Background

Adverse birth outcomes such as spontaneous preterm birth and low birth weight are major determinants of health from infancy through adulthood [1,2,3]. Understanding the molecular processes contributing to these outcomes has the potential to ameliorate the global burden of human disease. For example, analyzing proteomic signatures of birth outcomes may reveal proteins that enable early prediction of at-risk pregnancies and serve as targets for prevention research. Unlike mRNA expression or DNA methylation, which reflect upstream regulatory events or transient signals, proteins are the functional molecules that carry out the majority of biological processes [4]. Proteomics may offer distinct insights and be more clinically informative compared with other omics approaches, as suggested by the moderate correlations between mRNA and protein levels [5,6,7] and prior challenges in translating transcriptomic findings into clinical applications [8]. Ultimately, proteins found to be dysregulated in adverse birth outcomes could elucidate underlying mechanisms, serve as intervention targets, or act as markers when screening for at-risk individuals.

Untargeted proteomics characterizes the entire proteome without focusing on specific proteins or pathways. Label-free global quantitative proteomics uses mass spectrometry, in conjunction with an atlas of protein spectra for protein quantification without the use of labels or standards [9, 10]. Despite its potential for aiding in the discovery of new protein-disease associations [11], a large proportion of protein research examines a small number of well-characterized proteins [12]. In research on birth outcomes and pregnancy complications, modern, mass spectrometry-based untargeted proteomics has also been largely underutilized [13,14,15]. For example, a 2022 systematic review of proteomic studies of preterm birth included just 9 untargeted studies compared with 70 targeted studies [14].

An important consideration in proteomic research is selection of the most relevant tissue for the research objectives. In the context of pregnancy and birth outcomes, the placenta is a critical tissue, as it connects the mother and fetus, facilitating transfer of gases, nutrients, and waste products, in addition to coordinating hormonal and paracrine signaling regulating fetal growth and parturition [16]. Despite the placenta’s critical role in fetal development, few studies have investigated proteomic signatures in the placenta associated with birth outcomes. Notably, none of the 9 untargeted studies identified in the 2022 systematic review of proteomic studies of preterm birth discussed above utilized placental tissues, but instead evaluated associations of preterm birth with the proteomes of cervical vaginal fluid, amniotic fluid, and blood [14]. While these analyses may uncover diagnostic biomarkers predicting preterm birth in individual pregnancies, placental markers may better reflect the pathology and mechanistic processes related to adverse birth outcomes. Studies consistently show perturbed molecular features, including DNA methylation, mRNA, and microRNA expression in the placenta in association with maternal environmental conditions, adverse birth outcomes, and long term child development [17].

Only a small number of studies have examined the placental proteome in relation to birth outcomes, and these have primarily used gel-based or targeted analytical methods rather than modern label-free global approaches [18,19,20]. Early work using two-dimensional gel electrophoresis of placental membranes identified differential expression of structural and cytoskeletal proteins, endoplasmic reticulum lumen proteins, and proteins with anticoagulant properties in association with preterm labor [18]. A subsequent 2D gel-based study reported reduced placental alpha-1 antitrypsin (SERPINA1), a protease inhibitor, in association with spontaneous preterm birth [19]. More recently, a multiomic study integrating targeted proteomics with transcriptomic and metabolomic data found evidence of disrupted extracellular matrix organization and mitochondrial function in placentas from spontaneous preterm births [20]. Collectively, these findings implicate extracellular matrix remodeling, oxidative stress, and protease imbalance as potential pathways in preterm delivery, but the small sample sizes and targeted nature of these studies limit the broader discovery of proteomic biomarkers. We are unaware of studies investigating how the placental proteome may affect or be impacted by birthweight, although transcriptomic, epigenomic, and multiomic analyses utilizing placental tissues have identified candidate genes involved in cardiometabolic, immune, hormonal, protease, and extracellular matrix pathways [21,22,23,24].

Here, we explored untargeted placental proteomic signatures associated with birth outcomes including birthweight, gestational age at birth, and spontaneous preterm birth. Our aim was to identify molecular mechanisms underlying birth outcomes and to evaluate the potential of placental proteomic biomarkers for preterm birth prediction.

Methods

Study population

This study used data from the Conditions Affecting Neurocognitive Development and Learning in Early Childhood (CANDLE) cohort. CANDLE recruited women aged 16–40 years in the second trimester with singleton, low-medical-risk pregnancies from 2006 to 2011 in Shelby County (Memphis area), Tennessee, who were residents of Shelby County, could speak and understand English, and planned to deliver at one of four participating health care settings in the county [25]. We generated untargeted proteomics data in a subset of 99 CANDLE participants with complete exposure and covariate data. For this sub-analysis, in order to capture the subpopulation of spontaneous preterm and term deliveries unrelated to medical complications, we excluded participants with clinical chorioamnionitis, placental abruption, and preeclampsia, as well as births resulting from induced labor. We oversampled 19 spontaneous preterm births (< 37 weeks), then randomly selected 80 term samples (≥ 37 weeks) from the remaining births. This study was approved by the institutional review boards at the University of Tennessee Health Science Center and the University of Washington.

Birth outcomes

Birthweights (grams) and gestational ages at birth (weeks) were determined from medical record abstraction. Birthweight for sex and gestational age z-scores were calculated in accordance with INTERGROWTH standards [26]. Preterm birth was defined as birth before 37 completed weeks of gestation.

Covariates

Covariates included fetal sex (male; female), maternal age (years), maternal prepregnancy BMI (kg/m2), maternal tobacco use during pregnancy (yes; no), maternal alcohol use during pregnancy (yes; no), maternal education (< High School; High School completion; graduated college or technical school; some graduate work or graduate/professional degree), delivery method (vaginal; cesarean section), labor type (spontaneous; spontaneous, augmented), gravidity (number), and fetal genetic ancestry information derived using single nucleotide polymorphism (SNP) genotyping data as previously described [27]. To account for technical variability relating to sample collection, we also included an estimate of the blood contamination (percentage) based on the abundances of proteins positively correlated with albumin (blood marker) and proteins negatively correlated with aromatase (placental marker) [28]. Covariate differences between term and preterm births were evaluated via two-sample t-tests for continuous variables and chi-square goodness of fit tests for categorical variables (Table 1).

Table 1 Descriptive statistics of the study population (N = 99)
Full size table

Placental proteomics

One rectangular piece of placental villous tissue (approximate dimensions of 2 cm x 0.5 cm x 0.5 cm) was dissected from the middle of the parenchyma within 15 min of delivery. Tissue cubes were stored in a 50 mL tube with 20 mL RNAlater at 4 °C overnight, before longer term storage in RNAlater at −80 °C. Frozen placenta tissue in RNAlater (average weight ~ 100 mg) was washed with 1 mL cell wash buffer and drained with lint-free tissue paper. Samples were homogenized in lysis buffer consisting of 8 M urea, 0.1 M ammonium bicarbonate, and 1X protease and phosphatase inhibitor cocktail (final concentrations) before centrifugation at 16,000 xg for 15 min to remove non-homogenized tissue fractions. Supernatant total protein content was calculated using the BCA assay (Thermo Scientific, Waltham, MA). Protein (100 µg) was digested in a new set of tubes. At room temperature, samples were denatured by urea, reduced with 25 mM DTT (final concentration) for 1 h, and alkylated using 60 mM IAA (final concentration) for 30 min in the dark.

Samples were diluted fourfold with 0.1 M ammonium bicarbonate and digested with trypsin (trypsin: protein ratio of 1:40) at room temperature for 20 h. Digestion was halted by adding 5% formic acid until pH of 4–5. Peptides were desalted using Waters C18 1 cc columns (Milford, MA) according to the manufacturer’s protocol. The desalting procedure involved washing the cartridge with 0.1% TFA and eluting peptides with 60% acetonitrile containing 0.1% TFA. Finally, the peptides were dried and resuspended in 0.1% formic acid in water for LC-MS analysis.

Samples were analyzed using an EASY-nanoLC 1200 coupled to Q-Exactive HF Hybrid Quadrupole-Orbitrap MS (Thermo Fisher Scientific, Waltham, MA). Peptides were separated on a Thermo Scientific™ Acclaim™ PepMap™ 100 C18 HPLC column (75 μm × 250 mm) at 35 °C with a mobile phase composed of 0.1% formic acid in water (A) and 0.1% formic acid in 80% acetonitrile (B). Tryptic peptides from placental homogenates (1 µg total protein) were loaded onto the column, and separation was performed using the following LC gradient (%B) at a flow rate of 300 nL/min: 2–6% (0–10 min), 6–30% (10–110 min), 30–100% (110–115 min), and 100% (115–131 min). The spray voltage was set to 1.7 kV, and the heated capillary was maintained at 275 °C.

Proteomics data were acquired in data-dependent acquisition (DDA) mode using full MS scanning (m/z 375–1500) with an automatic gain control (AGC) target of 106. The ten most intense ions were selected for higher-energy collisional dissociation (HCD) fragmentation, with an AGC target of 105. Precursor isolation was performed using a 2.0 m/z window. MS1 spectra were acquired at a resolution of 120,000 at m/z 200, with a maximum ion injection time of 50 ms. HCD spectra were recorded at a resolution of 30,000 at m/z 200, with a maximum ion injection time of 60 ms. Normalized collision energy was 27. Precursor ions with single or unassigned charge states were excluded from fragmentation.

LC-MS data were processed through MaxQuant (version 1.6.8.0; Max Planck Institute of Biogeochemistry, Germany) [29] matching the MS/MS spectra against the Uniprot human proteome database via the Andromeda search engine. A MaxQuant decoy database was utilized to estimate the false discovery rate (FDR), with both peptide and protein FDR thresholds of 1%. Trypsin was set as the proteolytic enzyme, allowing a maximum of two missed cleavages. Variable modifications considered were N-terminal acetylation and methionine oxidation, while carbamidomethylation of cysteine was set as a fixed modification. Unique peptides were used for protein identification, while mass tolerances for peptide and fragment searches were set at 20 ppm and 4.5 ppm, respectively. Perseus (Max Planck Institute of Biogeochemistry, Germany) [30] was used to remove contaminants, decoy proteins with reversed sequences, and proteins identified only by site.

Statistical analysis

The units of analysis were Label-Free Quantification (LFQ) intensities for protein groups. Protein groups with more than 20% missingness were removed. Missing values for the remaining protein groups were imputed via random forest single imputation using the missForest R package [31], and then log2 transformed. Studies have shown that random forest-based single imputation of proteomic data outperforms other popular imputation methods [32, 33]. Among 3,494 protein groups that passed quality control and were detected in any samples, we quantified 1,221 protein groups that were detected in at least 80% of placental samples and included in downstream analyses.

Differential protein abundance analysis was conducted using the Differential Expression analysis of quantitative Mass Spectrometry data (DEqMS) Bioconductor package [34]. This method combines the limma empirical Bayes approach to moderate standard errors towards a common prior [35] with a correction based on the observed trend between protein variance and the number of peptides per protein in label-free proteomics data. DEqMS was utilized to fit linear models for each protein in separate models for each birth outcome, while adjusting for all covariates described above. Proteins were considered statistically significant with a false discovery rate (FDR) < 0.05 using the Benjamini-Hochberg method [36]. Following the complete-case methodology for handling missing covariate data, individuals with missing data for any covariate were excluded from these analyses. A heatmap of proteins significantly associated with preterm birth was generated using hierarchical clustering with Euclidean distance via the pheatmap package [37], with protein abundances scaled and centered across participants.

Overlapping genes between significant proteins observed in this study and significant mRNAs associated with preterm birth in a prior placental transcriptomic analysis including CANDLE participants and an additional cohort [38] were evaluated. While the majority (~ 95%) of protein groups can be mapped to a single gene symbol, some protein groups map to multiple genes. To compare overlap, we expanded the 1,221 quantified proteins into 1,278 unique corresponding gene symbols, and further restricted to 1,140 gene symbols that were included in the placental transcriptomic analysis. Among these 1,140 genes, we computed the percent that were significantly associated with preterm birth in both proteomic and transcriptomic analyses and compared the direction of association for overlapping genes.

Gene set enrichment analysis can support mechanistic discovery by linking individual protein changes to shared molecular and biological processes. Rotation gene set testing (ROAST) [39] with 99,999 rotations was performed to identify KEGG pathways (excluding KEGG human diseases) [40] that were differentially regulated in relation to birth outcomes. For the approximate 5% of protein groups that corresponded to more than one gene, only the first mapped gene was retained. KEGG pathways with fewer than 5 genes represented in our data were excluded, and results were considered statistically significant with FDR < 0.05.

Omics datasets often include thousands of correlated features. Principal Component Analysis (PCA) performs unsupervised dimensionality reduction, capturing the dominant sources of variation without outcome bias. This method can improve power to detect proteomic associations with health outcomes, and elucidate the contribution of individual proteins to total variance. PCA was performed, using Horn’s parallel analysis to determine the number of Principal Components (PCs) to retain [41,42,43]. The top 12 loadings for each PC were visualized, and linear and logistic regression models were fit to estimate associations of each birth outcome with each retained PC while adjusting for all covariates described above.

Beyond identifying individual associations, another common objective of omics studies is to assess whether molecular profiles can accurately classify diseases or predict clinical outcomes. Sparse Partial Least-Squares Discriminant Analysis (sPLS-DA) implemented in the Bioconductor mixOmics package [44] was utilized in a classification analysis to predict preterm versus term birth using the placental proteomics data. The ‘tune.splsda’ function with 5-fold cross-validation over 100 repeats was used to determine the optimal number of variables to load onto each component and the optimal number of components to retain based on one-sided t-tests. The top 20 loadings for these components were plotted, and cross validation error was reported in terms of the overall and balanced error rates (BER). While overall error reflects the proportion of wrongly classified individuals, BER is the average proportion of wrongly classified samples in each class, weighted by the number of samples in each class.

Results

Covariate and birth outcomes data are described in Table 1. Among 99 participants with placental proteome data, 19 (19.2%) underwent spontaneous preterm birth, the average [range] gestational age at birth was 38.8 [30.7 to 40.8] weeks, and the average birthweight was 3,279.5 [1,705.0 to 4,363.0] grams (Table 1). Mean maternal prepregnancy BMI was higher in term compared with preterm births (27.0 vs. 23.8, P = 0.039). Other covariates showed no significant differences, although maternal age tended to be higher in term pregnancies (27.9 vs. 25.1, P = 0.055), and reports of prenatal alcohol (15.0%) and tobacco (≤ 5.1%) use occurred only among term births. Complete covariate data were available for 89 participants that were included in downstream analyses.

Differential protein expression analysis with DEqMS (Supplemental Table 1) revealed that 295 (24%) of 1,221 placental proteins were significantly differentially expressed (FDR < 0.05) in association with preterm birth (Fig. 1a) and 367 (30%) were differently expressed in association with gestational age at birth (Fig. 1b). No proteins were significantly associated with birthweight for gestational age z-score (Fig. 1c). Among the 367 proteins significantly differentially expressed in association with gestational age, 264 (72%) were also significantly associated with preterm birth. Six of the top ten significant proteins (according to fold-change magnitude) for both birth outcomes overlapped. Two of these proteins were upregulated in association with preterm birth and downregulated per week increase in gestational age: Steryl-sulfatase (STS; preterm birth: LogFC = 2.08, FDR < 0.0001; gestational age at birth: LogFC= −0.57, FDR < 0.000001), and Endophilin-B2 (SH3GLB2; preterm birth: LogFC = 1.33, FDR < 0.001; gestational age at birth: LogFC= −0.35, FDR < 0.00001). The remaining four overlapping proteins were downregulated in association with preterm birth and upregulated per week increase in gestational age: Surfeit locus protein 4 (SURF4; preterm birth: LogFC= −1.99, FDR < 0.0001; gestational age at birth: LogFC = 0.48, FDR < 0.00001), Collagen alpha-2(I) chain (COL1A2; preterm birth: LogFC= −2.11, FDR < 0.001; gestational age at birth: LogFC = 0.49, FDR < 0.0001), Antithrombin-III (SERPINC1; preterm birth: LogFC= −1.70, FDR < 0.0001; gestational age at birth: LogFC = 0.34, FDR < 0.001), and Fibrillin-2 (FBN2; preterm birth: LogFC= −1.82, FDR < 0.001; gestational age at birth: LogFC = 0.41, FDR < 0.001).

Fig. 1
Fig. 1
Full size image

Differential protein abundance analysis: One point per protein depicts fold-changes and P-values for 1,221 proteins in differential protein abundance analysis (N = 89) for preterm birth (A), gestational age (B), and birthweight for gestational age z-score (C). Dashed line at P = 0.05, and gene symbols shown for top ten significant (FDR < 0.05) proteins with the largest fold-change magnitudes. Models adjusted for fetal sex (male; female), maternal age (years), maternal prepregnancy BMI (kg/m2), maternal tobacco use during pregnancy (yes; no), maternal alcohol use during pregnancy (yes; no), maternal education (< High School; High School completion; graduated college or technical school; some graduate work or graduate/professional degree), delivery method (vaginal; cesarean section), labor type (spontaneous; spontaneous, augmented), gravidity (number), and fetal genetic ancestry principal components. Heatmap (D) of proteins significantly associated with preterm birth depicts hierarchical clustering of protein abundances across individual proteins (rows) and participants (columns)

Steryl-sulfatase (STS) and Endophilin-B2 (SH3GLB2) were upregulated in preterm births and at lower gestational ages, while Surfeit locus protein 4 (SURF4), Collagen alpha-2(I) chain (COL1A2), Antithrombin-III (SERPINC1), and Fibrillin-2 (FBN2) were upregulated in term births and at higher gestational ages (Fig. 1a, b).

The 295 proteins significantly associated with preterm birth mapped to 314 unique gene symbols, of which 9 (2.9%) were also significantly associated with placental mRNA expression in a prior analysis [38]. In proteomic and transcriptomic analyses, 8 of the 9 genes showed consistent changes in expression levels, either both upregulated or both downregulated (Supplemental Table 2). Protein and mRNA expression levels were significantly correlated for only 4 of these 8 genes.

In gene set enrichment analysis with ROAST, 15 pathways (4 downregulated and 11 upregulated) were significantly (FDR < 0.05) associated with preterm birth, and 28 pathways (20 downregulated and 8 upregulated) were associated with gestational age at birth (Fig. 2). As expected, pathways downregulated in association with preterm birth were typically upregulated in association with higher gestational age at birth and vice versa.

Fig. 2
Fig. 2
Full size image

Rotation gene set testing for KEGG pathway enrichment: Rotation gene set testing with 99,999 rotations for KEGG pathway enrichment in the CANDLE proteome data (N = 89), excluding KEGG pathways with fewer than 5 genes, for preterm birth (A) and gestational age (B). KEGG pathways with FDR < 0.05 shown. Models adjusted for fetal sex (male; female), maternal age (years), maternal prepregnancy BMI (kg/m2), maternal tobacco use during pregnancy (yes; no), maternal alcohol use during pregnancy (yes; no), maternal education (< High School; High School completion; graduated college or technical school; some graduate work or graduate/professional degree), delivery method (vaginal; cesarean section), labor type (spontaneous; spontaneous, augmented), gravidity (number), and fetal genetic ancestry principal components

In the PCA analysis, ten PCs were retained which together explained 54% of the variance in the proteome data (Supplemental figure). Despite the fact that PCA is agnostic with respect to covariate and outcome data, many of the top PC loadings were proteins associated with preterm birth and gestational age (Fig. 3). For example, the top two protein loadings for PC1 were SURF4 and COL1A2, which were among the top proteins associated with both preterm birth and gestational age. COL1A2 was also the top loading for both PC3 and PC4. The SERPINC1 protein, which was strongly associated with preterm birth and gestational age, was among the top loadings for PC2, PC5 and PC7, while the MARCKS protein, which was strongly associated with gestational age, was a top loading for PC3. Thus, proteins strongly linked to preterm birth and gestational age were frequently among the most significant in explaining total variance within the proteome data.

Fig. 3
Fig. 3
Full size image

Principal component analysis (PCA) of the placental proteome in CANDLE: Principal component analysis (PCA) of the placental proteome in CANDLE using Horn’s parallel analysis to determine the number of PCs to retain. Top 12 loadings and variance explained for each retained PC shown only for PCs significantly associated with any birth outcome (P < 0.05)

PC1, PC3, PC5, and PC7 were significantly associated with both preterm birth and gestational age, while PC2 was associated with gestational age but not preterm birth (Table 2). Associations between the PCs and birth outcomes were generally consistent with individual protein-birth outcome associations. For example, placental SURF4 protein abundance was negatively associated with preterm birth in individual protein analyses (Fig. 1a). Consistent with the positive SURF4 loading on PC1, each one-unit increase in PC1 was associated with 25.4% [8.7% to 48.0%] lower odds of preterm birth (Table 2).

Table 2 Associations of proteome principal components with birth outcomes
Full size table

Three components were retained in sPLS-DA which together achieved a cross-validation error rate of 0.049 and a BER of 0.123 (Fig. 4, Supplemental Table 3). In other words, weighted by the number of term and preterm individuals, 12.3% of participants were wrongly classified, while sPLS-DA was able to correctly predict whether an individual was born term or preterm based on their proteome data for 87.7% of individuals. All 30 proteins in sPLS-DA component 1 were significantly associated with preterm birth in the DEqMS differential protein abundance analysis. Only 4 out of the 50 proteins in component 2 and 7 out of the 40 proteins in component 3 were significantly associated with preterm birth in the DEqMS analysis. Although one-sided t-tests indicated that the 3-component solution was optimal, preterm versus term birth prediction with sPLS-DA only achieved minor improvements with each successive component added to the model (Supplemental Table 3). For example, overall error and BER were 0.071 and 0.131 when including component 1 alone to predict preterm birth.

Fig. 4
Fig. 4
Full size image

sparse Partial Least Squares Discriminant Analysis (sPLS-DA) predicting preterm birth using placental proteome data: Separability of term and preterm participants (A) and top 20 loadings (B) for each component retained in sparse Partial Least Squares Discriminant Analysis (sPLS-DA). Orange is preterm birth and blue is term birth

Conclusions

In this placental proteomic study including 99 participants, timing of birth was strongly associated with the placental proteome, with 24% of proteins showing differential abundance in relation to preterm birth and 30% with gestational age. These widespread associations extended to molecular pathways, with 15 KEGG pathways differentially regulated in relation to preterm birth and 28 in relation to gestational age. Proteins most significantly associated with birth timing, including SURF4 and COL1A2, also contributed strongly to overall placental proteomic variance in PCA. Finally, proteomic profiles could accurately predict preterm versus term birth in sPLS-DA, highlighting their potential utility as early biomarkers of risk.

To the best of our knowledge, this is the first placental proteomic study of birth outcomes utilizing a label-free global quantitative proteomics approach, and one of the only studies to adjust for key factors that may confound associations between the proteome and preterm birth, such as maternal pre-pregnancy BMI, maternal tobacco and alcohol use during pregnancy, maternal socioeconomic status, and genetic ancestry. According to several meta-analyses and reviews [13,14,15], there have been three prior proteomic studies of preterm birth focused on placental tissues [18,19,20] (and one proteomic study of preterm premature rupture of membranes [45]). Two of these studies [18, 19] utilized two-dimensional (2D) differential gel-based electrophoresis analyses (e.g., 2D-PAGE and 2D-DIGE) while the third utilized multiple omics technologies, including a targeted proteomics panel [20]. Finally, while several studies have explored protein expression changes associated with preeclampsia [46] and preterm birth [13, 14, 19], less is known about how the proteome may affect or be impacted by birthweight. Here, we found no individual proteins or molecular pathways that were differentially regulated in association with birthweight for gestational age z-score.

Many of the top proteins associated with preterm birth and gestational age here play critical roles in placental biology, pregnancy complications, and birth outcomes. SURF4, COL1A2, STS, SH3GLB2, SERPINC1, and FBN2 were strongly associated with birth timing and contributed substantially to proteomic variance. SURF4, an integral membrane protein that mediates protein transfer from the endoplasmic reticulum to the golgi [47], has been implicated in early onset preeclampsia in humans [48] and shown to be essential for embryonic development in mice [49]. COL1A2 and FBN2 (Fibrillin-2) are components of connective tissues contributing to the amniotic membrane’s tensile strength [50, 51]. COL1A2 mutations have been implicated in an increased risk of preterm birth associated with preterm premature rupture of membranes [52, 53], while Fibrillins are suspected candidate genes for preterm birth risk owing to their roles in extracellular matrix components of the fetal membranes and cervix [52]. STS catalyzes the hydrolysis of estrone(E1)-sulfate and dehydroepiandrosterone(DHEA)-sulfate, the major circulating precursors of estrogens and androgens [54]. Elevated placental STS and steroid precursors have been previously implicated in early onset preeclampsia [55], polycystic ovary syndrome prior to pregnancy [56], and preterm birth [57]. SERPINC1 encodes antithrombin III, a serine protease inhibitor that regulates coagulation. Lower placental expression of SERPINC1 associated with preterm birth here aligns with the placental dysfunction seen in antithrombin deficiency-related pregnancy and birth complications, including intrauterine growth restriction and preterm birth [58]. A related serine protease inhibitor, SERPINA1 (alpha-1 antitrypsin), has also been linked to spontaneous preterm birth in placental proteomic studies, though it acts through a distinct mechanism, functioning mainly in protease-mediated inflammatory control rather than coagulation [19]. SH3GLB2, although not previously linked to birth outcomes, participates in endocytosis [59], a process central to placental nutrient and signal transfer [60]. Together, these proteins span multiple domains of placental and vascular biology, including steroid metabolism, vesicular transport, extracellular matrix organization, and vascular homeostasis, processes that are commonly disrupted across pregnancy complications and adverse birth outcomes.

KEGG pathways that were differentially regulated in association with preterm birth and gestational age here point to mechanisms that may increase preterm birth risk. The placenta has high energy needs compared with many other tissues and may modulate energy production and use in response to stressful conditions [61]. Here, preterm birth was associated with the upregulation of several energy production pathways. The AMPK signaling pathway regulates energy homeostasis, promoting energy (adenosine triphosphate [ATP]) production and inhibiting anabolic pathways. The fatty acid degradation pathway breaks down fatty acids into acetyl-coenzyme A (CoA), and propanoate metabolism produces succinyl-CoA, two key molecules in the production of ATP via the citric acid cycle. Several directional KEGG pathway changes in association with preterm birth also implicate elevated inflammation and oxidative stress. The pentose phosphate pathway was downregulated in association with preterm birth. This pathway generates NADPH, which contributes to oxidative stress mitigation by converting oxidized to reduced glutathione [62]. Additionally, reactive oxygen species may activate proinflammatory transcription factors resulting in enhanced chemokine signaling [63], which was upregulated here in association with preterm birth.

Some of the genes with differential placental protein abundance associated with preterm birth in this study also displayed differential mRNA expression in a prior analysis in this cohort [38] and may relate to higher preterm birth risk. Eight of the nine genes with differential protein and mRNA expression associated with preterm birth were differentially regulated in the same direction. Protein and mRNA expression levels were significantly correlated for only half of these genes, consistent with prior reports of weak transcript-protein concordance which may reflect post-transcriptional and post-translational regulation that influences protein abundance beyond mRNA expression [5,6,7]. Our results highlight that even in the absence of strong correlation, molecular features originating from the same gene can have directionally consistent associations with health outcomes. Among the overlapping genes, the CD81 protein was the most strongly differentially regulated in this proteomic analysis according to its P-value and fold-change. CD81 is a cell surface protein widely accepted as a marker for isolating extracellular vesicles [64, 65]. Lower protein and mRNA expression of this gene in association with preterm birth is consistent with prior studies reporting a link between lower placental derived extracellular vesicles and preterm birth, fetal growth restriction, and small for gestational age [65, 66]. Additionally, CD81 modulates immune and infection-related processes, influencing both B- and T-cell signaling pathways [67,68,69], while infection/inflammation are well known factors contributing to preterm birth [70]. It remains unclear whether altered CD81 expression contributes independently to preterm birth or is elevated because both CD81 and preterm birth are linked to infection/immune activation. In KEGG analysis, the only overlapping proteomic and transcriptomic pathways were regulated in different directions. Proteomic signatures supported upregulation of the VEGF and chemokine signaling pathways, while transcriptomic analyses found downregulation of those pathways in association with preterm birth [38].

Placental proteins and molecular pathways implicated in preterm birth here have also been shown to be differentially regulated in association with preterm birth risk factors including maternal stress and environmental exposures during pregnancy. For example, in a prior placental transcriptomic analysis including a larger number of CANDLE participants, maternal stressful life experiences during pregnancy and during the mother’s own childhood (preconception stress) were associated with upregulation of mTOR signaling [71], which was also upregulated in association with earlier gestational age at birth. Another transcriptomic study in CANDLE found that prenatal exposure to an endocrine disrupting plasticizer, mono (carboxyisooctyl) phthalate (MCIOP), was also associated with upregulation of placental mTOR signaling [72]. These overlapping findings support the utility of multi-omic signatures in identifying mechanisms linking prenatal exposures with birth outcomes.

Highly accurate prediction of term versus preterm birth using proteomic data in sPLS-DA suggests a potential clinical application for identifying at-risk pregnancies. While the low cross-validation error rates in sPLS-DA here indicate effective classification [73], predictive schemes relying on metabolomic or proteomic data often fail to replicate across multiple populations, possibly owing to heterogeneity in specimen collection and processing, omics methodologies, or population characteristics [74]. Replication in more accessible biospecimens, such as blood or urine, is also essential for clinical translation. For example, a recently developed serum proteomic screen based on the ratio of insulin-like growth factor-binding protein 4 (IGFBP4) to sex hormone-binding globulin (SHBG) has shown promising predictive performance for early preterm birth [75, 76]. However, its real-world effectiveness may be more limited, as a recent clinical trial found that among those identified as high-risk by the screen, only 2% delivered preterm [77]. Proteomic markers identified in the placenta, including those reported here, could guide the development of blood-based assays with improved predictive performance or greater mechanistic relevance.

Our results should be considered in the context of several limitations. First, it remains unclear whether these proteomic signatures precede or arise as a consequence of birth outcomes. As with any observational study, unmeasured confounding may influence exposure-outcome associations, and generalizability may be limited, particularly to populations with different racial or ancestral compositions. Although several proteins and pathways identified here have been observed in prior placental or proteomic studies of preterm birth, these birth outcome-protein associations have not been formally replicated in independent cohorts and should be interpreted with caution. In addition, our sample size did not permit stratified analyses across extremely, moderate, and late preterm groups. Different biological pathways may underlie these subgroups, and future studies with larger sample sizes will be needed to determine whether proteomic signatures vary across degrees of prematurity. Gene set enrichment analysis may be biased toward well-characterized proteins more likely to appear in curated gene sets. Dimensionality reduction methods like PCA offer an unsupervised, unbiased characterization of the proteome but can be difficult to interpret. Notably, however, many of the top proteins in PCA here were linked to preterm birth and gestational age. Finally, the predictive accuracy of sPLS-DA requires validation in external populations, and future studies with larger sample sizes are needed.

In conclusion, the placental proteome was strongly connected with birth timing, with consistent patterns linking protein and biological pathway expression to preterm birth and gestational age. Underscoring the strong connection between birth timing and the placental proteome, proteins most significantly associated with preterm birth and gestational age were also among the top contributors to variance explained in PCA. Proteins associated with birth timing in this study may shed light on the mechanisms underlying adverse birth outcomes and serve as targets in future studies aimed at preventing preterm birth. Finally, the measurement of 30 proteins as a screen for preterm birth risk via sPLS-DA represents a promising clinical application, but requires replication and needs to be evaluated in more easily sampled biospecimens during pregnancy.