Maternal metabolomic signatures for risk of preterm and early term birth in African American women

Abstract

Metabolic signals identifying pregnancies at risk for preterm birth (<37 weeks) and early term birth (37–38 weeks) remain limited. Here we used high-resolution metabolomics to characterize metabolic features associated with spontaneous and medically indicated early birth. We conducted metabolome-wide association studies among 279 (discovery) and 251 (internal validation) pregnant women from the Atlanta African American Maternal-Child Cohort (2014–2018), with serum collected in early (8–14 weeks) and later (24–30 weeks) pregnancy. Distinct metabolic profiles differentiated spontaneous and medically indicated preterm birth and early term birth across pregnancy windows. Perturbations in amino acid pathways, including arginine, proline, aspartate, glutamate, methionine and cysteine metabolism, were observed. Valine, leucine and tyrosine were inversely associated with spontaneous early birth, while acylcarnitines and aldosterone were associated with medically indicated early birth. Thirteen metabolites were validated, supporting their potential as markers of pregnancies at elevated risk.

Metabolites associated with preterm and early term birth were identified in a cohort of African American women, 13 of which were validated in an independent internal dataset, with the potential to be used as markers of at-risk women.

Main

Preterm birth (PTB), defined as birth before 37 weeks gestation, affects approximately 10% of infants in the United States¹ and is a leading contributor to infant morbidity and mortality². Surviving preterm infants face an elevated risk of health complications across the lifespan³. PTB is a heterogeneous outcome and can be phenotypically classified as spontaneous preterm birth (sPTB), if preceded by spontaneous labour or spontaneous pre-labour rupture of membranes, or as medically indicated preterm birth (miPTB), if a maternal or fetal complication necessitated early delivery by labour induction and/or pre-labour C-section. sPTB and miPTB account for approximately two-thirds and one-third of all PTBs, respectively⁴. Importantly, racial and ethnic disparities in rates of PTB have been intractable over past decades⁵, with non-Hispanic Black individuals experiencing PTB at rates nearly double that of non-Hispanic White individuals¹. Racial and ethnic minorities also experience higher rates of early term birth (ETB), defined as birth occurring between 37 completed weeks and less than 39 weeks⁶. In the United States, approximately 27% of infants are born early term and these infants are at higher risk for morbidities and mortality during childhood relative to their full-term counterparts⁷. Epidemiological studies have identified maternal and fetal factors associated with increased risk of early birth in the preterm and early term gestational age categories, including maternal socioeconomic status (SES), age, nutritional status, pregnancy complications, social stress, infection and genetic markers⁴. However, the underlying aetiology of the various early birth phenotypes remains imperfectly understood^4,8, and biomarkers (single or multiple) that serve to identify pregnancies at elevated risk for early birth and/or direct targeted preventive strategies remain elusive^9,10.

Biological fluids including amniotic fluid, cervical mucus, urine, blood and saliva have served as sources of biomarkers for the prediction of PTB^11,12. In these studies, numerous analytes have been associated with PTB such as biomarkers related to inflammation and oxidative stress, including cytokines, chemokines, oestriol and isoprostanes^13,14. However, none (singly or in combination) have proven suitable for the clinical prediction of PTB given lack of sensitivity, specificity and reproducibility^15,16. More recently, high-resolution metabolomics has emerged as an innovative tool to comprehensively analyse endogenous metabolites in biological samples to identify and validate specific metabolic features and pathways that could respond to exposures, stressors, diseases or therapeutic treatments¹⁷. Recent prenatal metabolomics studies have revealed potential metabolic signals and metabolic pathways perturbed in pregnancies ending in PTB, including pathways related to energy metabolism¹⁸, choline and amino acid/gut microflora metabolism¹⁹, lipid, pyruvate, tryptophan and vitamin metabolism²⁰, and xenobiotics²¹. A study from amniotic fluid of African American women found associations between sPTB and changes in fatty acid metabolism, coenzyme A (CoA) metabolism, histidine metabolism and cytochrome P450-related activity²². However, most studies have observed inconsistent associations with specific metabolic perturbations in relation to sPTB^23,24, and less is known about the pathways and molecules that may be involved in miPTB and ETB, underscoring the need for additional metabolomic studies to discover early indicators of PTB and ETB.

To fill these knowledge gaps, we performed a comprehensive metabolome-wide association study (MWAS) using advanced untargeted high-resolution metabolomics on maternal serum samples collected during early and later pregnancy from participants in the established Atlanta African American Maternal-Child Cohort. The objective was to identify novel metabolites and metabolic pathways associated with spontaneous and medically indicated early birth (PTB and/or ETB).

Results

A total of 279 individuals were included, comprising 64 PTB (19.3%), 54 ETB (16.3%) and 161 healthy full-term birth (HFTB; referent group; Table 1) participants. Among HFTB participants, the mean age at enrolment was 25.2 ± 4.8 years; over half were married or cohabitating (54.0%) and had more than a high school education (54.0%), and most were covered by Medicaid during pregnancy (72.7%). In contrast, a majority of ETB and PTB participants had a high school education or less (62.9% and 65.7%, respectively) and were covered by Medicaid (85.2% and 95.3%, respectively). Alcohol and tobacco use in the month before pregnancy were reported by 7.4% and 16.7% of ETB participants and 6.2% and 18.8% of PTB participants, respectively. Pregnancy complications included gestational hypertension (14.8% and 12.5%), pre-eclampsia (13.0% and 7.8%), gestational diabetes (3.7% and 1.6%) and intrauterine growth restriction (14.8% and 14.1%) among ETB and PTB groups, respectively. Significant differences across groups were observed for prior PTB, prenatal complications and socioeconomic indicators, including education, pregnancy insurance and hospital site (Table 1). Participant selection is shown in Extended Data Fig. 1, and characteristics were comparable to the larger cohort (Supplementary Table 1).

Table 1.

Characteristics of pregnant participants from the Atlanta African American Maternal-Child Cohort by birth outcome category, 2014–2018

Characteristics	HFTB (N = 161)	ETBa (N = 54)	PTBb (N = 64)	P valuec
Age (years)
Mean ± s.d.	25.24 ± 4.77	25.15 ± 4.84	24.25 ± 4.39	0.35
Married or cohabiting
Yes	87 (54.0%)	24 (44.4%)	39 (60.9%)	0.18
No	74 (46.0%)	30 (55.6%)	25 (39.1%)
Education level
Less than high school	21 (13.0%)	14 (25.9%)	12 (18.8%)	0.02
High school	53 (32.9%)	20 (37.0%)	30 (46.9%)
Some college	54 (33.5%)	15 (27.8%)	18 (28.1%)
College and above	33 (20.5%)	5 (9.3%)	<5 (6.2%)
Income-to-poverty ratio (%)
<100	70 (43.5%)	31 (57.4%)	37 (57.8%)	0.16
100–150	30 (18.6%)	10 (18.5%)	14 (21.9%)
150–300	35 (21.7%)	7 (13.0%)	9 (14.1%)
≥300	26 (16.1%)	6 (11.1%)	<5 (6.3%)
Pregnancy insurance
Private	44 (27.3%)	8 (14.8%)	<5 (4.7%)	<0.01
Medicaid	117 (72.7%)	46 (85.2%)	61 (95.3%)
Hospital site
Grady Memorial	82 (50.9%)	36 (66.7%)	46 (71.9%)	0.01
Emory Midtown	79 (49.1%)	18 (33.3%)	18 (28.1%)
Alcohol use in month before pregnancy
Yes	13 (8.1%)	<5 (7.4%)	<5 (6.2%)	0.90
No	148 (91.9%)	50 (92.6%)	60 (93.8%)
Tobacco use in month before pregnancy
Yes	19 (11.8%)	9 (16.7%)	12 (18.8%)	0.35
No	142 (88.2%)	45 (83.3%)	52 (81.2%)
Nulliparous
Yes	73 (45.3%)	16 (29.6%)	27 (42.2%)	0.13
No	88 (54.7%)	38 (70.4%)	37 (57.8%)
First prenatal BMI (kg/m2)
Mean ± s.d.	28.19 ± 7.17	28.08 ± 8.17	27.25 ± 6.96	0.39
Gestational age at 1st study visit (weeks)
Mean ± s.d.	11.36 ± 2.14	11.63 ± 2.30	11.62 ± 2.32	0.62
Gestational age at 2nd study visit (weeks)
Mean ± s.d.	26.74 ± 2.82	26.72 ± 2.40	26.48 ± 1.96	0.78
Gestational age at delivery (weeks)
Mean ± s.d.	39.79 ± 1.09	37.59 ± 0.51	33.55 ± 3.83	<0.01
Sex of infant
Male	76 (47.2%)	26 (48.1%)	39 (60.9%)	0.16
Female	85 (52.8%)	28 (51.9%)	25 (39.1%)
Prior PTB
Yes	6 (3.7%)	9 (16.7%)	9 (14.1%)	<0.01
No	153 (95.0%)	43 (79.6%)	48 (75.0%)
Intrauterine growth restriction
Yes	0 (0.0%)	8 (14.8%)	9 (14.1%)	<0.01
No	161 (100.0%)	46 (85.2%)	55 (85.9%)
Gestational hypertension
No	161 (100.0%)	39 (72.2%)	55 (85.9%)	<0.01
Gestational hypertension without pre-eclampsia	0 (0.0%)	8 (14.8%)	8 (12.5%)
Pre-eclampsia	0 (0.0%)	7 (13.0%)	5 (7.8%)
Gestational diabetes
Yes	0 (0.0%)	<5 (3.7%)	<5 (1.6%)	0.04
No	161 (100.0%)	52 (96.3%)	63 (98.4%)

^aAmong ETBs, 40 participants had sETB, and 14 participants had miETB.

^bAmong PTBs, 48 participants had sPTB, and 16 participants had miPTB.

^cP values were calculated using one-way analysis of variance for continuous variables and two-sided chi-squared (χ²) tests or Fisher’s exact tests (when expected cell counts <5) for categorical variables. All tests were two-sided. Exact P values are reported.

Extended Data Fig. 1. Derivation of the analytic sample from all births within the Atlanta African American Maternal-Child Cohort, 2014-2018.

Abbreviations: sPTB=Spontaneous preterm birth; miPTB=Medically indicated preterm birth; sETB=Spontaneous early term birth; miETB=Medically indicated early term birth; sPETB=Spontaneous preterm/early term birth; miPETB=Medically indicated preterm/early term birth.

Source data

Metabolic profiling according to early birth phenotype

We extracted 13,616 and 11,900 metabolic features from serum samples using hydrophilic interaction liquid chromatography (HILIC) with positive electrospray ionization (ESI) and C18 hydrophobic reversed-phase chromatography with negative ESI, respectively. Significant features associated with early birth phenotypes varied by gestational timing, analytical mode and phenotype (Table 2). We assessed overlapping significant features across these two prenatal time points of measurement (denoted as ‘early–late overlap’), with the following overall findings shown in Table 2.

Table 2.

Number of metabolites in the early (8–14 weeks) and later (24–30 weeks) pregnancy maternal metabolome associated with the early birth phenotypes under different q-value and P-value cut-offs^a, the Atlanta African American Maternal-Child Cohort (2014–2018)

Early birth phenotype	Prenatal timing of metabolome measurement	Number of associated metabolites from HILIC column					Number of associated metabolites from hydrophobic reversed-phase chromatography (C18) columns
		FDR q < 0.2	Raw P < 0.0005	Raw P < 0.005	Raw P < 0.01	Raw P < 0.05	FDR q < 0.2	Raw P < 0.0005	Raw P < 0.005	Raw P < 0.01	Raw P < 0.05
PTB
sPTB	Early	0	6	32	91	625	0	4	41	100	552
	Late	2	4	40	79	547	2	4	27	57	497
	Early–late overlap	0	0	0	0	32	0	0	1	2	33
miPTB	Early	1	2	41	80	572	0	1	29	78	479
	Late	3	3	26	67	603	2	3	21	56	546
	Early–late overlap	0	0	0	0	27	0	0	0	1	28
Overlap between sPTB and miPTB	Early	0	0	0	2	40	0	0	0	0	34
	Late	0	0	0	0	27	0	0	0	0	32
	Early–late overlap	0	0	0	0	0	0	0	0	0	0
ETB
sETB	Early	0	10	95	178	896	0	2	60	126	619
	Late	0	5	70	167	860	1	3	37	104	605
	Early–late overlap	0	0	1	2	53	0	0	0	1	37
miETB	Early	3	3	41	105	643	2	2	27	73	531
	Late	5	5	21	43	487	0	1	26	61	458
	Early–late overlap	0	0	0	0	23	0	0	1	3	28
Overlap between sETB and miETB	Early	0	0	0	1	52	0	0	0	1	45
	Late	0	0	1	1	47	0	0	0	4	33
	Early–late overlap	0	0	0	0	0	0	0	0	0	0
Early birth (combined preterm and early term birth)
sPETB	Early	0	10	87	167	860	0	13	84	142	665
	Late	0	4	51	101	700	0	0	27	68	521
	Early–late overlap	0	0	0	1	55	0	0	0	0	35
miPETB	Early	0	2	45	100	651	1	1	58	107	539
	Late	5	6	29	67	557	0	11	56	98	419
	Early–late overlap	0	0	0	2	28	0	0	0	0	19
Overlap between sPTB and sETB	Early	0	0	0	9	76	0	0	6	10	82
	Late	0	0	2	3	32	0	0	0	0	29
	Early–late overlap	0	0	0	0	0	0	0	0	0	0
Overlap between miPTB and miETB	Early	0	0	0	2	40	0	0	0	0	28
	Late	0	0	0	0	25	0	0	0	0	20
	Early–late overlap	0	0	0	0	0	0	0	0	0	1

^aAssociations were estimated using two-sided logistic regression models. P values are reported as exact values. Both Benjamini–Hochberg FDR procedure (q value) and raw P values were used to identify a reasonable number of significant metabolic features.

A number of metabolites in the early and later pregnancy metabolome were significantly associated with each early birth phenotype considering an unadjusted P < 0.05, with substantially fewer significant features when applying multiple comparisons (false discovery rate (FDR) of 20%) for P values.

For each of the four early birth phenotypes (sPTB, miPTB, spontaneous early term birth (sETB) and medically indicated early term birth (miETB)), there was a difference in the number of significant metabolites associated with the phenotype in the early and later maternal metabolome with a limited number of overlapping significant metabolites found in both the early and later maternal metabolome (‘early–late overlap’) for a given early birth phenotype.

In examining the overlap between the spontaneous and medically indicated births within a gestational age category (that is, overlap between sPTB and miPTB, overlap between sETB and miETB), there were relatively few significant metabolites that overlapped across labour types within the early and later pregnancy metabolome and none that overlapped across labour types and across the early and later pregnancy metabolome.

In examining the composite early birth categories (combining sPTB and sETB and miPTB and miETB), there were a similar number of significant metabolites in the early and later maternal metabolome with some overlapping significant metabolites in both the early and later maternal metabolome (‘early–late overlap’), but far fewer significant metabolites overlapped between the early birth phenotypes (that is, overlap between sPTB and sETB, overlap between miPTB and miETB) within the early and later maternal metabolomes, with no overlapping significant metabolites across the early and later pregnancy metabolome.

For example, for sPTB (raw P < 0.05), 625 and 552 metabolites were identified in early pregnancy and 547 and 497 in later pregnancy (HILIC and C18, respectively), with only 32 and 33 overlapping features across the early and later maternal metabolome (‘early–late overlap’). Overlap between sPTB and miPTB, as well as between sPTB and sETB, was limited within each time window and absent across both time points. Similar patterns were observed for medically indicated phenotypes, with modest within-time overlap but no early–late overlap.

Metabolic pathway enrichment analysis

Using pathway enrichment analysis, we observed that approximately 60% of enriched pathways overlapped between early (Fig. 1a) and later (Fig. 1b) pregnancy, with 22 shared pathways among 37 and 39 significant pathways identified at each time point, respectively. These overlapping pathways were primarily enriched in amino acid metabolism, including histidine; glycine, serine, alanine and threonine; alanine and aspartate; and glutamate metabolism, as well as fatty acid and nucleotide pathways. The top enriched pathways in early pregnancy included butanoate metabolism, vitamin C metabolism and histidine metabolism, whereas in later pregnancy, lysine metabolism, urea cycle/amino group metabolism and glutamate metabolism were prominent. Overall, pathway associations were stronger (that is, smaller P values) in later pregnancy compared with early pregnancy, particularly in the C18 column.

Fig. 1. Metabolic pathways in the early (8–14 weeks) and later (24–30 weeks) pregnancy maternal metabolome associated with the early birth phenotypes in the two technical columns (HILIC, C18).

a, Metabolic pathways in the early pregnancy (8–14 weeks) metabolome associated with spontaneous and medically indicated PTB, ETB and early birth (combined preterm and early term birth). b, Metabolic pathways in the later pregnancy (24–30 weeks) metabolome associated with spontaneous and medically indicated PTB, ETB and early birth (combined preterm and early term birth). The size and colour of the bubbles represent the pathway significant level based on −log₁₀(P value). Pathway enrichment significance was evaluated using mummichog with a permutation-based approach and the EASE score, a conservative modification of Fisher’s exact test. Adjusted P values were calculated from the permutation-derived null distribution and used to assess pathway significance. All tests were two-sided, and pathways with adjusted P < 0.05 were considered significant. Only significant associations (P < 0.05) were indicated by bubbles in the plots. The percentages of the number of significant putative metabolites (overlap size) to the number of metabolites within each pathway (pathway size) less than 30% were excluded. For HILIC-positive ion mode, only the following adducts were considered: M¹⁺, M + H¹⁺, M-H₂O + H¹⁺, M + Na¹⁺, M + K¹⁺, M + 2H²⁺ and M(C13) + 2H²⁺. For C18-negative ion mode, only the following adducts were considered: M-H⁻, M + Cl⁻, M + ACN-H⁻, M + HCOO⁻, M(C13)–H⁻, M-H₂O–H⁻ and M + Na–2H⁻. sPTBH, spontaneous preterm birth versus heathy full-term birth; miPTBH, medically indicated preterm birth versus healthy full-term birth; sETBH, spontaneous early term birth versus heathy full-term birth; miETBH, medically indicated early term birth versus healthy full-term birth; sPETBH, spontaneous preterm/early term birth versus healthy full-term birth; miPETBH, medically indicated preterm/early term birth versus healthy full-term birth.

Source data

Distinct pathway patterns were observed between spontaneous (sPTB, sETB and spontaneous preterm/early term birth (sPETB)) and medically indicated (miPTB, miETB and medically indicated preterm/early term birth (miPETB)) early birth phenotypes. Spontaneous early birth was associated with a greater number of significant pathways and stronger statistical evidence compared with medically indicated outcomes. Notably, amino acid metabolism pathways, including arginine and proline; aspartate and asparagine; and glutamate metabolism in early pregnancy, and methionine and cysteine metabolism in later pregnancy, were consistently associated with spontaneous early birth phenotypes.

Metabolite identification and confirmation

We identified 39 level 1 confirmed metabolites associated with early birth phenotypes across early and later pregnancy (Table 3). Consistent with pathway findings, most metabolites differ between spontaneous and medically indicated early birth. Valine, itaconate, 4-methyl-2-oxovaleric acid, leucine, methionine and tyrosine were inversely associated with sPTB and sPETB (P < 0.05), but not with miPTB, indicating potentially sensitive metabolic signals (Table 3 and Fig. 2a). Receiver operating characteristic (ROC) analyses indicated modest discrimination (area under the curve (AUC) > 0.5), with the highest performance for 4-methyl-2-oxovaleric acid (AUC = 0.74) and leucine (AUC = 0.72; Fig. 3). A multi-metabolite panel of these six metabolites showed slightly improved predictive performance (AUC = 0.77) compared with individual metabolites (Extended Data Fig. 2). In complementary random forest analyses, these metabolites consistently ranked among the most influential predictors (Extended Data Fig. 3).

Table 3.

Chemical identity of metabolites in the early (8–14 weeks) and later (24–30 weeks) pregnancy maternal metabolome associated with the early birth phenotypes (P < 0.05)

m/z	RT(s)	Metabolite	Adduct	Column	Metabolic pathway	Prenatal timing of metabolome measurement	Adjusted odds ratio (95% CI)c
							PTB		ETB		Early birth
							sPTBH	miPTBH	sETBH	miETBH	sPETBH	miPETBH
101.0246	22.5	2-Oxobutyrate	M-H	C18	Glycine, serine and threonine metabolism/cysteine and methionine metabolism/valine, leucine and isoleucine biosynthesis	Early	0.96 (0.51, 1.80)	2.31 (0.83, 6.44)	0.81 (0.66, 0.99)	1.08 (0.36, 3.26)	0.68 (0.40, 1.16)	1.73 (0.76, 3.91)
104.1072	37.7	Choline	M+	HILIC	Glycerophospholipid metabolism	Early	0.68 (0.25, 1.85)	2.92 (0.55, 15.45)	0.33 (0.12, 0.92)	0.42 (0.06, 3.06)	0.50 (0.23, 0.10)	1.40 (0.37, 5.28)
116.0718	24.7	L-Valine	M-H	C18	Valine, leucine and isoleucine degradation	Early	0.26 (0.09, 0.75)	0.59 (0.13, 2.66)	0.26 (0.09, 0.79)	1.28 (0.24, 6.72)	0.28 (0.12, 0.64)	0.89 (0.27, 2.87)
		5-Aminopentanoate			Lysine degradation/arginine and proline metabolism
126.022	57.7	Taurine	M + H	HILIC	Primary bile acid biosynthesis/taurine and hypotaurine metabolism	Early	0.78 (0.39, 1.55)	0.78 (0.24, 2.52)	0.42 (0.18, 0.97)	0.66 (0.20, 2.23)	0.61 (0.34, 1.09)	0.74 (0.30, 1.83)
128.0355	20.9	5-Oxo-D-proline	M-H	C18	D-Amino acid metabolism	Late	0.43 (0.22, 0.85)	2.24 (0.65, 7.73)	0.80 (0.40, 1.60)	0.49 (0.20, 1.24)	0.55 (0.32, 0.96)	0.90 (0.40, 1.99)
129.0195	22.1	Itaconatea	M-H	C18	C5-branched dibasic acid metabolism (TCA cycle)	Early	0.53 (0.27, 1.03)	0.58 (0.20, 1.65)	0.57 (0.25, 1.27)	0.42 (0.15, 1.19)	0.55 (0.31, 0.98)	0.48 (0.21, 1.08)
						Late	0.46 (0.23, 0.94)	0.94 (0.19, 4.67)	0.67 (0.30, 1.48)	0.54 (0.15, 1.94)	0.53 (0.29, 0.95)	0.65 (0.23, 1.84)
129.0558	21.1	4-Methyl-2-oxovaleric acid	M-H	C18	Valine, leucine and isoleucine degradation	Early	0.23 (0.09, 0.63)	0.86 (0.21, 3.58)	0.41 (0.15, 1,14)	1.13 (0.24, 5.26)	0.31 (0.15, 0.68)	0.99 (0.33, 2.93)
130.0874	23.9	Leucine	M-H	C18	Valine, leucine and isoleucine degradation	Early	0.20 (0.07, 0.57)	0.58 (0.13, 2.59)	0.18 (0.06, 0.57)	1.71 (0.38, 7.59)	0.20 (0.09, 0.47)	0.99 (0.33, 3.00)
132.0656	65	N-Acetylalanine	M + H	HILIC	Alanine, aspartate and glutamate metabolism	Early	0.79 (0.49, 1.27)	0.66 (0.31, 1.44)	0.54 (0.34, 0.88)	0.55 (0.23, 1.29)	0.66 (0.46, 0.95)	0.67 (0.37, 1.20)
		Hydroxyproline			Arginine and proline metabolism
134.0448	69.4	L-Aspartate	M + H	HILIC	Arginine biosynthesis/alanine, aspartate and glutamate metabolism/glycine, serine and threonine metabolism	Late	0.94 (0.61, 1.46)	0.90 (0.42, 1.92)	1.73 (1.11, 2.70)	1.17 (0.54, 2.54)	1.20 (0.84, 1.71)	1.01 (0.57, 1.77)
135.0304	19.7	Hypoxanthine	M-H	C18	Purine metabolism	Late	0.44 (0.20, 0.99)	2.32 (0.61, 8.78)	1.03 (0.46, 2.28)	0.89 (0.29, 2.74)	0.66 (0.36, 1.24)	1.31 (0.53, 3.23)
143.1077	25.4	Caprylic acid	M-H	C18	Fatty acid biosynthesis	Late	0.79 (0.54, 1.16)	0.54 (0.32, 0.90)	1.09 (0.76, 1.57)	1.16 (0.63, 2.12)	0.93 (0.70, 1.23)	0.76 (0.51, 1.14)
149.0232	22.3	Phthalic anhydride	M + H	HILIC	Metabolism of xenobiotics by cytochrome P450	Late	1.06 (0.62, 1.79)	1.28 (0.57, 2.85)	0.57 (0.38, 0.84)	0.59 (0.27, 1.26)	0.70 (0.51, 0.97)	0.82 (0.50, 1.34)
150.0583	47.1	L-Methionine	M + H	HILIC	Cysteine and methionine metabolism	Early	0.42 (0.19, 0.95)	0.55 (0.51, 1.98)	0.45 (0.19, 1.09)	1.45 (0.35, 5.92)	0.44 (0.23, 0.86)	0.89 (0.33, 2.43)
153.0193	21.6	2,3-Dihydroxybenzoate	M-H	C18	Benzoate degradation/xenobiotics biodegradation and metabolism	Early	0.96 (0.70, 1.33)	0.96 (0.59, 1.58)	1.04 (0.76, 1.41)	0.46 (0.26, 0.84)	1.00 (0.79, 1.28)	0.72 (0.49, 1.06)
153.0396	115.1	Xanthine	M + H	HILIC	Purine metabolism	Early	0.92 (0.69, 1.22)	0.75 (0.48, 1.17)	0.76 (0.56, 1.03)	0.57 (0.37, 0.88)	0.85 (0.68, 1.06)	0.66 (0.48, 0.92)
164.0716	25	L-Phenylalanine	M-H	C18	Phenylalanine metabolism/phenylalanine, tyrosine and tryptophan biosynthesis	Late	1.34 (0.31, 5.72)	0.56 (0.04, 8.63)	4.85 (1.05, 22.47)	2.62 (0.21, 32.52)	2.30 (0.76, 6.96)	1.39 (0.20, 9.45)
165.0546	49.7	Phenylpyruvate	M + H	HILIC	Phenylalanine metabolism	Early	0.51 (0.17, 1.49)	0.81 (0.16, 4.25)	0.27 (0.09, 0.82)	0.82 (0.15, 4.61)	0.37 (0.15, 0.89)	0.84 (0.24, 3.03)
167.0208	18.8	Urate	M-H	C18	Purine metabolism	Late	0.62 (0.20, 1.90)	18.59 (2.04, 168.99)	1.89 (0.59, 6.06)	0.53 (0.10, 2.73)	1.02 (0.42, 2.47)	2.21 (0.60, 8.10)
174.0885	20.9	Citrulline	M-H	C18	Arginine biosynthesis	Late	0.36 (0.15, 0.85)	2.08 (0.35, 12.50)	0.73 (0.31, 1.75)	0.55 (0.13, 2.29)	0.49 (0.25, 0.97)	0.95 (0.30, 2.93)
175.0248	18.7	Vitamin C	M-H	C18	Ascorbate and aldarate metabolism	Early	0.84 (0.67, 1.07)	0.73 (0.52, 1.01)	0.72 (0.57, 0.90)	0.79 (0.55, 1.14)	0.79 (0.65, 0.94)	1.05 (0.70, 1.58)
		D-Glucuronolactone
180.0665	23.3	L-Tyrosineb	M-H	C18	Tyrosine metabolism/phenylalanine metabolism/ubiquinone and other terpenoid-quinone biosynthesis	Early	0.20 (0.06, 0.60)	1.21 (0.29, 5.07)	0.31 (0.10, 0.92)	1.35 (0.28, 6.58)	0.27 (0.12, 0.63)	1.28 (0.41, 3.97)
182.0812	50.3	(L)-Tyrosineb	M + H	HILIC	Tyrosine metabolism/phenylalanine metabolism/ubiquinone and other terpenoid-quinone biosynthesis	Early	0.42 (0.18, 0.96)	1.13 (0.32, 3.94)	0.28 (0.12, 0.66)	0.96 (0.25, 3.70)	0.35 (0.18, 0.68)	1.08 (0.40, 2.86)
187.1338	24	10-Hydroxydecanoate	M-H	C18	Fatty acid metabolism	Early	0.56 (0.34, 0.94)	1.21 (0.53, 2.89)	1.14 (0.67, 1.92)	0.80 (0.32, 2.02)	0.80 (0.55, 1.18)	1.04 (0.54, 1.99)
191.0196	21.5	Citric acida	M-H	C18	Citrate cycle (TCA cycle)/alanine, aspartate and glutamate metabolism/glyoxylate and dicarboxylate metabolism	Early	0.78 (0.62, 0.99)	0.81 (0.32, 2.04)	0.65 (0.34, 1.26)	0.70 (0.51, 0.97)	0.82 (0.69, 0.99)	0.60 (0.32, 1.14)
						Late	0.64 (0.41, 1.00)	1.47 (0.47, 4.63)	0.80 (0.49, 1.31)	0.69 (0.33, 1.46)	0.69 (0.47, 1.00)	0.88 (0.47, 1.68)
197.0434	22.3	3-Methoxy-4-hydroxymandelate	M-H	C18	Tyrosine metabolism	Early	0.36 (0.16, 0.80)	1.37 (0.39, 4.83)	0.56 (0.24, 1.33)	1.02 (0.27, 3.90)	0.44 (0.23, 0.84)	1.23 (0.47, 3.23)
200.1282	39.2	Methyl ecgoninea	M + H	HILIC	Tropane, piperidine and pyridine alkaloid biosynthesis	Early	1.61 (1.11, 2.34)	1.04 (0.71, 1.51)	1.02 (0.79, 1.33)	0.83 (0.59, 1.17)	1.22 (0.97, 1.54)	0.92 (0.70, 1.20)
						Late	0.97 (0.62, 1.53)	2.99 (1.00, 9.58)	0.91 (0.64, 1.29)	0.21 0.05, 0.90)	0.93 (0.70, 1.24)	1.08 (0.70, 1.65)
260.1855	28.2	Hexanoylcarnitine	M + H	HILIC	Fatty acid metabolism	Late	0.92 (0.63, 1.34)	1.08 (0.59, 1.96)	1.12 (0.79, 1.58)	0.48 (0.25, 0.93)	1.03 (0.78, 1.34)	0.72 (0.48, 1.09)
267.0732	22.9	Inosine	M-H	C18	Purine metabolism	Late	1.49 (1.03, 2.15)	2.59 (1.22, 5.49)	0.89 (0.68, 1.16)	1.09 (0.59, 2.00)	1.10 (0.88, 1.38)	1.63 (1.00, 2.66)
315.2331	24.6	Progesteronea	M + H	HILIC	Steroid hormone biosynthesis	Early	1.22 (0.70, 2.12)	0.68 (0.30, 1.57)	1.88 (1.03, 3.43)	1.39 (0.55, 3.54)	1.54 (0.97, 2.42)	0.92 (0.48, 1.74)
						Late	1.85 (0.83,4.15)	1.51 (0.37, 6.09)	2.26 (1.04, 4.89)	0.70 (0.17, 2.82)	2.10 (1.14, 3.90)	1.01 (0.36, 2.77)
316.248	24.3	Decanoylcarnitine	M + H	HILIC	Fatty acid degradation	Late	0.84 (0.67, 1.04)	3.21 (1.25, 8.20)	1.50 (1.07, 2.10)	1.00 (0.60, 1.67)	1.01 (0.84, 1.21)	1.41 (0.92, 2.17)
344.2792	23.7	Lauroylcarnitine	M + H	HILIC	Fatty acid metabolism	Late	1.12 (0.81, 1.54)	2.18 (1.13, 4.22)	1.21 (0.88, 1.67)	0.99 (0.56, 1.77)	1.13 (0.89, 1.43)	1.42 (0.90, 2.24)
347.2215	24.5	Cortexolonea	M + H	HILIC	Steroid hormone biosynthesis	Early	1.50 (1.04, 2.15)	2.48 (1.30, 4.73)	0.72 (0.53, 0.98)	0.69 (0.44, 1.08)	0.98 (0.77, 1.24)	1.15 (0.81, 1.62)
						Late	0.89 (0.59, 1.34)	0.88 (0.44, 1.78)	2.42 (1.37, 4.29)	0.76 (0.46, 1.26)	1,31 (0.92, 1.86)	0.82 (0.54, 1.25)
357.3003	23.6	1-Oleoyl-RAC-glycerola	M + H	HILIC	Fatty acid biosynthesis	Early	1.41 (0.78, 2.56)	0.94 (0.36, 2.42)	2.08 (1.10, 3.93)	1.39 (0.99, 1.95)	1.67 (1.04, 2.70)	1.54 (0.75, 3.18)
						Late	1.71 (0.90, 3.26)	2.50 (0.78, 8.00)	1.64 (0.92, 2.92)	2.62 (0.93, 7.37)	1.60 (1.00, 2.55)	2.64 (1.17, 5.95)
361.201	25	Aldosterone	M + H	HILIC	Steroid hormone biosynthesis	Early	1.07 (0.61, 1.87)	0.44 (0.22, 0.87)	1.18 (0.64, 2.16)	0.60 (0.32, 1.12)	1.12 (0.71, 1.77)	0.59 (0.36, 0.94)
385.3466	21.8	25-Hydroxycholesterol	M-H2O + H	HILIC	Primary bile acid biosynthesis	Late	1.63 (0.77, 3.49)	0.52 (0.13, 2.04)	2.93 (1.33, 6.47)	1.16 (0.33, 4.04)	2.31 (1.27, 4.21)	0.83 (0.33, 2.09)
391.2842	22	BIS(2-ethylhexyl)phthalate	M + H	HILIC	Xenobiotics biodegradation and metabolism	Early	2.51 (0.87, 7.26)	1.26 (0.66, 2.40)	1.49 (0.94, 2.37)	1.15 (0.57, 2.32)	1.63 (1.07, 2.49)	1.23 (0.77, 1.96)

Chemical identity of metabolic features was confirmed by matching peaks via accurate mass-to-charge ratio (m/z) and retention time to authentic reference standards under the same conditions using tandem mass spectrometry. Associations were estimated using two-sided logistic regression models. Exact P values are reported, with P < 0.05 considered statistically significant and indicated in bold.

C18, C18 column; CI, confidence interval; RT, retention time.

^aMetabolites were identified in both early and later pregnancy.

^bMetabolites were identified in both HILIC and C18 column.

^cAdjusted for maternal age, education, first prenatal BMI, parity, infant sex and substance use in pregnancy.

Fig. 2. The metabolic mechanisms for confirmed metabolites significantly associated with early birth phenotypes versus healthy full-term birth.

a, sPTB versus HFTB. b, miPTB versus HFTB or miETB versus HFTB. c, sPTB versus HFTB, sETB versus HFTB or sPETB versus HFTB. Box plots display metabolite concentrations (μmol l⁻¹) or normalized intensities measured in maternal serum. Each point represents an individual participant (biological replicate) among HFTB (n = 161) and sPTB (n = 48). No technical replicates were used for statistical analysis; each measurement corresponds to a unique biological sample. Box plots show the median (centre line), interquartile range (box), mean (centre dot) and whiskers extending to 1.5 times the interquartile range from the quartiles; data points beyond this range are shown as outliers. Metabolites in blue denote a negative association, with light blue identified in early pregnancy and dark blue identified in later pregnancy. Metabolites in orange denote a positive association, with light orange identified in early pregnancy and dark orange identified in later pregnancy. The three cells under each metabolite represent the association between each metabolite with each of the three early birth phenotypes versus healthy full-term birth: (from left to right: sPTBH, sPETBH, sETBH), and the colours were shaded according to the P value of this association. An asterisk denotes the metabolites identified in both early and late pregnancy and the caret denotes metabolites identified in both HILIC and C18 column. Serum concentrations (μmol l⁻¹) of metabolites were measured via reference standardization of metabolomic signal intensities. Among these metabolites, leucine concentrations were missing in NIST reference materials, so we reported the metabolite intensity from liquid chromatography coupled with high-resolution mass spectrometry. NADPH, nicotinamide adenine dinucleotide phosphate; AMP, adenosine monophosphate; ATP, triphosphate.

Source data

Fig. 3. The ROC curves of six sensitive metabolic signals for sPTB.

In the following birth phenotype models, preterm birth versus healthy full-term birth, spontaneous preterm birth versus healthy full-term birth and spontaneous preterm/early term birth versus healthy full-term birth, six metabolites showed significant associations across these models and are considered potentially sensitive markers for spontaneous preterm birth. FPR, false positive rate.

Source data

Extended Data Fig. 2. The receiver operating characteristic (ROC) curve of six promising metabolic signals for spontaneous preterm birth.

All six metabolites were simultaneously fitted in logistic regression classifier and assessed discrimination. In following birth phenotype models: preterm birth (N = 64) vs heathy full term birth (N = 161), spontaneous preterm birth (N = 48) vs heathy full term birth (N = 161), and spontaneous preterm/early term birth (N = 88) vs healthy full term birth (N = 161), six metabolites showed significant associations across these models and are considered promising markers for spontaneous preterm birth. Abbreviations: AUC: area under the curve; FPR: false positive rate.

Source data

Extended Data Fig. 3. Feature importance ranking from random forest model using level 1 metabolites.

All level 1 metabolites meeting the p < 0.05 threshold in the discovery analysis were included in the random forest model. (A) Feature importance ranking for spontaneous preterm birth (sPTB) in early pregnancy. (B) Feature importance ranking for sPTB in late pregnancy.

Source data

Also, four exogenous metabolites from xenobiotics metabolism, including methyl ecgonine and 2,3-dihydroxybenzoate, were exclusively associated with miPTB/miETB in later pregnancy, and miETB in early pregnancy, respectively (Table 3). Other metabolites exclusively associated with medically indicated early birth included hexanoylcarnitine and lauroylcarnitine, which are involved in fatty acid metabolism (Fig. 2b). A large number of identified metabolites were significantly associated with either sPTB or sETB (Fig. 2c), most of which are closely connected within a network of amino acid metabolism, steroid biosynthesis and purine metabolism. In particular, citric acid, progesterone and cortexolone were found to be consistently associated with sPTB and sETB in both early and later pregnancy.

Internal validation

In total, 13 level 1 metabolites identified in the discovery analysis were consistently significant (P < 0.05) in the internal validation dataset, with 6 also passing FDR q value < 0.20. Associations of leucine, L-aspartate, itaconate and L-tyrosine were reproduced with the same early birth types and the same direction of effect. Hexanoylcarnitine, lauroylcarnitine and decanoylcarnitine were consistently associated with medically indicated early birth, while 10-hydroxydecanoate and 2-oxobutyrate were consistently associated with spontaneous early birth (Extended Data Fig. 4). These consistent findings across the discovery cohort and internal validation samples support the reproducibility of key metabolite–birth phenotype associations and strengthen confidence in the robustness of the discovery-phase results.

Extended Data Fig. 4. Independent validation results for metabolites (Level 1) identified in metabolome-wide association study (MWAS) in current study (p value < 0.05).

^* It donates the significant metabolites identified using FDR q < 0.2. Abbreviations: sPTBH=Spontaneous preterm birth (N = 48) vs Heathy full term birth (N = 161); miPTBH=Medically indicated preterm birth (N = 16) vs Healthy full term birth (N = 161); sETBH=Spontaneous early term birth (N = 40) vs Heathy full term birth (N = 161); miETBH=Medically indicated early term birth (N = 14) vs Healthy full term birth (N = 161); sPETBH=Spontaneous preterm/early term birth (N = 88) vs Healthy full term birth (N = 161); miPETBH=Medically indicated preterm/early term birth (N = 30) vs Healthy full term birth (N = 161).

Source data

Sensitivity analysis

First, we assessed whether specific medical indications influenced the findings (Supplementary Table 2 and Extended Data Fig. 5). The number of significant metabolic features (raw P < 0.05) was similar between all medically indicated births and those due to maternal cardiometabolic or fetal growth restriction indications across both columns and time points. Overlap between these indication groups was limited (33 features in early pregnancy; 20 and 13 in later pregnancy for HILIC and C18, respectively), with no overlap across early and late pregnancy. Pathway patterns differed by indication type, with more enriched pathways observed for maternal cardiometabolic indications. Several pathways were uniquely associated with each indication group, including fatty acid oxidation/peroxisome for cardiometabolic indications and keratan sulfate degradation for fetal growth restriction (Extended Data Fig. 5).

Extended Data Fig. 5. Metabolic pathways in the early (8-14 week) and later (24-30 week) pregnancy maternal metabolome associated with the medically indicated early birth categories in the two technical columns (HILIC, C18).

A. Metabolic pathways in the early pregnancy (8-14 weeks) metabolome associated with medically indicated early birth (combined preterm and early term birth) due to any indication; medically indicated early birth due to fetal or placental condition; and medically indicated early birth due to maternal cardiometabolic condition. B. Metabolic pathways in the later pregnancy (24-30 weeks) metabolome associated with medically indicated early birth (combined preterm and early term birth) due to any indication; medically indicated early birth due to fetal or placental condition; and medically indicated early birth due to maternal cardiometabolic condition. 1. The size and color of the bubbles represent the pathway significant level based on -log10(P-value). Only significant associations (p < 0.05) were indicated by bubbles in the plots. The percentages of the number of significant putative metabolites (overlap size) to the number of metabolites within each pathway (pathway size) less than 30% were excluded. 2. Among 30 subjects with medically induced early birth, 21 was due to maternal cardiometabolic condition (preeclampsia, gestational hypertension, and gestational diabetes mellitus), and 9 was due to fetal or placental condition (fetal growth restriction). 3. For HILIC positive ion mode, only the following adducts were considered: M [1 + ], M + H [1 + ], M-H2O + H [1 + ], M + Na [1 + ], M + K [1 + ], M + 2H [2 + ], and M(C13) + 2H [2 + ].For C18 negative ion mode, only the following adducts were considered: M-H [-], M + Cl [-], M + ACN-H [-], M + HCOO [-], M(C13)–H [-], M-H2O–H [-], and M + Na–2H [-]. Abbreviations: HILIC = hydrophilic interaction liquid chromatography column; C18 = C18 column; miPETBH=Medically indicated preterm/early term birth (N = 30) vs Healthy full term birth (N = 161); miPETBH (Fetal)=Medically indicated preterm/early term birth due to fetal or placental condition vs Healthy full term birth; miPETBH (Maternal)=Medically indicated preterm/early term birth due to maternal cardiometabolic condition vs Healthy full term birth.

Source data

Second, using different percentile cut-offs of P values (P < 0.01 and P < 0.005), key pathways, including urea cycle/amino group, histidine and glutamate metabolism, remained consistently enriched, supporting robustness of pathway findings (Extended Data Figs. 6 and 7).

Extended Data Fig. 6. Sensitivity analysis of pathway enrichment in the early (8-14 week) and later (24-30 week) pregnancy maternal metabolome in the two technical columns (HILIC, C18), using p-value < 0.01.

Each cell was colored by the p-value of the association of each metabolic pathway and different birth types at raw p-value 0.01 as cut-off input. Abbreviations: sPTBH=Spontaneous preterm birth (N = 48) vs Heathy full term birth (N = 161); miPTBH=Medically indicated preterm birth (N = 16) vs Healthy full term birth (N = 161); sETBH=Spontaneous early term birth (N = 40) vs Heathy full term birth (N = 161); miETBH=Medically indicated early term birth (N = 14) vs Healthy full term birth (N = 161); sPETBH=Spontaneous preterm/early term birth (N = 88) vs Healthy full term birth (N = 161); miPETBH=Medically indicated preterm/early term birth (N = 30) vs Healthy full term birth (N = 161).

Source data

Extended Data Fig. 7. Sensitivity analysis of pathway enrichment in the early (8-14 week) and later (24-30 week) pregnancy maternal metabolome in the two technical columns (HILIC, C18), using p-value < 0.005.

Each cell was colored by the p-value of the association of each metabolic pathway and different birth types at raw p-value 0.005 as cut-off input. Abbreviations: sPTBH=Spontaneous preterm birth (N = 48) vs Heathy full term birth (N = 161); miPTBH=Medically indicated preterm birth (N = 16) vs Healthy full term birth (N = 161); sETBH=Spontaneous early term birth (N = 40) vs Heathy full term birth (N = 161); miETBH=Medically indicated early term birth (N = 14) vs Healthy full term birth (N = 161); sPETBH=Spontaneous preterm/early term birth (N = 88) vs Healthy full term birth (N = 161); miPETBH=Medically indicated preterm/early term birth (N = 30) vs Healthy full term birth (N = 161).

Source data

Third, sensitivity analyses using alternative SES adjustments showed consistent results. Top enriched pathways remained stable across models (Extended Data Fig. 8), and 37 of 39 level 1 metabolites retained significant associations with similar effect direction and magnitude (Extended Data Fig. 9). Pregnancy timing and phenotype-specific associations were also consistent across models.

Extended Data Fig. 8. Sensitivity analysis of pathway enrichment in the early (8-14 week) and later (24-30 week) pregnancy maternal metabolome in the two technical columns (HILIC, C18), using different SES indicators.

Each cell was colored by the p-value of the association of each metabolic pathway and different birth types at raw p-value 0.05 as cut-off input. A total of 5 different models were presented. “m1” denotes the primary model used in the main analysis, adjusted for infant sex, maternal age, parity, prenatal body mass index in early pregnancy, substance use (any tobacco, alcohol, or marijuana use in the month prior to pregnancy), and maternal education. “s1” further adjusted for all additional socioeconomic status (SES) indicators, including prenatal insurance, income-to-poverty ratio, and hospital site. “s2,” “s3,” and “s4” each adjusted for a single SES indicator: hospital site, income-to-poverty ratio, and prenatal insurance, respectively. Abbreviations: sPTBH=Spontaneous preterm birth (N = 48) vs Heathy full term birth (N = 161); miPTBH=Medically indicated preterm birth (N = 16) vs Healthy full term birth (N = 161); sETBH=Spontaneous early term birth (N = 40) vs Heathy full term birth (N = 161); miETBH=Medically indicated early term birth (N = 14) vs Healthy full term birth (N = 161); sPETBH=Spontaneous preterm/early term birth (N = 88) vs Healthy full term birth (N = 161); miPETBH=Medically indicated preterm/early term birth (N = 30) vs Healthy full term birth (N = 161).

Source data

Extended Data Fig. 9. Sensitivity analysis for metabolites (Level 1) identified in metabolome-wide association study (MWAS) in current study (p value < 0.05), using different SES indicators.

A total of 5 different models were presented. “m1” denotes the primary model used in the main analysis, adjusted for infant sex, maternal age, parity, prenatal body mass index in early pregnancy, substance use (any tobacco, alcohol, or marijuana use in the month prior to pregnancy), and maternal education. “s1” further adjusted for all additional socioeconomic status (SES) indicators, including prenatal insurance, income-to-poverty ratio, and hospital site. “s2,” “s3,” and “s4” each adjusted for a single SES indicator: hospital site, income-to-poverty ratio, and prenatal insurance, respectively. The light blue denotes the association during early pregnancy and dark blude denotes the association during later pregnancy. Abbreviations: sPTBH=Spontaneous preterm birth (N = 48) vs Heathy full term birth (N = 161); miPTBH=Medically indicated preterm birth (N = 16) vs Healthy full term birth (N = 161); sETBH=Spontaneous early term birth (N = 40) vs Heathy full term birth (N = 161); miETBH=Medically indicated early term birth (N = 14) vs Healthy full term birth (N = 161); sPETBH=Spontaneous preterm/early term birth (N = 88) vs Healthy full term birth (N = 161); miPETBH=Medically indicated preterm/early term birth (N = 30) vs Healthy full term birth (N = 161).

Source data

Discussion

In this comprehensive MWAS, we applied high-resolution metabolomics to characterize serum metabolic perturbations in early and later pregnancy associated with spontaneous and medically indicated preterm and early term birth among African American pregnant women. To our knowledge, this is the first untargeted MWAS to examine these distinct early birth phenotypes with internal validation. We observed that the maternal metabolome during early versus later pregnancy was differentiated by unique combinations of metabolites and pathways representing biological processes essential for maintaining homeostasis during pregnancy and achieving healthy birth outcomes, including amino acids, vitamins, bioenergetics, steroids and lipids, and acylcarnitines. Together, these findings support the potential of early and later pregnancy metabolomic profiles to identify pregnancies at elevated risk for early birth. Associations with spontaneous early birth were more consistent, while findings for medically indicated outcomes are more suggestive and warrant confirmation in larger, pooled studies of high-risk populations.

We identified six potentially sensitive metabolic signals in early and/or later pregnancy that were consistently inversely associated with spontaneous early birth (sPTB and sPETB) relative to healthy full-term birth. These signals were enriched in valine, leucine and isoleucine metabolism, tyrosine metabolism, and cysteine and methionine metabolism, and demonstrated predictive utility in classification models for sPTB, warranting independent replication and validation to inform the future development of sensitive clinical biomarkers. Valine, 4-methyl-2-oxovaleric acid, leucine, methionine and tyrosine were identified in early pregnancy (8–14 weeks), whereas itaconate was identified in later pregnancy (24–30 weeks). Itaconate has been reported to have multiple immunoregulatory and antioxidative effects, including antibacterial and pro-tumour effects²⁵. Tyrosine, measured at the end of the first trimester, has also been associated with augmented placental oxidative stress, increased risk of intrauterine growth restriction and subsequent small for gestational age in the newborn²⁶. Methionine, an oxidative biomarker, can regulate the innate immune system and multiple metabolic processes, including decreasing DNA damage, counteracting oxidative stress and activating endogenous antioxidant enzymes²⁷. Valine and leucine are branched-chain amino acids that are involved in regulating protein synthesis, signalling pathway activation and cell proliferation²⁸. Particularly, valine is an important player in gluconeogenesis, and leucine can be converted into acetyl-CoA to feed the tricarboxylic acid (TCA) cycle. In metabolic profiling of second-trimester amniotic fluid, decreased levels of these amino acids are suggested to indicate a tendency for hyperglycaemia²⁹. In contrast, circulating concentrations of tyrosine, valine and itaconic acid have been positively associated with gestational diabetes mellitus^30,31, which may be related to the increased risk of sPTB. In our cohort, however, these metabolites were inversely associated with sPTB, suggesting a potentially protective role. This may reflect the compensatory mechanisms of amino acid metabolites against inflammation and oxidative stress, as sPTB is largely driven by stress-related mechanisms. Moreover, since the cross-sectional designs only capture associations at a single time point, the direction of associations can be temporal based on the timing of sample collection. During pregnancy, increased metabolic demands or increased exposures to reactive oxygen species (ROS) can promote oxidative stress, leading to DNA damage in the fetal membranes. These changes may trigger premature cellular ageing, endoplasmic reticulum stress, inflammatory activation and release of DNA fragments into amniotic fluid, and thus increasing the risk of PTB, especially sPTB³². Collectively, the amino acid metabolites identified in our study suggest underlying inflammatory activity in women with sPTB and may serve as potential markers for future investigation.

Several amino acid pathways were consistently identified as the top pathways that significantly associated with PTB and ETB phenotypes, particularly for sPTB and sETB. These pathways include aspartate and asparagine metabolism, arginine and proline metabolism, lysine metabolism glutamate metabolism, histidine metabolism, vitamin B₃ metabolism, vitamin C metabolism and tryptophan metabolism. Our findings are supported by previous human studies, which reported similar perturbed amino acid pathways associated with any PTB or sPTB^{19,20,22,26,33}. These amino acid pathways are also closely involved in oxidative stress, acute inflammatory response, and recognized as critical pathways in determining potential intrauterine growth retardation and adverse birth effects³⁴. During pregnancy, amino acids, especially those involved in glutamate metabolism, aspartate and asparagine metabolism, arginine and proline metabolism, are important precursors for fetal development and growth³⁴. Histidine and tryptophan metabolism are aromatic amino acids and are susceptible to attack by ROS, which induce oxidative stress³⁵. Additionally, vitamin B₃ and vitamin C may act as antioxidants for fetal growth improvement and pre-eclampsia treatment³⁶.

Consistently, several metabolites enriched in these amino acid pathways showed inverse associations with spontaneous early birth, where the strongest effects were observed for metabolites within arginine and proline metabolism. Specifically, hydroxyproline may serve as a potential ROS scavenger to regulate the cellular redox state, thus contributing to the modulation of cell metabolism, growth, development, and responses to nutritional and physiological changes³⁷. Prior work suggested that women with cervical incompetence, an important risk factor for sPTB, have decreased hydroxyproline concentration in the first trimester³⁸. In support of these findings, we also observed decreased hydroxyproline among spontaneous early birth groups in early pregnancy. Previous studies also found lower levels of citrulline among preterm neonates³⁹, and lower levels of arginine and citrulline have been linked to decreased umbilical blood flow, impaired transfer of nutrients to the fetus and intrauterine growth restriction through the arginine nitric oxide synthase pathway³⁹. Consistently, we observed decreased citrulline levels in later pregnancy serum samples among women with PTB.

We observed perturbations in steroid biosynthesis and purine metabolism across both early and later pregnancy, with increased steroid-related metabolites and decreased purine metabolites in spontaneous early birth phenotypes. A prior study reported that pregnancies ending in PTB had a different maternal serum metabolic profiling in progesterone metabolites and sterol lipids in the third trimester compared to the healthy pregnancies⁴⁰, as well as changes in energy metabolism and decreased hypoxanthine in the second-trimester amniotic fluid metabolome¹⁸. Progesterone is widely used in the clinical setting to prevent sPTB⁴¹. In pregnant women, progesterone exerts significant anti-inflammatory effects and maintains uterine quiescence and cervical competence⁴². In the present study, we observed an increased level of progesterone in women with sPTB in both the early and later pregnancy metabolome, suggesting a compensatory mechanism to reduce inflammatory effects.

We also identified four metabolites in the early and/or later pregnancy metabolome related to purine metabolism associated with either spontaneous or medically indicated early birth. Hypoxanthine has the capability to induce lipid peroxidation, reduce total radical-trapping antioxidant parameter and alter antioxidant enzymes⁴³. In the present study, medically indicated early births were positively associated with inosine and urate and negatively associated with xanthine. Previous work has demonstrated that pregnancies complicated by pre-eclampsia, one of the indications for induced labour, have increased placental xanthine oxidase activity. This enzyme catalyses the conversion of hypoxanthine to xanthine and subsequently to uric acid while generating reactive oxidants that can contribute to endothelial damage⁴⁴.

We found that three medium-chain acylcarnitines, hexanoylcarnitine, decanoylcarnitine and lauroylcarnitine, were associated with medically indicated early birth (Fig. 2). Altered levels of acylcarnitines were found in infants born with low birth weight and preterm⁴⁵. Also, many medium-chain acylcarnitines have been reported to serve as diagnostic markers for inherited disorders of fatty acid metabolism, insulin resistance and cardiovascular disease, as well as inflammation in metabolic syndrome. In particular, decanoylcarnitine is elevated in the blood of individuals with obesity during adolescence⁴⁶. Lauroylcarnitine is decreased in the sera of pregnant individuals with coronary heart disease⁴⁷ and may also serve as a mediator of obesity-induced inflammation⁴⁸. The implications of chronic medical conditions related to these acylcarnitines may explain the medically indicated labour in our study population. Acylcarnitines may contribute to mitochondrial dysfunction, and the resulting inflammatory and oxidative stress responses can impair placental function or disrupt fetal membranes, thereby increasing the risk of PTB⁴⁹. Taken altogether, elevated levels of medium-chain acylcarnitines may reflect a defect in lipid transport or beta-oxidation, and create an adverse in utero environment for the fetus, which negatively impacts their metabolic health. Due to the limited study power, we could not evaluate the confirmed metabolites characterized by medically induced early birth subgroups. Nonetheless, the predominant pathways closely relate to oxidative stress, and acute and systemic inflammatory responses, demonstrating consistent associations with both maternal cardiometabolic conditions and fetal growth restriction. Notably, fatty acid oxidation and peroxisomes, which are crucial for the elimination of ROS and maintenance of metabolic homeostasis⁵⁰, were uniquely associated with maternal cardiometabolic conditions at both early and later pregnancy metabolomes. Thus, perturbation of these pathways may indicate potential cardiometabolic changes during pregnancy.

We identified four exogenous metabolites associated with early birth phenotypes. Notably, methyl ecgonine, a cocaine-related metabolite, was observed in association with sPTB in the early pregnancy metabolome and with miPTB/miETB in the later pregnancy metabolome. Importantly, detection of this metabolite in untargeted metabolomics does not confirm substance use and should not be interpreted as evidence of a causal behavioural risk factor. Two organic compounds, bis(2-ethylhexyl)phthalate, a phthalate, and phthalic anhydride, are industrial chemicals commonly used in plasticizers, and phthalate metabolites have been widely reported in relation to PTB risk⁵¹. These findings may reflect broader environmental co-exposures rather than isolated substance use. The identification of xenobiotic metabolites highlights the ability of untargeted metabolomics to capture components of the exposome, including exogenous chemical signals from environmental, dietary or other non-endogenous sources. These associations should be interpreted within the exploratory and hypothesis-generating framework of MWAS and do not provide confirmatory evidence of specific substance use or causal risk factors for early birth.

This study provides insight into metabolic perturbations across heterogeneous early birth phenotypes during early and later pregnancy. The identification of phenotype-specific maternal metabolic profiles helps address gaps in understanding the complex aetiology of early birth and may inform future biomarker development. Strengths include rigorous outcome ascertainment through standardized medical record abstraction and early pregnancy dating, enabling precise phenotyping and appropriate selection of a healthy full-term comparison group without cardiometabolic complications. Exclusion of participants with chronic conditions further supports valid comparisons. Use of prenatal serum samples minimizes confounding from medications administered during labour. The study focuses on an African American cohort, a population disproportionately affected by PTB and often under-represented in research, providing critical insights into pregnancy and birth outcomes in this high-risk population. Samples were collected within defined gestational windows, improving temporal consistency. Finally, by incorporating a two-stage design with rigorous internal validation and multiple sensitivity analyses, we strengthened the robustness and reliability of our findings. These steps provide greater confidence in the reproducibility and stability of key metabolite–birth phenotype associations and the consistency of pathway-level signals across pregnancy windows. This framework provides a foundation for future hypothesis-driven studies, external replication and translational research aimed at developing clinically meaningful biomarkers.

Our study does have several limitations. First, the sample size for medically indicated births was relatively small, and the limited study power may have resulted in fewer detected pathways in MWAS analysis. In addition, results for medically indicated births were more sensitive to model specification, particularly with respect to SES covariate adjustment, likely due to the small sample size. These findings highlight the importance of careful selection and specification of SES covariates, especially in smaller subgroups. Second, the associations with exogenous metabolites may be influenced by co-exposure confounding, limiting the interpretation. Third, while our study focused on the serum metabolome, additional sensitive and specific markers may exist in amniotic fluid or cervicovaginal samples. Further studies are warranted to compare the metabolic profiling between different biospecimens. In addition, we used raw P values instead of multiple-testing-corrected q values for MWAS analysis, which may have increased the risk of false positive results. However, pathway enrichment analysis by mummichog has shown stable results when varying the cut-offs for significance⁵². The sensitivity analysis on the 0.5th and 1st percentiles of raw P values showed similar results relative to our primary analysis. Meanwhile, using a stringent P -value cut-off would increase the type II error, which may also exclude true positive findings. Furthermore, non-fasting samples and lack of dietary data may introduce measurement variation. However, we followed a comprehensive metabolomics workflow, which has shown to successfully analyse many non-fasting samples previously⁵³. Finally, although this study identifies potentially sensitive metabolic signals associated with early birth outcomes, the predictive analyses are exploratory and constrained by sample size and covariate availability. Additional factors, including detailed maternal and fetal complications, diet and environmental exposures, were not fully captured. As many identified signals are amino acid related, future studies integrating complementary omics, such as lipidomics, may provide a more comprehensive understanding of underlying pathways and metabolic signatures. Independent replication in larger and more diverse populations will be essential to evaluate the generalizability and clinical utility of these metabolomic signatures and to support the development of robust clinical biomarkers for spontaneous and medically indicated early birth.

Methods

This study was conducted in accordance with all relevant ethical regulations and was approved by the Emory University Institutional Review Board. All participants provided written informed consent before participation.

Participants in the present analysis were enrolled between 2014 and 2018 in the Atlanta African American Maternal-Child Cohort⁵⁴ with a prenatal serum sample available for metabolomics assay whose pregnancy ended in an early birth (PTB or ETB) or a HFTB. Participants (N = 331) were recruited from two prenatal clinics in Atlanta, Georgia, spanning a socioeconomically diverse array of patients. Inclusion criteria were: (1) being born in the United States; (2) self-identifying as African American or Black; (3) presenting with a singleton pregnancy between 8 and 14 weeks gestation; (4) ability to communicate in English; (5) between 18 and 40 years; and (6) absence of chronic medical conditions and not using prescription medications.

Questionnaire data

Sociodemographic information was obtained through maternal self-report and review of prenatal administrative records, including maternal age at study entry, educational attainment (less than high school, high school or GED, some college, college graduate) and type of prenatal health insurance (low-income Medicaid, Right-from-the-Start Medicaid or private insurance). Substance use in the month before enrolment (alcohol, tobacco, marijuana and other drugs) was assessed using a health survey based on the Timeline Follow-Back method^55–57.

Venous blood

Venous blood specimens for research assays were obtained during routine prenatal blood draws into serum tubes, which were held upright at room temperature for 60 min and then centrifuged to obtain serum. The resulting serum was stored at –80 °C until metabolomics assays.

Medical record data

After delivery, trained study personnel conducted standardized medical record abstraction to collect information on parity (categorized as prior birth versus none), body mass index (BMI) at the first prenatal visit (calculated from measured height and weight at 8–14 weeks gestation), pregnancy and labour/delivery complications (for example, pre-eclampsia, intrauterine growth restriction) and birth outcome phenotypes defined by gestational age at delivery and type of labour. Gestational age was determined using the best obstetric estimate, incorporating the delivery date relative to the estimated date of confinement established at the 8–14-week visit. All participants underwent early pregnancy dating based on last menstrual period and/or early ultrasound, following standard clinical practice⁵⁸. Data on labour type (spontaneous, induced or none), mode of delivery (vaginal or caesarean) and the clinical indications for induction or C-section were used to further classify PTB and ETB as spontaneous or medically indicated. Births following spontaneous labour or spontaneous pre-labour rupture of membranes were categorized as spontaneous, whereas those initiated due to maternal or fetal complications requiring induction or pre-labour C-section delivery were classified as medically indicated. Considering gestational age at birth and type of labour, early births were classified into the following phenotypic categories: PTB (between 22-0/7 and 36-6/7 weeks gestation) further classified as sPTB or miPTB⁵⁹; ETB (between 37-0/7 and 38-6/7 weeks gestation) further classified as sETB or miETB. HFTBs, the referent category, were births 39-0/7 weeks or greater for which the pregnancy, labour and delivery were not complicated by pre-eclampsia, gestational hypertension, gestational diabetes or intrauterine growth restriction.

High-resolution metabolomic profiling

Metabolomic profiling was completed using non-fasting serum samples collected during the early (8–14 week) and later (24–30 week) pregnancy study visits using established protocols^53,60–62. Each sample was run in triplicate using liquid chromatography–high-resolution mass spectrometry (Thermo Scientific Q-Exactive HF). To increase metabolite coverage, both polar and nonpolar separation methods were applied, including HILIC with positive ESI and C18 reversed-phase chromatography with negative ESI. The detected signals, referred to as metabolic features, were processed using a standardized pipeline that performs peak detection, m/z assignment, retention time alignment, feature quantification and data quality filtering^63,64. To reduce noise and improve data quality, metabolic features were retained only if they were detected in more than 15% of serum samples, had a median coefficient of variation below 30% across technical replicates and showed a Pearson correlation coefficient (ρ) greater than 0.7. The final dataset consisted of features characterized by m/z, retention time and ion intensities. Replicate measurements with at least one non-zero value were averaged, and the resulting data were log₂-transformed for downstream analyses.

Statistical analyses

To identify metabolic signals associated with the early birth phenotypes of interest, we conducted a series of MWAS using logistic regression models for features detected from serum samples collected at each of the two prenatal time points (early pregnancy, 8–14 weeks; later pregnancy, 24–30 weeks). Based on existing literature and using directed acyclic graphs (Extended Data Fig. 10), we controlled for infant sex, maternal age, parity, prenatal BMI at early pregnancy, substance use (any use of tobacco, alcohol, marijuana one month prior) and education level in these MWAS models as given by equation (1):

$$\begin{array}{c}\mathrm{Logit}\left(P\left(\text{Birthoutcomes}\right)\right)={\beta }_0+{\beta }_{1j}{\text{log}}_2{\mathrm{Intensity}}_i+{\gamma }_{1j}{\mathrm{InfantSex}}_i\hfill \\ +{\gamma }_{2j}{\mathrm{MaternalAge}}_i+{{\gamma }_{3j}\mathrm{MaternalEducation}}_i\hfill \\ +{\gamma }_{4j}{\mathrm{BMI}}_i+{\gamma }_{5j}{\mathrm{Substanceuse}}_i+{{\gamma }_{6j}\mathrm{Parity}}_i\hfill \end{array}$$ 1

where intensity denotes the intensity of each metabolic feature, ${\beta }_0$ represents the intercept and ${\beta }_{1-6}$ are the coefficients corresponding to each covariate. We used separate logistic models to examine metabolic features and pathways associated with each of the four early birth phenotypes of interest (PTB classified as sPTB or miPTB; ETB classified as sETB or miETB), as well as for two composite early birth categories in which we combined spontaneous early births (sPTB and sETB as sPETB) and medically inciated early births (miPTB and miETB as miPETB), with HFTB as the referent category.

Extended Data Fig. 10. Directed acyclic graph (DAG) shows the association of maternal metabolomics with early birth.

Green lines represent causal paths, and red lines were biasing paths. Minimal sufficient adjustments for estimating the association between maternal metabolomics and early birth outcomes include maternal age, education, prenatal BMI during early pregnancy, parity, and substance use. Infant sex was adjusted as a priori biological variable.

Source data

To better understand when in pregnancy a given metabolic feature might serve as a useful marker for a particular early birth phenotype, we examined the metabolome in early and later pregnancy separately as well as the overlap in metabolic features measured in the early and later pregnancy metabolome associated with a particular early birth phenotype (‘early–late overlap’). To better understand whether a given metabolic feature may serve as a useful marker for the gestational age category overall or for a spontaneous or medically indicated birth within that gestational age category, we examined the spontaneous and medically indicated gestational age at birth categories separately and the overlap between them.

Finally, to evaluate the performance of specific metabolites significantly associated with specific early birth phenotypes as potential predictive signals, we conducted ROC analyses to quantify model sensitivity and specificity. For the primary analyses, we implemented a binary logistic regression classifier for each candidate metabolic signal using a generalized linear model with a binomial link function, consistent with the primary MWAS framework. The dataset was randomly partitioned into training (80%) and independent testing (20%) sets using a fixed random seed to ensure reproducibility. As a supplementary analysis, we fit a multi-metabolite logistic regression model including all candidate metabolic signals simultaneously and evaluated discrimination using the area under the ROC curve to assess panel-based performance. In addition, we implemented a random forest classifier as a complementary nonparametric machine learning approach suitable for high-dimensional, nonlinear metabolomics data⁶⁵. Using all level 1 confirmed metabolites as input features, we examined variable importance rankings to determine whether the highlighted metabolites consistently emerged as influential predictors beyond the parametric logistic regression framework.

Metabolic pathway enrichment analysis and metabolite annotation

To characterize the biological functions of metabolic features, we conducted pathway enrichment and metabolite annotation analyses. Pathway enrichment was implemented using mummichog (v2.0), which predicts pathway-level activity from high-resolution mass spectrometry data without requiring prior metabolite identification⁵². Mummichog maps statistically significant m/z features onto known metabolic networks and evaluates pathway enrichment through permutation testing. The primary user-controlled parameter is the statistical cut-off for selecting significant features. Two criteria were applied to identify metabolic features for pathway analysis: a nominal P-value threshold of 0.05 and a FDR q-value threshold of 0.2, estimated using the Benjamini–Hochberg procedure to account for multiple comparisons. To minimize the chance of false positive discovery, we excluded pathways for which the proportion of significant putative metabolites mapped to the pathway was less than 30%, calculated as the overlap size divided by the pathway size and those identified by mummichog with a P value > 0.05. The overall significance of pathway enrichment was performed using a permutation-based approach with the EASE score⁶⁶, a conservative modification of Fisher’s exact test. Adjusted P values were derived from the permutation null distribution and reported for pathway significance.

Metabolic features that were both significantly associated with early birth phenotypes and enriched within relevant pathways were annotated by matching their m/z values to common adducts using databases including the Human Metabolome Database, METLIN and the Kyoto Encyclopedia of Genes and Genomes, with a mass error tolerance of 10 ppm. To reduce false positive identifications, candidate matches were further evaluated based on retention time, isotope distribution and spectral peak quality through inspection of extracted ion chromatograms. Finally, a subset of metabolites was subsequently confirmed with level 1 evidence⁶⁷ by comparing their m/z, retention time and fragmentation patterns to authentic standards from an in-house reference library analysed under the same experimental conditions. The reported associations for level 1 metabolites were based on raw P < 0.05, as none remained significant after FDR correction at the feature-wide level. We additionally calculated FDR restricted to level 1 metabolites, regardless of their significance in MWAS, and reported metabolites with FDR q < 0.20.

Internal validation

Findings were first generated in a discovery dataset and subsequently evaluated in an independent metabolomics dataset from the same cohort. The discovery dataset included 279 participants with deliveries ≥22 weeks gestation who had available maternal serum samples in early pregnancy (8–14 weeks) and later pregnancy (24–30 weeks). The internal validation dataset consisted of 251 participants with their metabolomics profiles analysed using the same platform and laboratory protocols within the same study population (that is, the Atlanta African American Maternal-Child Cohort; Supplementary Table 3). The same analytical pipeline was applied to the level 1 metabolites identified in the discovery analysis. The purpose of this validation step was to assess the reproducibility and stability of metabolite–birth outcome associations in this high-dimensional MWAS framework and to reduce the likelihood that observed signals arose from multiple testing, technical variability or cohort-specific effects.

Sensitivity analysis

First, to assess whether particular medical indications might influence our conclusions around metabolites associated with medically indicted early births as a category, we conducted a stratified analyses in which we stratified the miPETB group (n = 30) by specific indication, to include maternal cardiometabolic indication (pre-eclampsia, gestational hypertension and gestational diabetes mellitus; n = 21) and fetal growth restriction indication, and compared the findings within and across strata. Second, we conducted additional pathway enrichment analyses using the 0.5th and 1st percentiles of raw P values as alternative significance thresholds and examined whether the significant pathways would be largely different under different raw P values⁶⁸. This allows us to evaluate the robustness of the pathway-level results to potential uncertainty or error at the individual-feature level. Third, the current African American pregnancy cohort included multiple indicators of SES, many of which differed significantly across birth outcome groups (Table 1). These SES indicators were moderately to strongly correlated with one another (Spearman’s ρ: 0.42–0.61; Supplementary Table 4), except for substance use. To reduce multicollinearity, we adjusted for education only in the main analysis. To further address potential residual confounding by SES and to assess the robustness of our findings, we conducted additional analyses in which each SES indicator was adjusted for separately, as well as jointly in a single model. Results from these models were then compared with those from the main analysis.

All statistical tests were two-sided unless otherwise specified. All analyses were conducted using R (version 4.5.1, https://www.r-projects.org/) or Python (version 3.9, https://www.python.org/).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

D.L. and A.L.D. had full access to all of the data in the study and took responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: D.L., Y.T. and A.L.D. Acquisition, analysis or interpretation of data: D.L., Y.T., S.M.E., M.R., D.I.W., X.H., Z.L., K.R.T., E.C., D.P.J. and A.L.D. Drafting of the paper: D.L., Y.T., S.M.E. and A.L.D. Critical revision of the paper for important intellectual content: D.L., Y.T., S.M.E., M.R., D.I.W., X.H., Z.L., K.R.T., E.C., C.C.H., D.P.J. and A.L.D. Statistical analysis: D.L., Y.T., S.M.E., Z.L. and A.L.D. Obtained funding: D.L., E.C., D.P.J. and A.L.D. Administrative, technical or material support: D.L., Y.T., S.M.E., M.R., D.I.W., X.H., Z.L., K.R.T., E.C., C.C.H., D.P.J. and A.L.D. Supervision: D.L. and A.L.D.

Peer review

Peer review information

Nature Health thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Ben Johnson, in collaboration with the Nature Health team.

Funding

This work was supported by the Environmental Influences on Child Health Outcomes (ECHO) programme, Office of the Director, National Institutes of Health (NIH), under award numbers 5U2COD023375-05/A03-3824 (to D.L.) and NIH research grants (R21ES032117 and R01ES035738 to D.L.; R01NR014800, R01MD009064, R24ES029490 and R01MD009746 to A.L.D.; P30ES019776 to D.L. and D.P.J.; UH3OD023318 to A.L.D.).

Data availability

The raw and processed metabolomics data generated in this study have been deposited in the Metabolomics Workbench under study ID ST002351 by Emory Clinical Biomarker Laboratory. The clinical outcome data are available under restricted access to protect the privacy of the study participants; access can be obtained by emailing corresponding authors D.L. and A.L.D. with reasonable requests. The demographic covariates data are protected and are not available due to data privacy laws. Source data are provided with this paper.

Code availability

Specific code is available from the corresponding author D.L., and the request will be responded to within 10 business days.

Competing interests

The authors declare no competing interests.

Extended data

is available for this paper at https://doi.org/10.1038/s44360-026-00150-w.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s44360-026-00150-w.