A blood-based miRNA signature for early non-invasive diagnosis of preeclampsia

This study was approved by the human ethics committee at the Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University (approval number: LCKY2020-242), and informed consent was obtained from all patients and their families. And the study was performed in accordance with the regulations and guidelines established by this committee.

This study was carried out in three phases: discovery, training, and validation. The details of the overall study are shown in Fig. 1. In the discovery phase, we conducted miRNA sequencing in a retrospective cohort (WMU cohort 1) for genome-wide screening of candidate miRNA biomarkers and then verified the expression levels of the candidate miRNA biomarkers in 20 PE patients and 20 healthy controls from WMU cohort 2 using qRT-PCR for identifying robust miRNA biomarkers. In the training phase, a miRNA signature for early diagnosis of PE (miR2PE-score) was developed in WMU cohort 1. In the validation phase, the performance of miR2PE-score was firstly evaluated in two independent multinational cohorts (Canada cohort and USA cohort) from the Gene Expression Omnibus (GEO) database (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​). The Canada cohort (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​acc=​GSE114349) contained 21 normotensive women and 20 preeclamptic women [27], and the USA cohort (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi?​acc=​GSE84260) [28] included 16 normotensive women and 16 preeclamptic women, respectively. To further validate the non-invasive diagnostic performance, the miR2PE-score was further tested in two prospective plasma cohorts, including WMU cohort 3 with 10 PE patients and 8 healthy controls and WMU cohort 4 with 19 PE patients and 18 healthy controls, respectively.

We collected the samples at the Department of Obstetrics, the Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University (Zhejiang, China) between June 2020 and December 2021. Blood samples were obtained from pre-cesarean women with overnight (≈10 h) fasting in 10-ml EDTA-coated Vacutainer tubes before any medical interventions. Five milliliters of blood for obtaining plasma was separated by centrifugation at 3000 rpm for 15 min at 20℃ within 30 min after each collection, and then the supernatants were extracted to a new tube.

Leukocytes were purified from another 5 ml blood using red blood cell lysis solution (Solarbio, China) following the manufacturer’s protocol. In brief, 5 ml blood was incubated with a threefold volume of red cell lysate on ice for 15 min and mixed evenly. The mixture was centrifuged at 450xg for 10 min at 4℃, the supernatant was discarded, and the leukocyte precipitate was obtained. The experiment described above was repeated three times. Finally, red cells were dissolved and leukocytes were washed with PBS.

Placental samples were rapidly processed as previously described [29]. In brief, the tissue was sectioned transversally (~ 5 cm) near the cord insertion site. Both the decidual layer along with the basal plate and the chorionic surface and membranes were removed by dissection. Each sample was rinsed thoroughly in cold physiological saline solution and frozen in liquid nitrogen.

All samples were stored at − 80 °C freezer, where aliquots were stored until assay.

A total amount of 1 μg total RNA per sample was used as input material for the small RNA library. Sequencing libraries were generated using TruSeq® miRNA Sample Prep Kit for Illumina®. Manufacturers’ recommendations and index codes were added to attribute sequences to each sample. The first-strand cDNA was synthesized using SuperScript II (ThermoFisher), and library quality was assessed on the Qubit system (Invitrogen). Raw data preparations were sequenced on a Nova6000 platform (Illumina). Fastp (0.20.1) was applied to remove low-quality reads and adapters of the raw sequencing reads. Using the miRdeep2 mapper.pl command, the filtered clean sequencing reads were mapped to human genome GRCh38 obtained from ENSEMBL (release 95). Then, the miRDeep2.pl command was executed to quantify known miRNAs, with miRNA reference files containing mature miRNA sequences and their corresponding precursor sequences downloaded from the miRBase database (https://​www.​mirbase.​org/​).

qRT-PCR analysis was used to verify miRNA expression. RNA was isolated from tissues and plasma using the miRcute miRNA Isolation Kit (Tian Biotech, China) and reverse transcribed using PrimeScript RT Master Mix for RT-PCR (Roche Diagnostics, USA) according to the manufacturer’s protocol. For miRNA expression analysis, miRNA quantification was determined by using Bulge-Loop™ miRNA qRT-PCR Primer Sets (one RT primer and a pair of qRT-PCR primers for each set) specific for miR-196b-5p and miR-584b-5p, designed by RiboBio (RiboBio, China). All reagents for stem-loop RT mature miRNA assays were obtained from RiboBio (RiboBio, China). Real-time PCR was performed using SYBR Select Rremix Ex Taq II (TaKaRa Bio, Japan) on the Bio-Rad CFX96 Real-Time Detection System (Bio-Rad, USA). Data were normalized to levels of U6 (tissue) or cel-miR-39-3p (plasma) as appropriate, respectively. All analysis was carried out using the 2^−ΔCt method.

The target genes of miRNAs were extracted from miRTarBase (https://​mirtarbase.​cuhk.​edu.​cn) [30], with at least one experimental verification as the standard. Gene Ontology (GO) term and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was performed using Metascape (https://​metascape.​org/​) [31]. All genes of the human genome were used as the enrichment background. Terms and pathways with p-value < 0.05, the number of genes enriched greater than 2, and enrichment factor > 1.5 were considered enriched terms and pathways. According to semantic similarity, similar enriched GO terms or KEGG pathways with kappa score > 0.3 were clustered to form a functional network. Each node in the functional network represents a GO term or KEGG pathway. The node size is proportional to the number of overlapping genes between the genes of interest and the genes belonging to the term or pathway, and the node color represents a cluster annotation. Similar term or pathway nodes are marked with the same color and entries of the representing nodes are used as cluster annotations.

All statistical analysis and generation of figures were performed with the statistical software R, version 4.0.5 (https://​www.​r-project.​org), together with R packages “ggplot2,” “ggrepel,” “ggpubr,” “pheatmap,” “corrplot,” “pROC,” “logistf,” “caret,” and “PRROC.” The Mann–Whitney U test was applied to compare outcomes between two independent groups. The area under the receiver operating characteristic curve (AUROC), precision–recall (PR) curve, sensitivity (TPR), specificity (TNR), positive predictive value (PPV), negative predictive value (NPV), and overall accuracy (ACC) were used to evaluate the diagnostic performance.

The clinical and demographic data were obtained during routine visits and were recorded using standard datasheets. Gestational time was calculated with an algorithm based on a participant’s last normal menstrual period confirmed by early ultrasound. Patients with PE were diagnosed according to the technical bulletin of the American College of Obstetricians and Gynecologists [32]. All PE patients were nulliparous or multigravida (range 1–2). Subjects were excluded if they had abnormal 1-h glucose tolerance tests or pre-existing medical conditions such as hypertension, cardiovascular disease, diabetes mellitus, renal disease, or other chronic systemic diseases.

Maternal plasma samples and placental tissues were collected as part of four cohorts. The clinical characteristics of each cohort according to women with and without PE are presented in Table 1. Maternal age and BMI were statistically similar between healthy controls and PE patients (all p > 0.05). As expected, in all cohorts, preoperative systolic blood pressure (PreSBP), preoperative diastolic blood pressure (PreDBP), and urine protein/urine CERA (mg/mmol) in patients with PE were significantly higher than those in healthy controls (all p < 0.05). There was an earlier median gestational age of delivery and lower mean birthweight in PE patients compared to healthy controls. Except for cohort 1, despite large variations among samples, albumin and blood urea nitrogen were statistically significant in the analysis with the other three cohorts included, respectively.

In the discovery stage, we first compared the expression profiles of 2888 miRNAs derived from miRNA sequencing between placental tissue samples of 5 PE patients and 5 healthy controls in WMU cohort 1 and identified five significantly differentially expressed miRNAs (DEmiRNAs) with the screening strategy of fold change > 2.0 or lower than 0.5 and FDR-adjusted p < 0.05 between the PE and the normal group (Fig. 2A). Unsupervised hierarchical clustering analysis showed that the expression patterns of five DEmiRNAs were able to discriminate PE samples from healthy controls (Fig. 2B). Among the five DEmiRNAs, two miRNAs (miR-19a-3p and miR-520f-5p) were found to be up-regulated and three miRNAs (miR-584-5p, miR-196b-5p, and miR-1299) were down-regulated in PE placental samples compared with normal placental samples (Fig. 2C). Functional enrichment analysis of GO and KEGG showed that the target genes of DEmiRNAs tended to be enriched in biological processes and pathways related to fetal growth and development and the pathogenesis of PE (Fig. 2D–G).

To further examine the association between expression levels of DEmiRNAs and PE-associated clinical characteristics (including age, BMI during pregnancy, PreSBP, PreDBP, and the content of platelet, ALB, GLOB, ALT, BUN, and BUN/CREA), we calculated Pearson correlation coefficients and found that four of the five DEmiRNAs were associated with at least one PE-associated clinical characteristics (Fig. 3A, B). Thus, we further verified the expression of these four DEmiRNAs in placental tissue samples of 20 PE patients and 20 healthy controls from WMU cohort 2 using qRT-PCR and validated miR-196b-5p and miR-584-5p as reliable and robust biomarkers (Fig. 3C).

To establish a miRNA signature for the detection of PE, Firth’s bias-reduced logistic regression method was used to construct the predictive model based on the expression levels of miR-196b-5p and miR-584-5p in WMU cohort 1. Finally, a miR2PE-score was calculated for each sample using the observed weights from the regression model and the expression value of miR-196b-5p and miR-584-5p with the equation:

When miR2PE-score was applied to WMU cohort 1, the miR2PE-score and predicted labels of 5 PE placental samples and 5 normal placental samples were calculated and visualized in Fig. 4A. In WMU cohort 1, with the risk cutoff point of 0.5, five of five PE samples (100% sensitivity) and four of five normal samples (80.0% specificity) were correctly classified with an overall accuracy of 90.0% (nine of ten). The AUROC was 0.920 (95% confidence interval (CI) 0.736–1.000; Fig. 4A) and the area under the precision–recall curve (AUC-PR) was 0.919 (Additional file 1: Fig. S1A). To preliminarily verify cross-platform compatibility of miR2PE-score in PE diagnosis, the miR2PE-score was subsequently tested based on qRT-PCR data from WMU cohort 2. Figure 4B shows the expression levels of the two identified miRNAs, miR2PE-scores, and predicted labels and true labels of 20 PE and 20 control samples in WMU cohort 2. As expected, the miR2PE-score still works well in the qRT-PCR data and successfully identified 33 of 40 samples with an overall accuracy of 82.5%, an AUROC of 0.848 (95% CI 0.720–0.975; Fig. 4B), and PR-AUC of 0.815 (Additional file 1: Fig. S1B). These results initially confirmed that the miR2PE-score is reliable and compatible.

To further evaluate the cross-platform compatibility and multinational robustness of the miR2PE-score, we tested the performance of the miR2PE-score in two external cohorts, including the Canada cohort with the miRNA sequencing platform and the USA cohort with the microarray platform. In the Canada cohort, with the same formula and the risk cutoff (0.5), 16 of 20 PE samples (80.0% sensitivity) and 15 of 21 normal samples (71.4% specificity) were correctly classified by the miR2PE-score with an overall accuracy of 75.6% (31 of 41; Fig. 5A). The two identified miRNA biomarkers, miR2PE-scores, and predicted labels and true labels of placental tissue samples in the Canada cohort are shown in Fig. 5B. As shown in Fig. 5C, the AUROC was 0.864 (95% CI 0.757–0.972; Fig. 5C) and the AUC-PR was 0.874 (Additional file 1: Fig. S2A). In the USA cohort, the miR2PE-score successfully identified 11 of 16 PE samples (68.8% sensitivity) and 12 of 16 normal samples (75.0% specificity) with an overall accuracy of 71.9% (23 of 32; Fig. 5D). The two identified miRNA biomarkers, miR2PE-scores, and predicted labels and true labels of placental tissue samples in the USA cohort are shown in Fig. 5E. Similarly, in the USA cohort, the miR2PE-score demonstrated an AUROC of 0.812 (95% CI 0.661–0.964; Fig. 5F) and an AUC-PR of 0.786 (Additional file 1: Fig. S2B). These results further confirmed that the predictive performance of miR2PE-score is reliable and universal across different technology platforms and countries.

To further examine whether the miR2PE-score can be detected in blood samples and used as a potential minimally invasive diagnostic biomarker, the expression levels of two identified miRNA biomarkers in peripheral blood leukocytes from WMU cohort 3 (including 10 PE and 8 normal samples) and WMU cohort 4 (including 19 PE and 18 normal samples) were analyzed with small RNA sequencing and qRT-PCR, respectively. As shown in Fig. 6A and B, two miRNA biomarkers (miR-196b-5p and miR-584-5p) revealed lower expression levels in peripheral blood leukocytes from PE patients than in healthy controls in both WMU cohort 3 and WMU cohort 4, which is consistent with observations from placental tissue-based validation. Using the miR2PE-score, 8 of 10 PE samples (80.0% sensitivity) and 5 of 8 normal samples (62.5% specificity) in WMU cohort 3 were correctly classified with an overall accuracy of 72.2% (13 of 18; Fig. 6C). The two identified miRNAs, miR2PE-scores, and predicted labels and true labels of placental tissue samples in WMU cohort 3 are shown in Fig. 6D. The AUROC was 0.787 (95% CI 0.565–1.000; Fig. 6D) and the AUC-PR was 0.823 (Additional file 1: Fig. S3A). In WMU cohort 4, 17 of 19 PE samples (89.5% sensitivity) and 15 of 18 normal samples (83.3% specificity) were correctly classified by the miR2PE-score with an overall accuracy of 86.5% (32 of 37; Fig. 6E). The two identified miRNAs, miR2PE-score, and predicted labels and true labels of placental tissue samples in WMU cohort 4 are shown in Fig. 6F. The AUROC was 0.933 (95% CI 0.844–1.000; Fig. 6F) and the AUC-PR was 0.953 (Additional file 1: Fig. S3B). Taken together, the above results suggest that miR2PE-score could serve as a minimally invasive diagnostic biomarker for the diagnosis of PE in the clinic.

The performance of the miR2PE-score was further compared with five recently proposed PE miRNA biomarkers (miR-210-3p, miR-19b-1-5p, miR-92a-1-5p, miR-486-5p, and miR-18a-5p) in different cohorts. As shown in Fig. 7A–C, the miR2PE-score and miR-210-3p performed very well in all three placental tissue-based cohorts compared with the other four previously published miRNA biomarkers. However, the miR2PE-score achieved the best predictive performance in blood-based WMU cohort 3 compared with the other five previously published miRNA biomarkers (Fig. 7D). These results demonstrated that the miR2PE-score holds great potential to become a more effective and robust non-invasive biomarker for the early diagnosis of PE.