Predictive Metabolomic Markers in Early to Mid-pregnancy for Gestational Diabetes Mellitus: A Prospective Test and Validation Study

Abstract

Gestational diabetes mellitus (GDM) predisposes pregnant individuals to perinatal complications and long-term diabetes and cardiovascular diseases. We developed and validated metabolomic markers for GDM in a prospective test-validation study. In a case-control sample within the PETALS cohort (GDM n = 91 and non-GDM n = 180; discovery set), a random PETALS subsample (GDM n = 42 and non-GDM n = 372; validation set 1), and a case-control sample within the GLOW trial (GDM n = 35 and non-GDM n = 70; validation set 2), fasting serum untargeted metabolomics were measured by gas chromatography/time-of-flight mass spectrometry. Multivariate enrichment analysis examined associations between metabolites and GDM. Ten-fold cross-validated LASSO regression identified predictive metabolomic markers at gestational weeks (GW) 10-13 and 16-19 for GDM. Purinone metabolites at GW 10-13 and 16-19 and amino acids, amino alcohols, hexoses, indoles, and pyrimidine metabolites at GW 16-19 were positively associated with GDM risk (false discovery rate <0.05). A 17-metabolite panel at GW 10-13 outperformed the model using conventional risk factors, including fasting glycemia (area under the curve: discovery 0.871 vs. 0.742, validation 1 0.869 vs. 0.731, and validation 2 0.972 vs. 0.742; P < 0.01). Similar results were observed with a 13-metabolite panel at GW 17-19. Dysmetabolism is present early in pregnancy among individuals progressing to GDM. Multimetabolite panels in early pregnancy can predict GDM risk beyond conventional risk factors.

Figure 1

Study flowchart of the discovery set and validation set 1 (A) and validation set 2 (B). RCT, randomized controlled trial. *Case-control ratio of one to two, with two GDM cases; each had only one matched control with biospecimens available.

Figure 2

Volcano plots showing the associations of individual metabolites at GW 10–13 (A) and 16–19 (B) with risk of GDM in the PETALS nested case-control discovery set. Odds ratios (OR) were adjusted for maternal age at delivery, race/ethnicity, family history of diabetes, chronic hypertension, history of GDM, prepregnancy BMI, and gestational age at blood collection. Horizontal dashed line indicates the value of FDR-corrected level of significance.

Figure 3

Multivariate ChemRICH enrichment plots depicting the pathways and metabolite hits within each pathway at 10–13 (A) and 16–19 (B) weeks of gestation significantly associated with risk of GDM in the PETALS nested case-control discovery set. OR, odds ratio. *P value for pathway <0.05 after FDR adjustment. †Pathways significantly associated with risk of GDM at both gestational periods (P < 0.05).

Figure 4

Incremental prediction value of multimetabolite panels at GW 10–13 (A) and 16–19 (B) beyond conventional risk factors for GDM in the PETALS nested case-control discovery set. Receiver operating characteristic curves and AUC statistics were estimated by 10-fold cross-validation for GDM risk prediction using conventional risk factors (age at delivery, family history of diabetes, chronic hypertension, history of gestational diabetes, prepregnancy BMI, and fasting serum glucose values; model 1, red curves); a multimetabolite panel (model 2, green curves) selected by LASSO regression at 10–13 weeks (1,5-anhydroglucitol, 1-monoolein, 2,3-dihydroxybutanoic acid, 2-hydroxyglutaric acid, 5,6-dihydrouracil, alanine, α-aminoadipic acid, β-alanine, β-sitosterol, cellobiose, citramalic acid, citric acid, lactic acid, N-acetylputrescine, β-tocopherol, uric acid, and urea) (A) and 16–19 weeks of gestation (1,5-anhydroglucitol, 2,3-dihydroxybutanoic acid, 2-aminobutyric acid, α-aminoadipic acid, arachidic acid, aspartic acid, citric acid, hydrocinnamic acid, lauric acid, oleic acid, quinic acid, uracil, and uridine) (B); and the selected multimetabolite panel in addition to conventional risk factors (model 3, blue curves). P values for differences in AUC statistics between models were derived by DeLong test.