Machine Learning-Driven Decoding of Maternal Immune Signatures in Repeated Pregnancy Loss

RPL decidua shows elevated overall immune activation, and this signal originates mainly from maternal-origin decidual NK subsets dNK2 and dNK3. Prior understanding of RPL immunology came largely from peripheral blood or bulk tissue; this work uses single-cell resolution with genotype-based origin assignment to localize the activation signal to specific subsets and to the maternal host. Based on 62,588 high-quality single cells (57,976 with confidently inferred origin); CRM scores were significantly elevated in RPL (Mann-Whitney U test, p = 1.27E-144), with dNK2 (p = 7.55E-13) and dNK3 (p = 1.23E-17) showing the most significant elevation, with Benjamini-Hochberg correction applied.

A hierarchical machine learning framework identified maternal T cells as the population carrying the most distinct and generalizable RPL signatures across internal and external datasets. Unlike conventional single-gene mean-difference testing, this approach learns combinatorial transcriptomic patterns to classify cell states and tests generalization in independent cohorts. Internal test set weighted F1 was 0.87 (first level) and 0.97 (second level); external RPL cohort (Du) was 0.81 and 0.93; external normal cohort (Vento) second-level F1 was 0.32, but T cells, granulocytes, and NK CD16+ cells retained second-level F1 of 0.86, 0.94, and 0.89 in Vento.

SHAP-derived T-cell RPL signature genes transfer across model architectures, with markedly improved classification after scGPT fine-tuning. An independent Transformer architecture, distinct from devCellPy, was used to validate the signature genes, reducing the chance of overfitting to a single modeling framework. Pre-trained scGPT zero-shot weighted F1 was 0.61, improving to 0.86 after fine-tuning; T cells and granulocytes achieved F1 > 0.94 on the test set; T cells and granulocytes achieved F1 > 0.96 in Vento; T cells maintained a weighted average F1 of 0.99 in Du, whereas NK CD16+ cells failed to maintain robust classification.

Network centrality analysis and origin-controlled expression filtering converge to nominate CXCR4 and JUN as candidate druggable molecules associated with T-cell dysregulation. Both genes emerged from independent PPI network hub ranking and from differential expression screening that removed maternal-versus-fetal origin confounding; neither had previously been highlighted as a distinguishing feature of decidual T cells in RPL. The initial 100 SHAP-ranked genes were reduced by two filtering steps (excluding 42 maternal-fetal DEGs, then 38 genes not significant within maternal T cells) to a final panel of 20 genes; CXCR4 and JUN were significantly upregulated in RPL maternal T cells (Wilcoxon rank-sum test with Benjamini-Hochberg FDR correction, adjusted p < 0.001); ASGARD predicted five compounds significantly reversing RPL-associated expression in at least three tissues, three of which (prednicarbate, sirolimus, niclosamide) have prior links to RPL-relevant contexts.

This work applies to first-trimester (roughly 9-11 weeks) decidual tissue from karyotype-confirmed euploid RPL; its analytical framework can be reused by other researchers who have matched maternal genotypes and single-cell data. The conclusions are positioned as nomination of candidate markers and targets for subsequent functional experiments (such as organoid models or in vivo edits) and for clinical stratification research.

Readers should still watch whether CXCR4 and JUN upregulation is a primary driver of pregnancy loss or a secondary consequence of fetal demise, which requires functional perturbation to determine; cell-proportion differences are sensitive to pre-analytical factors such as sampling site, biopsy depth, and dissociation efficiency, and the authors treat them as contextual observations; recorded gestational age differed between groups (Norm 9.5 weeks, RPL 10.3 weeks), and in RPL the recorded age reflects clinical diagnosis rather than the actual timing of embryonic demise; NK CD16+ cells showed reduced generalization in external datasets and granulocyte validation in the Du dataset was limited by extreme scarcity (n = 9), so the stability of conclusions for these cell types awaits more data.