Identifying serum amino acid as biomarkers of gestational diabetes mellitus in second-trimester among Chinese pregnant women: a machine learning-based cross-sectional study

Abstract

Background

Levels of plasma branched-chain and aromatic amino acids in pregnancy have been associated with gestational diabetes mellitus (GDM), but the metabolic role of serum amino acid (AA) profiles in its pathogenesis remains insufficiently elucidated.

Objective

This study evaluated the diagnostic potential of second-trimester serum AA profiles, including Cys, Met, Val, Lys, Cit, Tau, Asp, Ile and Ala, for distinguishing GDM patients from healthy controls.

Methods

A total of 189 women with GDM and 189 healthy women at 24–28 weeks of gestation were enrolled in the study, recruited from 2019 to 2022. Serum levels of 21 amino acids were precisely measured using automatic amino acid analyzer. Three machine learning methods were employed to select the most significant variables. Generalized linear models (GLMs) were established to evaluate the association between Serum AAs and GDM.

Results

Serum cysteine (Cys) and lysine (Lys) were inversely associated with GDM risk, whereas methionine (Met) and citrulline (Cit) showed positive associations. Notably, Met demonstrated an inverted U-shaped relationship, with an inflection at 239.9 µmol/L. The adjusted model achieved higher discrimination than the crude model. Sensitivity and subgroup analyses confirmed robust associations for Cys, while associations for Met, Lys, and Cit varied by pre-pregnancy body mass index (BMI).

Conclusions

Mid-pregnancy serum AAs, particularly Cys, Lys, Met, and Cit, were associated with GDM risk. These findings highlighted the heterogeneity of GDM metabolic signatures and support AAs as potential biomarkers for diagnosis of GDM.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12978-026-02272-6.

Introduction

GDM is a common pregnancy complication, affecting approximately 14% of pregnancies worldwide in 2021 and has become a significant global health concern [1]. The total incidence of GDM in mainland China was 14.8% ranged from 12.8 to 16.7% [2]. It increases maternal and infant risks, including preterm birth, macrosomia, and future type 2 diabetes mellitus (T2DM) [3–5]. However, prior epidemiologic studies have placed greater focus on lifestyle risk factors linked to GDM risk, and the metabolic alterations underlying its hyperglycemic phenotype remain poorly defined. Therefore, identifying new biomarkers for GDM is vital to deciphering its causes, accurately pinpointing high-risk women, and paving the way for better interventions.

Accumulating evidence suggests that some amino acids could regulate various metabolic processes, including glucose and lipid metabolism. Disrupted branched-chain amino acid (BCAA) homeostasis represents a metabolic signature of type 1 diabetes mellitus (T1DM) and T2DM in mice [6], and higher intake of aromatic AAs has been associated with increased T2DM risk [7]. Given the shared pathophysiology of GDM and T2DM [8, 9], altered AA metabolism may similarly contribute to GDM. Lower mid-pregnancy glycine levels have been inversely associated with GDM risk, whereas isoleucine (Ile), phenylalanine (Phe), and tyrosine (Tyr) show positive associations [10]. Women with GDM exhibit elevated alanine (Ala), glutamic acid, and serine (Ser) in early pregnancy, implicating glucose-related pathways [11]. Additional metabolites, including glycine and serine, have also been linked to elevated blood glucose later in gestation [12]. Recent studies highlight the potential role of plasma AA profiles in GDM development [10, 13], although metabolite signatures differ substantially between serum and plasma [14]. These amino acids may serve as GDM biomarkers and therapeutic targets, since their altered levels reflect underlying metabolic disturbances. Therefore, a more comprehensive understanding of amino acid profiles involved in GDM is essential.

In this study, to investigate the association between serum amino acid profiles and GDM, we conducted a cross-sectional study enrolled 378 participants from a tertiary hospital in Henan Province, China. Three machine learning (ML) algorithms were applied for feature selection, followed by construction of a GLM rigorously evaluated through multiple performance metrics and sensitivity analyses. Therefore, elucidating these associations provide deeper insights into the metabolic mechanisms underlying GDM, potentially informing future preventative strategies.

Methods

Study design and population

This cross-sectional study was conducted at the Department of Obstetrics, a tertiary hospital in Henan Province, China, between January 2019 and October 2022. Singleton pregnant women aged 18–45 years who had undergone routine antenatal examinations were eligible. Participants were classified according to the results of a 75-g oral glucose tolerance test (OGTT) at 24–28 gestational weeks. A total of 189 women with GDM and 189 without GDM were included. Controls were frequency-matched to cases by age (± 3 years).

Exclusion criteria included: (1) pre-pregnancy diabetes or hypertension; (2) use of medications affecting hormone secretion or glucose/lipid metabolism before pregnancy; (3) comorbidities during pregnancy; (4) bad lifestyle habits such as alcohol consumption and smoking; and (5) assisted reproduction and multiple pregnancies.

Anthropometric measurement

Data on sociodemographic characteristics such as age, was collected from questionnaires conducted by professional trained researchers. Detailed information including pregnancy times, family history of diabetes, gravidity, parity, and blood glucose was obtained from medical records. The weight and height of participants were measured by the InBody J30 (Biospace, Seoul, South Korea) and a stadiometer in the hospital without shoes and wearing light clothing, respectively. Pre-pregnancy body mass index (BMI) was calculated by weight (kg)/height² (m²). Blood pressure was measured twice after 5 min of rest using an appropriate cuff according to arm size. The average of two blood pressure measurements was recorded as the final value.

Diagnostic criteria

A standard 75-g OGTT was performed after an overnight fast of at least 8 h at 24–28 gestational weeks. Plasma glucose concentrations were measured at fasting, 1 h, and 2 h. GDM was diagnosed if one or more of the following thresholds were met: fasting plasma glucose (FPG) ≥ 5.1 mmol/L, 1-h plasma glucose ≥ 10.0 mmol/L, or 2-h plasma glucose ≥ 8.5 mmol/L, according to the International Association of Diabetes and Pregnancy Study Groups (IADPSG) criteria [15]. Women whose FPG was ≥ 5.6 mmol/L (100 mg/dL) and/or 2-h PG was ≥ 7.8 mmol/L (140 mg/dL) were defined as GDM by National Institute for Health and Care Excellence (NICE) criteria [16].

Laboratory measurements

Fasting venous blood samples were collected between 24 and 28 weeks of gestation, immediately prior to the OGTT. Samples were centrifuged at 2,500 rpm for 10 min at 4 °C, and serum was aliquoted and stored at − 80 °C. Amino acid concentrations were quantified using the HITACHI L8900 automatic amino acid analyzer at 20–30 °C. Twenty-one amino acids were measured, including glycine, alanine, aspartic acid, arginine, glutamic acid, threonine, serine, valine, leucine, isoleucine, phenylalanine, tyrosine, methionine, histidine, lysine, phosphorylated serine, citrulline, ornithine, cystine, taurine, and hydroxylysine.

In addition, serum triglycerides (TG), total cholesterol (TC), high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C) were measured using the Excellence 360 automated biochemical analyzer (Shanghai Kehua Bioengineering Co., Ltd.).

Statistical method

All analyses were performed using SPSS version 27.0 and R version 4.5.0. Continuous variables were expressed as mean ± SD for normally distributed data or median (interquartile range) for skewed data, and compared using Student’s t test or Mann–Whitney U test. Categorical variables were compared using the χ² test. All tests were two-sided, and the significance level was α = 0.05. The Benjamini-Hochberg (BH) procedure was employed to adjust P-values for multiple comparisons, controlling the false discovery rate (FDR). The required sample size was calculated using G*Power software version 3.1 with effect size d = 0.5, α = 0.05, power (1-β) = 0.95, and allocation ratio = 1; results indicated that a minimum of 210 participants were required. The statistical power for each outcome was also assessed with the same software.

To identify potential diagnostic biomarkers for GDM, we employed support vector machine–recursive feature elimination (SVM-RFE), extreme gradient boosting (XGBoost), and Boruta algorithm. The SVM-RFE method outperforms linear discriminant analysis and the mean squared error approach in its ability to efficiently select relevant features and eliminate redundancy [17]. XGBoost identifies key features through iterative optimization and regularization, thereby enhancing model robustness and feature selection reliability while reducing overfitting [18]. The Boruta algorithm identifies important variables by comparing the Z value of each true feature with that of its corresponding “shadow features.” Features with significantly higher Z values than their shadow counterparts were classified as “important” (green area), whereas those without significant differences were considered “unimportant” (red area) [19]. Moreover, Boruta effectively controls false positives and captures nonlinear relationships within the data. We employed 10-fold cross-validation to prevent overfitting. The training dataset was partitioned into 10 subsets, and in each iteration, the model was trained on 9 subsets and validated on the remaining one. The top 10 variables identified by each of the three algorithms were subsequently incorporated into a generalized linear model (GLM) to ensure robust feature selection.

GLM were applied to assess linear associations between amino acids and GDM. Nonlinear associations were explored using generalized additive models (GAMs). For variables with nonlinear trends, inflection points were identified using a recursive algorithm, and threshold effect analysis compared GLMs with piecewise linear models. Discriminative ability was evaluated using receiver operating characteristic (ROC) curves, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Calibration was assessed using calibration plots. Clinical utility was evaluated using decision curve analysis (DCA) and clinical impact curves (CIC). We performed a series of sensitivity and subgroup analyses to evaluate the robustness of our primary findings. First, we redefined GDM according to the NICE criteria. Second, to ensure that our results were not unduly influenced by outliers, we truncated all amino acid values used in the GLMs to the 5th and 95th percentiles and repeated the main analyses. Third, we conducted stratified analyses by pre-pregnancy BMI to examine whether it modified the association between GDM and mid-trimester serum amino acid levels.

Results

Study participant characteristics

The clinical characteristics of the 378 pregnant women at 24–28 weeks gestation were presented in Table 1. No significant differences were observed for the family history of diabetes, gravidity, parity, systolic blood pressure (SBP), diastolic blood pressure (DBP) and LDL-C between patients with GDM and controls (all P>0.05). Notably, compared with the women without GDM, women with GDM were older and had higher age, pre-pregnancy BMI (both P < 0.05). TG, TC and HDL-C were significantly different between case and control groups in the second trimester (all P<0.05). The levels of fasting blood glucose, 1-hour OGTT, and 2-hour OGTT in the GDM group were significantly higher than those in the control group (all P < 0.05). The biochemistry characteristics were summarized in Table S1. In the mid-term serum amino acid levels of pregnant women, the levels of Val, and Met in the case group were higher than those in the control group, whereas Cit and Cys were lower (all P < 0.05). All other amino acids did not differ significantly between the two groups (all P>0.05). Among the four amino acid categories, nonpolar amino acids exhibited the highest proportion of significant associations (42.9%), followed by uncharged polar amino acids (16.7%). No significant associations were observed for charged polar or chemically modified amino acids. These findings suggest that nonpolar amino acids, as a group, may play a more prominent role in the pathogenesis of GDM.

Table 1.

Characteristics of GDM and controls

characteristics	Control (n = 189)	GDM (n = 189)	P value
Age, year	30.0 (27.0, 32.0)	31.0 (29.0, 33.0)	< 0.05
Pre-pregnancy BMI (kg/m2)	20.80 (19.57, 23.15)	22.89 (20.96, 24.57)	< 0.001
Family history of diabetes, n (%)			0.72
Yes	5 (2.6)	3 (1.6)
No	184 (97.4)	186 (98.4)
Gravidity (%)			0.41
1	88 (46.6)	79 (41.8)
≥ 2	101 (53.4)	110 (58.2)
Parity (%)			0.76
0	106 (56.1)	102 (54.0)
≥ 1	83 (43.9)	87 (46.0)
SBP (mmHg)	111.54 ± 9.88	110.87 ± 10.58	0.53
DBP (mmHg)	65.00 (59.00, 70.00)	67.00 (60.00, 71.00)	0.15
TG (mmol/L)	2.39 (1.82, 2.98)	2.53 (2.05, 3.21)	0.05
TC (mmol/L)	6.09 (5.33, 6.89)	5.86 (5.30, 6.46)	< 0.05
HDL-C (mmol/L)	1.740 (1.53, 2.00)	1.636 (1.42, 1.84)	< 0.05
LDL-C (mmol/L)	3.280 (2.75, 3.90)	3.210 (2.78, 3.59)	0.38
FBG (mmol/L)	4.64 (4.37, 5.12)	4.830 (4.600, 5.170)	< 0.001
OGTT-1 h (mmol/L)	7.29 (6.40, 8.25)	10.24 (9.14, 10.88)	< 0.001
OGTT-2 h (mmol/L)	6.83 (6.05, 7.59)	9.04 (8.39, 9.80)	< 0.001

Feature selection and linear association analyses of serum amino acids with GDM

We constructed two GLMs to investigate the associations between serum amino acid concentrations and the risk of GDM. Covariates for both the crude and adjusted models were selected using SVM-RFE, XGBoost, and Boruta algorithms (Fig. 1 and Table S2). The factors included in each model were as follows:

Fig. 1.

Feature selection result graphs after computerized process through machine learning methods involving, SVM-RFE (A), XGBoost (B) and Boruta (C, D) algorithms. A The chart illustrates the final 10 features eliminated by the SVM-RFE algorithm. The y-axis lists the feature names in descending order of importance, while the x-axis displays their ranking scores. B The top 10 most important features identified by XGBoost, ordered by mean absolute SHAP value on the y-axis with their corresponding importance scores shown on the x-axis. C After 600 iterations based on Boruta’s algorithm, a subset of features was identified with the minimum cost function. The horizontal axis represents the input variables, and the vertical axis denotes the feature importance (Z-scores). The blue boxes represents the minimum, average and maximum Z-scores of shadow features, respectively. And the green boxes, yellow boxes, red boexes correspond to confirmed features, tentative features and rejected features, accordingly. D This plot illustrates the iterative feature selection process of the Boruta algorithm, showing features on the x-axis in descending order of their median importance, with the corresponding importance scores displayed on the y-axis

1. Crude model: Nine amino acids were included Cys, Met, Val, Lys, Cit, Tau, Asp, Ile, P-Ser, and Ala without adjustment for confounding factors.

2. Adjusted model: The same nine amino acids were included, and confounding factors age, BMI, HDL-C, CHOL, TG, SBP, and DBP were adjusted for.

Both crude and adjusted GLM analyses indicated that Cys, Met, Lys, and Ile were significantly associated with the risk of GDM (Fig. 2). Notably, the association between Ile and GDM reached significance only in the adjusted model. After adjusting for potential confounders, 1-each increment in Cys and Lys concentrations was associated with a reduction in the log odds of developing GDM by 0.88 (β = -0.88, 95% CI: -1.25 to -0.51) and 0.85 (β = -0.85, 95% CI: -1.38 to -0.33), respectively. This corresponds to a reduction in the odds of GDM by factors of 2.72 and 2.34, respectively. Conversely, 1-µmol/L increment in Met and Cit levels was associated with an increased risk of GDM, reflected by changes in the log odds of 0.69 (β = 0.69, 95% CI: 0.31 to 1.07) and 0.09 (β = 0.09, 95% CI: 0.03 to 0.16), respectively. These changes translate into increased odds of GDM by factors of 1.99 and 1.09, respectively. Notably, Lys concentrations did not differ significantly between the GDM and control groups (P = 0.20, Table S1), whereas Cit levels were higher in the control group (P > 0.05).

Fig. 2.

Association between the serum amino acid levels and GDM in the second trimester. Crude model was unadjusted. Adjusted model was adjusted for age, BMI, HDL-C, CHOL, TG, SBP, DBP

Model performance evaluation

Calibration curves were used to evaluate the discriminative ability of the crude model and the adjusted model between normal pregnant women and pregnant women with GDM. The mean absolute error (MAE) values for the two model were 4.2% and 5.6%, respectively, indicating strong agreement between estimated and actual observations in adjusted model (MAE < 0.05, Fig. 3A). Compared with the crude model, the adjusted model demonstrated higher accuracy and superior discriminative ability for GDM (AUC: 0.792 vs. 0.753; P < 0.05; Fig. 3B). The NRI and IDI for adjusted model vs. crude model (NRI = 0.4180, 95% CI: 0.2351–0.6672; IDI = 0.0635, 95% CI:0.0390–0.0880) were all greater than 0, suggesting that the reclassification ability of the adjusted model was superior to that of the crude model. (Table 2).

Fig. 3.

A Calibration curve. The two curves represent the net benefits of the crude model and adjusted model, respectively. B ROC curves of different generalized linear models exploring the association between mid-trimester serum amino acid profiles and GDM. C DCA curves reflect the net benefit of different models for GDM. D Clinical impact curve of the adjusted model

Table 2.

Model performance evaluation of different models. NRI Net reclassification improvement, IDI Integrated discrimination improvement

	NRI		IDI
	β (95%CI)	P	β (95%CI)	P
Crude Model	Ref		Ref
Adjusted Model	0.4180 (0.2351, 0.6672)	< 0.001	0.0635 (0.0390, 0.0880)	< 0.001

The potential clinical utility of the models was explored using decision curve analysis (DCA). These analyses suggested that both models may offer a net benefit relative to either the treat-all or no-treatment scenarios when the estimated probability of GDM was below 40% (Fig. 3C). To further visualize their hypothetical clinical impact, clinical impact curves (CICs) were derived from the DCA. As shown in Fig. 3D and Figure S1, the curves compare the number of women classified as high risk by the model with the actual GDM cases at each threshold. At the 40% threshold, the number designated as high risk closely aligned with the observed number of true cases.

Non-linear association analyses of serum amino acid levels with GDM

We performed GAM and smooth curve fitting to detect potential non-linear relationships between serum amino acid levels and the risk of GDM, aiming to further validate the results. Although GAMs suggested potential non-linear associations for Met (Fig. 4A), and Cit (Fig. 4B), only Met exhibited a statistically significant improvement in model fit when a multiple-piecewise linear regression was applied, as indicated by a significant log-likelihood ratio test (Table S3, P < 0.05). For Cit, the multiple-piecewise linear models did not provide a significantly better fit than simple linear models (all P > 0.05). Overall, serum Cys concentrations exhibited a non-linear positive association with the prevalence of GDM. Moreover, both Cys and Lys demonstrated significant linear negative associations with the risk of GDM (edf ≈ 1; both P-values for association < 0.05; Figure S1B–C). No significant non-linear associations were identified for Val, Tau, Asp, Ile, or Ala in relation to GDM (all P-values for association > 0.05; Figure S1D–H).

Fig. 4.

A-B Smooth curve fitting using GAM to evaluate the nonlinear relationship between serum Met and Cit levels and the risk of GDM. The blue solid line represents the trend of GDM risk with changes in serum amino acid concentrations, and the purple dotted line represents the 95% confidence interval

A non-linear, inverted U-shaped association was observed between Met and GDM (Fig. 5A), with an inflection point at 240.43 identified through threshold effect analysis (Table S3). When Met concentrations were < 240.43, higher Met levels were significantly associated with an increased risk of GDM (β = 0.89, 95% CI: 0.47 to 1.31). In contrast, when Met concentrations were ≥ 240.43, higher Met levels were significantly associated with a decreased risk of GDM (β = -2.00, 95% CI: -3.51 to -0.50).

Sensitivity analyses

We observed similar significant associations between serum Cys and Met levels and the risk of GDM in further analyses, including excluding participants reclassified according to the NICE criteria. However, significant associations between serum Cys and Met levels and GDM were not observed (Figure S2A–B). The significance of the associations between Cit, Lys, and GDM varied depending on the dataset and the diagnostic criteria used.

Discussion

Main findings

In this study, we examined the association between mid-pregnancy serum amino acid profiles and GDM. Serum Cys and Lys were inversely associated with GDM risk, whereas Met and Cit were positively associated. Model evaluation showed that adjusted models provided superior accuracy compared to crude models. Echoing our findings, a case-control study indicates that the accurate diagnosis of metabolic and reproductive disorders can be improved by employing comprehensive multimarker panels alongside appropriate confounder adjustment [20]. However, it is crucial to emphasize that due to the cross-sectional design of our study, the observed associations between serum amino acids and GDM do not imply causality. The altered amino acid levels could either contribute to the development of GDM or be a consequence of the metabolic dysregulation inherent to the disease state.

Amino acid metabolism in pregnancy and GDM

The metabolism of key amino acids undergoes significant adaptation during normal pregnancy. For instance, maternal plasma Cys levels decrease towards the third trimester [21]. However, in the context of GDM, an exaggerated deficiency or an inability to maintain adequate Cys levels may occur, potentially compromising its crucial roles in antioxidant synthesis and redox signaling, which are vital for placental function and counteracting pregnancy-associated oxidative stress. In contrast, in pregnancies complicated by GDM, the homeostasis of other amino acids appears disrupted. Studies have observed elevated levels of methionine Met in the umbilical circulation of GDM pregnancies, suggesting alterations in placental transport or fetal metabolism [22]. Similarly, perturbations in the urea cycle, indicated by altered ratios of citrulline Cit to ornithine, have been associated with an increased risk of GDM, particularly in certain ethnic groups [23]. Our findings of altered mid-pregnancy serum levels of Cys, Met, and Cit align with this paradigm of dysregulated amino acid metabolism in GDM.

Protective amino acids in GDM

Within the constraints of our cross-sectional design, which precludes causal inference, our study identified Cys and Lys as potential protective factors against GDM. Cys is a sulfur-containing amino acid essential for synthesis of the antioxidant glutathione [24, 25]. Our results are consistent with previous findings from both South Asian and Chinese cohorts, which reported inverse associations with GDM risk [23, 26]. As a precursor of glutathione (GSH), Cys plays a critical role in redox homeostasis. Reduced Cys may increase oxidative stress, impair insulin signaling, and contribute to GDM pathogenesis [25, 27, 28]. While the role of Lys appears more complex and somewhat inconsistent across general populations, its potential protective effect in our study may be interpreted within the specific physiological context of pregnancy. Gestation is characterized by progressive insulin resistance and heightened metabolic demand, requiring adaptive β-cell expansion and sustained insulin sensitivity. Lys, through its metabolite 2-aminoadipic acid (2-AAA), has been shown to modulate insulin secretion and sensitivity in experimental models [29–32]. Although moderate exogenous 2-AAA was found to stimulate the transcription of key genes involved in glucose uptake (Glut2) and glucose metabolism (Gck, Pcx), thereby forming a feedforward regulatory loop that sustains β-cell function under metabolic stress [33]. High levels of 2-AAA may impair insulin signaling in insulin-responsive tissues such as liver, skeletal muscle, and adipocytes [34]. Collectively, our findings suggest Lys could exert protective effects, particularly in the context of gestational insulin resistance, though further clinical and mechanistic validation is warranted.

Adverse amino acids in GDM

Two amino acids, Met and Cit, were identified as potential risk biomarkers for GDM in our study. Met is a sulfur-containing amino acid, is central to transsulfuration and one-carbon pathways, which regulate oxidative stress and inflammation. Elevated Met levels, observed in early pregnancy and umbilical circulation in GDM [26], could disrupt redox balance and promote inflammation [35]. Interestingly, our analysis revealed a threshold effect at approximately 240.43 µmol/L. Met levels below this threshold were associated with increased GDM risk, whereas higher concentrations showed an inverse association. This suggesting that adequate Met may support enhanced conversion to Cys and GSH, thereby bolstering antioxidant defenses crucial for counteracting the oxidative stress inherent to GDM [36, 37]. Similarly, the positive association between Cit and GDM risk observed here warrants mechanistic interpretation with caution. Our study identified Cit as a potential risk factor for GDM. This finding can be interpreted through its role as the primary endogenous precursor for Arg and, consequently, nitric oxide (NO) synthesis [38]. In normal pregnancy, nitric oxide (NO) supports vasodilation and vascular insulin sensitivity. In GDM, however, elevated Cit may drive excessive NO production. Within the oxidative environment of GDM, this excess NO is scavenged to form peroxynitrite, a damaging oxidant that thus exacerbates insulin resistance and vascular dysfunction [39, 40]. Thus, the positive association we observed may be explained by a dual mechanism: although citrulline-derived arginine can stimulate insulin secretion, its subsequent conversion to harmful oxidants (e.g., peroxynitrite) in the oxidative milieu of GDM appears to dominate, worsening insulin resistance. This shift also clarifies why Cit supplementation, though beneficial in T2DM models [41, 42], fails in GDM and instead correlates with increased risk, given the pronounced endothelial dysfunction characteristic of this condition.

Stratified analyses reveal heterogeneous associations

Sensitivity analyses confirmed Cys and Met as a robust protective factor across diagnostic criteria. No significant association was found between serum Cys and Met concentrations and GDM risk using truncated data. Notably, the limited sample size (n = 160) likely constrained statistical power to detect a robust association. Consequently, larger prospective cohort studies are warranted to validate or refute these findings and to clarify the temporal relationship and potential causal role of these amino acids in GDM pathogenesis.

Strengths and limitations

Strengths of this study include targeted serum AA profiling in mid-pregnancy, robust ML-based feature selection, and rigorous model evaluation. Novel associations were identified, including protective roles of Cys and Lys, risk enhancing effects of Met and Cit, and a non-linear effect of Met. Limitations include the single-center design, modest sample size. Most importantly, the cross-sectional design precludes any causal inference. As noted, the observed alterations in serum amino acid levels could be a consequence of GDM-related metabolic changes rather than a contributing cause, a key direction for future longitudinal research to unravel. Based on our findings, future studies should track the longitudinal changes of these amino acid biomarkers in at-risk populations throughout pregnancy. It would also be valuable to further examine the role of redox status indicators, such as glutathione, glutathione peroxidase, and malondialdehyde, in the metabolic dysregulation associated with GDM. Additionally, expanding biomarker panels to incorporate glucose metabolism‑related markers and inflammatory cytokines could offer deeper insight into the metabolic‑reproductive interface in GDM pathophysiology.

External validation and integration with metabolomics and mechanistic studies are warranted, and residual confounding by diet, lifestyle, or genetics cannot be excluded.

Abbreviations

GDM

Gestational diabetes mellitus

AA

Amino acid

GLM

Generalized linear models

Gly

Glycine

Ala

Alanine

Asp

Aspartic acid

Arg

Arginine

Glu

Glutamic acid

Thr

Threonine

Ser

Serine

Val

Valine

Leu

Leucine

Ile

Isoleucine

Phe

Phenylalanine

Tyr

Tyrosine

Met

Methionine

His

Histidine

Lys

Lysine

Cit

Citrulline

Orn

Ornithine

Cys

Cystine

P-Ser

Phosphorylated serine

Hylys

Hydroxylysine

Tau

Taurine

ML

Machine learning

T2DM

Type 2 diabetes mellitus

BCAA

Branched-chain amino acid

T1DM

Type 1 diabetes mellitus

OGTT

Oral glucose tolerance test

FPG

Fasting plasma glucose

IADPSG

International Association of Diabetes and Pregnancy Study Groups

NICE

National Institute for Health and Care Excellence

TG

Triglycerides

TC

Total cholesterol

HDL-C

High-density lipoprotein cholesterol

LDL-C

Low-density lipoprotein cholesterol

SVM-RFE

Support vector machine–recursive feature elimination

XGBoost

Extreme gradient boosting

GAM

Generalized additive model

ROC

Receiver operating characteristic

NRI

Net reclassification improvement

IDI

Integrated discrimination improvement

DCA

Decision curve analysis

CIC

Clinical impact curves

SBP

Systolic blood pressure

DBP

Diastolic blood pressure

BMI

Body mass index

MAE

Mean absolute error

GSH

Glutathione

2-AAA

2-aminoadipic acid

NO

Nitric oxide

Glut2

Glucose transporter 2

Gck

Glucokinase

Pcx

Pyruvate carboxylase

Authors’ contributions

**Lingling Cui: ** Conceptualization, Writing – review & editing, Data curation, and Methodology. **Ruijie Sun: ** Data curation, Writing – original draft, Investigation. **Xiaoli Fu: ** Formal analysis, Software, Validation. **Yibo Wang** and **Linpu Ji: ** Formal analysis, Investigation, Visualization. **Qiaorui Liu** , **Xiyue Zheng** and **Xinqian Li: ** [Visualization](http:/credit.niso.org/contributor-roles/visualization) , and Validation. **Mengru Song, ** and **Haojie Zhao: ** Investigation and [Writing – review & editing](http:/credit.niso.org/contributor-roles/writing-review-editing) . **Dongmei Xu** and **Hua Ye** : Resources, Funding acquisition and Supervision. **Luying Qin** : Conceptualization, Methodology, Project administration and Supervision.

Funding

This study was supported by the project “Research and development of key technology for personalized assessment of nutritional genetic risk and precise nutritional intervention assessment based on the Central Plains population” (231111311200) as well as by research on “The association between red blood cell folate concentrations in early/mid-pregnancy and maternal and child health outcomes” (252102311111).

Data availability

Some or all datasets generated during and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in strict accordance with the ethical principles of the World Medical Association Declaration of Helsinki and relevant institutional guidelines, and was approved by the Ethics Committee of the Third Affiliated Hospital of Zheng Zhou University (No.2022-143-01). Informed consent was provided by all participants before they were recruited for the study, and data were analyzed anonymously.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.