Metabolomic Profile of Low–Copy Number Carriers at the Salivary α-Amylase Gene Suggests a Metabolic Shift Toward Lipid-Based Energy Production

Salivary α-amylase (sAA), encoded by the AMY1 gene, is produced under the control of the autonomic nervous system by the salivary glands. It initiates starch digestion in the mouth, yielding a mixture of maltose, isomaltose, and glucose. Therefore, it was expected that individuals with high levels of sAA would show higher postprandial glycaemia, compared with low sAA counterparts. A recent study (1) showed, surprisingly, the opposite effect. Lower serum amylase levels have been associated with susceptibility to metabolic abnormalities, including nonalcoholic fatty liver disease, obesity, and diabetes (2–4).

Salivary amylase level is strongly correlated with the copy number (CN) at the AMY1 gene (5). Higher AMY1 CN and sAA protein levels were observed in populations traditionally consuming starch-rich diets compared with those consuming a low-starch diet (5). We recently confirmed AMY1 gene expression and protein levels in serum to be inversely correlated with obesity and suggested CN variation at AMY1 to be an important driver of this relationship, showing an association between low amylase CN and increased BMI and obesity risk in both adults (6) and children (7). On the other hand, a novel study has suggested that CNV might not be the main factor driving the association between amylase levels and obesity susceptibility (8). The mechanisms by which sAA contributes to an increased risk of metabolic diseases remain elusive (8,9). To further investigate the role of genetic variation at the AMY1 gene on the metabolism, we used untargeted ¹H-nuclear magnetic resonance (NMR) and mass spectrometry (MS)–based metabolomics, and compared the metabolomic profiles of a carefully selected sample of 100 healthy and normal-weight, age-matched women carrying either low–AMY1 CNs (LAs) or high–AMY1 CNs (HAs).

Only women were included to provide a more homogenous study population to maximize power, given our small study population. The characteristics of the women carrying HAs (n = 50) or LAs (n = 50) are shown in Supplementary Table 1. The women were selected from the prospective D.E.S.I.R. (Data from an Epidemiological Study on the Insulin Resistance Syndrome) cohort (9) from among those for whom data on the inferred CN at the AMY1 gene were already available (6). To control for the effects of potential confounders, we carefully selected two subsamples from the top and bottom 20% of the CN distribution (four or fewer CNs and eight or more CNs). Exclusion criteria included the following: 1) BMI >25 or <18.5 kg/m²; 2) fasting glycemia >6 mmol/L; and 3) development of any metabolic disorders during the 9-year follow-up period. Food and caloric intake were not significantly different between the HA and LA groups.

All participants signed a written informed consent, and the study was approved by the Ethics Committee for the Protection of Subjects for Biomedical Research of Bicêtre Hospital (Le Kremlin-Bicêtre, France).

Serum pancreatic and total amylase levels were measured by enzymatic colorimetric assays with an auto-Roche/Hitachi Cobas 8000 c701 (AMY-P-20766623322 and AMYL2–03183742122; Hoffman-La Roche). The sAA levels were obtained by subtracting the pancreatic amylase level from the total amylase level.

MS-based metabolomics profiling was carried out at Metabolon (Durham, NC). The processing and analysis of samples was performed as previously described (10) (Supplementary Data).

Spectra of serum samples were measured at 600.13 MHz on a Bruker Avance II 600-MHz spectrometer (Bruker BioSpin GmbH, Rheinstetten, Germany) following the study by Beckonert et al. (11). NMR spectra were phased and baseline corrected. Spectra were aligned using the CluPA algorithm (12), and the alanine signal was used as a reference (1.4784 parts per million [ppm]). Residue water signal was removed. Data were reduced to 513 bins by intelligent bucketing and were normalized by total sum normalization. Mean centering and Pareto scaling were applied prior to performing the data analysis.

We used multivariate statistical data analyses based on projection methods. Principal component analysis was used for the initial exploratory data analysis. Metabolomics signatures distinguishing HA and LA women were identified through projection to latent structures discriminant analysis (PLS-DA), based on variable importance in projection (VIP) selection (13). Given the limited sample size, we used a careful and conservative approach to validate our results.

First, results were validated through N-fold full cross-validation, using different values of N (N = 6, 7, and 8; we report only results for sevenfold as the prediction error estimated by cross-validation [cross-validated R²] [Q²] and area under the receiver operating characteristic curve estimated by cross-validation [AUC_CV], two measures to estimate the predictive ability of the model), as well as through 1,000 permutation testing on the HA/LA classes (13).

Second, to avoid overfitting we performed stability selection by Monte Carlo subsampling: real differences should be present consistently and therefore should be found even under perturbation of the data by subsampling (14). We generated 500 random data sets by Monte Carlo subsampling (with prior probability of 0.70), and then VIP-based PLS-DA was applied to each subsample, obtaining a set of 500 discriminant models. Each model was additionally validated by N-fold full cross-validation and 1,000 permutations of the class response. Metabolites discriminating between HA and LA classes were identified as those that were selected by >50% of all the models.

Both in the original and in the randomly generated data sets, the VIP threshold for the selection of metabolites discriminating between HA and LA classes was determined by maximizing Q², whereas the number of latent variables of the PLS model was calculated by maximizing Q² under the constraint to pass the permutation test. To clarify model interpretation, PLS-DA VIP-based models were post-transformed into the equivalent orthogonal PLS-DA models (OPLS-DA), where the predictive and the orthogonal parts of the models were identified.

HA subjects showed significantly higher levels of serum sAA compared with LA counterparts (Wilcoxon test, P = 6.04 × 10⁻⁸). A smaller difference was detected for the pancreatic amylase levels (Wilcoxon test, P = 0.025), which is likely to be a reflection of both physiological and structural correlations between the two enzymes (15) (Supplementary Table 1 and Supplementary Fig. 1).

The analysis of the NMR and MS data sets by principal component analysis detected the presence of 15 outliers that were excluded from further analysis. The major patterns within the original data were captured by a small number of components, three for both the NMR and MS data sets. The score scatter plots in Fig. 1A and B show the separation of the HA and LA samples based on the PLS-DA analyses of the MS and NMR data. The goodness of fit and the predictive ability estimated by sevenfold cross-validation were R² = 0.45 and AUC_CV = 0.79 for the NMR data set and R² = 0.71 and AUC_CV = 0.92 for the MS data set. Figure 1C and D compare the R² and Q² obtained through random permutation of the class response with what was observed in the original data set (empirical P < 0.001 for both R² and Q²).

All data sets generated through Monte Carlo sampling successfully passed both permutation testing and N-fold full cross-validation, thus suggesting that model fitting in the data was not driven by a small number of observations. On the basis of the stability selection, we identified two signatures from MS and NMR, encompassing 41 metabolites and 15 signals, respectively (Tables 1 and 2), which provided the largest contribution to the discrimination of the HA and LA groups. To investigate the role of each metabolite in the discrimination of the two groups, we plotted the median of the VIP calculated by stability selection versus the correlation between the predictive latent variable of the OPLS-DA model and the measured variables. Since the VIP and the correlation coefficient estimate the strength and the direction of the effect, respectively, metabolites in the right quadrant of the plot showed higher levels in the HA group, whereas those in the left quadrant showed higher levels in the LA group (Fig. 2A and B).

Metabolites from MS data were mapped into the Ingenuity Pathway Analysis Knowledge Base database (accessed October 2015; QIAGEN, Redwood City, CA). The reference set included the endogenous chemicals; we chose to focus only on experimentally observed interactions with no restrictions on cell type or species. The top five most significantly enriched molecular and cellular functions included carbohydrate metabolism, energy production, and lipid metabolism (Supplementary Table 2). Their most significant functional annotations were as follows: oxidation of glucose-6-phosphate (P = 3.12 × 10⁻⁸), oxidation of lipid (P = 1.32 × 10⁻⁶), and accumulation of acylglycerol (P = 6.19 × 10⁻⁸) and triacylglycerol (P = 2.85 × 10⁻⁶). Of the 41 MS compounds, 28 (68%) were fatty acids (FAs), 17 and 11, respectively, with higher and lower levels in LA sera, compared with HA sera. Similarly, 9 of the 15 NMR signals (60%) corresponded to FAs, and their levels were higher in LA sera.

Despite ¹H-NMR experiments being highly reproducible, different lipids cannot be easily distinguished from each other, but rather are assigned to multiple macro-categories, characterized by the abundance of specific aliphatic chains and their magnetic shielding. Therefore, since the signals of the lipids detected by NMR can be characterized by only the three main functional groups (CH2-CH = CH-, -CH2-, and -CH3), we investigated their correlation with the identified MS metabolites. The strongest correlations (r > 0.7) were observed among the four dicarboxylic FAs (DFAs) sebacate, azelate, suberate, and dodecanedioate, which were identified by MS and the -CH2- NMR bucket in the region from 1.15 and 1.19 ppm. Interestingly, the same DFAs were also strongly correlated (r > 0.8) with the glucose NMR bucket at 3.40 ppm. The BMI distribution was not different between the HA and LA groups (Wilcoxon test, P = 0.986). BMI was independent of HA/LA group definition and did not have any effect on the multivariable PLS models (r² < 0.01, P > 0.20). We further excluded the association between BMI and each single discriminant metabolite through linear regression by considering both BMI and HA/LA level as independent variables (P for BMI > 0.10).

In this study, we aimed to explore the effect of CN variation at AMY1 on metabolism by comparing the metabolomic signature in carefully matched healthy women carrying LA or HA, as reflected in a significant difference in enzyme levels between the two groups (P = 6.04 × 10⁻⁸) (Supplementary Fig. 1). Although this study design facilitates the identification of discriminant metabolites associated with AMY1 CN, the results might be affected by spectrum bias because women were selected from the two ends of the CN distribution.

The best-fitting models (empirical P < 1 × 10⁻³) identified through multivariate analyses of the MS and NMR metabolomics data concordantly highlighted differences in lipid metabolism between LA and HA women. Analysis of the MS-discriminant metabolites using the Ingenuity Knowledge Base annotations confirmed the significant enrichment of functional categories for “lipid metabolism,” “energy production,” and “carbohydrate metabolism.” In particular, the MS data showed that the levels of four DFAs, suberate, sebacate, azelate, and dodecanedioate, were higher in the sera from LA women compared with HA women (Table 2). This increase in DFA levels is suggestive of an upregulation of the ω-oxidation. In normal conditions, ω-oxidation plays a minor role in overall FA oxidation. However, in certain physiological states in which mitochondrial β-oxidation is blocked or overwhelmed, free FAs are ω-oxidized in the reticulum, by a CYP4A-dependent pathway, to DFAs (16), which can then enter mitochondrial β-oxidation.

To corroborate the proposed interpretation, alongside the increase in DFAs, women with LA also showed lower serum concentrations of several medium- and long-chain FAs (Table 2). This reduction in FA levels is coherent with a more active uptake and oxidation, both via β- and ω-oxidation.

Interestingly, LA women also showed increased levels of 2-hydroxybutyrate, a biomarker for insulin resistance in humans (17) (Table 2). Taken together, these results suggest a pattern of reduced cellular glucose uptake, and a consistent metabolic shift toward lipid exploitation, in LA women. Interestingly, NMR findings suggested a slight increase of glucose levels in LA sera.

In a similar line, Mandel and Breslin (1) recently published an experiment in which they measured the dynamics of postprandial glucose and insulin after controlled starch or glucose meals in healthy LA and HA carriers. Although LA and HA carriers did not show any differences after the glucose-based meal, LA carriers showed significantly higher glucose levels after the starch-based meal, suggesting glucose malabsorption after starch ingestion in people with constitutionally low salivary amylase levels. Mandel and Breslin (1) advocated that this effect might be mediated by delayed preabsorptive insulin response in the LA group after the starch meal.

In summary, our results support the hypothesis that sAA CN might play a role in glucose uptake, offering novel evidence to support further research on its mechanisms of action.

The D.E.S.I.R. Study Group’s participating institutions (sites and investigators) are as follows: INSERM-U1018 (Paris: B. Balkau, P. Ducimetière, E. Eschwège); INSERM-U367 (Paris: F. Alhenc-Gelas); CHU d’Angers (A. Girault); Bichat Hospital (Paris: F. Fumeron, M. Marre, R. Roussel); CHU de Rennes (F. Bonnet); CNRS UMR-8199 (Lille: S. Cauchi, P. Froguel); Medical Examination Services (Alençon, Angers, Blois, Caen, Chartres, Chateauroux, Cholet, Le Mans, Orléans, and Tours); Research Institute for General Medicine (J. Cogneau); general practitioners of the region; and the Cross-Regional Institute for Health (C. Born, E. Caces, M. Cailleau, N. Copin, O. Lantieri, J.G. Moreau, F. Rakotozafy, J. Tichet, S. Vol).

Acknowledgments. The authors thank all participants in this study. The authors would like to express special gratitude to Phillipe Froguel and his team, particularly Dr. Loic Yengo, for their insightful discussions on the whole project and for facilitating access to the samples.

Funding. This study was supported by an intramural grant from the Qatar Biomedical Research Institute, a member of the Qatar Foundation. The D.E.S.I.R. study has been supported by INSERM contracts with CNAMTS, Lilly, Novartis Pharma, and Sanofi and by INSERM (Réseaux en Santé Publique, Interactions entre les déterminants de la santé, Cohortes Santé TGIR 2008), the Association Diabète Risque Vasculaire, the Fédération Française de Cardiologie, La Fondation de France, ALFEDIAM, ONIVINS, Société Francophone du Diabète, Ardix Medical, Bayer Diagnostics, Becton Dickinson, Cardionics, Merck Santé, Novo Nordisk, Pierre Fabre, Roche, and Topcon.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. A.A. and M.F. conceived and designed the study, helped to interpret the data, wrote the first draft of the manuscript, and contributed to the final version of the manuscript. M.S. helped to interpret the data, performed the statistical analysis and wrote that respective section of the manuscript, and contributed to the final version of the manuscript. N.C. performed the nuclear magnetic resonance data analysis and wrote that respective section of the manuscript, helped to interpret the data, and contributed to the final version of the manuscript. J.E.-S.M., B.B., and M.M. helped to interpret the data and contributed to the final version of the manuscript. J.T. conceived the D.E.S.I.R. protocol, participated in the conduct of the study over 10 years, and aided in financing the study. T.B. performed the measurement of the amylase levels. The D.E.S.I.R. Study Group provided the serum samples and the demographic data. A.A. and M.F. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.