Article pubs.acs.org/jpr
Prediction of Gestational Diabetes through NMR Metabolomics of Maternal Blood Joana Pinto,† Lara M. Almeida,† Ana S. Martins,† Daniela Duarte,† António S. Barros,‡ Eulália Galhano,§ Cristina Pita,§ Maria do Céu Almeida,§ Isabel M. Carreira,∥ and Ana M. Gil*,† †
CICECO - Aveiro Institute of Materials, Department of Chemistry, University of Aveiro, 3810-193 Aveiro, Portugal QOPNA Research Unit, Department of Chemistry, University of Aveiro, 3810-193 Aveiro, Portugal § Maternidade Bissaya Barreto, Centro Hospitalar e Universitário de Coimbra − CHUC, 3000 Coimbra, Portugal ∥ Cytogenetics and Genomics Laboratory, Faculty of Medicine, and CNC.IBILI, University of Coimbra, Portugal and CIMAGO Center for Research in Environment, Genetics and Oncobiology, 3000 Coimbra, Portugal ‡
S Supporting Information *
ABSTRACT: Metabolic biomarkers of pre- and postdiagnosis gestational diabetes mellitus (GDM) were sought, using nuclear magnetic resonance (NMR) metabolomics of maternal plasma and corresponding lipid extracts. Metabolite differences between controls and disease were identified through multivariate analysis of variable selected 1H NMR spectra. For postdiagnosis GDM, partial least squares regression identified metabolites with higher dependence on normal gestational age evolution. Variable selection of NMR spectra produced good classification models for both pre- and postdiagnostic GDM. Prediagnosis GDM was accompanied by cholesterol increase and minor increases in lipoproteins (plasma), fatty acids, and triglycerides (extracts). Small metabolite changes comprised variations in glucose (up regulated), amino acids, betaine, urea, creatine, and metabolites related to gut microflora. Most changes were enhanced upon GDM diagnosis, in addition to newly observed changes in low-Mw compounds. GDM prediction seems possible exploiting multivariate profile changes rather than a set of univariate changes. Postdiagnosis GDM is successfully classified using a 26-resonance plasma biomarker. Plasma and extracts display comparable classification performance, the former enabling direct and more rapid analysis. Results and putative biochemical hypotheses require further confirmation in larger cohorts of distinct ethnicities. KEYWORDS: pregnancy, gestational diabetes mellitus (GDM), prediagnosis GDM, NMR, metabolomics, maternal plasma, lipid extracts
■
INTRODUCTION Gestational diabetes mellitus (GDM) consists of carbohydrate intolerance occurring with onset or first recognition during pregnancy.1 This condition affects 17% of live births worldwide2 and 2−6% in Europe,3 and, if untreated, it translates into significant risks of preeclampsia, macrosomia, fetal congenital anomalies, intrauterine fetal death, neonatal hypoglycaemia, and neonatal hyperbilirubinemia.4 In addition, women with a GDM history and offsprings of GDM pregnancies are known to have increased risk of developing type 2 diabetes later in life, along with an enhanced risk of obesity for children.4 GDM is routinely diagnosed by (a) measurement of fasting plasma glucose performed in the first prenatal medical appointment and, in the case of a negative result, and (b) the 75 or 100 g OGTT performed at 24−28 gestational weeks (g.w.).5 In 70− 85% of GDM cases, the condition is effectively managed by controlled diet and physical exercise, whereas the remaining subjects usually require insulin therapy.4 Given the important risks involved with GDM pregnancies and infants, there is still © XXXX American Chemical Society
scope for devising earlier and more complete biomarkers of the condition onset and development. In recent years, metabolomic studies have searched for metabolic biomarkers of GDM by profiling maternal biofluids (blood serum/plasma and urine) and fetal/infant samples (amniotic fluid, umbilical cord blood, newborn meconimum, and urine), to obtain fuller descriptions of GDM impact on maternal and fetal metabolisms, as reviewed recently.6 Some of these studies have sought biomarkers detectable prior to clinical diagnosis so that women at risk of developing GDM may be identified in time to allow improved disease management. A large number of studies may be found in relation to postdiagnosis cases, an initial GC−MS study revealing an increase in maternal serum total fatty acids (FAs) and several specific FAs (e.g., linoleic, arachidonic, docohexaenoic) from controls to slight hyperglycemia and to third trimester GDM, with differences in palmitoleic and docosahexReceived: March 26, 2015
A
DOI: 10.1021/acs.jproteome.5b00260 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Table 1. List of Maternal Plasma Samples and Lipid Extracts Obtained for Each Group of Subjects, with Corresponding Number of Samples (n); Maternal Age (years); Gestational Age (weeks) at Collection and, for the Pre-Diagnosis GDM Group, from Collection to Diagnosis; Gestational Age (weeks) at Delivery; Baby Weight at Birth (g); and Pre-Pregnancy Body Mass Index (BMI, kg·m−2)a n
maternal age/ years
g.w. at collection/ weeks
controls
49
25−42 (36)
16−24 (17)
prediagn. GDMa
32
30−44 (38)
16−21 (17)
postdiagn. GDMb
12
18−41 (33)
24−27 (26)
controls
15
28−42 (37)
16−22 (17)
prediagn. GDMc
14
36−42 (39)
16−19 (17)
postdiagn. GDMb
12
18−41 (33)
24−27 (26)
subjects group
g.w. to diagnosis/ weeks
g.w. at delivery/ weeks
baby weight (g)
prepregnancy BMI/kg· m‑2
Whole Blood Plasma 37−41 (39) 2−21 (13)
35−40 (38) 34−40 (38)
2405−3960 (3222) 2125−3805 (3115) 1745−3505 (2736)
20−33 (24) [for n = 37] 18−35 (26) [for n = 24] 18−36 (26) [for n = 12]
Plasma Lipid Extracts 38−40 (39) 2−19 (12)
36−40 (38) 34−40 (38)
2560−3960 (3288) 2310−3805 (3109) 1745−3505 (2756)
20−26 (22) [for n = 11] 18−35 (24) [for n = 14] 18−36 (26) [for n = 12]
a
Prediagnosis GDM includes cases of hypertension (n = 1), twin pregnancy (n = 2), preterm delivery (n = 3), macrosomia (n = 1), and intrauterine growth restriction (n = 1). bPostdiagnosis GDM includes cases of hypertension (n = 1), twin pregnancy (n = 2), and preterm delivery (n = 1). cThis group includes one case of hypertension. aValues in brackets correspond to average values.
enoic acid levels found between hyperglycemia and GDM.7 A subsequent study showed higher serum levels of triglycerides (TG), 3-hydroxybutyrate, and Ala, Pro, and branched-chain amino acids (BCAA) and lower 1,5-anhydroglucitol levels, in women with higher fasting glucose levels.8 GDM impact on lipid metabolism was confirmed by LC− and GC−MS of maternal plasma, where several lysophospholipids, taurine-bile acids, and long-chain polyunsaturated FA derivatives were found discriminative of GDM and indicative of low-grade inflammation and altered redox-balance.9 To our knowledge, no nuclear magnetic resonance (NMR) metabolomics characterization (a more holistic approach, compared with MS) of maternal blood related to postdiagnosis GDM has been reported. The profiling of maternal urine, however, has been performed by NMR, in a large cohort (n = 823) of pregnant women with GDM, concluding that no reliable GDM classification could be achieved and suggesting increased excretion of citrate, with increasing hyperglycemia.10 In addition, NMR profiling of cord serum of infants born from GDM mothers revealed decreased glucose and increased pyruvate, His, Ala, Val, Met, Arg, Lys, hypoxanthine, lipoproteins, and lipids,11 while LC−MS metabolic profiles of newborn meconimum and urine suggested disruptions in lipid, amino acid, and purine metabolisms affecting the newborn.12 The search for metabolic alterations preceding GDM diagnosis was initiated by NMR of amniotic fluid, unveiling higher glucose levels and decreases in amino acids, acetate, formate, creatinine, and glycerophosphocholine.13 In maternal plasma, NMR of a small cohort detected decreases in trimethylamine N-oxide (TMAO) and betaine, interpreted as changes in choline-related homocysteine-methionine conversion and possible alterations in renal function.14 Recently, GC− MS of maternal sera unveiled early increases in itaconic acid and cis-aconitate, possibly due to inflammation,15 and a subsequent LC−MS study unveiled increases in anthranilic acid, Ala, Glu, Ser, and alantoin and decreased creatinine.16 One maternal urine NMR study has revealed prediagnosis GDM changes in glucose (increase) and several changes in common with type 2 diabetes, namely, increased acetate, Nmethylnicotinamide and N-methyl-2-pyridone-5-carboxamide,
changes in creatine and creatinine, and decreased hippurate and phenylacetylglutamine.17 The present paper presents a NMR metabolomics study of maternal plasma and corresponding lipid extracts obtained for a group of pregnant women without clinical signs of the disease but who 2−21 g.w. later developed GDM (n = 32, prediagnosis group) and a group of pregnant women at the time of GDM diagnosis (n = 12, postdiagnosis group), in comparison with controls (n = 49). This adds to a previous smaller cohort NMR study of prediagnosis GDM (n = 14) reported by our own group14 and provides a first NMR study of postdiagnosis GDM, to the best of our knowledge. In the latter case, GDM and control cases differed in average gestational age, and hence the metabolites related to the normal underlying pregnancy evolution were identified prior to determining the postdiagnosis GDM metabolic profile. Furthermore, the complementary information provided by whole plasma and lipid extract analysis is discussed and sample sensitivity for GDM is compared.
■
EXPERIMENTAL SECTION
Samples
Blood samples were collected for healthy pregnant women, pregnant women who later developed GDM (prediagnosis GDM group, 2−21 g.w. prior to diagnosis), and women after GDM diagnosis (postdiagnosis GDM group, 1 to 2 weeks after the OGTT test, at 18−38 g.w., and before any clinically advised treatment) (Table 1). GDM was diagnosed according to the International Association of Diabetes and Pregnancy Study Groups criteria.5 Because it was not possible to obtain a control group matched with the postdiagnosis group for gestational age, the gestational age-dependent metabolites were identified and taken into account. All samples were collected under the approval of the Coimbra Hospital Centre ethical committee (refs 18/04 and 29/09), and individual informed consents were obtained. Subject groups were independent, and samples were collected nonfasting during routine and diabetes medical appointments. The effect of nonfasting on NMR-viewed plasma metabolome was found residual, as described elsewhere.18 Blood (9 mL) was collected into sodium heparin tubes B
DOI: 10.1021/acs.jproteome.5b00260 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Figure 1. Average 500 MHz CPMG 1H NMR spectra of plasma of (a) controls and (b) pre- and (c) postdiagnosis GDM groups, with indication of spectral variables (data points) selected in the corresponding PLS-DA models (gray dots under spectra). 1-CH3 lipids, 2-Val, 3-(CH2)n lipids, 4lactate, 5-Ala, 6-CH2CC lipids, 7-N-acetyl glycoproteins, 8-Gln, 9-pyruvate, 10-citrate, 11-creatine, 12-creatinine, 13-N(CH3)3 choline of PL, 14glucose, 15-urea, 16-Tyr, 17-His, 18-unknown (δ 7.15−7.35), 19-formate. Arrows indicate visible alterations. *: excluded spectral regions.
and centrifuged (1500g, 4 °C, 10 min) within 30 min of collection, followed by storage of supernatants at −80 °C. Sample preparation for NMR is described elsewhere.19 Plasma lipid extracts were prepared in duplicate according to the methyl-tert-buthyl ether method.20 Organic phases containing the lipid fraction were dried and stored at −80 °C. Before analysis, extracts were thawed at room temperature, followed by the addition of 600 μL of CDCl3 (99.96%) with tetramethylsilane (0.03%, v/v). The mixture (580 μL) was transferred to a 5 mm NMR tube.
and BMRB databases,21,22 and statistical total correlation spectroscopy (STOCSY).23 Chemometrics
The full-resolution 1D spectra were used to construct data matrices for blood plasma after exclusion of the water (δ 4.5− 5.0) and ethanol (δ 1.15−1.20 and δ 3.62−3.68) regions, the latter found randomly in blood samples possibly due to skin disinfection prior to collection. Full-resolution spectra were also used for plasma lipid extracts, excluding water (δ 1.60−1.97) and CDCl3 (and corresponding satellites) (δ 7.04−7.05, δ 7.21−7.32, and δ 7.46−7.47) peaks. 1D spectra were aligned using recursive segment-wise peak alignment (RSPA)24 and normalized using probabilistic quotient normalization (PQN).25 Principal component analysis (PCA) and partial least-squares discriminant analysis (PLS-DA) were performed on data scaled by unit variance, Pareto, and centered scaling (SIMCA-P 11.5, Umetrics, Umeå, Sweden), with the former providing the best results (therefore, all results presented refer to unit variance scaling). PCA and PLS-DA loadings were backtransformed, multiplying each variable by its standard deviation and colored according to variable importance to the projection (VIP) (Matlab 7.9.0). For PLS-DA models, a randomized (Monte Carlo) cross-validation approach (MCCV, in-house developed) was carried out with recovery of Q2 values and confusion matrices; simultaneously, a randomized classpermutation procedure assessed the null hypothesis. Classification rates, specificity (spec.), and sensitivity (sens.) were computed, and model predictive power was further assessed by a receiver operating characteristic map. PLS-DA models were considered to be robust for minimal overlap of true
NMR Spectroscopy
NMR spectra were recorded, at 300 K, on a Bruker Avance DRX 500 spectrometer operating at a proton frequency of 500.13 MHz and equipped with an actively shielded gradient unit with a maximum gradient strength output of 53.5 G/cm. For each plasma sample, three types of 1D 1H NMR spectra were recorded: a standard spectrum (1D NOESY-presat), a Carr−Purcell−Meiboom−Gill (CPMG) spectrum, and a diffusion-edited spectrum, all acquisition parameters being listed elsewhere.19 For lipid extracts, 1H NMR spectra were acquired at 298 K using a standard 90° pulse sequence (zg), with a 2.34 s acquisition time, 5 s relaxation delay (d1), 128 scans, 32 k data points, and 7002.801 Hz spectral width. Each free-induction decay was zero-filled to 64 k points and multiplied by a 0.3 Hz exponential function prior to Fourier transformation. Spectra were manually phased, baselinecorrected, and chemical-shift-referenced to α-glucose H1 (δ 5.23) or tetramethylsilane (δ 0.00), for plasma and extracts, respectively. Peak assignments were carried out using 2D NMR, the Bruker B-BIOREFCODE spectral database and the HMDB C
DOI: 10.1021/acs.jproteome.5b00260 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Table 2. MCCV Parameters Obtained When Considering Original (full dataset) and Variable Selected 1H NMR Spectraa original spectra subject group Controls vs Prediagn. GDM whole plasma/CPMG whole plasma/diffusion-ed. lipid extracts Controls vs Postdiagn. GDM whole plasma/CPMG whole plasma/diffusion-ed. lipid extracts
variable selection
LV
Q2
CR %
Sens %
Spec %
LV
Q2
CR %
Sens %
Spec %
% VS
1 1 1
0.1 0.3 0.3
68 73 64
53 60 63
77 81 64
2 1 1
0.5 0.2 0.5
81 72 83
69 57 75
89 82 90
36 38 33
1 1 4
0.4 0.5 0.6
86 93 84
43 80 80
96 97 87
1 1 2
0.6 0.7 0.7
90 92 79
59 77 69
97 96 87
32 38 29
a LV, no. of latent variables; Q2, most frequent value of predictive power obtained by MCCV; CR, classification rate; sens, sensitivity; spec, specificity; %VS, percentage of variables selected from the original spectra. Values in bold refer to robust PLS-DA models.
Figure 2. PLS-DA scores scatter plots and Volcano plots (effect size vs − Log (p value)) of (a) plasma CPMG 1H NMR spectra of controls (■, n = 49) versus prediagnosis GDM (□, n = 32) (R2X 0.29, R2Y 0.58), (b) plasma diffusion-edited 1H NMR spectra of controls (■, n = 49) versus prediagnosis GDM (□, n = 32) (R2X 0.49, R2Y 0.45), and (c) lipid extracts 1H NMR spectra of controls (▲, n = 15) versus prediagnosis GDM (△, n = 14) (R2X 0.37, R2Y 0.66). Bet, betaine; FA, fatty acids; FC, EC, free and esterified cholesterol; glyc, glyceryl; Lac, lactate; Pyr, pyruvate; TMAO, trimethylamine N-oxide; 1,5-AG, 1,5-anhydroglucitol; Un 3, unknown 3 (δ 5.91). Three-letter codes used for amino acids.
D
DOI: 10.1021/acs.jproteome.5b00260 J. Proteome Res. XXXX, XXX, XXX−XXX
Article
Journal of Proteome Research
Table 3. List of Metabolites/Resonances Selected in the NMR Spectra of Blood Plasma As Important for Discrimination of PreAnd Post-Diagnosis GDM from Controls prediagn. GDM (n = 32) vs controls (n = 49) compound
chemical shift (ppm)a
E.S.b
postdiagn. GDM (n = 12) vs controls (n = 49)
p valuec
E.S.b
p valuec
d
valine alanine pyruvate glutamine citrate creatine creatinine dimethyl sulfone TMAO betaine proline methanol glycine 1,5-anhydroglucitol lactate glucose urea unknown 1 C18H cholesterol CH3 lip. HDL CH3 lip. LDL+VLDL (CH2)n lip. HDL (CH2)n lip. LDL+VLDL CH2CH2CO lip. CH2CC lip. CH2CO lip. CCCH2CC lip. N(CH3)3 choline HDd,g N(CH3)3 chol. LDL+VLDL CH2-N(CH3)3 chol.g unknown 2 glyceryl-C1,3Hg glyceryl-C1,3H′ HCCH HDL HCCH LDL+VLDL (CH2)n/CH3 lip. HDL HCCH/CH3 lip. HDL (CH2)n/CH3 lip. LDL+VLDLg HCCH/CH3 lip. LDL+VLDL CH3 HDL/CH3 LDL+VLDL
1.03 (d) 1.47 (d) 2.36 (s) 2.44 (m) 2.52 (d) 3.03 (s) 3.04 (s) 3.14 (s) 3.27 (s) 3.29 (s) 3.34 (dt) 3.36 (s) 3.55 (s) 3.95−4.00 (dd) 4.11 (q) 5.23 (d) 5.77 (br) 7.15−7.35 (br) 0.67 (br) 0.79−0.85 (br) 0.85−0.91 (br) 1.18−1.25 (br) 1.25−1.37 (br) 1.45−1.62 (br) 1.90−2.02 (br) 2.17−2.26 (br) 2.65−2.84 (br) 3.19−3.205 (br) 3.23−3.26 (br) 3.62−3.68 (br) 3.84−3.92 (br) 4.02−4.10 (br) 4.21−4.32 (br) 5.24−5.28 (br) 5.28−5.37 (br)
Low Mw Compounds ↑ (0.38 ± 0.46)
2.6 × 10−2
↑ 0.49 ± 0.46 ↓
4.5 × 10−2 >0.05
↓
>0.05
↓ ↓ ↓ ↓ −0.53 ± 0.46
>0.05 >0.05 >0.05 2.1 × 10−2
↓ −0.48 ± 0.46 ↑ ↑ ↓ −0.58 ± 0.46
3.5 × 10−2 >0.05 >0.05 2.1 × 10−2
High Mw Compoundse ↑ 1.23 ± 0.49 ↑ 0.60 ± 0.46 ↑ 0.82 ± 0.47 ↑ 0.60 ± 0.46 ↑ 0.69 ± 0.46 ↑ 0.56 ± 0.46 ↑ 0.65 ± 0.46 ↑ ↑
1.0 × 10−6f 8.5 × 10−3 4.2 × 10−3 8.8 × 10−3 3.1 × 10−2 2.5 × 10−2 9.2 × 10−3 >0.05 >0.05
↑
>0.05
↑ 0.54 ± 0.46 ↑ 0.63 ± 0.46 ↑ ↑ ↑ ↑ ↓
2.4 × 10−2 1.0 × 10−2 >0.05 >0.05 >0.05 >0.05 >0.05
↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓
−0.65 ± 0.64 −0.74 ± 0.65 −1.27 ± 0.67
−1.69 ± 0.70 −0.86 −1.40 −1.33 −1.18
± ± ± ±
0.65 0.68 0.68 0.67
4.5 × 10−2 5.0 × 10−2 4.4 × 10−3 2.0 × 10−5f >0.05 >0.05 3.5 × 10−9f >0.05 4.8 × 10−3 3.5 × 10−8f 2.0 × 10−5f 9.4 × 10−6f 4.6 × 10−3
↓ ↓ −0.96 ± 0.66
>0.05 1.3 × 10−3f
↑ 1.62 ± 0.70
4.6 × 10−5f
↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↓ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↓
2.2 × 10−5f 1.3 × 10−6f 1.1 × 10−5f 4.0 × 10−5f 9.7 × 10−4f 4.0 × 10−5f 4.8 × 10−9f 1.1 × 10−4f 5.6 × 10−4f 1.9 × 10−3 1.7 × 10−7f 2.1 × 10−5f 1.2 × 10−5f 2.3 × 10−2 2.8 × 10−7f 3.8 × 10−5f 1.7 × 10−6f >0.05 1.2 × 10−3f 3.8 × 10−2 >0.05 2.4 × 10−3
1.35 ± 0.68 1.47 ± 0.68 2.54 ± 0.78 1.38 ± 0.68 1.87 ± 0.71 1.50 ± 0.68 2.37 ± 0.76 1.49 ± 0.68 1.29 ± 0.66 −1.15 ± 0.67 2.24 ± 0.75 1.29 ± 0.67 1.72 ± 0.70 1.00 ± 0.66 1.88 ± 0.71 1.71 ± 0.70 1.55 ± 0.69 1.30 ± 0.67 0.85 ± 0.65 −1.30 ± 0.67
a
Chemical shifts of signals used for integration: s, singlet; d, doublet; dt, doublet of triplets; q, quartet; m, multiplet; br, broad. bE.S.: effect size determined as described in ref 28; values in brackets correspond to high uncertainties. c95% significance level (p value 0.05 ↑ >0.05 (FC) and Esterified Cholesterol (EC) ↑ >0.05 ↑ >0.05 ↑ >0.05 ↑ >0.05 ↑ >0.05 ↑ >0.05 Triglycerides (TG) ↑ 0.56 ± 0.53 3.6 × 10−2 ↑ 0.61 ± 0.53 4.2 × 10−2 ↓ −0.69 ± 0.53 1.2 × 10−2
E.S.b ± ± ± ± ± ± ± ± ±
p valuec
↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑
1.25 0.94 1.21 1.21 0.68 1.26 0.62 0.98 0.74
↑ ↑ ↑ ↑ ↑ ↑
0.73 ± 0.55 0.72 ± 0.55 0.76 ± 0.56 0.69 ± 0.55 0.60 ± 0.55 (0.42 ± 0.52)
0.59d 0.57 0.58 0.58 0.55 0.59 0.55 0.57 0.55
↑ 1.12 ± 0.58 ↑ 1.07 ± 0.57 ↓ −0.73 ± 0.55
8.5 1.6 4.4 1.5 2.4 9.7 4.0 5.3 5.7
× × × × × × × × ×
10−5d 10−3d 10−4d 10−4d 10−2 10−5d 10−2 10−3 10−3
2.0 × 10−2 2.1 × 10−2 1.5 × 10−2 2.7 × 10−2 4.8 × 10−2 >0.05 7.2 × 10−5d 5.1 × 10−4d 1.0 × 10−2
a
Chemical shifts of signals used for integration: s, singlet; d, doublet; dt, doublet of triplets; q, quartet; m, multiplet; br, broad. bE.S.: effect size determined as described in ref 28; values in brackets correspond to high uncertainties. c95% significance level (p value
