Identification and evaluation of diagnostic signatures for VTE based on bioinformatics analysis and machine learning

Abstract

Abstract Background Venous thromboembolism (VTE) is a common life-threatening vascular disorder for which reliable diagnostic biomarkers remain limited. This study aimed to identify potential diagnostic signatures by integrating bioinformatics and machine learning, followed by model evaluation. Methods Both GSE19151 and GSE48000 were adopted to identify differentially expressed genes (DEGs) in VTE patients. Functional enrichment analyses were performed using GO and KEGG databases. Five machine learning algorithms (LASSO, RF, XGBoost, GBM, and KNN) were applied to screen diagnostic biomarkers. The expression levels and diagnostic performance of the candidate biomarkers were both evaluated in GSE19151, GSE48000 datasets and examined in a clinical cohort by ELISA. The models were evaluated using AUC, calibration curves, and decision curve analysis (DCA), and DeLong’s test was used to compare AUCs between models. Results Venn analysis identified 76 common DEGs in VTE, primarily enriched in erythrocyte homeostasis and oxidative stress pathways. Four candidate biomarkers (BFAR, CXorf21, PIP5K1A, and UBFD1) were identified. BFAR (535.7 ± 84.3 vs. 451.4 ± 24.2 pg/mL, p < 0.001) and CXorf21 (688.5 ± 89.8 vs. 598.8 ± 31.9 pg/mL, p = 0.006) protein levels were significantly higher in VTE patients than in controls. The GSE19151-derived four-molecule nomogram had an apparent AUC of 0.979 (95% CI: 0.962–0.997, p < 0.001) in the training cohort. Re-fitted four-biomarker models achieved AUCs of 0.775 (95% CI: 0.677–0.873, p < 0.001) in GSE48000 and 0.818 (95% CI: 0.701–0.935, p < 0.001) in the clinical cohort, respectively. DCA suggested a potential net benefit in the GSE19151, GSE48000 and clinical evaluation cohorts. Conclusion BFAR and CXorf21 are candidate VTE-associated biomarkers with directionally consistent transcriptomic and protein-level findings. The four-biomarker model remains exploratory and requires further validation.

The authors' abstract, as published at the source. Hereditas, 2026 · DOI ↗