Variation in human data can be generated from multiple sources. Identifying deviations between individual responses to dietary challenges and grouping individuals with similar phenotypic responses is one means of "metabotyping." Inter-individual variability in human studies can interfere with the identification of human metabotypes, but appropriate data transformations can facilitate data interpretation. These tools can reduce the volume and complexity of data sets and can quiet variance from both biological and analytical sources. Hence, pre-treatment of data has been widely used to organize data, enabling efficient analysis with improvement in statistical power . In the current study, we transformed data to an individualized response scale to reduce inter-individual variability, which enhanced the ability to identify distinct groups within the measured response profiles. The response profile of RG1 appears consistent with the expected differential response to a low and high glycemic meal challenge. RG2 and RG3 appear to indicate alternate physiological responses within this clinically healthy population. We hypothesize that these distinct "response profiles" are putative metabolic phenotypes that can be identified based on their postprandial response to a single meal challenge.
The range of a dataset is affected by both biologically significant inter-individual variability, and measurement or analytical errors. Range-scaling first centers the data matrix, followed by scaling each data point to the range of the data array. This transforms the dataset into one containing values bound by ± 50% of the mean. Centering reduces offsets of the data set, decreasing the variability introduced by large differences in magnitude, while shifting the scale of the resulting data set by the magnitude of the mean. Once variance is reduced, this cleaner dataset is more proficient at revealing trends that would likely have gone unnoticed in untreated data . When metabolites are analyzed together using a multivariate technique that identifies trends in data based on variance, often the variables that exhibit large changes under similar treatment or challenge conditions tend to dominate the results. In this study-insulin which is highly responsive to the glycemic index of meals could be considered such a metabolite, as opposed to leptin, which might not display such large variations in magnitude. Range-scaling controls for this, by placing each metabolite on a 0-100% scale of its own experimental response range . The data was nadir-corrected for each test day, and scaled to the range of the high glycemic index response in order to better highlight the individual's relative responses to the HGI and LGI challenges. Raw data and mean-centered data failed to show this discrimination. One drawback of range scaling is its sensitivity to outliers in the dataset , as is PCA, albeit a conservative multivariate analysis technique. When applying PCA, one must be cautious of outliers in the dataset that define the variance of the distribution instead of the inherent biological variation that is the target. It is also important to note that the PCA was performed on "auto-scaled" data (i.e. data transformed to unit variance) prior to the analysis, which allows each variable to have equal weight in the analysis . Performing the analysis on the adjusted-range scaled data without further transformation did reveal RG3 as distinct from RG1 and RG2 (p = 0.11), while RG2 subjects could not be excluded from RG1 using a test for statistical outliers. Mean centering the data improved the RG3 difference (p = 0.001), but increased the variance in the subjects making up the RG2 group. In this dataset, nadir-adjusted range scaling enabled the PCA to identify clustering of individuals with similar postprandial responses better than the raw dataset or the conventional range scaled dataset.
Circulating leptin exhibits a diurnal rhythm and increases about 4-h following a mixed meal . In humans, acute leptin responses to meals-high fat, high carbohydrate, and high GI-have been studied, yielding variable results. High carbohydrate as opposed to high fat meals have been shown to increase leptin secretion . Compared to a low GI meal, a high GI meal has been shown to result in lower postprandial leptin concentrations, but only over a 2-h period , a time frame too short to observe insulin-induced leptin secretion.
In vitro and animal model studies have suggested that the adipoinsular axis is a long term regulatory circuit that likely plays a role in the metabolic response to meals. The postprandial insulin response is modulated by circulating concentrations of leptin. The theory behind the adipo-insular axis posits that chronic leptin secretion, and its downstream insulin suppression is maintained in normal healthy individuals at an optimal level. Analyzing the insulin, glucose and leptin response to the meal challenges can give insight into the functioning of this axis. The adipo-insular axis can play an important role in establishing fasting and postprandial glycemia, and glucose clearance from circulation.
In this study, the response group designated RG1 (n = 18) is most likely representative of the general population, i.e. normal, healthy individuals with typical GI meal induced responses. On the other hand, RG2 displays the highest insulin AUC profile of the 3 groups associated with a lower leptin response. This relatively high insulin response is irrespective of the meal composition (seen both in HGI and LGI) since the 3 × 3 ANOVA indicated that RG2 had higher HGI insulin response than RG1 and RG3 challenged with either diet and at early- mid- and late-postprandial phases. The relatively elevated insulin response in RG2 could indicate reduced insulin sensitivity as compared to RG1, despite RG2 having a lower fat mass than RG1. Also, while RG1 and RG3 display different responses to the two meal challenges, RG2 displays similar insulin and glucose responses to both the meal challenges. This further is indicative of RG2's errant metabolic fine-tuning to meal composition.
In contrast, relative to RG1, RG3 displays a higher early postprandial leptin AUC, in association with higher postprandial glucose, and a similar insulin AUC response. Since RG3 insulin response is similar to or slightly lower than RG1, but ineffective at maintaining glycemia similar to that of RG1, could indicate either a reduced sensitivity or a higher postprandial suppression of insulin by leptin. It can be speculated that in RG3, the meal challenges elicited a higher leptin response, with stronger suppression of insulin secretion, resulting in lower circulating insulin, and higher glucose concentrations. In addition, this group has a significantly higher fat mass than RG2, although not different from RG1. Postprandial leptin secretion is likely a function of the insulin sensitivity of adipose tissue. Ideally, in response to a HGI meal, the higher insulin secretion will be regulated by leptin, but optimally so as to not disrupt glucose clearance. Thus, impairment in this axis can result in poor glycemic control as exhibited by RG3.
The number of subjects in each of the response groups (n = 3) is small. While individual variation or noise might explain this observation, the three clusters observed in the PCA scores plot of the transformed data could also indicate clustering of similar metabolic response profiles. Further mechanistic studies are necessary to evaluate if the identified groups are distinct phenotypes. Another limitation of the current study is that it is not possible to distinguish the influence of the run in diets on the postprandial meal challenge response that we have characterized. It is likely that the response groups that we have identified are a function of an acclimation to the 3 days of the run in diets, in addition to the meal challenge. Also, the subjects were challenged just once with each of the two meal challenges, and this is a limitation of this study.
Together, these findings suggest that the use of meal challenges and statistical exploratory tools could lead to metabolic phenotyping. This may enable better population sorting to identify health risks associated with specific postprandial phenotypic responses. Scaling, transformations, and non-parametric multivariate analyses have identified subsets from the study population of differing metabolic response profiles. Depending on how similar or dissimilar the individuals' response is to previous literature suggested postprandial metabolic events (which are likely also the majority of the normal, healthy population's responses) a metabolic fingerprint of this variation can be used to classify metabolic status at a finer gradient. Even the current clinically normal responses can be better characterized by using meal challenge responses and statistical tools to identify subgroups of metabolic characteristics that can form the basis of putative phenotypes. Such efforts may ultimately allow greater diagnostic finesse.