Bioinformatics identification and validation of maternal blood biomarkers and immune cell infiltration in preeclampsia: An observational study

Abstract

Preeclampsia (PE) is a pregnancy complication characterized by placental dysfunction. However, the relationship between maternal blood markers and PE is unclear. It is helpful to improve the diagnosis and treatment of PE using new biomarkers related to PE in the blood. Three PE-related microarray datasets were obtained from the Gene Expression Synthesis database. The limma software package was used to identify differentially expressed genes (DEGs) between PE and control groups. Least absolute shrinkage and selection operator regression, support vector machine, random forest, and multivariate logistic regression analyses were used to determine key diagnostic biomarkers, which were verified using clinical samples. Subsequently, functional enrichment analysis was performed. In addition, the datasets were combined for immune cell infiltration analysis and to determine their relationships with core diagnostic biomarkers. The diagnostic performance of key genes was evaluated using the receiver operating characteristic (ROC) curve, C-index, and GiViTi calibration band. Genes with potential clinical applications were evaluated using decision curve analysis (DCA). Seventeen DEGs were identified, and 6 key genes (FN1, MYADM, CA6, PADI4, SLC4A10, and PPP4R1L) were obtained using 3 types of machine learning methods and logistic regression. High diagnostic performance was found for PE through evaluation of the ROC, C-index, GiViti calibration band, and DCA. The 2 types of immune cells (M0 macrophages and activated mast cells) were significantly different between patients with PE and controls. All of these genes except SLC4A10 showed significant differences in expression levels between the 2 groups using quantitative reverse transcription-polymerase chain reaction. This model used 6 maternal blood markers to predict the occurrence of PE. The findings may stimulate ideas for the treatment and prevention of PE.

1. Introduction

Preeclampsia (PE) is a pregnancy-specific complication characterized by increased blood pressure after 20 weeks of gestation, which causes severe damage to multiple systems and organs.^[1] The incidence of hypertension during pregnancy is increasing each year, becoming an important risk factor for the health of pregnant women and their fetuses.^[2] However, the pathogenesis of PE is not fully understood, and effective early diagnostic methods are lacking.

Currently, the clinical diagnosis of PE mainly depends on the assessment of blood pressure and 24-hour urine protein levels.^[3] However, this diagnostic method is not sufficiently accurate or sensitive, which can lead to missed data, misdiagnosis, and delayed treatment. To better prevent and treat PE, it is important to understand the molecular mechanisms underlying the disease. Studies have shown that the development of preeclampsia is closely related to many biological processes and pathological changes, including vasoconstriction, endothelial cell dysfunction, coagulation abnormalities, and immune dysfunction.^[4–6] Therefore, in-depth bioinformatics analysis and mining can help identify key genes and molecular pathways associated with the pathogenesis of PE.

In this study, we aimed to screen core pathogenesis genes in patients with PE by mining a high-throughput bioinformatics database and using high-throughput bioinformatics analysis tools for data analysis. To achieve this goal, we first collected and organized publicly available PE-related gene expression data. Genes with potential diagnostic and therapeutic value were screened using bioinformatics techniques. These findings provide reliable research directions for further experimental studies and help elucidate the pathophysiological processes of PE.

2. Methods

2.1. Data collection

Using “preeclampsia” as the search term, the mRNA expression data for blood cells from PE patients and healthy pregnant patients were retrieved from the Gene Expression Omnibus (GEO) database. The gene expression microarray datasets GSE48424, GSE149440, and GSE166846 were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/). GSE48424 contained data for 18 healthy control individuals and 18 patients with preeclampsia,^[7] GSE149440 contained data for 282 healthy control individuals and 66 patients with preeclampsia,^[8] and GSE166846 contained data for 3 healthy control individuals and 3 patients with preeclampsia.^[9] The “sva” ComBat function in the R package was used to eliminate the batch effect of the 3 datasets and to merge the 3 datasets.^[10] We then used principal component analysis (PCA) to assess the distributions of the datasets before and after the merge.

2.2. Differential expression analysis

The limma package was used to identify differentially expressed genes (DEGs) between PE patient samples and control samples in the combined dataset.^[11] DEGs were screened based on an adjusted corrected P < .05 and |log fold change (FC) | > 0.2, after which a heatmap and volcano map of the DEGs were generated.

2.3. Intersecting gene screening

Intersecting genes from the DEGs were identified using 3 machine-learning algorithms.^[12,13] Least absolute shrinkage and selection operator (LASSO) is an established linear prediction method that uses regression coefficients to make predictions. The LASSO regression model is a regression analysis algorithm executed using the glmnet package in R to identify DEGs that significantly differentiate between PE patients and normal individuals.^[14] Another machine-learning technique is support vector machine (SVM), which is widely used in classification or regression analysis and is executed using the e1071 software package in R language.^[15] Through the mapping of a kernel function, an SVM projects input data into a higher-dimensional feature space, providing a more accurate classification than would be possible without it. First, the recursive feature elimination (RFE) algorithm was used to select the appropriate genes to avoid overfitting, and the SVM-RFE algorithm was used to select and identify the set of feature genes with the highest resolving power. A random forest (RF) model was constructed using the random forest package. The effectiveness and popularity of RF are based on combining decision trees using majority voting to achieve high precision and rapid auto-learning across diverse datasets.^[16] A large number of classification trees was randomly generated, and the classification outcome was obtained by iteratively scoring the DEG classification results of each tree. The classification results for all single trees were determined comprehensively. The Caret package was used to rank the importance of DEGs in the model, and the top 10 DEGs were selected. Finally, 3 algorithms were used to determine the overlapping portions of the screened DEGs via the “Venn” package in R.

2.4. Screening and evaluation of key genes

We used multivariate logistic regression analysis to identify key genes associated with PE from the overlapping DEGs. The receiver operating characteristic (ROC) curve, C-index, and GiViTi calibration band were used to evaluate the discriminant genes between patients with PE and healthy controls.^[17,18] Considering the potential clinical application value of the established genetic diagnostic model, decision curve analysis (DCA) was used to assess whether the decision based on the model was beneficial for diagnosing PE.^[19]

2.5. Gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses

Overlapping DEGs were identified using machine learning and multivariate logistic regression analyses. These key genes underwent GO analysis and KEGG pathway enrichment analysis using the “clusterProfiler” R package.^[20,21] In addition, the screening criteria for important functions and pathways were adjusted so that the P < .05.

2.6. Immune cell infiltration analysis

To determine the relative proportions of infiltrating immune cells in PE, the CIBERSORT (https://cibersortx.stanford.edu/) method for analyzing immune cell types in different tissues was used to analyze the combined expression data and calculate immune cell infiltration.^[22] The ratio was set to run at 1000. Suitable samples were screened based on P < .05. The percentage of each immune cell type in the sample was calculated and displayed as a bar graph. Comparisons and visualizations were performed on 22 immune cell samples from the PE and control groups.

2.7. Correlation analysis between the identified genes and infiltrating immune cells

The Spearman rank correlation coefficient was calculated to evaluate the correlation between the 22 infiltrating immune cells and the relationship between immune cells and key genes. Visualization was performed using the “linkET” package, P < .05 was considered to indicate statistical significance.

2.8. Collection and validation of human tissue specimens

Between October 2023 and December 2023, 10 blood specimens were collected at Shijiazhuang Fourth Hospital, including 5 blood specimens from patients with PE and 5 from healthy pregnant women. Written informed consent was obtained from all the participants. This study was approved by the Research Ethics Committee of Shijiazhuang Fourth Hospital (No. 20190046) based on the ethical requirements of the Declaration of Helsinki.

Blood samples were quickly frozen in liquid nitrogen and stored to preserve the RNA. The expression of these characteristic genes was verified using quantitative reverse transcription-polymerase chain reaction (qRT-PCR). The mRNA expression levels of FN1, MYADM, CA6, PADI4, SLC4A10, PROS1, PPP4R1L, and BEND2 were measured using a qRT-PCR kit (Servicebio, Wuhan, China) according to the manufacturer instructions. Gene-specific primers were purchased from Shanghai Sangon Biotech Co., Ltd. (Shanghai, China; Table 1).

Table 1.

Sequence of primers used for quantitative reverse transcription-polymerase chain reaction.

Gene	Primer sequence
GAPDH	F: 5′-ACAACTTTGGTATCGTGGAAGG-3′ R: 5′-GCCATCACGCCACAGTTTC-3′
MYADM	F: 5′-CCCTGTCTTGGCGCAACTT-3′ R: 5′-GGAACTGGACATAGGTGGTGG-3′
FN1	F: 5′-AGGAAGCCGAGGTTTTAACTG-3′ R: 5′-AGGACGCTCATAAGTGTCACC-3′
CA6	F: 5′-TTTGTGCTGGCAGATTTTGTCA-3′ R: 5′-CTGCGGTAATCGTTGTGGATG-3′
PADI4	F: 5′-CAGGGGACATTGATCCGTGTG-3′ R: 5′-GGGAGGCGTTGATGCTGAA-3′
SLC4A10	F: 5′-GTACGCCATAGGGTCCATGAG-3′ R: 5′-GGAACGAACAATGGGAATCCT-3′
PPP4R1L	F: 5′-CTGTGCAGAATGCTTCACGG-3′ R: 5′-GGGTCGCTGACAAGTCTGAT-3′

2.9. Statistical analysis

R software (version 4.3.1; https://www.r-project.org/) was used for all statistical analyses. Continuous variables are represented by the means, standard deviations, and differences between the 2 groups. The student t test was used for normally distributed variables, while the Mann-Whitney test was used for non-normally distributed variables.

3. Results

3.1. Merging of datasets and identification of DEGs

Three PE datasets from the GEO database were included in this study, namely, GSE48424, GSE149440, and GSE166846. Owing to differences in sample sources, detection platforms, and other factors, there was a batch effect. To eliminate the batch effect, the sva package was used for batch calibration of the samples. After merging the 3 datasets, PCA (Fig. 1A) revealed that cross-platform normalization successfully eliminated batch effects (Fig. 1B).

Figure 1.

Combination of datasets and screening of differentially expressed genes. (A) Before merging the principal component analysis datasets. (B) After the principal component analysis datasets were merged. (C) Differentially expressed genes between the preeclampsia (PE) and control groups. (D) Heatmap of differentially expressed genes between PE and control groups.

Seventeen DEGs were identified, of which 4 were significantly upregulated and 13 were significantly downregulated in the PE group compared with the healthy control group (Fig. 1C and D).

3.2. Identification of intersecting genes

Fifteen characteristic genes were identified using the LASSO regression algorithm (Fig. 2A and B), 8 characteristic DEGs were identified using the SVM-RFE algorithm (Fig. 2E), and the RF method identified the first ten genes, as shown in Figure 2C and D. The following 8 genes were found to overlap between the 3 calculation methods: FN1, MYADM, CA6, PADI4, SLC4A10, PROS1, PPP4R1L, and BEND2 (Fig. 2F).

Figure 2.

Identification of intersecting genes through the use of 3 machine-learning methods in merged datasets. (A, B) Least absolute shrinkage and selection operator regression analysis results. (C, D) Random forest analysis results. (E) Support vector machine-recursive feature elimination analysis results. (F) The intersecting genes are machine-learning-based.

3.3. Construction and evaluation of a nomogram model

The abovementioned 8 genes were further screened using multivariate logistic regression. Overall, 6 genes (FN1, MYADM, CA6, PADI4, SLC4A10, and PPP4R1L) were identified as key for the diagnosis of PE (Fig. 3A). An additional nomogram model was constructed to predict PE based on these 6 genes (Fig. 3B). The area under the curve and C-index (> 0.85) indicated that these 6 genes had good diagnostic value (Fig. 3C). The purpose of the GiViTi calibration band is to reveal the relationship between the predicted and observed probabilities by fitting a polynomial logistic function. The degree of fit of the predictive model was better when the diagonal bisector was used, and the 95% confidence interval was <45°. A P value > .05 for the GiViTi calibration curve indicated that the prediction model fit well (Fig. 3D). The DCA curve revealed that the red line was always above the gray line, and the black line ranged from 0 to 1, indicating that the decision based on the 4-gene model may be beneficial for PE (Fig. 3E). The clinical impact curves showed that the predictive ability of the model has potential clinical value (Fig. 3F).

Figure 3.

Construction and evaluation of a nomogram model for preeclampsia patients. (A) Potential diagnostic markers for preeclampsia (PE) were identified using multivariate logistic regression. (B) Construction of nomogram model for PE prediction. (C) Receiver operating characteristic curve, (D) GiViTi calibration curve, € decision curve analysis, and (F) clinical impact plot for the PE prediction model.

3.4. Functional enrichment analysis

Sixty hub genes were analyzed for functional enrichment. The enriched biological process items included substrate adhesion-dependent cell spreading (GO:0034446), pyramidal neuron development (GO:0021860), and negative regulation of heterotypic cell-cell adhesion (GO:0034115). The enriched cellular component items included apical dendrite (GO:0097440), apical plasma membrane (GO:0016324), and apical part of cell (GO:0045177). The enriched molecular function items included DNA-binding transcription factor binding, ubiquitin-like protein ligase binding, ubiquitin protein ligase binding solute: inorganic anion antiporter activity (GO:0005452), carbonate dehydratase activity (GO:0004089), and bicarbonate transmembrane transporter activity (GO:0015106) (Fig. 4A). The main KEGG pathways were bacterial invasion of epithelial cells, ECM-receptor interaction, and AGE-RAGE signaling pathway in diabetic complications (Fig. 4B).

Figure 4.

Functional enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis of key genes. (A) Functional enrichment analysis. (B) KEGG signaling pathway analysis.

3.5. Analysis of immune cell infiltration and its correlation with key genes

CIBERSORT was used to screen 390 blood samples from PE patients and healthy controls (P < .05). Different immune cells were counted in the blood samples of the control and PE groups (Figure S1, Supplemental Digital Content, http://links.lww.com/MD/M598). The results showed that the numbers of M0 macrophages and activated mast cells were significantly higher in the control group than in the PE group (P < .05) (Fig. 5A). The correlations between immune cells, including M1 macrophages, gamma delta T cells, activated mast cells, and resting natural killer cells, were >0.5. The expression levels of FN1, MYADM, and PADI4 were positively correlated with the number of M0 macrophages (r > .3, P < .05), while FN1 and MYADM levels were positively correlated with the number of activated mast cells (r > .2, P < .05) (Fig. 5B).

Figure 5.

Analysis of immune cell infiltration and its correlation with key gene expression levels. (A) Analysis of immune cell infiltration between the preecelampsia (PE) And control groups. (B) Correlation analysis between immune cell infiltration and key gene expression levels.

3.6. Validation of the expression levels key genes

Differences in the expression levels of the 6 genes in the 3 datasets are shown in Figure 6. In the GSE149440 dataset (Fig. 6A), which had the largest sample size, all 6 genes showed significantly different expression levels between the 2 groups, while in the GSE48424 dataset (Fig. 6B), only 3 genes showed significantly different expression levels between the 2 groups. In the GSE166846 dataset (Fig. 6C), which had a smaller sample size, there were no significant differences in gene expression levels between the 2 groups. After verification by qRT-PCR (Fig. 6D), all genes except SLC4A10, showed significant differences in expression levels between the 2 groups (P < .05).

Figure 6.

Expression levels of key genes in the preeclampsia (PE) and control groups. (A) GSE149440, (B) GSE48424, (C) GSE166846. (D) Validation of clinical blood samples.

4. Discussion

PE seriously affects the health of both mothers and their babies. The underlying mechanism is currently unclear, and there is a lack of early diagnostic methods.^[23] There is no effective treatment for PE, and the only strategy is to relieve symptoms after pregnancy.^[24] Therefore, in-depth exploration of the dysregulated molecules and pathways involved in PE is highly important for early diagnosis, prognosis, and treatment. The analysis of a single dataset is limited by its single-center nature and small sample size. Even in studies of the same disease, the results often cannot be completely matched in other similar datasets.^[7–9] This study used bioinformatics analysis methods to expand the sample size and increase the credibility of the study using public GEO database datasets. Three datasets were merged and 3 machine-learning and logistic regression analysis methods were used to identify 6 commonly explainable pathogenic genes. The expression levels of some of these 6 genes have been reported to vary in previous studies.

Fibronectin 1 (FN1) is a glycoprotein that is involved in cell adhesion and cell spreading and is widely distributed in healthy membranes, vascular structures, and smooth muscle cell layers.^[25–27] The plasma levels of FN1, which may be the best marker for vascular endothelial injury during PE, are increased in patients with PE.^[28] Multiple studies have shown that FN1 is related to the pathogenesis of PE, and FN1 can be used as a potential marker in the diagnosis of PE.^[29] This is consistent with the results of this study. Previous studies have reported that mice lacking peptidyllin deiminase 4 (Padi4) and with PAD4 deficiency have reduced levels of inflammation and susceptibility to abortion.^[30] The reason for these effects was that soluble fms-like tyrosine kinase 1 is an antiangiogenic protein that is pathogenically linked to abnormal placenta formation disorders in early pregnancy, causing miscarriage and the accumulation of neutrophils and neutrophil extracellular traps in the placenta of wild-type mice.^[30]

It has not been examined whether the other 5 genes (MYADM, CA6, PADI4, SLC4A10, and PPP4R1L) are related to PE, although 3 genes (MYADM, PADI4, and SLC4A10) have been linked to changes in blood pressure. As a progenitor cell marker, myeloid-associated differentiation marker (MYADM) is a transmembrane protein that is up-regulated during hematopoietic differentiation.^[31] Genetic studies have shown that MYADM is closely associated with hypertension.^[32,33] Furthermore, Massarenti et al found that PADI4 contains 5 single nucleotide polymorphisms that are associated with hypertension.^[34] Guo et al reported the relationship between common variants of the sodium-coupled bicarbonate transporter (NCBT) gene and the blood pressure response was affected by dietary sodium, and a mixed-effects model was used to evaluate the additive association between 5 common variants of the NCBT gene and the salt sensitivity phenotype, including SLC4A10.^[35] Carbonic anhydrase (CA) plays a complex role in vascular function and vascular tension regulation.^[36] Regulating of CA6 and vascular function, therefore, will be an important future research area. However, it appears that PPP4R1L is related to the regulation of prostate cancer.^[37] Hence, further research is needed to prove that these 5 genes play roles in PE.

The KEGG signaling pathways associated with PE were identified based on these 6 genes. Advanced glycation end products (AGEs) and other ligands interact with their receptor for advanced glycation endproducts (RAGE) and are located in various tissues, but mainly in endothelial cells and blood vessel wall cells.^[38,39] RAGEs are known to be expressed in placental bed tissues during pregnancy, but RAGEs and their ligands are also expressed in fetal membranes and found in amniotic fluid and maternal serum.^[40] Pregnant patients with diseases with important vascular involvement, such as preeclampsia and diabetes, have additional increases in AGE/RAGE expression levels.^[41] In addition, previous studies have reported that resveratrol treats preeclampsia by regulating the AGE-RAGE signaling pathway.^[42] To assess bacterial invasion through the epithelial cell signaling pathway, we investigated the possible correlation between Helicobacter pylori (Hp) infection and the severity of clinical manifestations of PE and determined the immune mechanism triggered by Hp infection that may lead to the pathogenesis of PE.^[43]

This study has certain limitations. First, the data for this study were mainly obtained from public datasets that lacked clinical information related to the samples. Second, the data used in this study from the 3 different datasets were collected from blood samples and there were large differences between the blood samples. Third, the number of patients in the validation cohort was relatively small. Therefore, the reliability of the results requires further validation. Finally, although this study inferred the functions of the 6 signature genes and the role of immune cell infiltration in PE through bioinformatics analysis, further prospective studies with larger sample sizes are needed for verification.

5. Conclusions

In summary, we screened genes related to the pathogenesis of PE and analyzed the infiltration of related immune cells using bioinformatics and machine learning methods. Six signature genes of great significance in the early diagnosis of PE, which can be used for the subsequent development of PE, were identified. The prevention and treatment strategies have provided new research targets.

Acknowledgments

Data on preeclampsia was obtained from the Gene Expression Omnibus (GEO) database, which the authors acknowledge.

Author contributions

Conceptualization: Hong Li, Yuanyuan Rong.

Data curation: Haijiao Wang, Hong Li.

Formal analysis: Hong Li.

Funding acquisition: Haijiao Wang.

Methodology: Hongmei He, Yi Wang, Yujiao Cui, Lin Qi, Chunhui Xiao, Hong Xu, Wenlong Han.

Project administration: Haijiao Wang.

Supervision: Hong Li.

Visualization: Haijiao Wang, Hong Li.

Writing – original draft: Haijiao Wang.

Writing – review & editing: Hong Li, Yuanyuan Rong.