To investigate the relationship between early second trimester serum lipidomic variation and maternal glycemic traits at 28 weeks and to identify predictive lipid biomarkers for gestational diabetes mellitus (GDM).
Prospective study of 817 pregnant women (discovery cohort, n = 200; validation cohort, n = 617) who provided an early second trimester serum sample and underwent an oral glucose tolerance test (OGTT) at 28 weeks. In the discovery cohort, lipids were measured using direct infusion mass spectrometry and correlated with OGTT results. Variable importance in projection (VIP) scores were used to identify candidate lipid biomarkers. Candidate biomarkers were measured in the validation cohort using liquid chromatography–mass spectrometry and tested for associations with OGTT results and GDM status.
Early second trimester lipidomic variation was associated with 1-h postload glucose levels but not with fasting plasma glucose levels. Of the 13 lipid species identified by VIP scores, 10 had nominally significant associations with postload glucose levels. In the validation cohort, 5 of these 10 lipids had significant associations with postload glucose levels that were independent of maternal age and BMI, i.e., TG(51.1), TG(48:1), PC(32:1), PCae(40:3), and PCae(40:4). All except the last were also associated with maternal GDM status. Together, these four lipid biomarkers had moderate ability to predict GDM (area under curve [AUC] = 0.71 ± 0.04, P = 4.85 × 10−7) and improved the prediction of GDM by age and BMI alone from AUC 0.69 to AUC 0.74.
Specific early second trimester lipid biomarkers can predict maternal GDM status independent of maternal age and BMI, potentially enhancing risk factor–based screening.
Introduction
Gestational diabetes mellitus (GDM) affects 9–26% of pregnancies (1). It is clinically important because it increases the risk of obstetric complications (e.g., pre-eclampsia and shoulder dystocia), as well as neonatal complications (e.g., hypoglycemia and hyperbilirubinemia). In the multicenter Hyperglycemia and Adverse Pregnancy Outcome (HAPO) study, a continuous linear relationship was shown between maternal glucose levels at 24–32 weeks of gestation and the odds of several adverse pregnancy outcomes, even for glucose levels well within the normal range (2). This has provided impetus for the International Association of the Diabetes and Pregnancy Study Groups (IADPSG) to recommend universal screening for GDM via the 75-g oral glucose tolerance test (OGTT) between 24 and 28 weeks gestation (3).
This recommendation by the IADPSG has not been uniformly adopted because of various concerns, including the implications for service provision and the large number of normoglycemic women who will have to undergo what can be an unpleasant and poorly tolerated test (4). Some countries, such as the U.K. (4) and Italy (5), perform risk factor–based screening, with only high-risk individuals undergoing the diagnostic 75-g OGTTs. Specified risk factors include obesity, previous macrosomic baby, previous GDM, family history of diabetes, and minority ethnic family origin with high prevalence of diabetes. In the U.S., the American Diabetes Association and American Congress of Obstetricians and Gynecologists continue to endorse a two-step approach with an initial universal nonfasted 50-g glucose load test at 24–28 weeks of gestation (6).
Another recognized limitation of screening for GDM at 24–28 weeks of gestation is the delay in detecting cases of GDM that developed in the first or early second trimesters. By the time of screening, significant increased fetal adiposity may have developed (4,7,8). Although performed in a high-risk ethnic minority population, Agarwal et al. (9) showed that >40% of cases of GDM could be diagnosed by a 75-g OGTT performed before 18 weeks of gestation. These limitations of universal 24- to 28-week OGTT support the potential value of predictive early gestational biomarkers for GDM. The biomarkers that have been studied include fasting plasma glucose (9) and, more recently, maternal metabolites (10,11), plasma proteins (12), and microRNA (13) using high-throughput technologies. Although these efforts have not yet informed clinical strategies, the results of metabolomic studies in particular have provided valuable insights into the pathophysiology of GDM (11,14).
Variations in lipid profiles have yet to be comprehensively studied, even though changes in maternal lipid metabolism are well described from the beginning of pregnancy (15). In early pregnancy, levels of plasma lipids, including triglycerides (TGs), phospholipids, and cholesterol, decrease before steadily increasing from week 8 onward (15). The rise in TGs is accompanied by an increase in VLDL, LDL, and HDL levels (16). Changes in lipid levels during GDM have also been observed, with women who have GDM having higher serum TG levels but lower LDL levels compared with normoglycemic pregnant women (17).
Therefore, we hypothesized that lipidomic variation in early second trimester maternal serum samples could be associated with later glucose tolerance measured at 28 weeks. Confirmation of this hypothesis would enable the identification of candidate lipid biomarkers that are predictive of GDM, so as to improve GDM screening and provide mechanistic insights into the pathophysiology of GDM.
Research Design and Methods
Recruitment and Sample Collection
The Cambridge Baby Growth Study (CBGS) is a prospective longitudinal study that has been described previously (18–20). Briefly, 2,212 women in early pregnancy were recruited between 2001 and 2009 from ultrasound clinics at the Rosie Maternal Hospital, Cambridge, U.K. Such dating scans are routinely offered to all pregnant women receiving antenatal care and are performed at 8–14 weeks of gestation.
Shortly after recruitment, at 15.2 ± 0.07 weeks of gestation, a nonfasting venous blood sample was collected if women consented. After clotting and within 2 h of sample collection, these samples were centrifuged at 3,000g for 10 min, and the serum was separated and stored at −80°C. They were maintained at −80°C until analysis, with the exception of a single freeze-thaw cycle to prepare the necessary aliquots for lipid analysis. A total of 1,260 serum samples were collected.
All participants were also invited to undergo a standard 75-g OGTT, which was performed at 28 weeks of gestation after an overnight fast. A total of 1,069 women underwent the OGTT. Plasma glucose levels were analyzed by the standard glucose oxidase method.
Cohort Selection
For this study, we excluded OGTT participants who 1) were missing fasting or 1-h postload venous plasma glucose level measurements (n = 10), 2) subsequently gave birth to twins (n = 17), or 3) did not provide an early second trimester serum sample (n = 219). A very small number of participants (n = 6) were also excluded for various other reasons (e.g., inadequate remaining serum samples, no paired DNA sample for use in other studies). This yielded a total of 817 women, who were assigned to a discovery cohort of 200 women and a validation cohort of 617 women. Women in the discovery cohort were selected because they had data on other genetic or phenotypic traits, which other ongoing studies in our group were interested in correlating with lipidomic variation.
There were two differences in clinical characteristics between the discovery and validation cohorts. First, 1-h postload glucose levels were 0.27 mmol/L higher in the validation cohort (Table 1). This result was of borderline significance on univariate testing (P = 0.05) and was nonsignificant when multiple testing was accounted for using the Benjamini-Hochberg method (P = 0.175). Importantly, there was no significant difference in the proportion of case patients with GDM in the two groups. Second, samples were obtained in the validation cohort at a slightly later point in gestation, ∼0.5 weeks later.
Clinical characteristics of participants in the lipidomics study
Clinical characteristics . | All OGTT . | Lipidomics total . | Discovery cohort . | Validation cohort . | P value (discovery vs. validation cohort) . | ||||
---|---|---|---|---|---|---|---|---|---|
Mean . | SD . | Mean . | SD . | Mean . | SD . | Mean . | SD . | ||
Total participants (n) | 1,069 | 817 | 200 | 617 | |||||
Age (years) | 33.39 | 4.19 | 33.27 | 4.11 | 33.38 | 3.83 | 33.22 | 4.21 | 0.65 |
Maternal prepregnancy BMI (kg/m2) | 24.10 | 4.46 | 24.19 | 4.48 | 23.90 | 4.13 | 24.30 | 4.60 | 0.31 |
Maternal height (m) | 1.66 | 0.07 | 1.66 | 0.07 | 1.66 | 0.07 | 1.66 | 0.07 | 0.22 |
Fasting glucose levels (mmol/L) | 4.33 | 0.55 | 4.33 | 0.49 | 4.33 | 0.53 | 4.32 | 0.47 | 0.93 |
1-h postload glucose levels (mmol/L) | 6.83 | 1.72 | 6.77 | 1.67 | 6.57 | 1.59 | 6.84 | 1.69 | 0.05 |
GDM (%) | 9.45 | 8.20 | 6.50 | 8.75 | 0.38 | ||||
Gestational age at serum sample collection (years) | 15.2 | 2.46 | 15.0 | 2.02 | 14.6 | 1.73 | 15.1 | 2.10 | 0.004 |
Clinical characteristics . | All OGTT . | Lipidomics total . | Discovery cohort . | Validation cohort . | P value (discovery vs. validation cohort) . | ||||
---|---|---|---|---|---|---|---|---|---|
Mean . | SD . | Mean . | SD . | Mean . | SD . | Mean . | SD . | ||
Total participants (n) | 1,069 | 817 | 200 | 617 | |||||
Age (years) | 33.39 | 4.19 | 33.27 | 4.11 | 33.38 | 3.83 | 33.22 | 4.21 | 0.65 |
Maternal prepregnancy BMI (kg/m2) | 24.10 | 4.46 | 24.19 | 4.48 | 23.90 | 4.13 | 24.30 | 4.60 | 0.31 |
Maternal height (m) | 1.66 | 0.07 | 1.66 | 0.07 | 1.66 | 0.07 | 1.66 | 0.07 | 0.22 |
Fasting glucose levels (mmol/L) | 4.33 | 0.55 | 4.33 | 0.49 | 4.33 | 0.53 | 4.32 | 0.47 | 0.93 |
1-h postload glucose levels (mmol/L) | 6.83 | 1.72 | 6.77 | 1.67 | 6.57 | 1.59 | 6.84 | 1.69 | 0.05 |
GDM (%) | 9.45 | 8.20 | 6.50 | 8.75 | 0.38 | ||||
Gestational age at serum sample collection (years) | 15.2 | 2.46 | 15.0 | 2.02 | 14.6 | 1.73 | 15.1 | 2.10 | 0.004 |
Fasting and postload glucose levels were measured using a standard OGTT at 28 weeks of pregnancy. GDM was diagnosed based on fasting and 1-h postload venous plasma glucose measurements (details in text). Comparisons were performed using Student t test and χ2 test for the proportion of participants with GDM. P values were not corrected for multiple testing.
Ethical Approval
The study protocol was approved by the local research ethics committee, Addenbrooke’s Hospital, Cambridge, U.K. Written informed consent was obtained from all participants.
Lipid Biomarker Analyses
Statistical Analysis
Partial least squares (PLS) regression and PLS-discriminant analysis (DA) were used to identify associations between lipidomics variables and OGTT results in the discovery cohort. Fitted models were considered significant if the value of Q2 (i.e., R2 of the model as estimated by cross-validation) was positive. The importance of individual lipid species was quantified via the variable importance in projection (VIP) score and used to identify candidate lipid biomarkers. The VIP score is a widely used method of variable selection. It takes into account the amount of Y variance explained by the projection and the loadings of each variable on this projection, while adjusting for the absolute magnitude of each X variable. As such, two variables with identical contributions to the explanatory power of the model will have identical VIP scores, regardless of which component they have a large influence on or their absolute magnitudes.
Standard linear and logistic regression techniques were used to assess the association between candidate lipid biomarkers and maternal OGTT results or GDM status. GDM was defined based on fasting and 1-h postload glucose levels using IADPSG thresholds (i.e., ≥5.1 and 10.0 mmol/L, respectively). The 2-h postload data were unavailable for most women and were omitted from our case definition for uniformity. This is acceptable because only 7% of U.K. women with GDM receive a diagnosis based on the 2-h measurement alone (1).
Logistic regression was used to combine the predictive ability of candidate lipid biomarkers, which was then assessed using receiver operating characteristic (ROC) plots. Where backward stepwise selection was used, a significance threshold of 0.10 for removal was used. Linear DA was used to ensure the robustness of these results.
The threshold for statistical significance was 0.05. For the discovery cohort, uncorrected P values were considered, whereas in the validation cohort P values were corrected for multiple testing using the Benjamini-Hochberg method. Values in the text are given as the mean ± SE, unless otherwise specified. Regression coefficients were standardized by the predictor (i.e., change in response variable for each SD increase in the predictor).
PLS and PLS-DA regressions were performed using SIMCA version 14 (MKS Umetrics AB, Umeå, Sweden). All other analyses were performed using SPSS version 21 (IBM, Armonk, NY).
Results
Association Between Early Second Trimester Lipidomic Variation and 28-Week Glucose Tolerance
Analysis of the discovery cohort was confined to 196 samples because 4 samples were found to be unsuitable for analysis due to hemolysis. In these samples, 189 lipid species were detected.
A PLS model was constructed to examine the extent to which early second trimester lipidomic variation explained postload venous glucose levels during an OGTT at 28 weeks of gestation. The resulting model yielded one fitted component, which used 15% of lipidomic variation (R2X) to explain 11% (R2Y) of variation in postload glucose levels. This was robust to internal cross-validation, yielding a Q2 value of 4%.
Because some lipid species may show nonlinear relationships with postload glucose levels, we divided participants into tertiles of OGTT levels and constructed a PLS-DA model to explain membership in the top tertile. The resulting model also yielded a single component, which used 15% of lipidomic variation to explain 9% of the variation in top tertile membership, with a Q2 value of 1.31%.
Similar PLS and PLS-DA models were also constructed to explore the relationship between lipidomic variation and fasting plasma glucose levels. However, the models were overfitted, with the PLS and PLS-DA models yielding Q2 values of −3% (R2X = 15%, R2Y = 9%) and −10% (R2X = 9%, R2Y = 8%), respectively. These models were not used in subsequent analyses.
Identification of Lipid Biomarkers of Postload Plasma Glucose Levels
From each of the two models considering postload glucose levels, ∼10 lipids with the highest VIP scores were selected, with the exact cut point selected using a graphic method (Fig. 1A and B). The PLS model yielded 9 lipid species, and the PLS-DA model yielded 10 lipid species, with 5 lipid species being identified in both models. One of the species annotated as diglyceride-water (DG-H2O[32:0]) was likely to be an in-source fragment of a different lipid species. As in-source fragments are artifacts of mass spectrometry, this species was disregarded, leaving a total of 13 lipid species identified.
Candidate lipid biomarkers identified from the discovery cohort. A and B: VIP scores of individual lipid species in a PLS model mapping early second trimester lipid profiles to late second trimester 1-h postload glucose levels. B: VIP scores of individual lipid species in a PLS-DA model mapping early second trimester lipid profiles to membership in the top tertile of late second trimester 1-h postload glucose levels. C: Correlation between 10 candidate lipid biomarkers taken forward to validation cohort. Underlined lipid species refer to species that had nominally significant associations with 1-h postload glucose levels (A) or the top tertile thereof (B) and were taken forward to the validation cohort.
Candidate lipid biomarkers identified from the discovery cohort. A and B: VIP scores of individual lipid species in a PLS model mapping early second trimester lipid profiles to late second trimester 1-h postload glucose levels. B: VIP scores of individual lipid species in a PLS-DA model mapping early second trimester lipid profiles to membership in the top tertile of late second trimester 1-h postload glucose levels. C: Correlation between 10 candidate lipid biomarkers taken forward to validation cohort. Underlined lipid species refer to species that had nominally significant associations with 1-h postload glucose levels (A) or the top tertile thereof (B) and were taken forward to the validation cohort.
The 13 lipid species were regressed against postload glucose levels or membership in the top tertile thereof. To ensure that they were not simply surrogates for known risk factors of GDM, maternal age and prepregnancy BMI were adjusted for. This resulted in 10 of the 13 lipid species having a nominally significant association with postload glucose levels and/or the top tertile thereof (Supplementary Table 1). The two lipid species showing the strongest association with postload glucose levels were TG(51:1) (0.40 mmol/L per SD increase, P = 8.88E-4) and the choline ether phospholipid (PCae) PCae(40:3) (−0.41 mmol/L per SD increase, P = 9.73E-4) (Supplementary Table 2).
Further examination of the correlations among these 10 lipid species revealed clustering into two large groups (Fig. 1C). The first group contained the PCaes, whereas the second group contained TGs and a phosphatidylcholine (PC).
Validation of Lipid Biomarkers of Postload Plasma Glucose Levels
These 10 candidate lipid biomarkers were measured in the validation set of 617 subjects using an LC-MS method. This provided additional chromatographic information, eliminating any possible artifacts introduced by the shotgun approach in the discovery set. With this method, we were unable to detect PCae(44:4) in our samples, suggesting that this putative signal in the discovery set was an artifact, and it was omitted from further analysis. We also omitted 20 samples in which none of the remaining nine lipid species were detectable, as this was likely indicative of poor sample quality. This left data on the nine candidate lipid biomarkers in 597 subjects for analysis.
Of the nine remaining lipid species, five showed significant associations with postload glucose levels, even after adjustment for maternal age and BMI and correction for multiple testing (Table 2). TG(51:1), TG(48:1), and PC(32:1) were positively associated with maternal postload glucose levels, whereas PCae(40:3) and PCae(40:4) were inversely associated with maternal postload glucose levels. Logistic regression against membership of the top tertile of postload glucose levels did not validate any additional candidate lipid species, nor did adjusting for gestational age at the time of serum sample collection (data not shown).
Relationship between candidate lipid biomarkers and second trimester 1-h postload glucose levels in the validation cohort
Lipid species . | Univariate analysis . | Adjusted for maternal age and BMI . | ||||
---|---|---|---|---|---|---|
Regression coefficient (mmol/L/SD) . | P value . | BH P value . | Regression coefficient (mmol/L/SD) . | P value . | BH P value . | |
TG(51:1) | 0.18 | 9.15E-03 | 0.02 | 0.18 | 0.03 | 0.05 |
TG(48:1) | 0.27 | 1.00E-04 | 3.82E-04 | 0.22 | 7.40E-03 | 0.02 |
TG(50:1) | 0.07 | 0.31 | 0.40 | −0.06 | 0.45 | 0.50 |
PC(32:1) | 0.33 | 1.45E-06 | 1.31E-05 | 0.21 | 8.94E-03 | 0.02 |
PCae(38:4) | −0.06 | 0.39 | 0.44 | −0.15 | 0.06 | 0.09 |
PCae(44:6) | 0.02 | 0.80 | 0.80 | −0.03 | 0.69 | 0.69 |
PCae(40:3) | −0.26 | 1.70E-04 | 3.82E-04 | −0.24 | 3.63E-03 | 0.02 |
PCae(40:5) | −0.16 | 0.02 | 0.03 | −0.14 | 0.08 | 0.10 |
PCae(40:4) | −0.27 | 1.31E-04 | 3.82E-04 | −0.29 | 2.60E-04 | 2.34E-03 |
Lipid species . | Univariate analysis . | Adjusted for maternal age and BMI . | ||||
---|---|---|---|---|---|---|
Regression coefficient (mmol/L/SD) . | P value . | BH P value . | Regression coefficient (mmol/L/SD) . | P value . | BH P value . | |
TG(51:1) | 0.18 | 9.15E-03 | 0.02 | 0.18 | 0.03 | 0.05 |
TG(48:1) | 0.27 | 1.00E-04 | 3.82E-04 | 0.22 | 7.40E-03 | 0.02 |
TG(50:1) | 0.07 | 0.31 | 0.40 | −0.06 | 0.45 | 0.50 |
PC(32:1) | 0.33 | 1.45E-06 | 1.31E-05 | 0.21 | 8.94E-03 | 0.02 |
PCae(38:4) | −0.06 | 0.39 | 0.44 | −0.15 | 0.06 | 0.09 |
PCae(44:6) | 0.02 | 0.80 | 0.80 | −0.03 | 0.69 | 0.69 |
PCae(40:3) | −0.26 | 1.70E-04 | 3.82E-04 | −0.24 | 3.63E-03 | 0.02 |
PCae(40:5) | −0.16 | 0.02 | 0.03 | −0.14 | 0.08 | 0.10 |
PCae(40:4) | −0.27 | 1.31E-04 | 3.82E-04 | −0.29 | 2.60E-04 | 2.34E-03 |
BH, Benjamini-Hochberg corrected. P values in bold type indicate a statistically significant result (P < 0.05).
Because fasting and postload glucose levels have common pathophysiological determinants and were moderately correlated in our cohort (r = 0.343, P < 0.001), we tested the nine putative lipid biomarkers for an association with fasting glucose levels (data not shown). TG(51:1) and PCae(40:4) showed significant associations with fasting glucose levels, even after adjusting for multiple testing. However, after adjusting for maternal age and BMI, only TG(51:1) remained significant (0.06 mmol/L per SD increase, P = 0.003; Benjamini-Hochberg test, P = 0.02).
Lipid Predictors of GDM
Of the 597 subjects in the validation cohort, 53 met the criteria for GDM. The five validated lipid species were tested for association with GDM. TG(51:1), TG(48:1), PC(32:1), and PCae(40:4) were significantly associated with GDM, even after adjustment for maternal age and maternal BMI and correction for multiple testing (Fig. 2A). While not reaching significance, PCae(40:3) nonetheless demonstrated a strong trend in the same direction as PCae(40:4) (Benjamini-Hochberg test, P value = 0.07).
Validated lipid biomarkers and GDM prediction within the validation cohort. A: Individual predictive power of each lipid, independent of maternal age and BMI. *P < 0.05, with Benjamini-Hochberg correction for multiple testing. B: Combined predictive power of four lipid species. C: Enhancement of predictive power of conventional risk factors.
Validated lipid biomarkers and GDM prediction within the validation cohort. A: Individual predictive power of each lipid, independent of maternal age and BMI. *P < 0.05, with Benjamini-Hochberg correction for multiple testing. B: Combined predictive power of four lipid species. C: Enhancement of predictive power of conventional risk factors.
To assess the combined predictive ability of these four lipid species, logistic regression was used to calculate the probability of the GDM status of each subject and the probability scores used to construct a ROC curve (Fig. 2B). This yielded a mean area under the curve (AUC) of 0.709 ± 0.040 (P = 4.85E-7). Similar results were obtained using linear DA.
Of the 597 subjects, 410 (including 37 GDM case patients) had available data on maternal age and BMI. This enabled us to assess the additional predictive power conferred by these four lipid biomarkers over maternal age and BMI alone. In this subgroup, maternal age and BMI produced a mean AUC of 0.689 ± 0.046 (P = 1.54E-4), and further inclusion of the four lipid biomarkers increased the AUC to 0.741 ± 0.045 (P = 1.33E-6) (Fig. 2C). Graphically, the improvement in AUC was most marked at stringent thresholds (i.e., enhancing the sensitivity at high levels of specificity). For instance, at 91.7% specificity, the sensitivity is 21.6% based on traditional risk factors but rose to 48.6% when lipid predictors were included. In addition, there was some improvement at high levels of sensitivity, for instance at 97.3% sensitivity, where the inclusion of lipid predictors raised specificity from 9.9%, based on traditional risk factors alone, to 24.9%.
Finally, we sought to identify the most parsimonious model from these six potential predictive variables. Using a backward stepwise selection algorithm, the only terms left in the model were maternal BMI, TG(48:1), and PCae(40:4). This yielded a mean AUC of 0.732 ± 0.045 (P = 3.28E-6), which is similar to the model including all six predictors.
Conclusions
In this unselected cohort of predominantly Caucasian pregnant women, using an unbiased lipidomics approach and a preassigned validation cohort, we show for the first time that specific lipid species in the maternal early second trimester lipid profile are associated with maternal glycemic traits assessed by standard 75-g oral glucose tolerance testing at 28 weeks. We identified four lipid biomarkers, TG(51:1), TG(48:1), PC(32:1), and PCae(40:4), that predict later onset of GDM, independent of maternal age and BMI, and could potentially enhance the performance of existing risk factor–based screening approaches used in many countries.
The performance of clinical risk factor–based screening has been examined in many different populations. A recent study (24) in an Australian population compared the performance of the National Institute for Health and Care Excellence, the American Diabetes Association, and the Australasian Diabetes in Pregnancy Society risk factor–based screening guidelines, yielding sensitivities of 92%, 100%, and 99% and specificities of 32.4%, 3.9%, and 13.7%, respectively. This and other studies reveal limited test performance with the need for low levels of specificity to achieve the high levels of sensitivity.
The four lipid biomarkers that we identified have moderate predictive performance, with an estimated AUC of 0.709. This is comparable to other early pregnancy biomarkers, including conventional biomarker fasting levels of plasma glucose (estimated AUC = 0.579) (9), HbA1c (AUC in high risk population = 0.67) (25), TGs (AUC = 0.55–0.61), and TG-to-HDL ratio (AUC = 0.62) (26), as well as novel biomarkers, such as second trimester serum microRNA (AUC = 0.669) (13). Furthermore, the lipid biomarkers that we derived were specifically identified to predict GDM independent of maternal age and BMI, and thus can enhance the predictive performance of existing risk factor–based approaches. Indeed, the inclusion of the four lipid biomarkers with maternal age and BMI increased the AUC from 0.689 to 0.741.
The enhancement of sensitivity at high levels of specificity was particularly marked. Although for the purpose of predicting GDM our data on the clinical performance of lipid biomarkers must be considered very preliminary, lipid biomarkers may potentially have a role in identifying high-risk women who should undergo an immediate/early second trimester OGTT.
The lipid biomarkers we identified can be divided into two groups. TG(51:1), TG(48:1), and PC(32:1) are associated with increased postload glucose levels and GDM risk and are moderately correlated. PCae(40:3) and PCae(40:4) are associated with decreased postload glucose levels and/or GDM risk and are strongly correlated.
The association of TG(51:1), TG(48:1), and PC(32:1) with maternal glucose levels is consistent with previous investigations into lipidomic changes associated with type 2 diabetes. TG(48:1) has been implicated with type 2 diabetes risk in the Framingham cohort (27), and PC(32:1) levels were raised in Australian Diabetes, Obesity and Lifestyle Study (AusDiab) subjects with type 2 diabetes (28). These three lipid species are notable for the presence of a single double bond, which implies the presence of a monounsaturated fatty acid, predominantly palmitoleate and, to a lesser extent, oleate, on closer inspection of LC-MS spectra in our analysis. In one study (29), palmitoleate content within circulating phospholipids was found to be associated with increased insulin resistance. Because circulating palmitoleate is principally synthesized in the liver in humans, this may reflect hepatic insulin resistance, in line with our finding that TG(51:1) is associated with fasting plasma glucose levels (30). Mechanistically, palmitoleate and oleate are produced from palmitate and stearate by the action of steroyl-CoA desaturase 1 (SCD1), and SCD1 activity has recently been linked in a large cohort to type 2 diabetes risk and hepatic steatosis (31). Indeed, SCD1−/− mice display increased insulin sensitivity (32).
The association of TG(51:1), primarily comprising TG(18:1/17:0/16:0) (or positional isomers thereof) in our cohort, with maternal glucose levels was also surprising because odd-chain fatty acids, including heptadecanoic acid, have been associated with reduced risk of type 2 diabetes (28,33,34). However, our finding is in keeping with results of an untargeted metabolomic screen using fasted serum samples from women at 28 weeks of gestation who were enrolled in the HAPO study (35), in which heptadecanoic acid was raised in subjects with fasting plasma glucose levels in the 90th percentile but with BMI values similar to those of control subjects. This may be due to genuine differences in the pathophysiology of GDM and type 2 diabetes but may also reflect the fact that the studies of type 2 diabetes measured the fatty acid content in phospholipids, whereas the latter study (35) of pregnant women measured free fatty acids. This underscores the advantage of intact lipid studies as opposed to fatty acid profiling (30), which is a strength of our study.
The other group of lipid biomarkers identified, PCae(40:3) and PCae(40:4), were inversely associated with maternal glucose levels. This is consistent with an earlier report from the AusDiab cohort (28) in which ether phospholipid levels were inversely related to postload glucose levels and were reduced in patients with diabetes. The physiological function of ether phospholipids remains largely unknown (36), but they have been implicated as physiological ligands of peroxisome proliferator–activated receptor γ (37). Intriguingly, SCD1 is a target of peroxisome proliferator–activated receptor γ, potentially providing a mechanistic link between low levels of ether phospholipids and high levels of palmitoleate- and oleate-containing lipids (32).
The broad overlap between our lipid biomarkers and those identified in studies of type 2 diabetes may be a result of the fact that lipid profiles were derived from samples obtained early in pregnancy, reflecting the contribution of preexisting insulin resistance to the development of GDM. Indeed, it will be ideal to obtain a sample before conception as well as one obtained during pregnancy to identify lipid biomarkers that reflect the pathophysiological contribution of pregnancy itself. Nevertheless, this explanation of our findings is made less likely by the fact that genome-wide association studies have revealed a broadly shared genetic architecture between GDM and type 2 diabetes and metabolomic studies from later in pregnancy, including one using samples from 28 weeks of gestation, have shown overlapping metabolic signatures between GDM and type 2 diabetes (11,14,35).
There are several limitations to our study. First, although we validated our candidate lipid biomarkers using a predefined subset, these biomarkers have not been externally validated, for example in ethnic populations at high risk. Second, our study was not designed to demonstrate the superiority of a lipid biomarker–based and risk factor–based approach compared with a risk factor–based screening alone. Thus, we lacked data on other conventional risk factors (e.g., family history of diabetes and personal history of GDM). For similar reasons, we also lacked data on other traditional biochemical risk factors, such as HbA1c, TGs, and HDL cholesterol, and are thus unable to directly compare the performance of the lipid biomarkers to these alternatives within this study. Third, because we selected only 13 lipids from the 189 lipids measured for univariate analysis, we might have been overly conservative in our approach. Finally, because serum samples for lipidomic analysis were obtained in the nonfasting state without controlling for meal time and meal content, this would have added additional lipidomic variability that was not related to variation in OGTT results at 28 weeks of gestation, thus reducing study power. However, this additional variability might be small compared with existing intersubject and intrasubject variation (38). Our reanalysis of data from Begum et al. (38) suggests that variance is partitioned among intersubject differences, intrasubject differences not due to meal time, and the effect of meal time in the proportion 62%, 31%, and 7%, although the population that they studied was less heterogeneous than that in the CBGS.
In summary, we report for the first time an association between maternal early second trimester lipid species and glycemic traits at 28 weeks of gestation, as assessed by a standard OGTT. We further show that four lipid biomarkers, TG(51:1), TG(48:1), PC(32:1), and PCae(40:4), are able to predict maternal GDM status independent of maternal age and BMI and have the potential to improve the performance of clinical risk factor–based screening. The lipid biomarkers identified also revealed marked similarities between the pathophysiology of GDM and that of type 2 diabetes (14). In particular, we highlight the established role of monounsaturated fatty acids (especially palmitoleate) and the emerging role of ether phospholipids, as well as the potential pathological role of odd-chain fatty acids, which might indicate a divergence in the pathophysiologies of GDM and type 2 diabetes.
Article Information
Acknowledgments. The authors thank Dianne Wingate, Rachel Seear, Katrin Mooslehner, and Radka Platte (University of Cambridge, Cambridge, U.K.) for excellent laboratory assistance. The authors also thank the families who participated in this study; the research nurses, including Suzanne Smith, Ann-Marie Wardell, and Karen Forbes (University of Cambridge, Cambridge, U.K.), for their important contributions; colleagues at the Addenbrooke's Wellcome Trust Clinical Research Facility; and the midwives at the Rosie Maternity Hospital for data collection.
Funding. This part of the Cambridge Baby Growth Study was funded by Wellbeing of Women (grant RG1644) and Diabetes UK (grant 11/0004241). The lipidomics assays were supported by the Medical Research Council (UD99999906) and Cambridge Lipidomics Biomarker Research Initiative (grant G0800783). Core funding was also obtained through the Medical Research Council (grants 7500001180 and UD99999906), European Union Framework 5 (grant QLK4-1999-01422), World Cancer Research Fund (2004/03), Mothercare Foundation (grant RG54608), and the Newlife Foundation for Disabled Children (grant 07/20). There has also been support from National Institute for Health Research Cambridge Biomedical Research Centre.
Duality of Interest. No potential conflicts of interest relevant to this article were reported.
Author Contributions. L.L. designed the study, interpreted and analyzed the data, and drafted and edited the manuscript. A.K. designed the study, collected and interpreted the data, and edited the manuscript. C.J.P. and D.B.D. designed the study, interpreted the data, and edited the manuscript. B.J. and L.M. collected the data. I.A.H., C.L.A., and K.K.O. contributed to discussion and reviewed the manuscript. L.L. and D.B.D. 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.