Participant characteristics

In total, we collected 2,590 plasma samples from pre- and full-term pregnancies. Plasma samples were collected once per pregnant participants, between 12 and 28 weeks of gestation from three independent hospitals in China: Jiangmen maternal and child healthcare hospital (JM), Foshan maternal and child healthcare hospital (FS), and Nanfang hospital of southern medical university (NFY). Of the 2,590 participants, 518 women experienced a preterm delivery, while the remaining 2,072 women delivered at full-term (Table 1). Participants from JM were collected between January 1, 2017 and December 30, 2020. Participants enrolled at FS were recruited between December 1, 2018 and December 1, 2020. Participants from NFY were enrolled between May 1, 2016 and May 1, 2020. The institutional ethics committees of all hospitals approved this retrospective analysis, and informed written consent was obtained from all participants (IRB number: 2019053).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Clinical characteristics of pregnancies in four cohorts.

https://doi.org/10.1371/journal.pmed.1004571.t001

Pregnancies were excluded if they met any of the following criteria: (1) Multiple pregnancies; (2) Uterine fibroids or malformation; (3) Chorioamnionitis; (4) Chromosomal or congenital abnormalities; (5) Pregnancies with assisted reproductive technology. After exclusion, samples were retrospectively assigned to a birth outcome group based on their subsequent delivery time, with spontaneous premature delivery defined as birth before 37 weeks of gestation. Gestational age was determined based on the last menstrual period and ultrasonography. PTB was classified as spontaneous if a woman presented with either cervical dilation and/or preterm premature rupture of membranes (PPROM) and delivered before 37 weeks of gestation. Spontaneous labor with intact membranes accounted for 53.1% of spontaneous preterm birth (sPTB), while PPROM accounted for 46.9%. Delivered for maternal indications (e.g., preeclampsia) or fetal indication (e.g., fetal distress) were excluded from the sPTB category. Full-term controls included pregnancies with a gestational age exceeding 38 weeks, excluding those with pregnancy complications. Following the case-control ratios from previous case-control studies [14–16], four full-term controls were selected for each preterm pregnant case, matched by maternal age, gestational age at sampling, and body mass index (BMI). The clinical characteristics of preterm pregnancies and their corresponding controls were well matched across four cohorts (Table 1; all P-value >  0.05, Mann–Whiney U test).

DNA-Seq processing and promoter profiling analysis

The procedure for sample collection, cfDNA isolation, and DNA sequencing are detailed in the S1 Text. We estimated the fetal fraction using the proportion of all sequencing reads from the Y chromosome or the seqFF method [17]. Gene information was obtained from the RefSeq of the University of California Santa Cruz Genome Browser Database [18]. For each transcript, the promoter region, spanning from −1 to +1 kb around the transcriptional start site, was defined as pTSS regions. pTSS regions overlapping with the Duke blacklist region were removed (http://hgdownload.cse.ucsc.edu/goldenpath/hg19/encodeDCC/wgEncodeMapability/). After sequencing, the raw reads were aligned to the human reference genome (hg19) using bwa-mem (ver. 0.7.4). Polymerase chain reaction (PCR) duplicates were removed using the rmdup function of SAMtools (ver. 1.2). GC-bias correction was implemented using deeptools (ver. 3.5.0) with default settings. Read coverage for each pTSS region was extracted using bedtools (ver. 2.17.0). We normalized the read coverage data using the following formula:

Since the pTSS region was defined as −1 to +1 kb around the TSS, the pTSS length of each gene is equivalent to 2,000. Read depth refers to the number of cfDNA reads overlapping specific genomic regions. cfDNA coverage around TSS refers to the reads of cfDNA overlapping with the pTSS region, which reflects the extent of nucleosomal footprints at the promoter.

Gene expression profile analysis and gene information acquisition

Placenta and whole blood expression profiles for preterm pregnancies (GSE73685) [19] were downloaded from the Gene Expression Omnibus (GEO) database. Additional placenta (GSE148402 [20] and GSE174415 [21]) and whole blood expression profiles of preterm pregnancies (GSE46510 [20] and GSE59491 [22]) were also downloaded from GEO and normalized using GEOquery (ver. 3.3.1). The top 500 highly expressed and 500 least expressed genes, representing the highest and lowest mean expression levels in the placenta and whole blood, were identified by analyzing their expression profiles in GSE73685 (S1 Table). Placenta- and blood-enriched genes were obtained from the study by Gong and colleagues [23] and PaGenBase [24], respectively (S2 Table). Genes with expression levels below 0.1 in all tissues according to FANTOM5 were defined as unexpressed genes. Housekeeping genes, defined as those widely expressed across various tissues, along with unexpressed genes were obtained from previous studies (S3 Table) [9].

Genes with significant differential promoter coverages

At the discovery stage, we selected 20 PTB cases and 20 full-term pregnancies (S4 Table) and then performed whole-genome sequencing of their cfDNA. After data processing and normalization, pTSS coverages between pre- and full-term samples were compared to calculate the P-value using the Wilcoxon rank-sum test. The raw P-values were adjusted to the false discovery rate (FDR) using the Benjamini–Hochberg procedure. Genes with significant differential coverages in the pTSS regions were identified by log₂ | fold change | ≥  1 and FDR ≤  0.05 (S5 Table).

Sample clustering and gene function annotation

Principal component analysis (PCA) was performed using the prcomp function of the stats package, and then the results were plotted using the ggord package (ver. 1.1.7). Hierarchical clustering of the coverage data with complete linkage clustering algorithms was implemented using the pheatmap package (ver. 1.0.2). Enrichment analysis of gene ontology (GO) was completed using Metascape (ver. 20220101) [25], and Kyoto Encyclopedia of Genes and Genomes (KEGG) was completed using clusterProfiler (ver. 3.18.1) [26] with their default settings.

Gene correlation network construction

To construct the gene correlation network with genes exhibiting differential coverages, functional relationships were obtained from the String database (ver. 11.5) [27], which served as a reference for known protein-protein interactions and guided the connections displayed in the network. The network was then visualized with Cytoscape (ver. 3.8), which depicted gene interactions by illustrating interactions according to correlation strengths. Cytohubba (ver. 0.1) [28] was then used to assess centrality and importance within the network, aiding in the identification of hub genes. Finally, 10 genes with the highest gene degrees as the top 10 hub genes were selected.

Predictive classifier construction and validation

To develop classifiers for predicting spontaneous PTB, we preformed whole-genome sequencing of cfDNA on 2,590 pregnant women, including 518 preterm and 2,072 full-term pregnancies from three independent hospitals: JM, FS, and NFY. The 20 preterm and 20 full-term pregnancies used in the discovery stage were randomly selected from JM, with well-matched gestational age, maternal age, and BMI (S4 Table). These samples were also included in the training stage. Samples collected from JM (n = 1,310) were randomly divided into training (n = 915, training cohort [n = 915]: 183 cases and 732 controls) and internal validation cohorts (validation1 cohort [n = 395]: 79 cases and 316 controls) at a ratio of 7:3. Samples from FS (validation2 cohort [n = 930]: 186 cases and 744 controls) and NFY (validation3 cohort [n = 350]: 70 cases and 280 controls) were used as external validation cohorts. The clinical characteristics of the pre- and full-term pregnancies were well-matched across the four cohorts (Table 1). The workflow for classifier construction is illustrated in S1 and S2 Figs. Since many studies have reported that discrete data may improve the predictive performance [29–31], we discretized the read coverage of each pTSS identified in the discovery cohort according to the optimal cut-off point with the largest AUC value in the training cohort before building the classifiers. A total of 277 genes exhibiting differential coverages were subsequently used for classifier construction. The detailed workflow of classifier construction has been illustrated S1 Fig. Previous studies have demonstrated that discretization enhances data interpretability, uncovers non-linear relationships, harmonizes mixed-type datasets, and facilitates the derivation of count data from continuous variables [32]. Therefore, the read coverage for each promoter in each subject was set to one when it was larger than the corresponding optimal cut-off; otherwise, it was set to zero. Then, the sigFeature package (ver. 1.8.0) was used to evaluate the importance of the pTSS regions [33]. Pearson correlation coefficients were calculated for all pairs of genes with differential coverages. Highly correlated genes with higher importance were retained (S5 Table, | r | >  0.5). To minimize collinearity in the model predictors, 49 highly correlated variables were removed, resulting in a final set of 228 genes for further analysis (S5 and S6 Tables).

We then evaluated four predictive models, including support vector machine (SVM), logistic regression (LR), random forest (RF), and XGBoost (XGB) as the basis for developing a novel predictive classifier for PTB, referring to published studies [29,34]. The SVM model was constructed using the linear kernel in the e1071 package (ver. 1.7.9) with default settings. The LR model was developed using the glm functions of the stats package. The RF model was constructed using the randomForest package. The XGBoost was constructed using the xgboost package. The grid search method was employed to determine the optimal hyperparameters for the random forest and XGBoost models. Each of these predictive models was implemented independently using both backward and lasso algorithms for feature selection. For backward-like feature selection, one feature was deleted at a time, and the classifier with the maximum AUC after 10-fold cross-validation (CV) among all classifiers was selected. The feature to be deleted was chosen according to the maximum AUC increase after deletion. For lasso feature selection, the best lambda was identified using 10-fold CV with the cv.glmnet function from glmnet package (ver. 4.1), using parameters: type.measure = “AUC” and family = “binomial”, and then the features with non-zero coefficients were selected. These selected features were independently used as inputs for the LR, SVM, random forest, and XGBoost models to develop classifiers for PTB prediction. The robustness of the trained classifiers was assessed using the leave-one-out cross-validation (LOOCV) method. Briefly, each subject in the training cohort was withheld in turn, and the remaining subjects were used to train the classifier. The trained classifier was then used to predict the class of the withheld subject. This process continued until all subjects in the training cohort had been judged. Based on the AUC after LOOCV, the classifier with the highest AUC, named PTerm, was selected for further validation. PTerm is based on SVM model with the backward feature selection algorithm, ultimately including 83 genes (S7 Table).

To further assess the performance of PTerm, whole-genome sequencing data from one internal validation (JM, validation1 cohort) and two external validation cohorts (FS, validation2 cohort; NFY, validation3 cohort) were used. The performance of PTerm was further evaluated using data from these internal and external cohorts.

Statistical analysis

The power calculation of the sample size in the discovery stage was around 80% with 20 premature delivery and 20 full-term pregnancies (S1 Text). Two-sided Wilcoxon rank-sum test was used to compare the continuous variables between the pre- and full-term groups, while two-sided Pearson’s χ² and Fisher’s exact tests were used for comparisons of categorical variables. Clinical data was presented as mean ±  standard deviation (SD). The Wilcoxon rank-sum test was used to identify genes with differential read coverages within the pTSS regions and P-values of <  0.05 in two-sided tests were considered to be statistically significant. The mean coverage levels of the top 500 highly expressed and the bottom 500 least expressed genes were first calculated for each individual in preterm pregnancies. Subsequently, the differences in coverage between these two groups of genes were analyzed using the Wilcox Signed-Rank test. A similar approach was applied to evaluate the coverage differences between housekeeping genes and unexpressed genes. The R software (ver.4.2) was used for statistical analysis. ROC curves and the significant differences in their AUC, sensitivity, and specificity were plotted and calculated using the pROC package in R.

cfDNA carries information about its origin in pregnant women

Previous studies have reported that cfDNA carries information regarding its tissues of origin [8–10,35], making it suitable for evaluating PTB. Thus, we designed experiments to characterize the cfDNA profiles of pregnancies resulting in pre- and full-term births (Fig 1). To this end, we collected whole-genome sequencing data from 20 preterm and 20 full-term pregnancies (S4 Table). Additionally, we collected the RNA expression profiles of both the placenta and whole blood from preterm pregnancies (GSE73685) [19].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic overview of predicting preterm birth using promoter profiling of plasma cfDNA.

During pregnancy, plasma cell-free DNA (cfDNA) is mainly derived from placental trophoblasts and maternal hematopoietic cells, released by their apoptotic cells. Exposed DNA not bound to nucleosomes is digested, while nucleosome-bound DNA escapes digestion and enters into the maternal circulation. Through whole-genome sequencing, we found that read coverages at pTSS regions (−1 to +1 kb around the transcription start site [TSS]) reflect the gene expression patterns of their tissues of origin. Since premature delivery is closely associated with dysfunction and changes in the placenta and maternal immune system, we proposed that cfDNA coverages at pTSS regions could predict preterm birth at early gestational ages. We tested this hypothesis using high-throughput whole-genome sequencing of plasma cfDNA from 2,590 pre- and full-term pregnancies across three independent hospitals. By comparing their promoter profiling, we found differences between pre- and full-term pregnancies. We then used genes with differential promoter coverage and four machine-learning models to develop predictive classifiers for preterm birth. Nucleosome-depleted regions are typically found upstream of the TSS. To show greater differences, all nucleosomes in the promoter regions of highly expressed genes are depicted as depleted in the figure.

https://doi.org/10.1371/journal.pmed.1004571.g001

We first compared the read coverage at the pTSS (primary transcription start site, defined as the region ranging from −1 to + 1 kb around the transcriptional start site) between the 500 most highly and 500 least expressed genes in the preterm placenta. The 500 most highly expressed genes showed reduced depth at the pTSS regions compared to the 500 least expressed genes (Fig 2A and 2B; P-value =  1.9e−06, Wilcoxon Signed-Rank test). Additionally, the housekeeping genes with highly expressed levels exhibited lower read depth, whereas the unexpressed genes exhibited higher read depth (Fig 2C and 2D; P-value =  1.9e−06, Wilcoxon Signed-Rank test). Similar patterns were observed in maternal whole blood data (S3 Fig). To ensure the robustness of our findings, we extended our analysis by incorporating additional high/low expression data from four other datasets representing diverse ethnic backgrounds, including samples from South Korea (GSE148402), China (GSE174415), Canada (GSE59491 and GSE46510), and the USA (GSE73685). These results are consistent with those in Figs 2B, 2D, and S4, supporting the reliability of our methodology. Therefore, we confirmed that the coverage of plasma cfDNA at the pTSS regions closely correlates with the expression profiles of its original tissues, suggesting that promoter profiling could reflect the expression status of its original tissues. Next, we focused on the cfDNA profiles of placenta-enriched genes, which were closely related to placental functions. The results revealed that placenta-enriched genes were characterized by reduced coverage at the pTSS regions in PTB pregnancies compared to full-term pregnancies (Fig 2E; P-value =  7.386e−11, Wilcoxon rank-sum test), implying that there may be broad differences in the promoter profiling of these two groups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. cfDNA profiles of promoter regions reflect nucleosome positioning in pregnant women. (

a) Average expression of the 500 most- (Top500, red) and least-expressed (Bottom500, blue) genes in the placenta of preterm birth pregnancies. (b) Read depth of whole-genome sequencing across pTSS regions (spanning from −1 to +1 kb around TSS) for the 500 most- (Top500, red line) and least-expressed (Bottom500, blue line) genes. The read depth of the Top500 genes was lower than that of the Bottom500 genes (P-value =  1.9e−06, Wilcoxon Signed-Rank test). (c) Average expression levels of housekeeping (red) and unexpressed (blue) genes in the placenta. (d) Read depth of whole-genome sequencing at the pTSS region of the housekeeping genes (red line) was lower than that of unexpressed genes (blue line) in the placenta (P-value =  1.9e−06, Wilcoxon Signed-Rank test, including 2,985 housekeeping and 670 unexpressed genes). (e) Sequencing read depth of placenta-enriched genes was more depleted in preterm (yellow line) pregnancies than in full-term (green line) pregnancies (P-value =  7.386e−11, Wilcoxon rank-sum test). The RNA expression profiles of the placenta from preterm pregnancies were downloaded from GEO (GSE73685). For the boxplot, the center line represents the median of the data distribution. The box limits denote the interquartile range (IQR), with the lower and upper edges corresponding to the first (Q1) and third quartiles (Q3), respectively. The whiskers extend to the smallest and largest values within 1.5 times the IQR from Q1 and Q3. The Top500 and bottom500 genes are based on the RNA-Seq data of placental tissues from the preterm pregnancies in GSE73685. Genes with expression levels lower than 0.1 in all tissues in FANTOM5 are defined as unexpressed genes. Placenta-enriched, unexpressed, top500, and bottom500 genes are shown in S1–-S3 Tables. The pTSS region is denoted by grey dashed lines. The areas with light colors along the mean lines represent standard error of mean (SEM). PTB,  preterm birth; TSS,  transcriptional start site; cfDNA,  cell-free DNA.

https://doi.org/10.1371/journal.pmed.1004571.g002

Promoter profiling of cfDNA revealed distinct patterns between PTB and controls

We then investigated whether cfDNA promoter profiling of pre- and full-term pregnancies demonstrated any deviations in patterns. By comparing their cfDNA promoter profiling, we identified 277 genes with differential coverages at the pTSS (Fig 3A; | Log₂ fold change | >  1 and FDR <  0.05). These genes included 146 genes with increased coverage and 131 genes with decreased coverage (S5 Table). Next, we used PCA on these genes and found that samples from the same group had similar promoter profiling (S5 Fig). More importantly, unsupervised clustering analysis produced a heatmap that revealed distinct differences in promoter coverage for pre- and full-term pregnancies (Fig 3B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Differences in pTSS profiles between pre- and full-term pregnancies.

(a) Volcano plots showing genes with differential read coverages within the pTSS regions (spanning from −1 to +1 kb around TSS) between 20 preterm birth (PTB) and 20 full-term pregnancies. A total of 277 transcripts with differential read coverages within pTSS regions were identified (|log₂ fold change | ≥  1 and false discovery rate [FDR] ≤  0.05). The red, blue, and grey dots indicate transcripts with increased, decreased, and non-differential coverage, respectively. The x- and y-axes represent the log fold change and P-value, calculated by the two-sided Wilcoxon rank-sum test, respectively (n = 40, including 20 preterm and 20 full-term pregnancies). The raw P-value was adjusted to the false discovery rate (FDR) using the Benjamini-Hochberg procedure. The top five up-regulated and down-regulated genes are marked in the volcano plot. (b) Heatmap showing the z-scores of genes with differential read coverages at pTSS, generated by the pheatmap package (ver. 1.0.2) using the complete-linkage clustering algorithm. (c) Gene Ontology enrichment analysis of transcripts with differential coverage between PTB and full-term groups using Metascape (ver. 20220101). (d) KEGG pathway enrichment analysis of transcripts with differential coverage between PTB and full-term groups using clusterProfiler (ver. 3.18.1). (e) Gene correlation network for transcripts with differential coverage between PTB and full-term groups, with gene correlations evaluated using the String database (ver. 11.5) and network visualization by Cytoscape (ver. 3.8). (f) Degrees and heatmap of hub gene interconnections within the correlation network. The degree represents the importance of the genes in the network, evaluated by cytoHubba (ver. 0.1). (g) Correlation network of hub genes.

https://doi.org/10.1371/journal.pmed.1004571.g003

GO and KEGG enrichment analyses were used to annotate the functions of the genes with differential coverages at pTSS. The results of GO enrichment showed that the terms associated with cell junction organization, response to mechanical stimulus, apoptosis, and development were closely related to embryonic development and premature delivery (Fig 3C). For instance, previous studies have shown that apoptosis of fetal membranes could plausibly contribute to the risk of PTB [36]. Additionally, KEGG enrichment analysis revealed that a large proportion of the enriched pathways were closely associated with embryonic development and PTB (Fig 3D).

Finally, we sought to find the potential key genes associated with PTB using a gene correlation network (Fig 3E). The analysis of gene functional connections allowed us to evaluate the degree of gene influence and importance, which may help identify essential genes in the occurrence and progression of PTB. According to gene degree, our evaluations identified the top 10 hub genes with the maximum degree (Fig 3F). These genes included ERBB2, ESR1, NFKBIA, HSPA5, PRKCB, RAF1, NFE2LE, SNAI1, GSN, and ATF3 (Fig 3F). The correlation network showed the close relationship among the hub genes (Fig 3G). Literature retrieval revealed that the 10 hub genes were associated with PTB, embryonic development, and pregnancy (S8 Table). For example, the ESR1 gene polymorphism is associated with premature delivery, with its DNA methylation patterns also showing distinct differences between pre- and full-term pregnancies [37,38]. Similarly, NFKBIA degradation could activate NF-κB, resulting in the production of proinflammatory IL-6, and inflammation is closely associated with PTB [39]. In particular, the expression of ATF3 is significantly decreased in preterm placentas, and ATF3 is the regulator of soluble fms-like tyrosine kinase 1 (sFlt-1) and soluble Endoglin (sEng), which are important markers of premature delivery and preeclampsia [40].

Promoter profiling of plasma cfDNA can predict PTB

To further validate the potential of cfDNA promoter profiling in predicting PTB, we established a large-scale, multi-center, case-control study that included 2,590 pregnant women, consisting of 518 preterm and 2,072 full-term pregnancies, from three independent hospitals (Fig 4). According to the rank of gene importance and gene correlation coefficient, highly related genes were filtered, and 228 genes were retained (S5 Table). Our training stage focused on the 228 genes with differential coverage at the pTSS identified in the discovery stage. We then used four predictive models (SVM, LR, RF, and XGB) and two feature selection algorithms (backward and lasso algorithms) to develop the optimal predictive classifier. We found that the AUC of the optimal classifiers for each model based on the backward feature selection algorithm was significantly higher than those of classifiers with the lasso algorithm (Fig 5A–5D and S9 Table; all P-values <  0.05, DeLong’s test). More importantly, we found that a classifier that relied on the SVM model and backward algorithm, named PTerm, performed well as the best predictor of PTB (AUC =  0.878 [0.852–0.904], accuracy =  87.3%, and recall =  88.5%; S10 Table). PTerm exhibited the largest AUC value after LOOCV (0.878 [0.852–0.904]), and its AUC value was higher than those of the optimal classifiers produced using the LR, RF, and XGB models (Fig 5A and 5B and S11 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Pipeline for preterm birth classifier construction.

In this study, 2,590 plasma cfDNA samples (518 preterm birth and 2,072 full-term pregnancies) were collected from three independent hospitals, including Jiangmen maternal and child healthcare hospital (JM)^a, Foshan maternal and child healthcare hospital (FS)^b, and Nanfang Hospital of Southern Medical University (NFY)^c. These samples were collected between 12 and 28 weeks of gestation. According to their subsequent delivery time, the pregnant women were categorized into preterm or full-term groups. The whole-genome sequencing data was then used to develop classifiers for predicting PTB via a three-step process: discovery, training, and validation. In the discovery stage, we identified 277 transcripts with differential coverage at pTSS regions (spanning from −1 to +1 kb around TSS) between the two groups. * The 20 preterm and 20 full-term pregnancies derived from JM used in the discovery stage were also included in the 915 samples of the training stage. In the training stage, we applied non-linear support vector machine (SVM), logistic regression (LR), random forest (RF), and XGBoost (XGB) models, independently augmented with backward and Lasso feature selection algorithms to develop a set of predictive classifiers. The predictive classifier, denoted by PTerm, achieved the largest area under the curve (AUC) was identified and its performance was further validated in the three validation cohorts, including one internal cohort (validation1, JM: n =  395) and two external cohorts (validation2 derived from FS: n =  930; Validation3 derived from NFY: n =  350). Additional details about participant definitions and classifier construction are provided in the Methods section and S1 Text. PTB, preterm birth; GA, gestational age.

https://doi.org/10.1371/journal.pmed.1004571.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Performance of the classifiers in predicting preterm birth.

(a) Receiver operating characteristic (ROC) curves for each of the predictive classifiers using the backward feature selection algorithm. (b) Performance of each of the predictive classifiers using the backward feature selection algorithm. (c) ROC curves for the predictive classifiers using the Lasso feature selection algorithm. (d) Performance of the classifiers using the Lasso feature selection algorithm. (e) ROC curves of the optimal classifier, PTerm. (f) Performance of PTerm across each cohort. (g) Performance of different combinations. (h) Area under the curves (AUCs) of different combinations. In this study, 2,590 plasma cfDNA samples (518 preterm births and 2,072 full-term births) were collected from three independent hospitals. In the training stage, the predictive classifiers were trained using support vector machine (SVM), logistic regression (LR), random forest (RF), and XGBoost (XGB) with backward and Lasso feature selection algorithms (n =  915). Leave-one-out cross-validation (LOOCV) was used to evaluate the performance of the selected classifiers. In the validation stage, the performance of the classifier was evaluated in three cohorts, including the internal cohort (validation1, n =  395) and two external cohorts (validation2, n =  930; validation3, n =  350). The significant differences in ROC curves were compared using the two-sided DeLong’s test. FF, fetal fraction; BMI, body mass index before pregnancy; P+BMI+FF^l, the combination of PTerm, BMI, and fetal fraction using the linear kernel of SVM model; P+BMI+FFⁿ, the combination of PTerm, BMI, and fetal fraction using the non-linear kernel (RBF kernel) of SVM model.

https://doi.org/10.1371/journal.pmed.1004571.g005

In PTerm, the classifier contains 83 genes with a linear SVM model (S7 Table). In addition, 4 of the 10 hub genes (ERBB2, NFKBIA, RAF1, and GSN) were retained in the classifier. Then, we evaluated the performance of PTerm across three validation cohorts, including one internal and two external validation cohorts. Consistent with the results of the training cohort, PTerm exhibited solid predictive capacity in all three cohorts. The AUC for the internal validation cohort was 0.845 (0.799–0.891), and the AUCs for external validation 2 and validation 3 cohorts were 0.833 (0.802–0.863) and 0.812 (0.761–0.863), respectively (Fig 5E and 5F). In addition, PTerm produced an AUC of 0.849 (0.831–0.866) across all cohorts when discriminating between pre- and full-term pregnancies, with a recall of 84.4% and an accuracy of 85.3% (Fig 5E and 5F and S10 Table).

PTB is a diverse etiology with some causes being iatrogenic and others being wholly spontaneous. Spontaneous labor with intact membranes and PPROM are the main types of PTB. In our study, we found that spontaneous labor with intact membranes accounted for 53.1% of spontaneous preterm pregnancies, with a prediction accuracy of 0.855 (S6 Fig). Meanwhile, PPROM accounted for 46.9% of spontaneous preterm pregnancies, with a prediction accuracy of 0.823 (S6 Fig). Additionally, the pathogenic factors of premature delivery in different birth weeks may be different. For our prediction, the accuracy for preterm pregnancies with birth before 35 weeks reached 0.866 (S12 Table), demonstrating the potential application value of our method.

PTerm combined with clinical features

Previous studies have reported that certain clinical features, such as fetal fraction (FF) and BMI before pregnancy, could be used to predict PTB. In our data, we found that the AUCs of BMI (0.527 [0.503–0.551]) and FF (0.526 [0.502–0.550]) were significantly lower than that of PTerm (Fig 5G and 5H; all P-values <  0.05, DeLong’s test). To improve the performance of our classifier, we combined the features of BMI and FF with PTerm. These clinical features were incorporated as variables alongside the 83 original variables in PTerm to construct the classifiers. The AUCs of the combined classifiers (PTerm + BMI, PTerm + FF, and PTerm + BMI + FF) were 0.842 (0.824–0.860), 0.837 (0.819–0.855), and 0.834 (0.815–0.852), which were also significantly lower than that of PTerm alone (Fig 5G and 5H; all P-value <  0.05, DeLong’s test). Previous studies have shown that the clinical features (FF and BMI) might exhibit a non-linear association with premature delivery. Therefore, we applied the non-linear kernel of the SVM model to integrate PTerm with BMI and FF (Integration of clinical features in S1 Text). We found that the AUC of the combined classifiers was 0.849 (0.831–0.866), which was significantly higher than that of the linear model (0.834 [0.815–0.852]; P-value =  0.0054, DeLong’s test) and equal to that of PTerm alone (Fig 5G and 5H). The performance of PTerm when combined with cfDNA and clinical variables (0.833 [0.814–0.851]) was significantly higher than that of clinical variables alone (0.527 [0.504–0.551]), highlighting the clinical utility of our novel model (P-value < 2.2e−16, DeLong’s test).