Five Placental Genes Flag Preeclampsia in Interpretable AI Model

Preeclampsia, a hypertensive disorder of pregnancy that affects roughly two to eight percent of pregnancies worldwide, remains one of the leading causes of maternal and fetal illness and death. Clinically, it appears as new-onset hypertension after twenty weeks of gestation, often accompanied by proteinuria and dysfunction of organs such as the kidneys and liver. If it is not treated in time, it can progress to eclampsia, HELLP syndrome, or multiple organ failure. A new study published in Frontiers in Genetics on 29 October 2026 now reports a five-gene signature tied to placental oxidative stress that distinguishes preeclamptic placentas from healthy ones, and an interpretable neural network built on those genes that performed consistently across independent datasets.

The research team, led by Leilei Tian and Xuyuan Ma of Hebei Medical University with senior author Hong Xin of the Second Hospital of Hebei Medical University, combined publicly available transcriptomic data with clinical placental specimens. They drew on nine datasets from the Gene Expression Omnibus, dividing them into three roles: GSE10588 and GSE75010 for screening differentially expressed genes, six datasets (GSE30186, GSE44711, GSE54618, GSE24129, GSE43942, and GSE147776) for training a machine learning model, and GSE25906 as a completely quarantined independent validation cohort. Batch effects among the training cohorts were adjusted with the ComBat method in an unsupervised manner, without exposing disease labels, and principal component analysis confirmed that the correction worked.

Differential expression analysis on the merged screening datasets, which comprised 103 normal samples and 97 preeclampsia samples, identified 87 differentially expressed genes: 65 upregulated and 22 downregulated. The screening criteria required a Benjamini–Hochberg adjusted P value below 0.05 and an absolute log fold change greater than 0.585. When these genes were intersected with a predefined list of 1,399 oxidative stress-related genes drawn from the GeneCards database with a relevance score of at least seven, 17 candidate differentially expressed oxidative stress-related genes emerged, including LEP, TGM2, BCL6, MME, MMP1, ENG, ABCG2, FLT1, NTRK2, CRH, NDRG1, HK2, ERO1A, LTF, FRZB, SERPINA3, and CP.

To narrow this list, the researchers applied three recursive feature elimination algorithms—support vector machine, random forest, and extreme gradient boosting—within a stratified five-fold cross-validation framework. Cross-validated AUC stayed stable between 0.88 and 0.90 as features dropped from 17 to 7, but fell sharply when fewer than seven features were used. Taking the top seven genes from each algorithm and computing the consensus intersection across all three methods yielded five core signature genes: LEP, HK2, NDRG1, MME (membrane metallo-endopeptidase), and SERPINA3. The authors note that this five-gene panel matched the discriminative performance of the larger seven-gene panels while keeping the model simpler.

A key strength of the study is that the gene findings were tested at the protein level in real tissue. The team collected placental samples from 50 women with preeclampsia and 50 women with normal pregnancies who delivered at the Second Hospital of Hebei Medical University between October 2023 and January 2025, all with singleton pregnancies and no history of smoking or alcohol use. Western blotting showed that all five signature proteins were significantly upregulated in preeclamptic placental tissue, with P values below 0.05. Immunohistochemistry on the same samples then localized the positive staining predominantly to placental villous trophoblasts, especially the syncytiotrophoblast layer, with weaker signals in cells around villous blood vessels and scattered cells in the villous stroma.

On the computational side, the group built an artificial neural network discriminant model using the five genes as inputs. In five-fold cross-validation across the six training cohorts, repeated over three random seeds, the model achieved an average AUC of 0.888 ± 0.010, with accuracy of 81.5 percent, sensitivity of 83.0 percent, and specificity of 79.7 percent. When applied to the independent validation cohort GSE25906, which contained 60 samples (23 preeclampsia cases and 37 controls), the model maintained an AUC of 0.819 ± 0.010, accuracy of 73.3 percent, sensitivity of 73.9 percent, and specificity of 73.0 percent. Head-to-head comparisons on the same cohort showed the five-gene neural network performed on par with seven-gene support vector machine, random forest, and XGBoost models, whose AUCs ranged from 0.814 to 0.821.

Because neural networks are often criticized as black boxes, the researchers used SHAP (SHapley Additive exPlanations), a post hoc interpretation method based on Shapley values from cooperative game theory, to decompose each prediction into feature-level contributions. The SHAP beeswarm analysis showed that LEP and MME had the greatest predictive impact on classification, followed by HK2, SERPINA3, and NDRG1. The authors are careful to point out that high SHAP attribution reflects algorithmic feature importance within the trained model, not biological causality in disease. Validation data also confirmed that all five genes were significantly differentially expressed between preeclampsia and control samples in GSE25906.

The study also probed the placental immune microenvironment using CIBERSORT deconvolution on the GSE75010 dataset, estimating the relative abundance of 22 immune cell types with 1,000 permutations. Compared with normal pregnancies, the preeclampsia group showed significantly increased resting mast cells and significantly decreased M2 macrophages, monocytes, neutrophils, and plasma cells, while NK cells did not change significantly. Correlation analysis linked the signature genes to these immune shifts: LEP, HK2, and NDRG1 were positively correlated with regulatory T cells and resting mast cells but negatively correlated with monocytes, neutrophils, and M2 macrophages; MME was positively correlated with eosinophils and negatively correlated with M2 macrophages; and SERPINA3 was mainly negatively correlated with monocytes and neutrophils.

The authors place these findings within the established two-stage model of preeclampsia, in which inadequate trophoblast invasion and impaired spiral artery remodeling reduce placental perfusion, triggering ischemia, reactive oxygen species production, and eventually systemic maternal endothelial dysfunction. Each signature gene has a plausible role in this framework: LEP is a placental hormone whose elevation in preeclampsia has been reported repeatedly and which may amplify pro-inflammatory signaling through JAK-STAT and PI3K/Akt pathways; HK2 is a rate-limiting glycolytic enzyme that oxidative stress mediators such as hydrogen peroxide can directly modify; NDRG1 is a hypoxia-responsive molecule that can impair trophoblast invasion and angiogenesis; and SERPINA3 shows epigenetically and hypoxically driven overexpression that impairs trophoblast adhesion in prior functional studies.

The researchers acknowledge important limitations. The model is built from postpartum placental tissue, so it characterizes local molecular alterations rather than serving as an antenatal screening tool, and prospective validation in maternal peripheral blood would be needed before any clinical use. Retrospective public cohorts also vary in gestational age reporting and in distinguishing early-onset from late-onset disease, which are potential confounders, and the immune infiltration analysis rested on a single platform. The team plans prospective cohorts with better-matched clinical covariates, single-cell RNA sequencing studies, and mechanistic experiments to clarify the functional roles of the five genes, and to explore whether the signature can eventually be developed into non-invasive blood-based biomarkers for earlier prediction of preeclampsia.

Subject of Research: Oxidative stress-related signature genes and an interpretable machine learning model for preeclampsia

Article Title: Screening of oxidative stress-related signature genes and construction of an interpretable discriminant model for preeclampsia

Article References: Tian, L., Ma, X., Zhang, Y., Hao, T., Wang, H., Meng, H., & Xin, H. (2026). Screening of oxidative stress-related signature genes and construction of an interpretable discriminant model for preeclampsia. Frontiers in Genetics, 17, Article 1855189. https://doi.org/10.3389/fgene.2026.1855189

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1855189

Keywords: preeclampsia, oxidative stress, placenta, machine learning, artificial neural network, SHAP, biomarkers, gene expression, immune infiltration, trophoblast, GEO datasets, Frontiers in Genetics