Integrative Transcriptomic and Bioinformatics Approaches Combined With Transformer Models Identify Key Gene Networks in Atherosclerosis.

This study aimed to prioritize candidate transcriptomic mediators associated with human atherosclerosis progression by integrating transformer-based deep learning with classical bioinformatics and experimental validation. Gene expression profiles from dataset GSE100927 (69 plaques, 35 controls) were normalized and analyzed for differential expression using the limma package (FDR < 0.01). A TabTransformer model employing multi-head self-attention was trained (80/20 split, 10-fold cross-validation) to predict disease status, and SHapley Additive exPlanations (SHAP) analysis quantified gene-level contributions. Network topology was examined using STRING and Cytoscape to identify hub genes through betweenness centrality, while pathway enrichment was assessed via GO and KEGG analyses. Among 638 differentially expressed genes (402 upregulated, 236 downregulated), the TabTransformer achieved a mean AUC of 0.964, surpassing the LASSO baseline by 3.8%. The top-ranked genes, TYROBP, TNF, PTPRJ, DHRS9, and COL1A1, were primarily involved in leukocyte activation, NF-κB signaling, and smooth muscle dysfunction. Experimental assays in oxidized LDL-treated human umbilical vein endothelial cells confirmed significant upregulation of TYROBP (6.3-fold) and TNF (4.8-fold), validating computational predictions. These findings support an immunoinflammatory axis linked to the suppression of smooth muscle identity in atherosclerosis, and highlight TYROBP and TNF as mechanistically plausible candidate biomarkers and potential therapeutic targets.