Integrating machine learning with transcriptome‐wide association studies to identify novel predictive biomarkers for venous thromboembolism

Venous thromboembolism (VTE) is a multifactorial disorder in which genetic factors play a critical role. Existing tools like polygenic risk scores rely on single nucleotide polymorphisms (SNPs) with limited biological interpretability, potentially reducing predictive accuracy. To address this limitation, we propose an integrative approach that combines transcriptome-wide association study (TWAS), patient-derived transcriptomic data and machine learning. A total of 577 candidate genes were identified through a TWAS leveraging large-scale genome-wide association study summary statistics. These genes were refined using transcriptomic data from VTE patients and prioritized through the least absolute shrinkage and selection operator (LASSO) and Boruta algorithms, resulting in four predictive genes: KLKB1, ATG16L1, SELL and GLRX2. Predictive models based on these genes, constructed with XGBoost, random forest and logistic regression, demonstrated consistently high performance in both training (area under the receiver operating characteristic curve [AUC] range: 0.913-0.970) and validation cohorts (AUC range: 0.916-0.968). Shapley additive explanations (SHAP) and regression coefficients further supported the contribution of these genes to model predictions. This approach may facilitate the identification of biologically interpretable predictors and contribute to improved VTE risk prediction.