Single-nucleus and machine-learning integration identifies HLA-DRA and FTL as immune–metabolic axes and traditional Chinese medicine–targetable hubs in calcific aortic valve disease

Figures

• 

  Fig. 1. Single-nucleus transcriptomic landscape and intercellular communication network in human calcific aortic valve disease (CAVD). A UMAP visualization of nuclei from five human aortic-valve donors (GSE220774) showing clustering by sample and donor. B UMAP colored by annotated cell type, identifying major populations including valve interstitial cells (VICs 1–9), valve endothelial cells (VECs 1–7), transitional endothelial–mesenchymal cells (EndMT 1–2), stromal fibroblasts, T lymphocytes and six monocyte/macrophage subsets. C Global ligand–receptor communication networks inferred by CellChat. Heatmaps display the number of interactions (left) and interaction strength (right)

• 

  Fig. 2. Pathway-specific intercellular signaling networks in calcific aortic valve disease. A Chord diagrams representing the top four enriched signaling pathways, APP, CD99, FN1 and COLLAGEN, identified by CellChat. Each edge denotes a predicted ligand–receptor interaction between source and target cell populations with line thickness proportional to communication probability. B Dot plots showing source–target communication probabilities for the same pathways. Dot size corresponds to interaction strength and color intensity to interaction probability

• 

  Fig. 3. Pseudotime trajectory analysis of cellular transitions in calcific aortic valve disease. A Monocle 3 trajectory plot showing pseudotime ordering of nuclei colored by pseudotime value, indicating progressive transcriptional transitions across immune, stromal, and endothelial lineages. B UMAP representation of pseudotime branches annotated by cell type, depicting the trajectory from monocyte/macrophage-derived origins to interstitial and endothelial states. C Expression trends of the top ten pseudotime-associated genes along pseudotime

• 

  Fig. 4. Consensus non-negative matrix factorization (cNMF) identifies latent transcriptional programs in calcific aortic valve disease. A Top 20 high-contribution genes defining each of the ten cNMF-derived gene programs, with dot size representing gene rank and color scale indicating contribution score. B Violin plots displaying program activity across annotated cell clusters, showing distinct activation profiles among cellular subsets. C UMAP projections illustrating the spatial distribution of each gene program’s activity across the single-nucleus dataset

• 

  Fig. 5. Machine-learning classification and SHAP-based interpretability analysis of calcific aortic valve disease. A Receiver operating characteristic (ROC) curves of twelve single machine-learning models showing high predictive accuracy (AUC > 0.80). B AUC values of all single and ensemble model combinations, highlighting the superior performance of the Ranger classifier. C Global SHAP feature-importance ranking identifying FTL, CXCL8, FTH1, HLA-DRA, and C1QA as major predictors. D SHAP dependence plots showing the relationship between gene expression and SHAP contribution for the top genes. E SHAP summary plot colored by normalized expression, illustrating overall impact of each gene on model output

• 

  Fig. 6. Gene Ontology (GO) and KEGG pathway enrichment of top SHAP-selected genes. A GO Molecular Function terms enriched for the top five SHAP-prioritized genes showing significant association with peptide and MHC class II complex binding activities. B GO Cellular Component enrichment indicating localization to MHC complexes and vesicular membranes involved in antigen processing. C KEGG pathway enrichment highlighting immune-inflammatory pathways such as antigen processing and presentation, rheumatoid arthritis, and Staphylococcus aureus infection

• 

  Fig. 7. External validation of the top five SHAP-selected hub genes across independent calcific aortic valve disease (CAVD) datasets. Boxplots showing Z-score–normalized expression of C1QA, CXCL8, FTH1, FTL and HLA-DRA between control and CAVD groups across five external GEO cohorts

• 

  Fig. 8. External validation of diagnostic performance of SHAP-prioritized genes across independent CAVD datasets. A Heatmap showing per-gene ROC–AUC scores for C1QA, CXCL8, FTH1, FTL and HLA-DRA across five independent datasets. B Representative ROC curves for top-performing gene–dataset combinations demonstrating high predictive accuracy (AUC > 0.80) for FTL and HLA-DRA

• 

  Fig. 9. Immune infiltration landscape and gene–immune correlations across external CAVD cohorts. A–E Estimated immune cell fractions derived using CIBERSORT across five external datasets (GSE148219, GSE12644, GSE51472, GSE235995, GSE83453) showing predominant macrophage and CD4⁺ T-cell infiltration in CAVD samples. F Spearman correlation heatmap between the five SHAP-prioritized genes and inferred immune cell proportions

• 

  Fig. 10. PPI and functional enrichment analysis of shortlisted CAVD biomarkers FTL and HLA-DRA. A High-confidence STRING PPI network of FTL and its interacting partners enriched in iron metabolism, with corresponding GO Biological Process enrichment showing strong overrepresentation of iron homeostasis and heme biosynthesis. B STRING PPI network of HLA-DRA and associated immune-regulatory proteins showing enrichment in antigen processing and presentation, T-cell activation and adaptive immune response

• 

  Fig. 11. Network pharmacology analysis and identification of phytochemical ligands targeting HLA-DRA and FTL. A TCMBank-derived herb–ingredient–target–disease network showing compound associations for FTL. B Corresponding network for HLA-DRA highlighting shared and specific compound interactions. C Summary table of shortlisted ligands and their properties including molecular formula and molecular weight: Chrysin and α-Tocopherol (targeting HLA-DRA) and Colchicine and Diallyl Trisulfide (targeting FTL)

• 

  Fig. 12. Structure-based docking of Traditional Chinese Medicine (TCM) compounds with FTL and HLA-DRA predicted by CB-Dock2. A Cavity detection on FTL (top) with docked poses of Colchicine (Vina = − 4.4 kcal/mol, cavity volume = 172 Å3) and Diallyl trisulfide (Vina = − 3.0 kcal/mol, volume = 173 Å3) in predicted binding pockets (bottom). B Cavity mapping on HLA-DRA (top) with docked poses of Chrysin (Vina = − 6.9 kcal/mol, volume = 304 Å3) and α-Tocopherol (Vina = − 7.1 kcal/mol, volume = 195 Å3) in identified surface cavities (bottom)