Explainable machine learning reveals ribosome biogenesis biomarkers in preeclampsia risk prediction

Figures

• 

  Figure 1. Technology roadmap.

• 

  Figure 2. Identification of RiboRDEGs. (A) Volcano plot of DEGs (|log2FC| > 0.1, P.adj < 0.05), red indicates upregulated genes and blue indicates downregulated genes (B) Venn diagram intersecting DEGs and RiboRGs,red indicates DEGs and blue indicates RiboRGs. (C) Heatmap of RiboRDEG expression across samples, red indicates high expression and blue indicates low expression. (D) 3D PCA plot demonstrating group segregation based on RiboRDEG profiles, red indicates PE samples, blue indicates normal samples.

• 

  Figure 3. Multi-modal enrichment analysis of PE-associated molecular pathways. (A) GSEA heatmap of combined dataset pathways (red: leading-edge enrichment; blue: trailing-edge enrichment). (B) Bubble plot of RiboRDEG-enriched GO terms (BP, CC, MF) and KEGG pathways, with term counts on the x-axis. (C) Bayesian pathway enrichment bar plot. (D) Bayesian network illustrating pathway interactions, with node size reflecting functional centrality and edge properties indicating interaction strength (thickness) and directionality (arrows). Solid/dashed lines distinguish established vs. hypothesized regulatory relationships.

• 

  Figure 4. Co-expression network construction and hub gene identification. (A) Soft-threshold selection for scale-free topology (left: fit index; right: mean connectivity). (B) Hierarchical clustering dendrogram with merged modules. (C) Module eigengene clustering tree. (D) Inter-module correlation heatmap. (E) Module-trait correlation heatmap highlighting MEblue-PE association. (upper: correlation coefficients; lower: p-values) (F) Venn diagram intersecting MEblue genes and RiboRDEGs. (G) Correlation heatmap of seven hub genes. In the heatmap, red indicates a positive correlation and blue indicates a negative correlation. |r|>0.95: significant correlation; |r|≥0.8: highly correlated; 0.5≤|r|<0.8: moderately correlated; 0.3≤|r|<0.5: weakly correlated; |r|<0.3: not correlated.

• 

  Figure 5. Development and validation of the machine learning-based predictive model. (A) Bar plot comparing mean AUCs of 113 algorithm combinations across training and validation cohorts. (B) ROC curves of the RF model (TPR, true positive rate; FPR, false positive rate; AUC>0.9 indicates high diagnostic value). (C) Decision curve analysis evaluating clinical utility. (D) Calibration curves assessing prediction accuracy. (E) Confusion matrices quantifying classification performance.

• 

  Figure 6. SHAP-based interpretation of key feature genes in the RF predictive model. (A) Feature importance bar plot ranked by mean absolute SHAP values. (B) Beeswarm plot showing SHAP value directionality (x-axis) versus gene expression (color scale: red = high, blue = low). (C–E). Interaction scatterplots: x-axis = gene expression, y-axis = SHAP value, color = interacting gene expression. (F) Waterfall plot for a PE case (predicted risk = 0.864). (G) Control case waterfall plot (risk = 0.131). Arrows indicate SHAP contributions (red: risk increase; blue: decrease).

• 

  Figure 7. Diagnostic validation and functional interaction networks of key genes. (A–F). Violin plots of key gene expression in PE (red) vs. controls (blue): GLUL (A), DDX28 (B), NCL (C), RIOK1 (D), SUV39H1 (E), RRS1 (F). G-L. ROC curves for training cohort: GLUL (G), DDX28 (H), NCL (I), RIOK1 (J), SUV39H1 (K), RRS1 (L). (M–R). Validation cohort ROC curves: GLUL (M), DDX28 (N), NCL (O), RIOK1 (P), SUV39H1 (Q), RRS1 (R). An AUC of 0.5–0.7 indicates low diagnostic utility, while an AUC of 0.7–0.9 suggests moderate diagnostic value. (S) PPI network (nodes: key genes and functionally associated partners; edges: physical interactions (red), co-expression (purple), geneticlnteractions (green); thickness: interaction confidence). ***p < 0.001; **p < 0.01.

• 

  Figure 8. Immune infiltration landscape and key gene correlations in PE. (A) Boxplots comparing immune cell enrichment scores between PE (red) and controls (blue). (B) Correlation network of immune cells and key genes. (C–H). Scatterplots of key gene-immune cell correlations: GLUL-MDSC (C), GLUL-Memory B cell (D), RRS1-Macrophage (E), RRS1-MDSC (F), NCL-Macrophage (G), RIOK1-MDSC (H). *p < 0.05; **p < 0.01; ***p < 0.001. red indicates a positive correlation and blue indicates a negative correlation. 0.3≤|r|<0.5: weakly correlated.

• 

  Figure 9. Validation of key gene expression in PE placental tissues. (A–F) mRNA expression levels of GLUL (A), DDX28 (B), NCL (C), RIOK1 (D), SUV39H1 (E), and RRS1 (F) in control (blue) versus PE (red) groups. ns, not significant; **p < 0.01; ***p < 0.001.