Patient cohort and dataset characterization

To characterize the transcriptomic landscape of euploid RPL, we collected first-trimester decidual biopsy specimens from women undergoing elective termination of uncomplicated pregnancies (Norm; n = 3; mean gestational age = 8.7 ± 1.5 weeks) and patients with RPL (n = 4; mean gestational age = 10.6 ± 2.5 weeks) (S1 Table). Given that fetal chromosomal abnormalities can significantly alter maternal immune responses and confound the computational analysis of immune-mediated mechanisms [12], we restricted our analysis strictly to cytogenetically normal cases. After excluding one RPL case identified with a chromosome 9 duplication (see Methods), the final cohort comprised matched maternal and fetal samples from euploid pregnancies. Matched peripheral blood samples were utilized to obtain maternal genotypes, which served as the ground truth for the subsequent computational demultiplexing of maternal versus fetal cells.

Single-cell transcriptomic profiling and genotype-based demultiplexing reveal altered cellular landscapes in the RPL decidua

To comprehensively characterize the cellular heterogeneity of the maternal-fetal interface during early pregnancy, we performed single-cell RNA sequencing on decidual tissues. Following quality control, one normal case was excluded due to low sequencing quality, resulting in a final dataset of 62,588 single cells from two normal (Norm; mean gestational age = 9.5 ± 0.7 weeks) and three euploid RPL samples (mean gestational age = 10.3 ± 2.9 weeks) (see Methods). To determine the origins of cells as either maternal or fetal cells, we employed a dual-verification strategy, previously described in our study [13]. We first clustered cells based on allele variations using Souporcell [14] and validated these clusters against ground-truth genotypes from matched maternal PBMCs, utilizing an average of 927.3 ± 435.3 informative alleles per sample. This approach retained 57,976 cells with confidently inferred maternal or fetal origins for downstream analysis. Specifically, the final dataset comprised 35,027 (60.4%) cells from RPL samples and 22,949 (39.6%) from Norm samples; 41,517 (71.6%) were classified as maternal, and 16,459 (28.4%) were classified as fetal.

Next, we performed unsupervised clustering and cell-type annotation to resolve 32 distinct decidual cell clusters spanning immune and non-immune compartments. To independently verify the reference-based annotations, the expression of canonical marker genes for each cell type is shown in S1 Fig. These data were visualized using Uniform Manifold Approximation and Projection (UMAP) [15], colored by cell type, clinical condition (Norm vs. RPL), and inferred origin (maternal vs. fetal) (Fig 1A–1C). The inferred cell origins were successfully aligned with established biological knowledge: trophoblast lineages (EVT, SCT, VCT) and Hofbauer cells (HB) were overwhelmingly classified as fetal in origin; for instance, > 99.8% of EVTs were assigned as fetal (Fig 1E). This consistency reinforces the validity of our classification approach.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Single-cell landscape of the human decidua from elective terminations of normal pregnancies and RPL.

(A) Uniform manifold approximation and projection (UMAP) embedding of cells from our study; each dot represents a single cell and is colored by cell type. Endo, endothelial cells; f, fetal; L, lymphatic; m, maternal; p, proliferative; Epi, epithelial glandular cells; SCT, syncytiotrophoblast; EVT, extravillous trophoblasts; VCT, villous cytotrophoblasts; DC, dendritic cells; dM, decidual macrophages; dS, decidual stromal cells; HB, Hofbauer cells; ILC, innate lymphoid cells; MO, monocytes; dNK, decidual natural killer cells; dP, decidual progenitor cells. (B) UMAP colored by clinical condition (Norm vs. RPL). (C) UMAP colored by inferred origin (maternal vs. fetal). (D) Stacked bar plots showing the proportions of maternal and fetal cells among immune cell types between Norm and RPL samples. (E) Stacked bar plots showing the proportions of maternal and fetal cells among fetal-derived cell types. (F) Stacked bar plots showing the proportions of cells from Norm and RPL samples among fetal-derived cell types. The red horizontal line indicates the expected baseline distribution based on the dataset composition (Normal: 0.4, RPL: 0.6).

https://doi.org/10.1371/journal.pcbi.1014806.g001

Based on the validated cell origin classification, we next compared the composition of major cell types between the RPL and Norm samples. Most immune cell types, such as decidual natural killer (dNK) cells and T cells, were primarily maternal in normal cohorts (Fig 1D). Specifically, approximately 81% of NK cells across six subtypes and 80% of T cells were classified as maternal (S2 Table). However, we observed a notable proportion of fetal-origin immune cells in RPL cohorts. Conversely, trophoblast lineages and Hofbauer cells—both of fetal origin—exhibited lower proportions of RPL-derived cells than the baseline RPL proportion in the dataset (0.6) (Fig 1F). This relative underrepresentation may reflect differences in the cellular composition of RPL decidual tissue, though reduced fetal viability prior to tissue collection may also contribute.

Distinct transcriptional landscapes and immune dysregulation in euploid versus aneuploid RPL

To assess the specificity of the immune impairment identified in our euploid cohort, we investigated whether the disease-associated transcriptional patterns were intrinsic to euploid RPL or shared with cases involving chromosomal abnormalities (aneuploid RPL). We employed the single-cell aneuploidy inference pipeline described in the Methods, which integrates dosage-sensitive gene expression (via scploid) and allelic imbalance at heterozygous SNP loci, to precisely identify fetal-derived decidual cells harboring chromosomal aberrations at the single-cell level within the samples included in this study. These identified cells were designated as the aneuploid subset, compared against euploid fetal cells, and were retained in all subsequent analyses.

We performed differentially expressed gene (DEG) analyses between RPL and Norm conditions separately within each subset to compare their regulatory landscapes (S2 Fig, S3 and S4 Tables). The analysis first revealed a set of immune-related genes that was consistently upregulated in RPL irrespective of ploidy. This group, which includes IL15, IGFBP1, IGFBP2, IGFBP7, and PRDM1, suggests a baseline level of stress-associated immune activation common to pregnancy loss [16–20]. However, the majority of signatures exhibited divergent regulation depending on chromosomal status. Key immune-activation genes associated with T/NK cell signaling (IL2RB, LCP2, ADGRE5, PTPRJ, PDE7A, RASA2) [21–27] were significantly upregulated specifically in euploid RPL but downregulated in aneuploid RPL. Conversely, genes related to mitochondrial metabolism and proteostasis (MRPL/MRPS subunits, COX5B, NDUFA11, VDAC1, PRDX3, PSMB5, CTSL) [28–33] exhibited the opposite trend, with strong upregulation restricted to aneuploid RPL. The biological distinction between immune-driven signatures in euploid cases and metabolic stress profiles in aneuploid samples highlights the distinct molecular pathologies of these conditions. These findings validate our strategy to specifically characterize the immune landscape within the euploid cohort, focusing on mechanisms distinct from the metabolic stress responses associated with aneuploid samples.

Altered immune response signatures in maternal dNK cells

To obtain an initial readout of immune activation in the decidua, we began with the Common Rejection Module (CRM) [34], an 11-gene module defined by meta-analysis across multiple tissues to capture the signal of activated infiltrating immune cells rather than tissue-specific programs. Its constituent genes include interferon-γ–induced chemokines (CXCL9, CXCL10), IFN-inducible antigen-processing machinery (PSMB9, TAP1), T-cell surface and signaling molecules (LCK, CD6), and a cytotoxic granule gene (NKG7).

A comparison across all profiled decidual cells revealed that RPL samples exhibited significantly elevated CRM scores compared to Norm samples (Mann-Whitney U test, p = 1.27E-144; Fig 2A). To identify the specific cell populations manifesting this signature, we examined CRM scores at the cell-type level. Notably, decidual NK cell subsets dNK2 (p = 7.55E-13) and dNK3 (p = 1.23E-17) showed the most significant elevation in RPL, surpassing the mean CRM score across all immune cell types (immune baseline, 0.36; indicated as a dotted line in Fig 2B). Crucially, although we observed a notable proportion of fetal immune cells in RPL (Fig 1D), the dNK2 and dNK3 populations remain overwhelmingly maternal in origin. This indicates that the elevated activation signatures originate primarily from the maternal host. In healthy pregnancies, dNK2 and dNK3 are essential for recruiting extravillous trophoblasts and regulating invasion [35]. Furthermore, they actively modulate the local immune microenvironment through the secretion of cytokines and chemokines, such as XCL1, CCL3, and CCL4 [36]. However, their high CRM scores in RPL suggest a functional shift from these physiological roles.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Altered immune response signatures in dNK cells from RPL patients.

(A) Comparison of common rejection module (CRM) scores between Norm and RPL samples across all cells, showing elevated CRM scores in RPL. (B) Comparison of cell type-specific CRM scores between Norm and RPL samples. Among all immune cell types, dNK2 and dNK3 had the most significantly increased CRM scores in RPL, indicating heightened immune activation. The dotted lines indicate three reference baselines: the mean CRM score across all cells (global), across all immune cell types (immune), and across all non-immune cell types (non-immune). (C–D) Volcano plots showing genes that are differentially expressed between RPL and Norm samples in dNK2 (C) and dNK3 (D) cells. DEGs were defined as those whose adjusted p values were < 0.05 and absolute log2 fold changes were > 2. (E) Bar plot summarizing genes commonly upregulated and downregulated in both the dNK2 and dNK3 subtypes. Proinflammatory mediators (e.g., CCL3, CCL4, and MIF) were markedly upregulated, whereas immune tolerance–related genes (e.g., IGF2, CGA, and SPP1) were consistently downregulated, indicating immune dysregulation in RPL-associated dNK cells. Significance for the CRM score comparisons in (A) and (B) was assessed using the Mann-Whitney U test (*** p < 0.001), with Benjamini–Hochberg correction applied for the multiple cell-type comparisons in (B).

https://doi.org/10.1371/journal.pcbi.1014806.g002

Guided by these scores, we performed DEG analysis within dNK2 and dNK3 to further elucidate the molecular basis of this shift (Fig 2C, 2D; S5 Table). To capture biologically robust transcriptional shifts, we specifically focused on genes exhibiting a greater than fourfold difference in expression levels between conditions. Using this criterion, we identified a consistent dysregulation pattern in both dNK2 and dNK3 (Fig 2E). Pro-inflammatory chemokines such as CCL3, CCL4, and CCL5 were upregulated, implying the recruitment of cytotoxic effectors (T cells and monocytes) via CCR5 signaling and the establishment of an inflammatory microenvironment [37]. The innate immune regulator MIF, known to amplify inflammatory responses and override glucocorticoid-mediated suppression, was also strongly induced [38]. In stark contrast, key tolerance-associated genes were suppressed. IGF2, a crucial growth factor promoting trophoblast development and fetal growth, was significantly downregulated [39]. Similarly, CGA, which encodes the alpha subunit of human chorionic gonadotropin (hCG) and plays a pivotal role in expanding regulatory T cells while suppressing NK cell cytotoxicity, was markedly reduced [40]. We also observed the downregulation of SPP1 (Osteopontin), a key cytokine that mediates cell-matrix interactions and promotes an anti-inflammatory Th2 shift at the maternal-fetal interface [41]. This concurrent upregulation of inflammatory mediators and downregulation of tolerance-inducing factors in maternal dNK cells provides transcriptomic evidence of the immune dysregulation characterizing RPL. A comparable analysis across all other immune cell types is provided in S3 Fig.

However, the marked disparity in maternal-to-fetal cell proportions between Norm and RPL samples (Fig 1D) acts as a critical confounding factor for standard statistical comparisons. Specifically, transcriptomic differences derived from population averages may predominantly reflect these compositional shifts rather than intrinsic disease-associated alterations. This limitation underscores the necessity for an alternative analytical strategy capable of disentangling true disease signatures from compositional biases.

Hierarchical machine learning framework identifies RPL-associated immune cell populations

To overcome the limitations posed by cellular data imbalance and to capture multivariate disease signatures, we implemented a supervised machine learning framework. The primary purpose of this framework was not cell-type identification but rather the prioritization of genes distinguishing Norm from RPL states within each immune population. Unlike conventional statistical methods that test for mean differences in individual genes, this approach learns combinatorial transcriptomic patterns to classify cell states. Specifically, to effectively disentangle the robust transcriptomic features defining cell identity from the more subtle signals associated with the RPL condition, we employed a stepwise, hierarchical classification architecture implemented via devCellPy [42] (Fig 3A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. The machine learning approach revealed distinct immune cell signatures linked to RPL.

(A) Schematic of the hierarchical classification framework using devCellPy. The model employs an XGBoost-based approach for multi-labeled, hierarchical cell identification. At the 1st level, cells are classified into 32 distinct lineage labels. Subsequently, at the 2nd level, the model distinguishes immune cell states based on pregnancy termination status (e.g., Tcells_Normal vs. Tcells_RPL). (B) Summary of the datasets used in this study. (C) Weighted average precision, recall, and F1 score across all cell labels in the held-out test set. The total number of classification labels in each level is also presented. The detailed performance of each cell label at each level is presented in S4 Fig. (D-E) Weighted average precision, recall, and F1 score for both 1st- and 2nd-level classifications across all cell labels in the Du (D) and Vento (E) datasets, respectively. (F) Weighted average precision, recall, and F1 score for each cell type at the 2nd level evaluated on Vento’s dataset. The performance at the 1st level is presented in S5 Fig.

https://doi.org/10.1371/journal.pcbi.1014806.g003

The first level distinguishes 32 distinct cell identities based on lineage-specific signatures, while the second level classifies the identified immune cells into “Norm” or “RPL” states. For instance, T cells identified at the first level are further classified at the second level as T cells from Norm samples or T cells from RPL samples. A hypothetical third level distinguishing maternal versus fetal origins was excluded due to the biological scarcity of fetal-derived immune cells in normal controls (e.g., only 0.04% of Norm T cells were fetal), which precluded effective model training (S2 Table).

To evaluate the generalization capability of our model, we assessed its performance on a held-out test set and two independent external datasets selected to match our study's biological context (Fig 3B). Du’s dataset [43] comprised 37,446 cells from six patients with recurrent miscarriage (5–8 weeks gestation; mean 6.83 ± 0.75 weeks). Crucially, these patients had ≥ 2 unexplained miscarriages and confirmed euploid karyotypes (no chromosomal abnormalities), ensuring alignment with our study’s focus on euploid RPL. Vento’s dataset [7] served as a healthy control benchmark, comprising 85,350 cells from uncomplicated elective terminations (6–12 weeks). These external datasets served as rigorous benchmarks to test whether the classifier could maintain predictive accuracy across distinct biological contexts and technical batches.

Performance metrics validated the efficacy of our hierarchical framework in resolving cell identities across diverse datasets and capturing RPL-associated patterns. In the internal test set, the model achieved high weighted average F1 scores at both first level (0.87) and second level (0.97) (Fig 3C), with high predictive accuracy maintained across the majority of individual cell types at both classification levels (S4 Fig). This high performance was replicated in the external RPL cohort (Du’s dataset), where the model achieved F1 scores of 0.81 (first level) and 0.93 (second level) (Fig 3D), again demonstrating consistent classification across most cell lineages (S5 Fig). These results confirm the model's sensitivity in distinguishing RPL-associated transcriptomic states. In the external normal cohort (Vento’s dataset), while first-level classification remained robust (F1 = 0.87), second-level performance showed variability (F1 = 0.32) (Fig 3E), likely reflecting technical batch effects or the inherent challenge of applying models trained to capture disease signatures to exclusively healthy control datasets.

Despite these variations, T cells, granulocytes, and NK CD16 ⁺ cells were the only immune populations that retained high second-level classification performance in both external datasets (weighted average F1 scores of 0.86, 0.94, and 0.89 in Vento; Fig 3F). Although several other subsets achieved comparable or higher recall in Du's dataset, their performance was not maintained in Vento's dataset. These results indicate that these populations retain the most reliable discrimination between Norm and RPL states across differing clinical contexts, capturing RPL signatures in the RPL cohort (Du) while correctly classifying healthy cells in the normal cohort (Vento). Consequently, our hierarchical approach prioritized these cell populations as the key immune populations characterizing the RPL immune landscape, providing a filtered focus for subsequent gene-level investigations.

Identification and cross-architecture validation of RPL immune signatures using transformer-based AI

To identify the specific genetic drivers underpinning our classification model, we performed level-specific Shapley Additive Explanations (SHAP) analysis [44]. This method quantifies the contribution of individual genes to the model's decision-making at each layer. For each gene, a feature importance value was computed as the mean absolute SHAP value across cells. A higher value indicates a greater contribution to distinguishing the classes. We therefore ranked genes by these values to identify the strongest contributors to each classification. As a proof of concept, applying SHAP to the first level (cell identity) correctly identified canonical markers as top contributors—for instance, KRT18 for villous cytotrophoblasts (VCT) [45]. This confirmation of biological validity justified extending our analysis to the second level to isolate genes distinguishing Norm from RPL states within immune subsets. The complete results of the SHAP analysis for each cell type and level, including the SHAP values for individual genes, are presented in S6 and S7 Tables.

To validate these SHAP-derived signatures and rule out potential overfitting to the devCellPy hierarchical framework, we employed scGPT, a transformer-based foundation model pre-trained on large-scale single-cell transcriptomes [46]. Here, scGPT was used not to discover or define novel immune cell states, but as an independent, cross-architecture validation of the SHAP-derived signatures. This validation focused on the three cell types whose classification performance remained robust across both external datasets: T cells, granulocytes, and NK CD16 ⁺ cells. We hypothesized that if the top SHAP-ranked genes identified from these cell types represent true biological signals rather than model-specific artifacts, they should enable high classification accuracy even in a fundamentally different model architecture.

For each target cell type, we selected the top 100 SHAP-ranked genes from each of the two classification levels. These gene sets were introduced sequentially during fine-tuning, mirroring the two-level hierarchy: the first-level genes (cell identity) in the first fine-tuning step and the second-level genes (Norm vs. RPL) in the second (Fig 4A). We refer specifically to the second-level genes as the “RPL signature,” as these are the genes that distinguish the disease state. Using only these selected genes as input, the fine-tuned scGPT model performed a unified, single-step classification into six distinct states (Tcell_Normal, Tcell_RPL, Granulocyte_Normal, Granulocyte_RPL, NK_CD16 ⁺ _Normal, and NK_CD16 ⁺ _RPL).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Fine-tuning the scGPT with SHAP-derived gene signatures robustly distinguished the RPL-associated immune cell subtypes.

(A) Schematic of the fine-tuning strategy using the transformer-based scGPT_human model. The model was fine-tuned on the top SHAP-ranked genes from both classification levels for T cells, granulocytes, and NK CD16 ⁺ cells, introduced sequentially across two fine-tuning steps. Fine-tuning was performed on expression values restricted to this gene set. (B–E) Model performance was evaluated using the weighted average precision, recall, and F1 score. All performance results shown are for the final (second) fine-tuned model depicted in (A), which classifies cells into the six distinct states. (B) Comparison between the pre-trained scGPT model and the final fine-tuned model on the test validation set. (C) Performance of the final fine-tuned model on the test validation set for each immune cell subtype. (D) Performance of the final fine-tuned model on the independent Norm validation set. (E) Performance of the final fine-tuned model on the independent RPL validation set.

https://doi.org/10.1371/journal.pcbi.1014806.g004

The results demonstrated the robustness of these signatures across model architectures. While the pre-trained scGPT model (zero-shot) achieved a baseline weighted F1 score of 0.61, fine-tuning with our identified signature genes improved performance to 0.86 (Fig 4B). This improvement confirms that the selected gene set contains robust and discriminatory information for RPL classification. Cell-type specific performance was particularly high; on the test set, T cells and granulocytes achieved high F1 scores (>0.94), whereas NK CD16 ⁺ cells showed distinctly lower recall compared to other cell types (Fig 4C).

We further validated the model on external datasets representing distinct biological conditions. In the Vento dataset (healthy controls), the model accurately classified T cells and granulocytes into their respective “Normal” subsets without misclassification, achieving high F1 scores (>0.96) (Fig 4D). Consistent with the internal test results, however, NK CD16 ⁺ cells exhibited poor generalization performance in this external cohort. This trend persisted in the Du dataset (RPL patients); while T cells maintained a weighted average F1 score of 0.99, NK CD16 ⁺ cells again failed to maintain robust classification accuracy (Fig 4E). Furthermore, granulocyte validation in the Du dataset was hindered by extreme sample scarcity (n = 9).

The observed results illustrate distinct validation outcomes for each cell type. While granulocytes showed high accuracy in healthy controls, their validation in RPL samples was limited by cellular scarcity. Conversely, NK CD16 ⁺ cells failed to generalize across external datasets. Therefore, T cells uniquely exhibited sustained high predictive accuracy across all validation cohorts, overcoming both the sample limitations and generalization challenges observed in other lineages. This consistent performance of T-cell-derived RPL signatures across distinct classification frameworks confirms their predictive power and highlights their potential for clinical diagnostic applications in RPL compared to other immune subsets. Given this robustness, we focused the subsequent functional and network analyses on the T-cell RPL signature, defined as the top 100 SHAP-ranked genes from the second-level (Norm vs. RPL) classification of T cells.

Integrative pathway and network analyses of SHAP-derived RPL signatures reveal immune-associated mechanisms and therapeutic targets of RPL

We assessed the potential clinical relevance of this T-cell RPL signature. Overrepresentation analysis of the signature revealed enriched cellular functions, with pathways showing adjusted P  <  0.05, which were considered significant. The top 20 pathways were ranked by odds ratio, defined as the proportion of query genes mapped to each pathway. This analysis revealed marked enrichment of RPL signatures in five biological categories: translation machinery, mRNA quality control, viral translation, ROBO signaling, and nutrient stress response (Fig 5A, S8 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Pathway analyses and drug repurposing reveal clinical implications of T-cell signature genes in RPL.

(A) Bar plot showing the results of overrepresentation analysis (ORA) of 100 RPL signatures using the Reactome pathway database. The top 20 pathways with the highest odds ratios, among those with adjusted P values ≤ 0.05, are displayed. Pathways are grouped according to functional categories. (B) Dot plot showing drug repurposing predictions for RPL based on RPL signatures using ASGARD. Each dot represents a predicted drug, with the dot size inversely proportional to the P value. Dots corresponding to drugs with P values ≤ 0.05 are color-coded according to the tissue type in which the prediction was made. (C) Network visualization of interactions between RPL signatures based on STRING protein–protein interaction data. Node color intensity reflects average centrality. The top 10 signature genes with the highest average centrality are labeled and highlighted in red. (D) Bar plot showing the centrality scores of the top 10 signatures in the protein–protein interaction network. Centrality values were min–max normalized to range between 0 and 1. (E) Visualization of the protein–protein interaction network of the top 10 RPL signatures and target proteins of the 5 repurposed drugs.

https://doi.org/10.1371/journal.pcbi.1014806.g005

Among the five functional groups, translation machinery exhibited the greatest number of enriched pathways, which was consistent with the T-cell origin of the RPL signatures. Upon antigen stimulation, T cells exit quiescence and undergo extensive transcriptional and metabolic reprogramming, processes closely linked to translational control during activation and immune regulation [47–49]. The mRNA quality control group was predominantly associated with nonsense-mediated decay (NMD), a post-transcriptional mechanism increasingly recognized for its role in shaping immune responses and emerging as a therapeutic target in autoimmunity [50,51]. ROBO signaling, which is implicated in early trophoblast function, may contribute to RPL through placental dysregulation [52,53]. Although the connections of viral translation and nutrient stress response pathways to immune tolerance are less clearly defined, their relevance is supported by established links between pregnancy loss and immune responses [54–56]. Together, these findings underscore the strong immunological component of the RPL transcriptomic signatures.

We next asked whether reversing RPL-associated transcriptional changes could mitigate RPL pathology. To address this, we applied ASGARD [57], a drug-repurposing algorithm that matches disease-associated gene expression profiles with compounds from the L1000 drug-response dataset predicted to counteract those signatures (see Methods). ASGARD identified five candidate compounds predicted to significantly reverse RPL-associated gene expression in at least three distinct tissues (Fig 5B, S9 Table). All five drugs showed predicted efficacy in lung and prostate tissues, with variable activity across other tissues. Notably, three compounds (prednicarbate, sirolimus, and niclosamide) have previously been implicated in contexts relevant to RPL [58–65], providing orthogonal support for their therapeutic potential.

Corticosteroids such as prednicarbate are typically used to treat autoimmune conditions and are administered to women with repeated early pregnancy loss [58,59]. Recent meta-analyses report a significant increase in ongoing pregnancy rates following corticosteroid treatment [60]. Sirolimus (rapamycin), which inhibits T- and B-cell activation and proliferation, is approved for use in transplantation and oncology [61,62]. A recent phase II randomized trial revealed that sirolimus significantly improved both pregnancy and live birth rates in patients with repeated implantation failure [63]. Niclosamide has been proposed as a therapy for immune-mediated disorders, including autoimmune diseases and infections, through the modulation of immune pathways [64]. Notably, its inhibition of follicular helper T cells has been shown to be effective in preclinical autoimmune models [65]. Although neither thioridazine nor amlodipine has been directly associated with RPL or immune tolerance, calcium channel blockers such as amlodipine exhibit known immunosuppressive effects that may warrant further investigation [66,67]. These findings support the validity of our drug-repurposing predictions and further implicate RPL gene signatures in the pathophysiology of RPL.

To identify candidate therapeutic targets from the RPL gene signatures, we constructed a protein–protein interaction (PPI) network by mapping the T-cell RPL signature genes to their encoded proteins and assembling them using STRING (Fig 5C). In such networks, hub proteins with high connectivity are more likely to modulate overall network behavior [68,69]. To prioritize central nodes, we computed four network measures (degree centrality, betweenness centrality, eigenvector centrality, and random walk) and ranked the proteins based on their average centrality scores. The top ten hub proteins were selected from these 100 signature proteins (Fig 5D, S10 Table). These hub proteins exhibited dense connectivity with targets of repurposed drugs identified earlier (Fig 5E). In particular, mTOR is the molecular target of sirolimus, one of our top repurposing candidates. Its dense connectivity with the hub proteins links our network analysis to an independently predicted compound, and is further supported by mTOR's established role in T-cell–mediated immunity and autoimmune disease.

A focused literature review revealed that seven of the ten top-ranked hub proteins are associated with RPL or immune regulation, further supporting their therapeutic potential. CXCR4, for instance, plays a critical role in coordinating both innate and adaptive immune responses [70]. Additionally, a recent systematic review and meta-analysis revealed a significant association between maternal APOE genotype and RPL risk in Asian populations [71]. Single-nucleotide polymorphism analysis involving more than 200 patients further suggested that APOE variants may predispose individuals to RPL [72]. Although JUN, HSPA1B, PPIA, and VIM have not been directly linked to RPL, these genes are functionally implicated in T-cell activity and activation and are known to contribute to immune modulation [73–79]. Together, these findings highlight the functional relevance of the identified RPL signature genes in the pathophysiology of RPL and support their potential as biomarkers and therapeutic targets for future translational studies.

Origin-controlled expression filtering identifies robust T-cell–associated transcriptomic signatures in RPL

Building upon the insights from the functional and network-based characterization of SHAP-derived T-cell signatures, we next sought to further refine the top-ranked genes derived from the second-level classification (Norm vs. RPL). We evaluated whether their expression patterns remained consistent after accounting for potential confounding effects, such as differences in cell origin or cellular composition. To this end, we performed an origin-controlled expression filtering analysis aimed at identifying genes that exhibit disease-specific transcriptional changes independent of sample-level heterogeneity.

In particular, nearly all the Norm T cells were maternal in origin, whereas a substantial proportion of the RPL T cells were fetal in origin (S2 Table). This imbalance raised the possibility that some SHAP-ranked genes might reflect origin-related differences (Maternal vs. Fetal) rather than RPL-associated changes (Norm vs. RPL). To eliminate this confounding effect, we implemented a two-step validation approach (Fig 6A). First, we excluded genes whose expression levels significantly differed between maternal and fetal T cells. Next, from the remaining pool, we retained only those genes that showed significant differential expression between RPL and normal T cells strictly within the maternal subset.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Confounding-controlled filtering identifies robust T-cell signatures in RPL.

(A) Schematic of the two-step filtering strategy used to refine SHAP-ranked T-cell genes by removing potential confounding effects. Genes differentially expressed between maternal and fetal T cells (n = 42) were excluded to control for cell origin-associated bias (confounding check). Among the remaining genes, those that were not significantly differentially expressed between RPL and Norm maternal T cells (n = 38) were further excluded. This yielded a final panel of 20 genes showing RPL-specific expression patterns. (B) Venn diagram illustrating the overlap among the top 100 SHAP-ranked genes, maternal vs. fetal T-cell DEGs, and RPL vs. Norm maternal T-cell DEGs. The red-boxed sector (n = 20) represents genes included in the SHAP top set and RPL-related DEGs but not in maternal–fetal DEGs—comprising the final RPL-specific candidates. (C) Dot plot showing the expression and cellular prevalence of the 20 final candidate genes in RPL and Norm T cells. The dot size indicates the fraction of cells expressing the gene; the color indicates the average expression. All 20 genes were significantly differentially expressed between RPL and Norm maternal T cells. (D) Violin plots of CXCR4 and JUN expression in RPL vs. Norm maternal T cells. Both genes were retained after filtering and showed elevated expression in RPL T cells. Statistical significance in (C) and (D) was assessed using the Wilcoxon rank-sum test with Benjamini–Hochberg false discovery rate (FDR) correction (*** adjusted p < 0.001).

https://doi.org/10.1371/journal.pcbi.1014806.g006

This filtering process distilled the initial candidate list (the top 100 SHAP-ranked T-cell genes, selected from the second-level SHAP results in S7 Table) down to a robust panel of 20 genes (Fig 6B), with 13 genes showing consistently elevated expression in RPL across wide cellular distributions (Fig 6C). These genes represent the candidates for RPL-specific immune markers, supported by a dual validation strategy: machine learning–based prioritization and disease-contextual expression filtering.

Significantly, this independent filtering process converged with our previous network findings. Among the 20 retained genes, CXCR4 and JUN were directly identified as top-ranked hub proteins in the preceding protein–protein interaction (PPI) network analysis (Fig 5D). This convergence between network centrality (potential for systemic regulation) and confounding-controlled expression (robustness against artifacts) strongly supports their biological relevance, highlighting them as the most reliable candidates implicated in RPL. The expression levels of both genes were significantly upregulated in RPL T cells compared to Norm cells (Fig 6D).

CXCR4 encodes a chemokine receptor that regulates leukocyte trafficking and inflammatory responses, particularly through the CXCL12–CXCR4 axis [80]. In normal pregnancy, CXCR4 promotes immune tolerance and tissue remodeling, including decidual NK cell migration [81,82]. However, dysregulated CXCR4 expression is linked to excessive T-cell infiltration and sustained inflammation in autoimmunity [83,84]. Similarly, JUN, a key component of the AP-1 transcription factor complex, integrates T-cell receptor (TCR) signals to activate T-cell programs [85,86]. Upon stimulation, JUN translocates to the nucleus and initiates chromatin remodeling and cytokine transcription [87,88]. While its role in normal pregnancy involves trophoblast invasion [89], JUN overexpression in pathological contexts can sustain inflammatory responses and prevent T-cell exhaustion, promoting persistent immune activation [74,90].

Together, these findings suggest that CXCR4 and JUN may contribute to pathological immune activation at the maternal–fetal interface in RPL. To our knowledge, neither gene has been previously highlighted as a distinguishing feature of decidual T cells in the context of RPL, underscoring their potential novelty and relevance as key immune signatures in pregnancy loss.

Ethics statement

All participants provided written informed consent, and the study was approved by the Institutional Review Board (IRB) of Konyang University Hospital (IRB approval number: KYUH 2022-08-009-005).

Sample collection

Decidual tissue (transformed endometrium during pregnancy) samples were collected from seven pregnant individuals, including three with normal pregnancies and four with RPL. In both groups, decidual tissue and peripheral blood were collected via dilation and curettage (D&C). Tissue was collected under ultrasound guidance, with sampling directed at the decidua basalis. Owing to the nature of the D&C procedure, a minor contribution from other decidual compartments cannot be fully excluded; however, the collected tissue consisted predominantly of decidua basalis. In the normal group, samples were obtained during the elective termination of pregnancy, whereas in the RPL group, two resulted from spontaneous miscarriage, and two were due to elective termination. To investigate immune tolerance failure in the absence of fetal chromosomal anomalies, we performed cytogenetic karyotyping on all samples. One RPL case exhibiting a chromosome 9 duplication was identified and excluded. Consequently, the remaining cohort consisted of three normal controls and three RPL cases, all of which were confirmed to have normal euploid karyotypes. One normal sample was subsequently excluded at the sequencing quality-control stage, so that the final analyzed cohort comprised two normal controls and three RPL cases. All the samples were collected under sterile conditions and immediately processed for further analysis, including single-cell RNA sequencing (scRNA-seq).

Analysis of droplet-based single-cell RNA sequencing data

Single-cell suspensions were prepared through standard tissue dissociation and cell processing procedures. Single-cell RNA libraries were constructed using the 10X Genomics Chromium Platform in accordance with the manufacturer’s instructions. Libraries were sequenced using both the Illumina HiSeq-x and NovaSeq 6000 platforms. The raw sequencing reads were aligned to the reference genome (GRCh38) using Cell Ranger (v5.0.0) to generate gene‒cell count matrices. Ambient RNA noise was minimized using SoupX [92], and potential doublets were removed with Scrublet [93]. Preprocessing and quality control were conducted in Scanpy (v1.10.1), filtering out cells with mitochondrial gene expression exceeding 20%, fewer than 200 detected genes, fewer than 1,000 UMI counts, or more than 7,000 total genes per cell. These quality control steps yielded a final dataset of 62,588 high-quality cells for downstream analysis. The merged data underwent normalization, scaling, and the identification of highly variable genes, followed by dimensionality reduction using principal component analysis (PCA) and cluster visualization with UMAP [15]. Batch effects were corrected using Harmony [94] after the clustering results were compared with those of BBKNN [95]. Cell types were identified using a logistic regression modeling approach implemented in sceleto2 (https://pypi.org/project/sceleto2/), in conjunction with CellTypist [96], based on reference labels for cells in the human decidua during the first trimester [7].

Identification of fetal and maternal cells

Given the nature of the samples, which were obtained from the decidua of pregnant women following pregnancy termination, the tissues contain a mixture of maternal and fetal cells. Therefore, to accurately determine whether the acquired cells originated from the mother or the fetus, we employed two independent analytical methods for cross-validation. This approach enabled us to reliably identify the origin of each cell. We first utilized Souporcell [14] to cluster the cells based on their genotypes, dividing them into two genetically distinct clusters. To verify the origin of each cluster, we subsequently compared the genotype profiles generated by Souporcell with the genetic variant data obtained from maternal peripheral blood mononuclear cells (PBMCs) using the Illumina GSA v3 genotyping array. The single-nucleotide polymorphisms (SNPs) observed in the two clusters were compared with the maternal PBMC genomic data to assess their similarity. The cluster with higher similarity was determined to consist of maternal cells and was defined as the “maternal cell cluster,” whereas the cluster with lower similarity was considered to consist of fetal cells. To achieve this goal, an average of 927.3 ± 435.3 alleles per sample were examined, allowing us to genetically distinguish the two clusters and identify the origin of each cell.

Identification of single cells with aneuploidy from scRNA-seq data

Chromosomal aneuploidy was investigated in each cell of the scRNA-seq dataset using the method previously described by Starostik et al [97]. Briefly, chromosomal aneuploidy can be inferred by integrating two types of information: chromosome-wide differential expression of dosage-associated genes and allelic imbalance at heterozygous SNP sites. For each single cell, the expression levels of dosage-associated genes were quantified using scploid, which provides a predefined set of dosage-associated genes along with benchmarking data from mosaic aneuploid mouse embryos [98]. Allelic imbalance was estimated per chromosome by measuring allelic read counts at each heterozygous SNP site.

To apply this method to our scRNA-seq dataset, we first generated aligned BAM files for every single cell using a custom python script that split the BAM files by cell barcode. Based on cell cluster information obtained from Souporcell (see above), maternal and fetal pseudobulk BAM files were generated by merging cells using SAMtools (version 1.3.1) [99]. Heterozygous SNPs were identified from the pseudobulk using GATK HaplotypeCaller [100]. Allelic read counts were then measured for each heterozygous SNP site using GATK ASEReadCounter for every cell. Owing to the sparsity of 10X Chromium scRNA-seq data, we included only SNP sites with at least three mapped reads per cell. Based on the measured read counts, we computed the allelic ratio for each chromosome–cell pair by dividing the sum of minor allele reads by the total number of mapped reads. Aneuploidy was determined if the scploid-derived p value was ≤ 0.05 and if the allelic ratio was within the top 10% across all chromosome–cell pairs.

Calculation of the common rejection module score

The common rejection module (CRM) score is calculated using the expression levels of 11 predefined CRM genes: BASP1, CD6, CXCL10, CXCL9, INPP5D, ISG20, LCK, NKG7, PSMB9, RUNX3, and TAP1 [34]. These genes were identified through meta-analyses as being consistently overexpressed during acute rejection (AR) across multiple transplanted organs. The score is derived by averaging the expression values of these CRM genes and subtracting the average expression of a dynamically selected set of control genes. This process ensures that the score reflects the relative activity of the module genes while accounting for background transcriptional noise. The calculations were performed using the Scanpy package. Statistical significance of CRM score differences between groups (e.g., Norm vs. RPL) was assessed using the Mann-Whitney U test, with Benjamini–Hochberg correction applied for multiple comparisons.

Differential gene expression analysis and origin-controlled filtering

Differential gene expression analysis was performed using the Wilcoxon rank-sum test as implemented in scanpy.tl.rank_genes_groups, with Benjamini–Hochberg correction applied for multiple testing. All analyses were conducted on log-normalized gene expression values. Genes were considered differentially expressed if the adjusted p value was less than 0.05 and if the absolute log2 fold change exceeded 1. Comparisons were made between biological conditions of interest, including maternal versus fetal origin and RPL versus normal samples, depending on the analysis context.

To identify robust disease-specific signatures independent of sample-level heterogeneity (e.g., cell origin imbalance), we implemented an origin-controlled expression filtering strategy within this statistical framework. First, we excluded genes showing significant differential expression between maternal and fetal cells (confounding markers). Second, from the remaining genes, we retained only those that were significantly differentially expressed between RPL and Norm conditions strictly within the maternal cell subset. This two-step filtering ensured that the identified signatures reflected true pathological changes rather than artifacts arising from cell origin proportions.

Separately, for an analysis focused on the aneuploid cell subset, we additionally employed the DESeq2 package in R [101]. This approach was used to compare normal versus RPL conditions exclusively within the fetal aneuploid cell population.

Machine learning model

We utilized devCellPy, a hierarchical machine learning framework built on XGBoost, to identify key features associated with the Norm and RPL conditions [42]. The dataset was stratified into training (90%) and testing (10%) sets to ensure robust performance evaluation. The classification model included two stages. In the first level, broad cell types were identified, and in the second level, immune cell types were distinguished based on the Norm and RPL conditions. We assessed the contribution of individual genes in the classification model by leveraging SHAP (Shapley Additive Explanations) analysis, which quantified the impact of each feature on the model's predictions by attributing a specific importance value to each gene.

Validation of prioritized RPL-associated genes using a transformer model of AI

To validate the SHAP-derived gene signatures for RPL immune signature identification, we employed scGPT, a transformer-based large language model specialized for single-cell RNA-seq data [46]. To determine the optimal feature set size, we evaluated model performance by varying the number of top SHAP-ranked genes (n) from both classification levels of devCellPy. Based on this optimization analysis, we selected the top 100 genes for each of the three immune cell types (T cells, granulocytes, and NK CD16 + cells). Using the pretrained scGPT_human model, we fine-tuned scGPT exclusively on these selected genes. We assessed both the zero-shot performance of the pre-trained model and the performance after fine-tuning. The fine-tuned models were subsequently used to classify cells into six categories: Tcell_Normal, Tcell_RPL, Granulocyte_Normal, Granulocyte_RPL, NK_CD16 ⁺ _Normal, and NK_CD16 ⁺ _RPL. This validation step ensured that the SHAP-ranked genes retained their predictive power across a different model architecture, supporting the applicability of the identified gene signatures to detect RPL-associated immune signatures.

Drug repurposing

Drug repurposing analysis was performed using ASGARD [57]. As input, we used the signature gene list derived from the integration of patient and cell type information from the previous step. Therefore, patient identity and cell type were not separately used in this step. To ensure adequate coverage across drugs, we selected six tissues (lung, prostate, skin, breast, kidney, and large intestine) for which experimental data were available for at least 700 compounds. Drugs were selected if they were predicted to have significant effects (FDR < 0.05) in at least three of these tissues. Unless otherwise specified, all procedures followed the original ASGARD pipeline.

Protein target and interaction data

Drug–target interactions were extracted from the DrugBank database [102]. Protein–protein interaction (PPI) data were obtained from the STRING database [103]. To ensure high-confidence interactions, we applied a confidence score threshold of 200 based on the STRING-provided scores. The optimal threshold was determined by varying the cutoff in increments of 100 and selecting the highest value that retained interactions involving all signature proteins and drug targets. To prioritize central hub proteins within the RPL network, we computed four distinct centrality measures using the NetworkX library (3.4.2) in Python [104]: degree centrality, betweenness centrality, eigenvector centrality, and random walk centrality. Proteins were ranked based on their average centrality scores to identify key regulatory nodes.