Identification of m5C-related genes and subclusters in recurrent pregnancy loss

Abstract

Purpose

Recurrent pregnancy loss (RPL), which occurs in 1–5% of couples and nearly half of the cases remain unexplained, is a complex condition influenced by multiple factors. Previous investigations have demonstrated the role of m5C-related genes (MRGs) in cancer prognosis and the significance of epigenetic modifications during pregnancy. However, the connection between MRGs and the pathogenesis of RPL remains elusive. This study endeavors to elucidate this relationship through bioinformatics approaches.

Methods

Data of 48 endometrial tissue samples were obtained through GEO query. Twenty-one MRGs were analyzed. Multiple machine learning (ML) methods were applied to identify biomarkers. A nomogram was constructed, and further analyses like GSEA and scRNA-seq were carried out using the R software.

Results

Five core biomarkers (DNMT1, SMUG1, ZBTB38, MBD4, and TDG) were pinpointed by ML methods, with the prediction model achieving an AUC of 0.953. Based on hub genes, 24 RPL samples were grouped into cluster A (n = 9) and cluster B (n = 15). The study revealed differences in immune cells and microenvironments, and the scRNA-seq analysis confirmed the connection between immune cells and m5C.

Conclusion

This study identified five key m5C-related genes, unraveled their link to immune cells, and developed an accurate RPL diagnostic model. The RPL patients are innovatively divided into two clusters, and the difference of their immune microenvironment is analyzed. This study offers a fresh perspective for examining biomarkers and potential therapeutic targets for RPL.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10815-025-03580-9.

Introduction

Pregnancy loss refers to the natural termination of pregnancy between conception and 28 weeks, with a prevalence of about 10–15% [1]. Recurrent pregnancy loss (RPL) can be considered for diagnosis after two or more pregnancy losses, affecting about 1–5% of couples [2, 3]. It is worth noting that almost half of RPL cases remain categorized as unexplained RPL (URPL) [4]. RPL can result from chromosomal abnormalities, uterine anatomical defects, autoimmune diseases, and endometrial dysfunction. Endometrial dysfunction, abnormal responses, and chronic endometritis could increase the risk of miscarriage [5]. Living with RPL, couples often fear additional losses and may never have a live birth. Offering evidence-based information about the likelihood of a successful pregnancy is crucial. Determining the predictive factors for RPL can assist patients in deciding whether to continue attempting pregnancy or adjusting risk factors to improve the chances of a live birth.

Previous studies have indicated that genes linked to 5-methylcytosine (m5C) can forecast the prognosis of diseases like prostate cancer, cervical cancer, and liver cancer and guide immunotherapy [6–8]. Recent research has confirmed that epigenetic modifications, mainly DNA methylation and non-coding RNA, notably influence decidualization through the modulation of target gene expression [9]. The initiation, development, and maintenance of stromal cell decidualization are essential for implanting embryo, forming placenta, and maintaining pregnancy [10–12]. Abnormal DNA methylation can result in imprinting disorders, gene expression disorders, sperm abnormalities, and immune imbalances, which might impact embryo implantation, growth, and development, ultimately contributing to RPL [13]. This research intended to clarify the role of m5C-related genes (MRGs) in RPL, reveal the link of m5C with RPL pathogenesis, and identify potential hub genes utilizing bioinformatics analysis to propose potential therapeutic targets.

This study leveraged the Gene Expression Omnibus (GEO) database to explore the potential mechanisms of RPL by examining the differentially expressed genes (DEGs) between normal and RPL samples. Differentially expressed MRGs are identified through the intersection of DEGs and MRGs. Various machine learning (ML) algorithms are then leveraged to determine hub genes and develop a nomogram prediction model. The model performance is verified by means of receiver operating characteristic (ROC) curves and a column line graph. Ultimately, according to the expression profiles of five MRGs, 24 RPL samples are grouped into two clusters. The variations in immune cells are further appraised between the two clusters. This study provides a novel view of the possible molecular mechanisms involved in RPL pathogenesis.

Materials and methods

Data preprocess

This study employed data from 48 samples, including 24 normal samples and 24 RPL samples, all obtained from endometrial tissues. These data were stored in the GEO database under the serial number GSE165004 [14]. The GSE22490 [15] (six normal samples, four RPL samples), GSE26787 (five control samples, five RPL samples) [16], and GSE111974 (24 normal samples, 24 RPL samples) [17] datasets were downloaded from the GEO database (http://www.ncbi.nlm.nih.gov/geo/) for external validation. All data were gathered through the R package GEOquery [18]. Gene probes were labeled with gene symbols, and those without a corresponding gene symbol or with multiple matches were excluded. For duplicate gene symbols, the gene expression value was determined as the maximum value. The downloaded gene expression matrix files of GSE22490 and GSE26787 were merged. Moreover, batch effects were eliminated by utilizing the “sva” packages for analysis. The data has been approved by the Institutional Review Board (IRB) of the Ethics Committee of Istanbul University School of Medicine (Istanbul, Turkey) (IRB approval number 2013/170).

This analysis included 21 MRGs that were previously identified in earlier research [19]. These genes consisted of three writers (DNMT3A, DNMT3B, and DNMT1), four erasers (TET1, TET2, TET3, and TDG), and fourteen readers (MBD1, MBD2, MBD3, MBD4, MECP2, NEIL1, NTHL1, SMUG1, UHRF1, UHRF2, UNG, ZBTB4, ZBTB38, and ZBTB33).

Alteration analysis of MRGs between normal and RPL samples

Pearson correlation analysis was executed to examine the expression link among the 21 MRGs in RPL samples only. The expression differences of MRGs between normal and RPL samples were appraised by the Wilcoxon rank sum test. Utilizing the transcriptional data of the GSE165004 dataset, the CIBERSORT [20] deconvolution algorithm (https://cibersort.stanford.edu/) was leveraged to qualify the abundance of specific immune cell types. ROC curve analysis was executed by means of the R package pROC (version 1.16.2) [21], to identify the specificity, sensitivity, likelihood ratios, positive predictive values, and negative predictive values for all possible thresholds. The diagnostic performance of MRGs was appraised utilizing the results from the ROC curve analysis.

Screening and validation of m5C-related diagnostic markers

New and critical biomarkers for RPL were determined by leveraging three ML algorithms: random forests (RF), least absolute shrinkage and selection operator (LASSO) logistic regression, and support vector machine-recursive feature elimination (SVM-RFE). The R package randomForest [22] was leveraged to implement the RF algorithm. The LASSO logistic regression was executed by means of the R package glmnet [23], and the optimal lambda was chosen based on its minimal value. The optimization parameters in this study were cross-validated with a tenfold method, and the partial likelihood deviation satisfied the minimum requirement. Genes identified as common across the above three classification models were then chosen for additional investigation. A nomogram model was created to forecast the recurrence of RPL by means of the rms [24] package. The “total score” was the cumulative score of all the genes above. The model was appraised by ROC curves. Additionally, the model performance was measured by the area under the curve (AUC) of ROC. The predictive power of the model was subsequently appraised by means of calibration curves. Ultimately, the clinical utility of the model was examined through decision curve analysis and clinical impact curves.

Identification of m5C modification pattern

To recognize distinct m5C-related clusters, an unsupervised cluster analysis of 21 MRG expressions was executed utilizing the ConsensusClusterPlus package [25]. The K-means algorithm was applied with Euclidean distance as the metric. Furthermore, 80% of the items were resampled with 1000 replications. The proportion of ambiguous clustering was leveraged to determine the optimal value of k. The expression patterns of the 21 MRGs in various modification patterns were further validated through principal component analysis (PCA) and heatmap analysis. The Wilcox test was leveraged to compare the infiltrating immunocyte abundance score and immune checkpoint gene expression across various modification patterns.

Biological enrichment analysis for various m5C modification patterns

Gene set enrichment analysis (GSEA) was performed to determine the key pathways and core genes between distinct m5C modification patterns. We conducted enrichment analyses to identify if a series of pre-defined biological processes were enriched. Enriched pathways were sorted by their normalized enrichment scores, and pathways with P < 0.05 were selected for further analysis. GSVA is a non-parametric and unsupervised method. It estimates gene set enrichment variation in gene expression data, frequently employed to explore variations in pathway and biological process activity across samples. To examine the differences in biological process terms, GSVA was carried out across distinct m5C modification patterns.

Single-cell RNA sequencing (scRNA-seq) analysis

ScRNA-seq analyses were carried out for the RPL single-cell dataset GSE214607 [26] utilizing the R package Seurat (version 4.0.2) [27]. The following criteria were applied for controlling qualities: (i) cells with a total RNA count below 200 or above 5000 were excluded; (ii) cells with more than 20% of mitochondria unique molecular identifier (UMI) rate were excluded; (iii) mitochondrial genes were filtered out of the expression table. The NormalizeData function in R was employed to normalize the unnormalized dataset. Principal component analysis was carried out on the scaled data, focusing on the 2000 genes with the highest variability. The top 15 principal components were leveraged in the uniform manifold approximation and projection (UMAP) algorithm. After cluster classification, the R package SingleR was leveraged to recognize and annotate different cell clusters [28]. The featureplot function in the R package seurat was leveraged to display the expression of a particular gene. The data has been approved by the Institutional Review Board(IRB) of the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University.

Statistical analysis

The R programming (https://www.r-project.org/, version 4.3.1) was leveraged to calculate and statistically analyze all data. Differences between normal and RPL samples were appraised by means of Wilcoxon tests (mean ± SD). Moreover, P < 0.05 demonstrated statistical significance (ns: no significance, *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001). For correlation analysis, the Pearson correlation coefficient was computed.

Results

Landscape of MRGs between normal and RPL samples

There were 21 MRGs involved in the study, including DNMT3A, DNMT3B, DNMT1, TET1, TET2, TET3, TDG, MBD1, MBD2, MBD3, MBD4, MECP2, NEIL1, NTHL1, SMUG1, UHRF1, UHRF2, UNG, ZBTB4, ZBTB38, and ZBTB33. By analyzing the expression of 21 MRGs in the GSE165004 dataset, the expression levels of most genes in the RPL group were significantly different from those in the normal group (Fig. 1A). The expression levels of NEIL1, NTHL1, SMUG1, TDG, TET1, and UHRF2 were notably elevated in the RPL group, while DNMT1, MBD1, MBD4, ZBTB33, and ZBTB38 were noticeably increased in the normal group. The PCA result revealed that the m5C-related expression patterns were diverse between the normal and RPL groups (Fig. 1B). According to the Pearson correlation analysis, the expression of UHRF1 was strongly positively tied to those of MBD3, DNMT1, and UNG. Moreover, the expression of NTHL1 was strongly positively linked to those of NEIL1 and TDG (Fig. 1C). The connections of MRGs with immune infiltration cells were examined to reveal their role in the RPL immune microenvironment. The difference analysis demonstrated notable differences in the abundance of some infiltrating immunocytes between RPL and normal samples. Based on the correlation analysis, MRGs exhibited a strong association with the expression of tumor-infiltrating immune cells (TIICs). For instance, the expression of ZBTB4 may be positively linked to the levels of NK.cells.activated and Mast.cells.resting, while negatively linked to the levels of Dendritic.cells.activated. Moreover, the levels of T.cells.CD4.memory.resting may be positively tied to the expression of NEIL1 (Fig. 1D). These findings demonstrated that the expression of MRGs varied notably between RPL patients and healthy individuals, suggesting that these abnormal expression patterns may play a role in RPL progression. According to Gene Ontology (GO) analysis, the 21 MRGs were significantly enriched in methyl-CpG binding, DNA modification, regulation of DNA methylation, DNA (cytosine-5-)-methyltransferase activity, and heterochromatin (Fig. 1E).

Fig. 1.

Interaction between abnormal m5C epigenetic regulation and immune microenvironment in RPL. A MRG expression in RPL and control groups; B PCA plot; C Pearson correlation analysis; D correlation analysis between m5C gene and immune cells; E: Gene Ontology (GO) analysis

Screening and validation of diagnostic markers

Using the three ML algorithms, RPL biomarkers with diagnostic significance were recognized. There were eight genes from the RF model (Fig. 2A), 16 genes from the SVM-RFE model (Fig. 2B), and 12 genes from the LASSO regression analysis (Fig. 2C, D). The Venn diagram was employed to illustrate the intersection of these genes, revealing five robust core biomarkers (DNMT1, SMUG1, ZBTB38, MBD4, and TDG) (Fig. 2E). Subsequently, the diagnostic performance of the five genes was estimated. The ROC curve analysis revealed that the AUC of most core biomarkers exceeded 0.7 (Fig. 2F). A nomogram for forecasting the risk of RPL was constructed by means of the rms package (Fig. 3A). A minimal discrepancy between the actual and predicted risks was observed in the calibration curve, suggesting that the model had favorable accuracy (Fig. 3B). Moreover, the decision curve analysis demonstrated that the model delivered superior clinical benefit (Fig. 3C). Furthermore, a high AUC value (0.953) was noted in the column line graph in the GSE165004 cohort (Fig. 3D). Moreover, RPL samples exhibited notably higher risk scores compared to normal samples (Fig. 3E). These findings supported the result of the column line graph.

Fig. 2.

Screening of diagnostic markers and validation of hub genes for RPL based on multi-algorithm machine learning. A RF model; B SVM-RFE model; C–D LASSO regression analysis; E Venn diagram; F ROC curve analysis

Fig. 3.

Construction of a nomogram for the diagnosis of RPL and validation of its clinical utility. A The nomogram for forecasting the risk of RPL. B The calibration curve. C The decision curve analysis. D The column line graph. E Risk scores

In addition, the diagnostic performance of the nomogram model in the external validation cohorts of RPL was consistent with that of the training group (Figure S1). Moreover, it exhibited higher diagnostic performance in the validation cohorts of the combined dataset (AUC = 0.879) and GSE111974 (AUC = 0.953) (Figure S1a, c). Then, RPL samples exhibited notably higher risk scores compared to normal samples in the combined dataset and GSE111974 (Figure S1b, d). In summary, various results confirmed the reliability of the nomogram model as a tool for diagnosing RPL.

Consensus clustering analysis of MRGs

Consensus clustering analysis was executed by means of the R package ConsensusClusterPlus. Based on the expressions of 21 MRGs, it was determined that k = 2 yielded the most stable clustering (Fig. 4A, B). Therefore, the 24 RPL samples were split into two distinct categories for consensus clustering analysis: cluster A (n = 9) and cluster B (n = 15). The PCA plot revealed distinct variations in gene expression patterns between the two clusters (Fig. 4C). These expression differences were comprehensively evaluated to explore the molecular characteristics between the two clusters. Distinct expression profiles of the m5C-related gene were observed across cluster A and cluster B (Fig. 4D). To explore the differences in immune microenvironment characteristics across these distinct clusters, infiltrating immunocytes and immune checkpoint gene sets were appraised. Noticeable differences in immune cell types were noticed across the two clusters (Fig. 4E). Cluster A exhibited greater levels of infiltrated T.cells.gamma.delta and Dendritic.cells.resting, while cluster B showed enrichment in T.cells.CD8 and NK.cells.activated. As for immune checkpoint, the expression levels of CD276, LGALS9, and NRP1 were evidently elevated, while CD44, CD200, IDO1, IDO2, KIR3DL1, LAG3, PDCD1, TMIGD2, and TNFSF9 were markedly downregulated in the cluster A, compared with the cluster B (Fig. 4F). These findings suggested distinct differences in the immune microenvironment among the m5C-related clusters of RPL.

Fig. 4.

Identification of m5C subtypes and heterogeneity of immune microenvironment in RPL. A–B Consensus clustering analysis. C PCA plot. D Distinct expression profiles of the MRGs across cluster A and cluster B. E Immune cell types across the two clusters. F Immune checkpoint across the two clusters

Biological characteristics of the two m5C modification patterns

Differential expression analysis was executed to delve into the functional differences across the two clusters. In total, 335 DEGs were recognized. Out of these genes, 174 exhibited upregulation and 161 showed downregulation. Figure 5A displays the distribution of these DEGs. We then carried out GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses on these DEGs to further understand the potential molecular processes and functions. According to the two enrichment analyses, these genes primarily exhibited enrichments in the mitotic cell cycle process, modulation of mitotic nuclear division, extracellular matrix, cell cycle, regulation of hormone levels, response to vitamin, receptor ligand activity, extracellular matrix structural constituent, epithelial cell differentiation, and extracellular matrix structural constituent (Fig. 5B, C). GSEA analysis showed that cluster A was mainly associated with base excision repair, DNA replication, Fanconi anemia (FA) pathway, and homologous recombination, whereas cluster B was enriched in complement and coagulation cascades, ovarian steroidogenesis, TNF signaling pathway, and necroptosis and Toll-like receptor signaling pathway (Fig. 5D, E). Through the GSVA analysis, several enriched GO biology processes with differential expressions between the two subtypes were identified (Fig. 5F). In cluster A, higher expression was noted in DNA replication, cell cycle DNA replication, mismatch repair, and positive regulation of cell cycle checkpoint, while cluster B demonstrated greater activity in response to interleukin (IL)−15, IL-1-mediated signaling pathway, and cellular responses to increased oxygen levels.

Fig. 5.

Functional heterogeneity and pathway regulation of the m5C subtype in RPL. A Volcano map. B–C GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. D–E GSEA analysis. F GSVA analysis

ScRNA-seq analysis

The expression of RPL-related diagnostic biomarkers in the RPL microenvironment was examined utilizing the dataset GSE214607. After integrating data from three RPL samples in the GSE214607 dataset, 25,980 cells were totally identified. Cells with fewer than 200 total RNAs or more than 5000 total RNAs were excluded, as well as those with a mitochondria UMI rate greater than 20%. Linear dimensionality reduction was executed after choosing the 2000 genes with the highest variability to determine the available dimensions of the dataset. To generate the UMAP visualization, the first 15 principal components were utilized. The RPL dataset contained 29 cell populations and five major cell types in the RPL dataset (Fig. 6A, B), including B cells, monocytes, NK cells, T cells CD4, and T cells CD8. Most of these were monocyte and NK cells. DNMT3A and TET2 were primarily expressed in NK cells, MBD2 in T cells CD8, DNMT1 and MBD4 in B cells, and ZBTB38 in T cells CD4 (Fig. 6C). Diagnostic biomarkers expressed in various cells are shown in Fig. 6D.

Fig. 6.

Single-cell immune microenvironment expression profile of RPL diagnostic markers. A–C The UMAP visualization. D–F Diagnostic biomarkers expressed in various cells

Discussion

Pregnancy loss refers to the natural termination of pregnancy between conception and 28 weeks, with a prevalence of about 10–15%. RPL diagnosis is typically considered after two or more pregnancy losses, impacting around 1–5% of couples [2]. It should be noted that almost half of RPL cases are classified as URPL. The causes of RPL include chromosomal abnormalities, anatomical uterine defects, autoimmune disorders, and endometrial dysfunction. Endometrial dysfunction, abnormal response, and chronic endometritis all elevate the risk of pregnancy loss [5]. However, other than plasma cells, few studies have shown altered immune cell infiltration in RPL. No experimental evidence has proven that altered endometrial immune cell infiltration is the cause of RPL in chronic endometritis. Immune tolerance plays a crucial role in pregnancy in pregnancy, driving research over decades into immune-related causes and treatment targets for RPL. According to several epigenetic studies, RPL is tied to abnormalities in DNA methylation, methylation modification, and other relevant epigenetic changes. The occurrence of RPL may be triggered by imprinting barriers, gene expression disorders, sperm defects, and immune imbalance caused by abnormal DNA methylation, which might impact embryo implantation, growth, and development [13]. This study attempts to explore the role of MRGs in RPL, reveal the link between m5C and the pathogenesis of RPL, identify potential key genes, and investigate potential therapeutic targets through bioinformatics analysis.

This study leverages the GEO database to examine gene expression levels in normal and RPL samples. The expression of MRGs varied notably between RPL patients and healthy individuals. RPL patients exhibit more abnormal expression of MRGs than the normal population, pointing to the noticeable role of m5C in RPL development. The GO analysis reveals that the 21 MRGs are markedly enriched in methylated CpG binding, DNA modification, DNA methylation regulation, DNA (cytosine-5)-methyltransferase activity, and heterochromatin.

Meanwhile, notable variations in the proportion of immune cells are noted across normal individuals and RPL patients. A prior study has demonstrated that the immune system is strongly linked to RPL [29]. Our study shows that the abundance of certain infiltrating immune cells in RPL samples is notably different from that in normal samples. Moreover, MRGs are strongly linked to the expression of tumor-infiltrating immune cells, like NK.cells.activated and Mast.cells.resting, whose levels may be positively tied to ZBTB4 expression. However, the levels of Dendritic.cells.activated may be negatively tied to ZBTB4 expression. The levels of T.cells.CD4.memory.resting may be positively tied to NEIL1 expression.

In addition, utilizing the three ML classifiers, five central genes (DNMT1, SMUG1, ZBTB38, MBD4, and TDG) are identified. The diagnostic capability is verified using the ROC curve and column line graph. Previous studies have identified DNMT1 as the first DNA methyltransferase with DNA methyltransferase activity. It is believed to play a role in methylating DNA to protect epigenetic memory during cell proliferation [30]. DNMT1 loss impairs embryonic development [31]. MBD4 collaborates with DNMT1 to mediate methyl-DNA inhibition and protect mammalian cells from oxidative stress [32]. Our scRNA-seq analysis shows that DNMT1 and MBD4 are mainly expressed in B cells. DNMT1 gene knockout can inhibit the proliferation of endometrial cancer cells and induce apoptosis of endometrial cancer cells by regulating the expressions of Bcl-2, Bax, NF-κB-inhibitor alpha, and Cyclin (CCN)D1/2, showing potential application value in the treatment of endometrial cancer [33]. MBD4 is essential for maintaining DNA methylation and repairing G/T mismatches. Genetic defects in MBD4 increase the risk of adenomatous polyposis and colorectal cancer, which are marked by the buildup of C > T switches caused by spontaneous deamination of 5′-methylcytosine [34]. ZBTB38 is directly involved in DNA damage repair or the regulation of DNA synthesis [35]. Heterozygosity loss of ZBTB38 leads to early embryo death by inhibiting Nanog and Sox2 expression [36]. SMUG1 plays a key role in base excision repair, RNA processing, and RNA components, including telomerase, although its biochemical and cellular mechanisms remain largely unknown [37]. In addition to uracil, SMUG1 primarily acts on 5-hmdU, a lesion that is known to be increased in tumor cells [38]. Loss of SMUG1 activity promotes cancer development. A study has shown that SMUG1 can be employed as a predictive biomarker for cervix intraepithelial neoplasia grade III and cervical squamous cell carcinoma [39]. TDG is the sole DNA glycosylase whose germ line deletion leads to embryo death [40, 41]. It is involved in the base excision repair phase of active DNA demethylation. It is widely accepted that TDG exhibits strong glycosylase activity on DNA substrates that contain 5-formylcytosine or 5-carboxycytosine [42]. In both Th2 differentiation and macrophage activation, TDG contributes a small role to active replication-independent DNA demethylation [43]. The correlation analysis indicates that there are noticeable synergistic or antagonistic interactions between these pivotal genes. Using the five key genes to establish a diagnostic model may be helpful to guide the clinical diagnosis of RPL.

There are two different categories in the consensus clustering analysis, namely, cluster A (n = 9) and cluster B (n = 15). Ultimately, 335 DEGs were identified. GO and KEGG enrichment analyses indicate that these DEGs show enrichments in the cell division process, mitosis nuclear division regulation, extracellular matrix, cell cycle, hormone level regulation, vitamin response, receptor ligand activity, extracellular matrix structural components, epithelial cell differentiation, and extracellular matrix structural components. This is consistent with the conclusions of prior studies. Lim et al. employ RRBS (bisulfite sequencing) and RNA-seq to analyze DNA methylation and gene expression in human placental tissue, showing that genes with significant changes are enriched in pathways associated with cell cycle and immune response [44].

We also find notable variations in the immune cells across the two distinct categories. Cluster A exhibits greater levels of infiltrated T.cells.gamma.delta and Dendritic.cells.resting, while cluster B shows enrichment in T.cells.CD8 and NK.cells.activated. In terms of immune checkpoints, cluster A shows notably higher expression levels of CD276, LGALS9, and NRP1 compared to cluster B, while the expression levels of CD44, CD200, IDO1, IDO2, KIR3DL1, LAG3, PDCD1, TMIGD2, and TNFSF9 are significantly lower. These results indicate that the immune microenvironment associated with distinct m5C epigenetic patterns differs notably. Previous studies have shown that CD276, also known as B7-H3, is a member of the B7 superfamily of immune costimulatory molecules. It is considered a T cell coinhibitor. However, the effect of CD276 on T cells is contradictory. It increases the proliferation of CD4 + and CD8 + T cells and enhances the activity of cytotoxic T cells. Lymphocyte activation gene 3 (LAG-3) is an immune checkpoint of the immunoglobulin superfamily. It enhances the immunosuppressive activity of Tregs, downregulates the activity of CD4 + T cells, and has a synergistic effect with CTLA-4. LAG-3 is expressed in immune cells and inhibits T cell proliferation, activation, and homeostasis. In addition, it can downregulate the activity of CD4 + T cells by binding to major histocompatibility complex II [45]. An article we reviewed shows that the gene expression of CD276 and the protein tissue level in the endometrium of the RPL group are noticeably lower than those in the control group, while both the expressions of CTL-4 and LAG-3 are notably elevated. Our study shows that the expression of CD276 in cluster A is higher than that in cluster B. Moreover, the expression of LAG-3 in cluster A is lower than that in cluster B. This may mean that cluster A is less likely to be affected by this pathway to develop RPL than cluster B, thereby confirming our conclusion and making our classification more persuasive. In addition, the uterine NK (uNK) cell population consists of CD56^brightCD16⁻NK cells and CD56^dimCD16⁺NK cells. CD56^dimCD16⁺NK cells are more cytotoxic and account for most peripheral blood NK cells. Hence, they are called peripheral blood NK cells, which are clearly present in the maternal blood that perfuses the placenta. Some women with URPL, especially those who miscarry chromosomally normal embryos, seem to have elevated peripheral NK cells. Nevertheless, the specific mechanism is unknown. Animal studies have shown two different uNK subsets in mice: CD44^bright and CD44^mid. In a mouse model of URPL, it is observed that there is an increase in the number of CD44^bright uNK cells, while the CD44^mid uNK subset remains unchanged. A new miscarriage mechanism tied to the increase in the CD44^bright subset and intravenous immunoglobulin has been proposed. This mechanism inhibits the increase in the CD44^bright subset. However, it remains unclear whether the expression of CD44 in humans is consistent with that observed in mice [46].

Recently, immune checkpoint inhibitors (ICIs) have evolved into first- and second-line treatments for various cancer types, notably improving patient outcomes, such as therapies that target T cell checkpoints (PD-1, LAG 3, and CTLA-4). However, adaptive resistance may reduce their efficacy, leading to mutations in genes that regulate MHC-1 neoantigen presentation. Therefore, for various reasons, ICI monotherapy still does not provide benefits to most patients. This emphasizes the requirement for new immunotherapy strategies, where targeting NK cells might help address some of these challenges. Moreover, identifying additional immune checkpoints to target could further amplify the anti-tumor immune response. As new immunotherapy approaches are developed, it is crucial to understand the impact of missing intracellular checkpoints in the tumor-infiltrating immune cell profile [47].

The GSEA analysis shows that cluster A is mainly associated with base excision repair, DNA replication, FA pathway, and homologous recombination, while cluster B is enriched in the complement and coagulation cascade, ovarian steroidogenesis, TNF signaling pathway, necroptosis, and Toll-like receptor signaling pathway. As a rare autosomal recessive genetic disorder, FA is marked by progressive bone marrow failure and is a major risk factor for cancers like head and neck squamous cell carcinoma and leukemia, as well as declined fertility and other congenital abnormalities [48]. Both women and men with FA experience notable fertility reduction, with women commonly presenting with primary ovarian insufficiency. In addition, the incidence of gynecological cancer in female patients is greatly increased [49]. To date, the TNF signaling pathway has proven to be a valuable target for treating inflammation-related diseases, and the safety and tolerability of anti-TNF-α therapy have been well established [50].

Based on the GSVA analysis, cluster A shows higher expression levels in positive regulation of DNA replication, cell cycle DNA replication, mismatch repair, and cell cycle checkpoint, while cluster B shows higher activity in response to IL-15, IL-1-mediated signaling pathways, and cellular responses to increased oxygen levels. According to a prior study, decidual macrophages are essential in regulating decidual NK (dNK) cells by secreting IL-15, which induces resting endometrial NK cells to differentiate into activated dNK cells [51]. IL-15 expression in tumors can be enhanced by activating the STING pathway. IL-15, along with type I IFN that modulates IL-15, is essential for maintaining regular levels of CD8 T cells and NK cells in tumors. The upregulation of IL-15 strengthens CD8 T cell response and enhances anti-tumor response. Substantial evidence indicates that IL-15 and its agonists are promising cancer therapeutics when administered systemically, due to their can target the strengthened response of cytolytic T cells and NK cells [52]. In a mouse model of collagen-induced arthritis (CIA), inhibition of IL-15 signaling with CRB-15 (an antagonistic IL-15 mutant/FC-2A fusion protein that blocks IL-15R) has been indicated to inhibit inflammatory responses to an extent [53]. In CIA, IL-15 inhibition decreases the synovial concentrations of some inflammatory cytokines like TNF-α, IL-1β, IL-6, and IL-17, which are already being targeted therapeutically in clinical settings. These treatments included anti-TNF-α (etanercept, adalimumab, cetuzumab, golimumab, and infliximab), IL-1R antagonists (anakinra), anti-IL-1β (kananumab), anti-IL-6 (tocilizumab and salizumab), and anti-IL-17 (secunumab and ixekinumab). Thus, targeting IL-15 signaling may offer greater effectiveness compared to solely targeting inflammatory cytokines [54].

The scRNA-seq analysis reveals that DNMT3A and TET2 are primally expressed in NK cells, MBD2 in CD8 T cells, DNMT1 and MBD4 in B cells, and ZBTB38 in CD4 T cells. Previous research has shown that the interaction between dNK cells and macrophages stimulates dNK cells to produce IFNγ, which in turn induces macrophages to produce IDO [55]. Macrophages interact with T cells via co-stimulatory molecules. This interaction is essential for establishing maternal–fetal immune tolerance. Decidual macrophages express B7-H1, B7-DC, and B7-1/−2. The B7-H1/PD-1 interaction likely acts as a critical factor in controlling local IFNγ production by activated T cells, thereby supporting the formation of a proper maternal immune response to the fetus in the early stages of human pregnancy [56]. The connection between B7-1/−2 and CTLA-4 transmits inhibitory signals and upregulates the expression of IDO in decidual macrophages by inducing IFNγ production [57]. dNK cells are crucial for regulating trophoblast invasion and modulating the immune response against trophoblasts [58]. The main pro-pregnancy function of dNK cells is compromised in URPL, leading to the impaired performance of TNK cells [59]. NK cells and macrophages are important immune cells that affect female pregnancy, and the migration inhibitory factor (MIF) signaling pathway is an important pathway for the interaction between NK cells and macrophages [60]. The MIF signaling pathway is an essential cellular signaling pathway that participates in various physiological and pathological processes. Produced by immune cells, MIF can regulate immune and inflammatory responses, cell proliferation, and apoptosis [61]. Moreover, it activates mitogen-activated protein kinase, phosphatidylinositol 3-kinase/Akt, and NF-κB, thereby affecting cell function [62].

A previous study has demonstrated that most embryos are arrested at the two-cell stage after the knockout of DNMT1 and DNMT3a. This suggests that DNMT1 in B cells is crucial for embryonic development, and DNMT1 gene silencing may cause failures in embryonic development and induce RPL in women during pregnancy [63]. Another study has pointed out that DNMT1 affects the pathway of progesterone synthesis. When DNMT1 is overexpressed, the expression of CYP11A1 and progesterone are reduced. Subsequently, RPL is induced in pregnant women [64]. ZBTB38 is essential for early embryonic development by inhibiting the expression of Nanog and Sox2. The abnormality of ZBTB38 in CD4 T cells might cause RPL in patients through this pathway [36]. The pathway required for replication and stability of the human genome consists of three components: E3 ubiquitin ligases, transcription inhibitors, and replication proteins. Moreover, the destruction of the stability of transcription inhibitor ZBTB38 negatively regulates the transcription and level of MCM10 replication factors on chromatin, resulting in the instability of human genome replication. This may be one of the mechanisms causing RPL [65]. The increased expression of MBD4 can enhance the proliferation, migration, and invasion of trophoblast cells. Relevant research has found that MALAT1 promotes the expression of MBD4 by recruiting CREBBP to enhance the proliferation, migration, and invasion of trophoblast cells, thereby inhibiting APL (antiphospholipid syndrome) positive RPL [66].

Certain limitations need to be highlighted in this study. Firstly, although the current study is based on a comprehensive bioinformatics analysis, it does not include additional experimental or clinical trial validation. Due to practical difficulties in sample collection, the experimental validation of the study results is temporarily unable to perform. We are working to collect samples from RPL patients in order to perform these validations as soon as possible. Secondly, the relatively small sample size in this study should be noted. To validate the reliability of the results, studies with larger sample sizes are necessary. Additionally, this study only involves an RPL group and a control group, neglecting the potential impact of factors like age, and race on the results. Furthermore, more comprehensive clinical data is required to confirm the model performance for predicting RPL risk. Moreover, further external validation cohorts are essential to confirm its stability. Finally, additional RPL samples are necessary to elucidate the accuracy of the results for the clustering related to m5C.

Conclusion

In conclusion, this study reveals the link between MRGs and invasive immune cells, identifies five characteristic genes associated with m5C, accurately evaluates RPL subtypes, and establishes a model to diagnose patients with RPL. In addition, we elucidate immune heterogeneity in patients with different types of RPL. This study first discovers the involvement of the m5C-related gene in RPL development, offering novel insights into the underlying pathogenic processes and therapeutic strategies of RPL.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

All authors contributed to the study conception and design. Writing—original draft preparation: Huanmin Luo; writing—review and editing: Huanmin Luo; conceptualization: Shuqing Li; methodology: Yuming Cao; formal analysis and investigation: Jinfeng Xu; resources: Li Wang; supervision: Li Wang, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Data availability

The datasets analyzed during the current study are available in the GEO database (https://www.ncbi.nlm.nih.gov/gds), under the serial number GSE165004, GSE214607. Zhejiang provincial Basic Public Welfare Research Plan Project (No. LGD21H040001), Zhejiang Medical and Health Science Project (No.2022KY442).

Declarations

Ethics approval and consent to participate

The data has been approved by the Institutional Review Board (IRB) of the Ethics Committee of Istanbul University School of Medicine (Istanbul, Turkey) (IRB approval number 2013/170) and the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.