohydrates. However, the magnitude of this slope varies greatly across the cohort, with the PPGR of some people correlating well with the carbohydrate content (i.e., carbohydrates “sensitive”) and that of others exhibiting equally high PPGRs but little relationship to the amount of carbohydrates (carbohydrate “insensitive”; Figure 4Figure 4B). This result suggests that carbohydrate sensitivity is also person specific.

Figure 4

Factors Underlying the Prediction of Postprandial Glycemic Responses

(A) Partial dependence plot (PDP) showing the marginal contribution of the meal’s carbohydrate content to the predicted PPGR (y axis, arbitrary units) at each amount of meal carbohydrates (x axis). Red and green indicate above and below zero contributions, respectively (number indicate meals). Boxplots (bottom) indicate the carbohydrates content at which different percentiles (10, 25, 50, 75, and 90) of the distribution of all meals across the cohort are located. See PDP legend.

(B) Histogram of the slope (computed per participant) of a linear regression between the carbohydrate content and the PPGR of all meals. Also shown is an example of one participant with a low slope and another with a high slope.

(C) Meal fat/carbohydrate ratio PDP.

(D) Histogram of the difference (computed per participant) between the Pearson R correlation of two linear regression models, one between the PPGR and the meal carbohydrate content and another when adding fat and carbohydrate^∗fat content. Also shown is an example of the carbohydrate and fat content of all meals of one participant with a relatively low R difference (carb alone correlates well with PPGR) and another with a relatively high difference (meals with high fat content have lower PPGRs). Dot color and size correspond to the meal’s PPGR.

(E) Additional PDPs.

(F) Microbiome PDPs. The number of participants in which the microbiome feature was not detected is indicated (left, n.d.). Boxplots (box, IQR; whiskers 10–90 percentiles) based only on detected values.

(G) Heatmap of statistically significant correlations (Pearson) between microbiome features termed beneficial (green) or non-beneficial (red) and several risk factors and glucose parameters.

See also Figure S5Figure S5.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

The PDP of fat exhibits a beneficial effect for fat since our algorithm predicts, on average, lower PPGR as the meal’s ratio of fat to carbohydrates (Figure 4Figure 4C) or total fat content (Figure S5Figure S5A) increases, consistent with studies showing that adding fat to meals may reduce the PPGR (Cunningham and Read, 1989xThe effect of incorporating fat into different components of a meal on gastric emptying and postprandial blood glucose and insulin responses. Cunningham, K.M. and Read, N.W. Br. J. Nutr. 1989; 61: 285–290
Crossref | PubMed | Scopus (49)See all ReferencesCunningham and Read, 1989). However, here too, we found that the effect of fat varies across people. We compared the explanatory power of a linear regression between each participant’s PPGR and meal carbohydrates, with that of regression using both fat and carbohydrates. We then used the difference in Pearson R between the two models as a quantitative measure of the added contribution of fat (Figure 4Figure 4D). For some participants we observed a reduction in PPGR with the addition of fat, while for others meal fat content did not add much to the explanatory power of the regressor based only on the meal’s carbohydrates content (Figure 4Figure 4D).

Figure S5

Additional Features Underlying the Prediction of Postprandial Glycemic Responses, Related to Figure 4Figure 4

(A) Shown are partial dependency plots (PDPs, as in Figure 4Figure 4), for additional features. (B) Shown are PDPs for all microbiome-derived features.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Interestingly, while dietary fibers in the meal increase the predicted PPGR, their long-term effect is beneficial as higher amount of fibers consumed in the 24 hr prior to the meal reduces the predicted PPGR (Figure 4Figure 4E). The meal’s sodium content, the time that passed since last sleeping, and a person’s cholesterol levels or age all exhibit non-beneficial PDPs, while the PDPs of the meal’s alcohol and water content display beneficial effects (Figures 4Figures 4E and S5S5A). As expected, the PDP of HbA1c% shows a non-beneficial effect with increased PPGR at higher HbA1c% values; intriguingly, higher PPGRs are predicted, on average, for individuals with HbA1c% above ∼5.5%, which is very close to the prediabetes threshold of 5.7% (Figure S5Figure S5A).

The 72 PDPs of the microbiome-based features used in our predictor were either beneficial (21 factors), non-beneficial (28), or non-decisive (23) in that they mostly decreased, increased, or neither, as a function of the microbiome feature. The resulting PDPs had several intriguing trends. For example, growth of Eubacterium rectale was mostly beneficial, as in 430 participants with high inferred growth for E. rectale it associates with a lower PPGR (Figure 4Figure 4F). Notably, E. rectale can ferment dietary carbohydrates and fibers to produce metabolites useful to the host (Duncan et al., 2007xReduced dietary intake of carbohydrates by obese subjects results in decreased concentrations of butyrate and butyrate-producing bacteria in feces. Duncan, S.H., Belenguer, A., Holtrop, G., Johnstone, A.M., Flint, H.J., and Lobley, G.E. Appl. Environ. Microbiol. 2007; 73: 1073–1078
Crossref | PubMed | Scopus (331)See all ReferencesDuncan et al., 2007), and was associated with improved postprandial glycemic and insulinemic responses (Martínez et al., 2013xGut microbiome composition is linked to whole grain-induced immunological improvements. Martínez, I., Lattimer, J.M., Hubach, K.L., Case, J.A., Yang, J., Weber, C.G., Louk, J.A., Rose, D.J., Kyureghian, G., Peterson, D.A. et al. ISME J. 2013; 7: 269–280
Crossref | PubMed | Scopus (94)See all ReferencesMartínez et al., 2013), as well as negatively associated with TIIDM (Qin et al., 2012xA metagenome-wide association study of gut microbiota in type 2 diabetes. Qin, J., Li, Y., Cai, Z., Li, S., Zhu, J., Zhang, F., Liang, S., Zhang, W., Guan, Y., Shen, D. et al. Nature. 2012; 490: 55–60
Crossref | PubMed | Scopus (1022)See all ReferencesQin et al., 2012). RAs of Parabacteroides distasonis were found non-beneficial by our predictor (Figure 4Figure 4F) and this species was also suggested to have a positive association with obesity (Ridaura et al., 2013xGut microbiota from twins discordant for obesity modulate metabolism in mice. Ridaura, V.K., Faith, J.J., Rey, F.E., Cheng, J., Duncan, A.E., Kau, A.L., Griffin, N.W., Lombard, V., Henrissat, B., Bain, J.R. et al. Science. 2013; 341: 1241214
Crossref | PubMed | Scopus (667)See all ReferencesRidaura et al., 2013). As another example, the KEGG module of cell-division transport system (M00256) was non-beneficial, and in the 164 participants with the highest levels for it, it associates with a higher PPGR (Figure 4Figure 4F). Bacteroides thetaiotaomicron was non-beneficial (Figure S5Figure S5B), and it was associated with obesity and was suggested to have increased capacity for energy harvest (Turnbaugh et al., 2006xAn obesity-associated gut microbiome with increased capacity for energy harvest. Turnbaugh, P.J., Ley, R.E., Mahowald, M.A., Magrini, V., Mardis, E.R., and Gordon, J.I. Nature. 2006; 444: 1027–1031
Crossref | PubMed | Scopus (3468)See all ReferencesTurnbaugh et al., 2006). In the case of Alistipes putredinis and the Bacteroidetes phylum, the non-beneficial classification that our predictor assigns to both of them is inconsistent with previous studies that found them to be negatively associated with obesity (Ridaura et al., 2013xGut microbiota from twins discordant for obesity modulate metabolism in mice. Ridaura, V.K., Faith, J.J., Rey, F.E., Cheng, J., Duncan, A.E., Kau, A.L., Griffin, N.W., Lombard, V., Henrissat, B., Bain, J.R. et al. Science. 2013; 341: 1241214
Crossref | PubMed | Scopus (667)See all References, Turnbaugh et al., 2006xAn obesity-associated gut microbiome with increased capacity for energy harvest. Turnbaugh, P.J., Ley, R.E., Mahowald, M.A., Magrini, V., Mardis, E.R., and Gordon, J.I. Nature. 2006; 444: 1027–1031
Crossref | PubMed | Scopus (3468)See all References). This may reflect limitations of the PDP analysis or result from a more complex relationship between these features, obesity, and PPGRs.

To assess the clinical relevance of the microbiome-based PDPs, we computed the correlation between several risk factors and overall glucose parameters, and the factors with beneficial and non-beneficial PDPs across the entire 800-person cohort. We found 20 statistically significant correlations (p < 0.05, FDR corrected) where microbiome factors termed non-beneficial correlated with risk factors, and those termed beneficial exhibited an anti-correlation (Figure 4Figure 4G). For example, higher levels of the beneficial methionine degradation KEGG module (M00035) resulted in lower PPGRs in our algorithm, and across the cohort, this module anti-correlates with systolic blood pressure and with BMI (Figure 4Figure 4G). Similarly, fluctuations in glucose levels across the connection week correlates with nitrate respiration two-component regulatory system (M00472) and with lactosylceramide biosynthesis (M00066), which were both termed non-beneficial. Glucose fluctuations also anti-correlate with levels of the tetrathionate respiration two-component regulatory system (M00514) and with RAs of Alistipes finegoldii, both termed beneficial (Figure 4Figure 4G). In 14 other cases, factors with beneficial or non-beneficial PDPs were correlated and anti-correlated with risk factors, respectively.

These results suggest that PPGRs are associated with multiple and diverse factors, including factors unrelated to meal content.

Next, we asked whether personally tailored dietary interventions based on our algorithm could improve PPGRs. We designed a two-arm blinded randomized controlled trial and recruited 26 new participants. A clinical dietitian met each participant and compiled 4–6 distinct isocaloric options for each type of meal (breakfast, lunch, dinner, and up to two intermediate meals), accommodating the participant’s regular diet, eating preferences, and dietary constraints. Participants then underwent the same 1-week profiling of our main 800-person cohort (except that they consumed the meals compiled by the dietitian), thus providing the inputs (microbiome, blood parameters, CGM, etc.) that our algorithm needs for predicting their PPGRs.

Participants were then blindly assigned to one of two arms (Figure 5Figure 5A). In the first, “prediction arm,” we applied our algorithm in a leave-one-out scheme to rank every meal of each participant in the profiling week (i.e., the PPGR to each predicted meal was hidden from the predictor). We then used these rankings to design two 1-week diets: (1) a diet composed of the meals predicted by the algorithm to have low PPGRs (the “good” diet); and (2) a diet composed of the meals with high predicted PPGRs (the “bad” diet). Every participant then followed each of the two diets for a full week, during which they were connected to a CGM and a daily stool sample was collected (if available). The order of the 2 diet weeks was randomized for each participant and the identity of the intervention weeks (i.e., whether they are “good” or “bad”) was kept blinded from CRAs, dietitians and participants.

Figure 5

Personally Tailored Dietary Interventions Improve Postprandial Glycemic Responses

(A) Illustration of the experimental design of our two-arm blinded randomized controlled trial.

(B) Continuous glucose measurements of one participant from the expert arm (top) and another from the predictor arm (bottom) across their “good” (green) and “bad” (red) diet weeks.

(C) Boxplot of meal PPGRs during the “bad” (red) and “good” (green) diet weeks for participants in both the predictor (left) and expert (right) arms. Statistical significance is marked (Mann-Whitney U-test, ^{∗∗∗}p < 0.001, ^{∗∗}p < 0.01, ^∗p < 0.05, † p < 0.1, n.s. not significant).

(D) As in (C), but for a grouping of all meals of all participants in each study arm (p, Wilcoxon signed-rank test).

(E) Boxplot of the blood glucose fluctuations (noise) of participants in both the “bad” (red) and “good” (green) diet weeks for both study arms. Blood glucose fluctuations per participant are defined as the ratio between the standard deviation and mean of his/her weeklong blood glucose levels (p, Wilcoxon signed-rank test).

(F) As in (E), but for the maximum PPGR of each participant.

(G) Subset of dominant food components prescribed in the “good” (green) diet of some participants and in the “bad” (red) diet of other participants. See also Figure S6Figure S6 for the full matrix.

(H) Dot plot between the CGM-measured PPGR of meals during the profiling week (x axis) and the average CGM-measured PPGR of the same meals during the dietary intervention weeks (y axis). Meals of all participants in both study arms are shown.

(I) As in (H), but when PPGRs in the dietary intervention weeks are predicted by our predictor using only the first profiling week data of each participant. Boxplots - box, IQR; whiskers 1.5^∗IQR.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

The second, “expert arm,” was used as a gold standard for comparison. Participants in this arm underwent the same process as the prediction arm except that instead of using our predictor for selecting their “good” and “bad” diets a clinical dietitian and a researcher experienced in analyzing CGM data (collectively termed “expert”) selected them based on their measured PPGRs to all meals during the profiling week. Specifically, meals that according to the expert’s analysis of their CGM had low and high PPGRs in the profiling week were selected for the “good” and “bad” diets, respectively. Thus, to the extent that PPGRs are reproducible within the same person, this expert-based arm should result in the largest differences between the “good” and “bad” diets because the selection of meals in the intervention weeks is based on their CGM data.

Notably, for 10 of the 12 participants of the predictor-based arm, PPGRs in the “bad” diet were significantly higher than in the “good” diet (p < 0.05, Figure 5Figure 5C). Differences between the two diets are also evident in fewer glucose spikes and fewer fluctuations in the raw week-long CGM data (Figure 5Figure 5B). The success of the predictor was comparable to that of the expert-based arm, in which significantly lower PPGRs in the “good” versus the “bad” diet were observed for 8 of its 14 participants (p < 0.05, 11 of 14 participants with p < 0.1, Figure 5Figure 5C).

When combining the data across all participants, the “good” diet exhibited significantly lower PPGRs than the “bad” diet (p < 0.05, Figure 5Figure 5D) as well as improvement in other measures of blood glucose metabolism in both study arms, specifically, lower fluctuations in glucose levels across the CGM connection week (p < 0.05, Figure 5Figure 5E), and a lower maximal PPGR (p < 0.05, Figure 5Figure 5F) in the “good” diet.

Both study arms constitute personalized nutritional interventions and thus demonstrate the efficacy of this approach in lowering PPGRs. However, the predictor-based approach has broader applicability since it can predict PPGRs to arbitrary unseen meals, whereas the “expert”-based approach will always require CGM measurements of the meals it prescribes.

Post hoc examination of the prescribed diets revealed the personalized aspect of the diets in both arms in that multiple dominant food components (as in Figure 2Figure 2F) prescribed in the “good” diet of some participants were prescribed in the “bad” diet of others (Figures 5Figures 5G and S6S6). This occurs when components induced opposite CGM-measured PPGRs across participants (expert arm) or were predicted to have opposite PPGRs (predictor arm).

Figure S6

Partition of Dominant Food Components into “Good” and “Bad” Diets of Participants across the Diet Intervention Weeks, Related to Figure 5Figure 5

Colored entries indicate a food component (x axis) that was prescribed in the “good” (green) or “bad” (red) diet week of a participant (y axis). Food components are shown in submatrices of foods that were only prescribed in the “good” diet week (left column block), food components prescribed only in the “bad” diet week (middle column block), and foods that were prescribed in the “good” diet of some participants and in the “bad” diet of other participants (right column block). Only dominant food component are shown, defined as foods whose carbohydrate content was more than 50% of the entire Caloric content of the meal in which they were prescribed.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

The correlation between the measured PPGR of meals during the profiling week and the average CGM-measured PPGR of the same meals during the dietary intervention was 0.70 (Figure 5Figure 5H), which is similar to the reproducibility observed for standardized meals (R = 0.71–0.77). Thus, as in the case of standardized meals, a meal’s PPGR during the profiling week was not identical to its PPGR in the dietary intervention week. Notably, using only the first profiling week data of each participant, our algorithm predicted the average PPGRs of meals in the dietary intervention weeks with an even higher correlation (R = 0.80, Figure 5Figure 5I). Since our predictor also incorporates context-specific factors (e.g., previous meal content, time since sleep), this result also suggests that such factors may be important determinants of PPGRs.

Taken together, these results show the utility of personally tailored dietary interventions for improving PPGRs in a short-term intervention period, and the ability of our algorithm to devise such interventions.

Finally, we used the daily microbiome samples collected during the intervention weeks to ask whether the interventions induced significant changes in the gut microbiota. Previous studies showed that even short-term dietary interventions of several days may significantly alter the gut microbiota (David et al., 2014xDiet rapidly and reproducibly alters the human gut microbiome. David, L.A., Maurice, C.F., Carmody, R.N., Gootenberg, D.B., Button, J.E., Wolfe, B.E., Ling, A.V., Devlin, A.S., Varma, Y., Fischbach, M.A. et al. Nature. 2014; 505: 559–563
Crossref | PubMed | Scopus (1119)See all References, Korem et al., 2015xGrowth dynamics of gut microbiota in health and disease inferred from single metagenomic samples. Korem, T., Zeevi, D., Suez, J., Weinberger, A., Avnit-Sagi, T., Pompan-Lotan, M., Matot, E., Jona, G., Harmelin, A., Cohen, N. et al. Science. 2015; 349: 1101–1106
Crossref | PubMed | Scopus (33)See all References).

We detected changes following the dietary interventions that were significant relative to a null hypothesis of no change derived from the first week, in which there was no intervention, across all participants (Figures 6Figures 6A and 6B ). While many of these significant changes were person-specific, several taxa changed consistently in most participants (p < 0.05, FDR corrected, Figure 6Figure 6C and S7S7). Moreover, in most cases in which the consistently changing taxa had reported associations in the literature, the direction of change in RA following the “good” diet was in agreement with reported beneficial associations. For example, Bifidobacterium adolescentis, for which low levels were reported to be associated with greater weight loss (Santacruz et al., 2009xInterplay between weight loss and gut microbiota composition in overweight adolescents. Santacruz, A., Marcos, A., Wärnberg, J., Martí, A., Martin-Matillas, M., Campoy, C., Moreno, L.A., Veiga, O., Redondo-Figuero, C., Garagorri, J.M...., and EVASYON Study Group. Obesity (Silver Spring). 2009; 17: 1906–1915
Crossref | PubMed | Scopus (198)See all ReferencesSantacruz et al., 2009), generally decrease in RA following the “good” diet and increase following the “bad” diet (Figure 6Figure 6C,D). Similarly, TIIDM has been associated with low levels of Roseburia inulinivorans (Qin et al., 2012xA metagenome-wide association study of gut microbiota in type 2 diabetes. Qin, J., Li, Y., Cai, Z., Li, S., Zhu, J., Zhang, F., Liang, S., Zhang, W., Guan, Y., Shen, D. et al. Nature. 2012; 490: 55–60
Crossref | PubMed | Scopus (1022)See all ReferencesQin et al., 2012; Figure 6Figure 6E), Eubacterium eligens (Karlsson et al., 2013xGut metagenome in European women with normal, impaired and diabetic glucose control. Karlsson, F.H., Tremaroli, V., Nookaew, I., Bergström, G., Behre, C.J., Fagerberg, B., Nielsen, J., and Bäckhed, F. Nature. 2013; 498: 99–103
Crossref | PubMed | Scopus (494)See all ReferencesKarlsson et al., 2013), and Bacteroides vulgatus (Ridaura et al., 2013xGut microbiota from twins discordant for obesity modulate metabolism in mice. Ridaura, V.K., Faith, J.J., Rey, F.E., Cheng, J., Duncan, A.E., Kau, A.L., Griffin, N.W., Lombard, V., Henrissat, B., Bain, J.R. et al. Science. 2013; 341: 1241214
Crossref | PubMed | Scopus (667)See all ReferencesRidaura et al., 2013), and all these bacteria increase following the “good” diet and decrease following the “bad” diet (Figure 6Figure 6C). The Bacteroidetes phylum, for which low levels associate with obesity and high fasting glucose (Turnbaugh et al., 2009xA core gut microbiome in obese and lean twins. Turnbaugh, P.J., Hamady, M., Yatsunenko, T., Cantarel, B.L., Duncan, A., Ley, R.E., Sogin, M.L., Jones, W.J., Roe, B.A., Affourtit, J.P. et al. Nature. 2009; 457: 480–484
Crossref | PubMed | Scopus (2698)See all ReferencesTurnbaugh et al., 2009), increases following the “good” diet and decreases following the “bad” diet (Figure 6Figure 6C). Low levels of Anaerostipes associate with improved glucose tolerance and reduced plasma triglyceride levels in mice (Everard et al., 2011xResponses of gut microbiota and glucose and lipid metabolism to prebiotics in genetic obese and diet-induced leptin-resistant mice. Everard, A., Lazarevic, V., Derrien, M., Girard, M., Muccioli, G.G., Neyrinck, A.M., Possemiers, S., Van Holle, A., François, P., de Vos, W.M. et al. Diabetes. 2011; 60: 2775–2786
Crossref | PubMed | Scopus (289)See all ReferencesEverard et al., 2011) and indeed these bacteria decrease following the “good” diet and increase following the “bad” diet (Figure 6Figure 6C). Finally, low levels of Alistipes putredinis associate with obesity (Ridaura et al., 2013xGut microbiota from twins discordant for obesity modulate metabolism in mice. Ridaura, V.K., Faith, J.J., Rey, F.E., Cheng, J., Duncan, A.E., Kau, A.L., Griffin, N.W., Lombard, V., Henrissat, B., Bain, J.R. et al. Science. 2013; 341: 1241214
Crossref | PubMed | Scopus (667)See all ReferencesRidaura et al., 2013) and this bacteria increased following the “good” diet (Figure 6Figure 6C).

Figure 6

Dietary Interventions Induce Consistent Alterations to the Gut Microbiota Composition

(A) Top: Continuous glucose measurements of a participant from the expert arm for both the “bad” diet (left) and “good” diet (right) week. Bottom: Fold change between the relative abundance (RA) of taxa in each day of the “bad” (left) or “good” (right) weeks and days 0–3 of the same week. Shown are only taxa that exhibit statistically significant changes with respect to a null hypothesis of no change derived from changes in the first profiling week (no intervention) of all participants.

(B) As in (A) for a participant from the predictor arm. See also Figure S7Figure S7 for changes in all participants.

(C) Heatmap of taxa with opposite trends of change in RA between “good” and “bad” intervention weeks that was consistent across participant and statistically significant (Mann-Whitney U-test between changes in the “good” and “bad” weeks, p < 0.05, FDR corrected). Left and right column blocks shows bacteria increasing and decreasing in their RA following the “good” diet, respectively, and conversely for the “bad” diet. Colored entries represent the (log) fold change between the RA of a taxon (x axis) between days 4–7 and 0–3 within each participant (y axis). Asterisks indicate a statistically significant fold change.

See also Figure S7Figure S7 for all changes.

(D) For Bifidobacterium adolescentis, which decreased significantly following the “good” diet interventions (see panel C), shown is the average and standard deviation of the (log) fold change of all participants in each day of the “good” (top) diet week relative to days 0–3 of the “good” week. Same for the “bad” diet week (bottom) in which B. adolescentis increases significantly (see panel C). Grey lines show fold changes (log) in individual participants.

(E) As in (D), for Roseburia inulinivorans.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

Figure S7

Dietary Interventions Induce Significant Alterations to the Gut Microbiota Configuration, Related to Figure 6Figure 6

Shown are all bacteria present in at least one stool sample in at least 10 different intervention weeks. Entries are colored according to the (log) fold change between the relative abundance of a bacteria (y axis) in days 4–7 and the relative abundance of that bacteria at days 0–3 within each participant (x axis). Asterisks indicate a statistically significant fold change. Left and right columns blocks correspond to changes in “bad” and “good” intervention weeks, respectively. Top and bottom row blocks correspond to a trend of decreasing and increasing abundance in the “good” intervention week, respectively, and conversely in the “bad” week. Taxa with statistically significant (Mann-Whitney U-test) patterns of consistent change following the “good” diet that is opposite to a consistent change following the “bad” diet are marked to the right of the heatmap (^{∗∗} p < 0.01, ^∗ p < 0.05, FDR corrected). P, phylum; C, class; O, order; F, family; G, genus; S, species.

View Large Image | View Hi-Res Image | Download PowerPoint Slide

These findings demonstrate that while both baseline microbiota composition and personalized dietary intervention vary between individuals, several consistent microbial changes may be induced by dietary intervention with a consistent effect on PPGR.