High-Coverage Targeted Lipidomics Reveals Novel Serum Lipid Predictors and Lipid Pathway Dysregulation Antecedent to Type 2 Diabetes Onset in Normoglycemic Chinese Adults

Diabetes has become a major cause of death and disability worldwide (1). The prevalence of diabetes in China has more than quadrupled in recent decades (2,3). An estimated 10.9% of Chinese adults had diabetes and 35.7% had prediabetes in 2013 (3). Identification of biomarkers or novel pathway dysregulation predictive of diabetes denotes an area of intense research interest.

The advent of omics in recent decades offers clinicians an additional avenue to examine disease-relevant metabolite changes with an unprecedented resolution and coverage (4). Metabolomics captures changes of both endogenous and exogenous origins (5) and thus confers further insights to the intricate pathophysiology of diseases such as diabetes, for which phenotypic manifestations integrate both genetic and environmental inputs (4).

Lipidomics has emerged as an independent subfield of metabolomics, and a growing number of studies have investigated the relationship of dysregulation in lipid metabolism and pathogenesis of type 2 diabetes mellitus (T2DM) (6–16). A lipid biomarker panel comprising triacylglycerols (TAGs) of lower total carbon number and carbon double bonds was found to associate with an increased risk of diabetes in the Framingham Heart Study (FHS) cohort (9). Odd-chain TAGs, after adjustment for total TAGs, were shown to inversely correlate with diabetes risk in the Prevención con Dieta Mediterránea (PREDIMED) trial cohort (17). Recent studies have also linked lyso-phosphatidylcholine (LPC), phosphatidylcholine (PC), phosphatidylethanolamine (PE), and diacylglycerol (DAG) with increased diabetes risk (10,14,17), whereas sphingomyelins, a major class of mammalian sphingolipids, were associated with decreased risk of diabetes (11,12,17). In addition, changes in acylcarnitines (18,19) and fatty acids (13) have also been reported to associate with diabetes risk.

Most previous studies on human blood (serum/plasma) did not achieve a satisfactory lipidomic coverage to render investigation of lipid pathway dysregulation (9,11,20). A recent omics study used only one internal standard to measure 207 plasma lipids, which could considerably compromise quantitative accuracy (17). Given the immense complexities of the human serum lipidome, limited lipidomic coverage may hinder an unbiased evaluation of the subtle lipid pathway perturbations essentially underlying T2DM onset. In the current study, we used a high-coverage targeted lipidomics approach constructed principally on high-performance lipid chromatography coupled to multiple reaction monitoring (HPLC-MRM), which simultaneously confers accurate quantitation and extensive coverage of essential lipid classes central to the homeostasis of endogenous lipid metabolism (21).

Moreover, preceding lipidomics studies were predominantly conducted using a mixture of individuals with normal glucose regulation (NGR) or impaired glucose regulation (IGR) at baseline (17), making it difficult to dissect whether these identified lipid changes are harbingers of early dysglycemia or dysglycemia preceded these lipid changes (15). A comprehensive evaluation of lipidomic changes in individuals with NGR could therefore identify new markers and pathways that define early T2DM pathogenesis and possibly improve prediction of incident T2DM beyond clinical risk factors.

Herein, we extensively investigated serum lipidome changes in two prospective Chinese cohorts using a nested case-control design. Our aims were to 1) evaluate incipient lipidome patterns of diabetes existing before dysglycemia, 2) uncover lipid markers that could improve prediction of T2DM beyond clinical risk factors, and 3) systematically investigate lipid pathway dysregulation in the prodromal stage of diabetes.

The discovery cohort is a prospective, community-based cohort initiated in Jiading district, Shanghai, China, in 2010. Cohort design and characteristics were previously reported in detail elsewhere (22–23). Briefly, 10,569 subjects aged ≥40 years were invited by telephone or door-to-door visits, and 10,375 were enrolled at the baseline survey. Participants with diabetes or IGR at baseline were excluded from the analysis. After a mean follow-up of 4.4 years, among 3,821 eligible subjects with NGR at baseline, 189 developed diabetes. Owing to funding constraints, our pilot study cohort comprised 100 randomly selected case subjects with diabetes and 100 sex-matched NGR control subjects chosen using propensity score matching (PSM) (24) with a logistic model that includes age, BMI, and fasting plasma glucose (FPG).

The validation cohort comprising 55,062 individuals was randomly selected from 5 of 25 centers of the REACTION (Risk Evaluation of cAncers in Chinese diabetic Individuals: a lONgitudinal) study, a nationwide population-based prospective cohort of 259,657 individuals (aged ≥40 years) enrolled between 2011 and 2012 (25–28). The five centers included two cohorts (Jiangxi and Hubei) from the central region, and three cohorts (Guizhou, Gansu, and Sichuan) from the western region of China. Participants with diabetes or IGR at baseline were excluded from the analysis. During a follow-up of 3.8 years, among 14,651 subjects defined as NGR based on an oral glucose tolerance test (75 g) at baseline, 364 developed diabetes. Two serum samples were not available at baseline, thus 362 individuals with diabetes and 362 NGR individuals were included into the validation study using the same selection and matching methods as described above for the discovery cohort (Supplementary Fig. 1).

Furthermore, in a preliminary study to test whether the lipid predictors for diabetes were also significantly associated with the risk of developing IGR from NGR, 87 participants who developed IGR and 87 who remained NGR during follow-up were randomly selected from the Shanghai cohort.

The study protocol was approved by the Institutional Review Board of Ruijin Hospital affiliated to the Shanghai Jiao-Tong University School of Medicine. All participants provided written informed consent.

In both discovery and validation studies at baseline and follow-up visits, all participants underwent a 75-g oral glucose tolerance test, and plasma glucose was obtained at 0 h and 2 h during the test. Blood specimens were processed within 2 h of blood collection at the field center, and sera were shipped by air on dry ice to the central laboratory located at Shanghai Institute of Endocrine and Metabolic Diseases, which is certified by the College of American Pathologists (26).

Incident diabetes was defined as FPG ≥126 mg/dL, or 2-h postload plasma glucose (2hPG) ≥200 mg/dL, or self-reported previous diagnosis of diabetes by physicians and the use of antidiabetic medications. IGR was defined as FPG levels between 100 and 125 mg/dL, and 2hPG levels between 140 and 199 mg/dL in participants without prior diabetes diagnosis.

Serum lipid profiles were measured by a high-coverage targeted lipidomics approach constructed principally on HPLC-MRM. Lipids were extracted from serum (20 µL) using a modified Bligh and Dyer extraction procedure (double rounds of extraction) and dried in the SpeedVac under OH mode. All lipidomic analyses were performed on an Exion LC-system coupled with a QTRAP 6500 PLUS system (Sciex), and individual lipids from various classes were quantitated relative to their respective internal standards, as described previously (29–31). Additional details, including quantitative accuracy, calibration, and annotation, are provided in section 2 of the Supplementary Data.

Concentrations for individual lipids (mol/L) were log-transformed and standardized to z scores. Because association of lipids with diabetes risk can differ based on acyl chain length and unsaturation degree (9,17), lipids were grouped and further analyzed based on carbon atom and double bond numbers (Supplementary Data sections 1.3 and 1.4).

To identify lipids predictors, odds ratios (ORs) of developing T2DM per log-transformed SD increase in each lipid species were calculated by conditional logistic regression models, after adjusting for age, BMI, smoking and drinking status, education, family history of diabetes, physical activity, systolic blood pressure (SBP), and FPG. The P value was corrected for multiple testing via false-discovery rate (FDR) using the Benjamini-Hochberg method. Lipids that showed significant associations with incident diabetes (P value <0.05 and FDR <0.05) in both discovery and validation cohorts were further tested in a preliminary IGR study.

To identify independent predictors, serum triglycerides (TGs) and 2hPG, which showed significant differences between case subjects and control subjects at baseline, were further adjusted in conditional logistic regression models. The predictive values for incident diabetes of the identified lipids panel were evaluated in discovery cohort and validation cohort. Model performance was presented as receiver operating characteristic (ROC) curves. The Delong test was used for comparing areas under the ROC curves (Supplementary Data section 1.5).

R package MEGENA was used to build correlation networks from differentially correlated lipid pairs. Differential correlation was calculated using R package DGCA. Only lipid pairs with differential correlation (empirical P < 0.05) were included for analyses (see Supplementary Data section 1.6 for details).

All statistical analyses were performed using SAS 9.3 (SAS Institute) and R 3.4.2 software. Two-sided P < 0.05 was considered as statistically significant.

In the discovery cohort, in addition to sex, age, BMI, and FPG matched under PSM, baseline fasting HDL cholesterol, LDL cholesterol, insulin level, SBP, family history of diabetes, and lifestyle factors including smoking, drinking, physical activity, and education status in case subjects and control subjects were also well-matched. In the discovery cohort, case subjects showed higher levels of 2hPG and fasting TG (Table 1). In the validation cohort, case subjects had a significantly higher SBP and fasting TGs than control subjects (Table 1).

A total of 667 lipid species (of a targeted library screening >800 lipids) spanning 24 individual lipid classes was identified and quantitated in the serum lipidome of the discovery study (n = 200), after removing peaks without satisfactory signal-to-noise ratios (<3). Statistical analyses revealed 122 candidate lipids for segregating subjects with incident diabetes from healthy control subjects (Supplementary Data section 1.2). Based on results from the discovery cohort, we used a streamlined method comprising a subset of 250 lipids of outstanding interest to analyze the serum lipidomes of the validation cohort (n = 724) (Supplementary Fig. 2).

In multivariable logistic regression analysis, a panel of 38 lipids (FDR <0.05 and P < 0.05) emerged significant between case subjects and control subjects in the discovery and validation cohorts after adjusting for age, sex, BMI, FPG, smoking status, drinking status, education, physical activity, family history of diabetes, and SBP (Fig. 1). The panel contained 34 TAGs spanning across a comprehensive range of total carbon atom numbers (C44–C58) and unsaturation degree (n [C = C] = 0–8) that were consistently associated with an increased risk of T2DM. In addition, short-chain cholesteryl ester (CE14:0), diunsaturated PCs (PC34:3), and lyso-phosphatidylinositol (LPI)16:1 were positively associated with risk of T2DM (Fig. 1). However, polyunsaturated plasmalogen PE (i.e., PE38:4p [18:0p/20:4]) was negatively associated with risk of incident diabetes (Fig. 1).

We systematically looked for lipid patterns associated with diabetes risk that are dependent on carbon atom numbers and double bonds from the various lipid classes investigated. Statistically significant changes (P < 0.05) based on carbon atom numbers and unsaturation degree were observed only for TAGs. The discovery cohort was chosen for this aim due to its comprehensive coverage of the serum lipidome in both the carbon atom numbers and double bonds (Supplementary Fig. 2). We found that all serum TAGs examined, regardless of carbon atom numbers (C44–C60) and double bonds (n [C = C] = 0–9), were positively associated (P < 0.05) with diabetes risk (Fig. 2A, filled red shapes). After adjusting for total TAGs, however, only TAGs with low carbon atom numbers (C48–50) and low double bond numbers (n [C = C] = 2–3) were significantly (P < 0.05) associated with elevated diabetes risk (filled red shapes), while TAGs of higher carbon atom numbers (C54–60) or higher double bond numbers (n [C = C] ≥ 4) became inversely associated (P < 0.05) with diabetes risk (Fig. 2B, filled green shapes). In accordance with Razquin et al. (17), odd-chain TAGs (C53, C55) were inversely associated (P < 0.05) with diabetes (filled green shapes) risk after adjusting for total TAGs.

Next, we investigated the association of the lipid panel in Fig. 1 with the development of IGR in a preliminary study of 87 individuals who developed IGR and 87 individuals who remained NGR during the 4.4-year follow-up, randomly selected from the discovery cohort (Supplementary Fig. 1). We found that 26 of 38 identified lipids showed significant associations (P < 0.05) with the risk of IGR (Supplementary Fig. 3). Most identified TAGs, as well as CE14:0 and PC34:3, were also positively associated with risk of IGR, whereas PE38:4p (18:0p/20:4) was negatively associated with IGR (Supplementary Fig. 3).

To identify independent lipid predictors for diabetes, lipids were included one-by-one into the conditional logistic regression model with further adjustment for TG and 2hPG. Of 38 lipids, 6 reached nominal significance (P < 0.05): LPI16:1, PC34:3, PE38:4p (18:0p/20:4), TAG50:2 (16:2), TAG51:0 (17:0), and TAG54:7 (22:6) (Fig. 3A).

We then assessed the predictive value of these six identified lipids using ROC curve analyses. Matching factors were removed from the reference model. In the discovery cohort, combining our lipid panel with the reference (Ref) model increased the C statistic from 0.664 (Ref) to 0.764 (Ref + lipids) (P = 0.004) and from 0.710 (Ref + 2hPG) to 0.781 (Ref + 2hPG + lipids) (P = 0.016) (Fig. 3B and Supplementary Table 1). In the validation cohorts, when the lipids were added to the established risk prediction models of diabetes, discrimination was significantly improved. The addition of these six lipids increased the C statistic from 0.693 (Ref) to 0.717 (Ref + lipids) (P = 0.037) and from 0.698 (Ref + 2hPG) to 0.722 (Ref + 2hPG + lipids) (P = 0.028) (Fig. 3B and Supplementary Table 1). Models using penalized logistic regression yielded similar results (Supplementary Fig. 4 and Supplementary Table 2).

To systematically evaluate the perturbed lipid coregulation underlying diabetes development, we performed multiscale embedded correlation analysis based on the validation cohorts. Multiscale embedded correlation networks illustrate differential correlation between various lipids in subjects with incident diabetes relative to control subjects (Fig. 4). Three notable modules (Fig. 4, lower panel) were identified from the global networks via multiscale clustering analysis as part of MEGENA (Supplementary Data section 1.6). Module I illustrates that polyunsaturated fatty acid (PUFA)–plasmalogen phosphatidylethanolamine (PUFA-PEp; i.e., PE40:5p [22:5]) (Fig. 4, purple hub, bolded) became negatively correlated with several TAGs (orange-red) in incident diabetes relative to control (pink lines, 0/–), indicating that PUFA-PEp became negatively coregulated with TAGs specifically in incident diabetes. Module II comprises predominantly non–PUFA-TAGs (n [C = C] ≤ 3) radiating from TAG48:1 (16:1) and TAG44:1 (16:0) as central hubs, whereas module III constitutes PUFA-TAGs (n [C = C] ≥ 3) with TAG56:5 (18:2) as the central hub. Each TAG pair was connected by bright green lines (+/++), indicating that individual non–PUFA-TAGs in module II and PUFA-TAGs in module III became increasingly coregulated within the respective modules in subjects with incident diabetes relative to control subjects. Module II and module III therefore cumulatively showed that intraclass coregulation of serum TAGs became increasingly segregated on the basis of fatty acyl unsaturation in subjects with incident diabetes relative to control subjects. Taken together, differential correlation networks revealed that even before diabetes onset, perturbations in intraclass (i.e., non–PUFA-TAG and PUFA-TAGs) and interclass (i.e., TAGs and PUFA-PEps) lipid coregulation already existed.

In this prospective study, using a high-coverage targeted HPLC-MRM lipidomics approach, we presented serum lipid predictors for T2DM in a discovery cohort and replicated the findings in separate prospective cohorts of normoglycemic Chinese adults. We showed that these lipids substantially improved T2DM prediction beyond conventional clinical risk factors. Moreover, our results uncovered that intraclass coregulation among TAGs of differing carbon atom numbers and unsaturation degree and the correlation between PUFA-PEps and TAGs were altered in antecedent diabetes before the initiation of glucose perturbance. To our knowledge, this is the largest and most comprehensive lipidomics study investigating the association between lipid profiles and risk of developing diabetes in individuals with NGR at baseline.

Our study contributed a systematic evaluation of serum lipid profile changes predictive of incident diabetes in individuals with NGR. Among the 38 lipids illustrated, 34 species belong to the class of TAGs. We identified far more TAGs associated with increased risk of T2DM than other studies, including FHS (9), European Prospective Investigation into Cancer and Nutrition (EPIC) (10), and PREDIMED (17) (Supplementary Table 3), which may be due to the differences in population characteristics, analytical platform used, and variables adjusted in statistical models. Remarkably, all TAGs identified were associated with increased risk of diabetes. Our results partially corroborated the previous observations that TAGs of lower carbon atom numbers and fewer double bonds were associated with increased risk of diabetes. TAGs of higher carbon atom numbers and more double bonds were, however, found to associate with decreased risk of diabetes in the FHS cohort (9). This discrepancy could be partially explained by the differences in the study design, including only NGR at baseline in the current study versus both NGR and IGR in FHS, and other cohort epidemiological features, including ethnicity-related genetic and lifestyle variations. Interestingly, after adjustment for total measured TAGs, we observed a comparable pattern to Rhee et al. (9) that long-chain PUFA-TAGs were inversely correlated with risk of diabetes, while shorter and more saturated TAGs were associated with increased risk. These observations imply that given a constant total TAG load, compositional changes in specific subtypes of TAGs are associated with differential risk of diabetes. In particular, preceding studies (9,17) and our work have consistently reported that in both Caucasian and Chinese cohorts, long-chain polyunsaturated and/or odd-chain TAGs are inversely associated with diabetes risk, whereas short-chain and saturated TAGs are associated with elevated diabetes risk.

Our preliminary IGR study revealed that the panel of 38 lipids displayed associations with the risk of diabetes and IGR in the same direction, suggesting that our identified lipid predictors are effective in revealing incipient diabetes even before the onset of dysglycemia. It is worth mentioning that early metabolic disturbances might be already present in individuals classified as normal in terms of glycemia. Compared with preceding studies that included IGR at baseline, our reported predictors already have added pathophysiological value in diabetes prediction relative to matched normoglycemic control subjects.

A novel and possibly pathologically important observation from our correlation network analyses is the association between serum PUFA-PEps and the perturbed coregulation between PUFA-TAGs versus non–PUFA-TAGs, in subjects with incident diabetes relative to control subjects, which confer useful lipid-centric insights on the etiology of diabetes. Plasmalogen phospholipids denote essential components of mammalian cellular membranes by virtue of their distinct biophysical and biochemical properties, and previous works had shown that the levels of diacyl PEs are closely adjusted to counter fluctuations in cellular contents of PEps to maintain an overall steady level of total PEs (32,33). Synthesized in the peroxisomes, cells deficient in PEps were reported to exhibit aberrations in peroxisome assembly that drastically reduced the number of peroxisomes, albeit a causal link has yet to be elucidated (32). Peroxisomes are exceptionally dynamic organelles that adapt their number, morphology, and activity in response to physiological needs and nutritional status, which play crucial roles in fatty acid oxidation by their unique ability to consume very-long-chain fatty acids (VLCFAs) and branched-chain fatty acids (BRCFAs) (equivalent to odd-chain fatty acids in mammals) that cannot be β-oxidized by the mitochondria (34). Our observations therefore point to a possibility of attenuated peroxisomal fatty acid oxidation in incident diabetes that may be associated with reduction in PUFA-PEps, which could subsequently skew the compositional profiles of serum TAGs toward species comprising VLCFAs and BRCFAs (i.e., long-chain TAGs and odd-chain TAGs) that may account for their inverse correlation with diabetes risk compared with the remaining TAGs. Nonetheless, the precise causal links between these observations await further mechanistic elucidation, and whether reduction in PUFA-PEps and altered composition of the TAG pool denote cumulating or compensatory events of diabetes onset remains an interesting open question.

Another main clinical finding of the current study is that a selected panel of six lipids substantially improved T2DM prediction beyond that achieved using conventional risk factors in individuals classified as NGR based on an oral glucose tolerance test. We wish to emphasize the pathological relevance of our identified lipids with regard to incipient diabetes, because prediction was performed relative to normoglycemic individuals at baseline. These metabolite predictors may be particularly useful because traditional risk factors, such as blood glucose, may not serve as effective predictors of incident diabetes in such healthy individuals compared with those with IGR. Indeed, a recent study based on the FHS Offspring cohort conducted on individuals with normal fasting glucose at baseline also revealed that a metabolite panel appreciably improved T2DM risk prediction over traditional clinical factors (15).

The major strengths in our current study lie in our adopted analytical approaches as well as the well-characterized study cohorts. Our targeted lipidomics approach constructed upon HPLC-MRM, with specific quantitative internal standards catered to each lipid class, allows unambiguous identification and accurate quantitation of serum lipids to unveil subtle changes underlying prodromal disease stage. Our targeted approach is also high coverage, which extensively covers several key lipid classes essentially controlling the overall homeostatic balance of endogenous lipid metabolism, rendering possible the systematic evaluation of lipid pathway coregulation via multiscale embedded correlation analyses. In terms of the study cohorts, our discovery and multicenter-derived validation cohorts spanned six different regions of China, which are expected to better represent the general population and increase the credibility and clinical significance of the findings. It is noteworthy that in addition to our novel lipid predictors and pathway dysregulation underlying antecedent diabetes put forth in this study, numerous of our identified lipid predictors also coincided with those reported (Supplementary Table 3) in the overseas Singapore Chinese (LPI16:1) (14), the FHS (TAG44:1 [16:0], TAG44:1 [16:1], TAG48:1 [16:0], TAG48:1 [16:1], and TAG50:0 [18:0]) (9), and the PREDIMED trial (TAG) (17). Acylcarnitines did not emerge as relevant lipids associated with incident diabetes as reported in a previous study (18), which might be related to our study population being normoglycemic at baseline.

Our study, however, also has several limitations. First, the duration of follow-up is only 4 years, which limits the predictive potential of our identified lipid panel to a relatively narrow time window. Second, food consumption habits may influence the baseline levels of circulating lipids and were not controlled in the current model because diet information from participants is lacking. Third, because BMI and age were PSM for the case subjects and control subjects, discrimination by the reference model must therefore be driven by the other included risk factors (e.g., family history, TGs), and it is thus likely that the predictive value may have been affected by PSM. Fourth, we recognize that insulin resistance and insulin secretion are not evaluated in these cohorts. Fifth, all participants in this study were Chinese, and further work is needed to determine whether our findings can be extrapolated to other races and ethnicities.

In conclusion, our foregoing data underscore the potential importance of lipid metabolism early in the pathogenesis of diabetes and suggest that lipid profiles could effectively improve diabetes risk assessment in normoglycemic Chinese populations beyond that achieved by conventional clinic indices. Furthermore, we put forth a conceptual framework that perturbed peroxisomal oxidation of VLCFAs and BRCFAs associated with a reduction in PUFA-PEps may possibly give rise to an altered composition of TAGs essentially underlying the pathogenesis of dyslipidemia that precedes the development of overt diabetes.

Acknowledgments. The authors thank the staff and participants of the REACTION Study for their important contributions.

Funding. This study was supported by the Ministry of Science and Technology of China under Award Numbers 2016YFC1305600, 2016YFC1305202, 2016YFC1304904, 2017YFC1310700, 2018YFC1311705, and 2018YFC1311800; the National Natural Science Foundation of China under Award Numbers 81621061, 81670795, 81622011, and 81561128019; the National Key R&D Program of China under Award Numbers 2018YFA0506900 and 2018YFA0506902; the Shanghai Science and Technology Commission under Award Number YDZX20173100004881; and the Strategic Priority Research Program of the Chinese Academy of Sciences under Award Numbers XDB13000000 and XDA12030200.

The sponsors had no role in the design or conduct of the study.

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

Author Contributions. J.L. and S.M.L. contributed to interpreting the data and writing the manuscript. J.L., S.M.L., Y.B., G.N., and G.S. conceived and designed the study. Q.W., L.S., Y.H., L.C., and X.T. contributed to data acquisition. B.L., X.W., and K.P. performed data processing and analysis. M.L., S.W., Y.X., M.X., Y.B., and W.W. revised the manuscript. G.N., G.S., and W.W. 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.