In todays issue of Cell, two groups led by Eran Elinav and Eran Segal have presented a stunning paper providing startling new insight into the personal nature of nutrition. The Israeli research teams have demonstrated that there exists a high degree of variability in the responses of different individuals to identical meals, and through the elegant application of machine learning, they have provided insight into the diverse factors underlying this variability.
Artist’s rendering of figure by Zeevi and Korem et al./Cell 2015
As genetic factors are known to modulate anindividual’s innate responses to diseases, medications, and blood metabolites, it may come as no surprise that individuals do not respond to identical foods in the same manner.
Following a meal, glucose levels increase according to the type of foods that are ingested. Currently, meal carbohydrates or derived glycemic indexes are used to estimate the postprandial (post-meal) glycemic responses (PPGR). These factors assume that PPGRs are solely dependent on the intrinsic properties of the ingested food, and this assumption is the basis of universal dietary recommendations.
While it is has been proposed that individual differences in PPGR may be influenced by diverse factors including genetics, lifestyle, and insulin sensitivity, as well as the activity levels of exocrine pancreatic and glucose transporters, the influence of gut microbiota on PPGR is relatively poorly understood.
In order to better understand these relationships, the Elinav and Segal teams sought to quantify individual PPGRs for a sample population, characterize the variability within their sample, and then identify factors associated with that variability. Though this could be achieved with a small set of a dozen, or perhaps a few dozen participants, this is not the approach that the teams took.Observations of 800 study participants.
Elinav and Segal reported instead on their observations ofa cohort of 800 healthy and pre-diabetic individuals, a sample population representative of the larger Western non-diabetic population.
Each member of the sample population was connected toa continuous glucose monitor thatmeasured their interstitial glucose levels every 5 min for a full week (using subcutaneous sensors). They collected more than 2,000 measurements per participant, for a total of more than 1.6 million measurements for the entire population.
In order to better understand the relationship between the measured glucose levels and the unique physiologies and lifestyles of their studys participants, the teams collected an extremely diverse set of information from each participant. Study participants kept diaries of their physical activity, food intake, and sleep using a smartphone application. This information was then complimented with comprehensive profiles that researchers collected from each study participant, including food frequency, lifestyle, medical background, anthropometric measures (body measurements, such as height, hip circumference), full panel blood tests results, and single stool sample results used for microbiota profiling.
When the researchers analyzed their collected results, their findings varied from theexpected, to those that were truly startling. As expected, the researchers were able to validate known associations of PPGRs with risk factors such as BMI, glycated hemoglobin, morning glucose levels, and age. The scientists did make some surprising observations, noting that these associations were not limited to extreme values; associations between these known risk factors occurred over the entire phenotypic spectrum, indicating that incremental differences in the glucose response may be clinically relevant for some risk factors.Nutrition is personal. A high degree of variability exists in the responses of different people to the same food.
The collected observations further revealed both an individuals responses to the same food were reproducible, and that there exists a high levels of variability in the responses of different individuals to the same foods. The researchers found that the food associated with an individuals highest glucose response varied greatly between individuals. Foods that induced a healthy response in one individual might induce an unhealthy response in another. In a particularly compelling figure, the researchers showed an example where two participants had opposite responses to cookies and bananas.
Participant A maintained a stable blood glucose level after eating a cookie but responded with elevated glucose levels after eating a banana. Conversely, participant B experienced an increase in blood glucose level after eating a cookie, but not after consuminga banana.
In another striking example one of the participants, a middle-aged woman who was obese and pre diabetic learned that her “healthy” eating habits may have actually been unhealthy for her. The researchers found that in her case, eating tomatoes resulted in an “unhealthy” blood sugar spike, as the woman ate tomatoes frequently over the course of the week long monitoring period, this “healthy” habit may have been undermining her health.
Within their data set Elinav and Segals research teams found multiple significant associations between participants PPGRs following standardized meals, and both their clinical and gut microbiome data, with positive correlations detected between BMI, HbA1c%, systolic blood pressure, and alanine aminotransferase activity. With the microbiome data, the researchers noted that a preponderance of Proteobacteria and Enterobacteriaceae was positively associated with PPGRs.
This graphic depicts the inputs used to develop a personalized nutrition predictor algorithm.CREDIT:Zeevi and Korem et al./Cell 2015Identifying personal nutritional response profiles using a machine-learning based algorithm.
Using their set of amassed data, the researchers then went a step further, applying machine-learning algorithm to their cohort of 800 participants and developing an algorithm capable of predictingindividualized PPGRs. This intricate algorithm incorporates 137 features representing meal content, daily activity, blood parameters, CGM-derived features, questionnaires, and microbiome features. This modelpredicts measured PPGRs with a significantly higher correlation (R=0.68) than solely relying on carbohydrate counting (R=0.38) or the meal caloric input (R=0.33).Validation of the algorithm using an independently recruited cohort.
Going even farther, the research team then recruited a separate 100 participant cohort and validated the predictive capacity of their algorithm. Finally, revealing the true utility of their approach a final set of participants was recruited for an intervention study. The intervention study used a set of 26 participants, distributed into two experimental groups, one that would apply predictions developedby an expert dietitian and researcher working together using continuous glucose monitoring data, and one that would use the algorithm the research team developed. Each participant was then assigned, in a doubly blind manner and in a random order, a one week long good diet, and a one week long bad diet.Application of the algorithm to an intervention study
The results of the intervention study revealedthat the diets developed by both the expert dietician/researcher team and thealgorithm resulted in lower blood glucose variability on the good diet, and the results indicated that the two methods had similar efficacy. Impressively,daily microbiome samples from these participants revealed that even short-term dietary interventions induced changes in the microbiome. The good diet was consistent with an increase in the beneficial bacteria, witha reciprocal decrease following the bad diet.