The current study aimed to define obesity beyond BMI to better capture the metabolic diversity behind this complex condition and to better assess the risk of developing future cardiometabolic disease. In order to do that, the K-means clustering algorithm was deployed and three metabolomics-based clusters were defined within a subset of the general population cohort [email protected], that only included individuals with obesity. These clusters were designed to maximize the differences of several well-known obesity associated predictors among individuals. On the one hand, the traditional lipid profiling consisting of TG, LDLC and HDLC was considered for clustering, since its association with obesity has been clearly described in several clinical guidelines [37–38]. On the other hand, glucose was added to the model to account for the association between obesity and IR [2, 3, 8]. In addition, novel H-NMR metabolomics-based predictors were also considered. First, ratios of small lipoproteins to the total circulating particle number (s/t-VLDLp, s/t-LDLp and s/t-HDLp) were computed to incorporate information about lipoprotein number and size into the clustering algorithm, since both have been associated with obesity [2, 39–40]. Furthermore, the functionality of the main lipoprotein classes (VLDL, IDL, LDL and HDL) was also considered, by including the TG/C ratios for each class in the clustering algorithm. The presence of dysfunctional TG-enriched HDL lipoproteins, in combination with high concentrations of TG-rich lipoproteins (TRL) have been found to be associated with a higher risk of CVD [25, 39–40]. GlycA was the last predictor incorporated into the clustering algorithm, since this parameter is a robust biomarker of systemic inflammation [33–35, 41]. Based on all the parameters described, three different clusters were defined. Even though statistically significant differences were not observed for BMI and abdominal obesity between clusters, several metabolic differences were found among them. Each of them seemed to be describing different subtypes of obesity, that had already been mentioned in the first section of this document: (1) MUO with predominant atherogenic dyslipidemia, (2) MUO with predominant HC and (3) MHO. However, the current analysis enabled a detailed description of several pathogenic mechanisms that were taking place in addition to dyslipidemia and evaluated the differential cardiometabolic risk derived from each of the profiles found.
On the one hand, atherogenic dyslipidemia was found to be present on the second cluster. All its classical signs, consisting of HTG, low HDLC accompanied by TG-enriched HDLP and a high amount of sLDLp despite having moderate LDLC values [2, 40–41] were observed in this cluster. This lipid profile has been often described in individuals with obesity in whom IR was also present [42]. Coherently, several markers of IR such as glucose, lactate, alanine and branched-chain amino acids (BCAA) [23] reached their highest concentration values in this cluster. In addition, the highest concentrations of remnant cholesterol (i.e. VLDLC and IDLC) were also observed in this cluster. A recent study attributed a fundamental role to remnant cholesterol in the development of type 2 DM, especially in individuals with moderate levels of LDLC as the ones included in this cluster [43]. On top of the described metabolic disturbances, the presence of a high inflammatory state was assessed by the elevated levels of GlycA and GlycB found in this cluster [33–35, 44]. All the pathogenic alterations described pointed in the direction of a predisposition of individuals with this specific obesity profile to develop type 2 DM. The analysis of the follow-up data showed an association between this cluster and the future development of type 2 DM, thereby confirming the previous hypothesis.
On the other hand, predominant HC was found to be present in the third cluster. This cluster displayed the highest TC and LDLC concentrations. Accordingly, this cluster was found to have the largest amount of LDLp. However, the lLDLp and mLDLp were the most increased fractions. It is well known that LDLC-driven HC is one of the major risk factors of cardiovascular disease, atherosclerosis and CHD [45, 46]. Coherently with this, our results showed this cluster to be the most associated one with future CV events, including CHD among other conditions. Another interesting characteristic of this cluster was the GlycA-driven inflammation, in absence of increased GlycB levels. Our previous work showed that increased values of GlycA in combination with moderately increased values of GlycB were associated with a higher risk of developing type 2 DM in hyperglycemic individuals with similar TC and LDLC concentrations [47]. Thus, GlycB-derived glycosylation might be associated with TG metabolism rather than with serum cholesterol levels, and thereby, might be more strongly contributing to IR and type 2 DM than to CV events.
The last cluster defined, the one named as cluster 1, was the closest to MHO. The lowest concentrations of TC and TG in all the lipoprotein classes were found in this cluster, except for HDL. Indeed, the highest HDLC together with the lowest HDLTG concentrations were displayed by this cluster. These findings reflected a healthy HDL function that has been associated with a better cardiovascular health [41]. Accordingly, the individuals within this cluster displayed the lowest amount of ApoB-containing atherogenic lipoproteins (VLDL and LDL) and the highest concentrations of apoA-containing lipoproteins (HDLP). These findings, together with the absence of IR, AHT and dyslipidemia have been described in MHO phenotype [8–11]. Interestingly, our results showed a higher amount sHLDp and the smallest HDLP diameter to be present in this cluster. This contradicts previous literature describing larger HDLP in MHO [40, 44]. It must be noted that the MHO is still an ambiguous phenotype, whose definition is not currently standardized [8–11] and that the particle size thresholds vary between different techniques. Although this cluster showed the lowest association with future cardiometabolic disease, this association was not negligible. Several studies have discussed the fluctuant character of MHO, arguing that the lower association of this phenotype with type 2 DM could just be a matter of time [9, 11]. Moreover, specific associations between MHO and future type 2 DM have been found and attributed to abnormalities in Bromodomain and extra terminal (BET) proteins which promote pancreatic β-cell function and proliferation [8, 48].
All that has been discussed above made it clear that obesity is a complex disease that cannot be explained using just anthropometric measures, such as BMI or WHR. Nevertheless, it is not our aim to state that these traditionally used methods are useless but to demonstrate that complementing them with information derived from H-NMR metabolomics can improve the characterization of this complex disease. We showed that including H-NMR-metabolomics information together with BMI and WHR improved the prediction of the development of future cardiovascular disease in a 12% and a 6%, respectively. Moreover, the inclusion of metabolomics data increased the specificity of the models by 24% and 8%, when it was used together with BMI and WHR. This showed that the incorporation of H-NMR metabolomics into the predictive models enabled a better identification of those individuals being at risk of developing cardiometabolic disease in the future. Although further validation of the clusters defined in this study is needed, the current results suggested potential benefits of using H-NMR metabolomics as a complementary tool of the currently available clinical measures of obesity.
This study presents several strengths. First of all, a large sample that comprised men and women with obesity in a wide age range, was used for the definition of the clusters. Treated individuals were also included, thereby reflecting in a more accurate way the general population. Although different treatments might have a different impact on metabolism, the exclusion of treated individuals would bias the results since the population obtained would not be representative of the general population. In addition, H-NMR was used to conduct the metabolomics measurements. This is a non-destructive technology that requires a minimum sample processing to quantify the most abundant metabolites and macromolecules present in different biological matrices, even if these analytes have identical molecular masses [49]. In addition, NMR spectroscopy is unbiased, fast, very reproducible and highly automatable [14]. However, some limitations must be also acknowledged in this study. First, the validation of the results in a completely independent large cohort was not possible due to the lack of another cohort equally profiled. In addition, the number of individuals developing different kinds of cardiometabolic diseases in the follow-up was limited. To overcome this issue, the different conditions were grouped under the broader category of cardiometabolic disease, so balanced groups could be obtained for the predictive models.