A cost-effective machine learning-based method for preeclampsia risk assessment and driver genes discovery

Area of Science:

• Computational biology
• Genomics
• Reproductive medicine

Background:

• The placenta is crucial for pregnancy; placental dysfunction causes preeclampsia (PE), a major cause of maternal and infant mortality.
• Current PE diagnostic methods have high misdiagnosis rates, necessitating improved identification strategies.

Purpose of the Study:

• To develop a computational biology approach using single-cell transcriptome data to identify pathological placental cell subpopulations.
• To predict preeclampsia (PE) risk and improve diagnostic accuracy for clinical decision-making.

Main Methods:

• Employed single-cell RNA sequencing (scRNA-seq) on healthy and early-onset PE placentas.
• Utilized machine learning (TURF_XGB, LASSO) and feature selection for cell subpopulation classification and biomarker identification.
• Developed an ensemble model-based risk stratification card and an online web server for accessibility.

Main Results:

• The TURF_XGB model achieved 92.61% accuracy in classifying healthy placental cell subpopulations.
• Identified 110 marker genes for placental heterogeneity and 497 biomarkers for PE.
• Dendritic cells and genes C1QB/C1QC were strongly associated with early-onset PE, potentially via inflammation.
• The risk stratification model demonstrated high predictive power (AUC 0.99).

Conclusions:

• Single-cell transcriptome analysis with machine learning offers a robust method for preeclampsia risk assessment.
• C1QB and C1QC may play a role in early-onset PE pathogenesis by influencing inflammation through complement and coagulation pathways.
• The developed online tool enhances clinical accessibility for preeclampsia diagnosis and management.