Integrating machine learning with transcriptome-wide association studies to identify novel predictive biomarkers for...

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Venous thromboembolism (VTE) is a complex disorder with significant genetic contributions.
• Current risk prediction tools using single nucleotide polymorphisms (SNPs) lack biological interpretability.
• There is a need for improved VTE risk prediction models with enhanced accuracy.

Purpose of the Study:

• To develop a novel, integrative approach for VTE risk prediction.
• To identify biologically interpretable genetic predictors of VTE.
• To enhance the accuracy of VTE risk assessment using machine learning.

Main Methods:

• Utilized transcriptome-wide association study (TWAS) to identify candidate genes.
• Integrated patient-derived transcriptomic data for gene refinement.
• Employed machine learning algorithms (LASSO, Boruta, XGBoost, random forest, logistic regression) for model development and validation.

Main Results:

• Identified 577 candidate genes via TWAS, refined to four key predictive genes: KLKB1, ATG16L1, SELL, and GLRX2.
• Developed predictive models with high performance in training (AUC 0.913-0.970) and validation (AUC 0.916-0.968) cohorts.
• Confirmed gene contributions to model predictions using SHAP and regression coefficients.

Conclusions:

• The proposed integrative approach enhances VTE risk prediction through biologically interpretable genetic markers.
• This method offers a promising strategy for identifying novel VTE susceptibility genes.
• Improved VTE risk prediction can facilitate personalized prevention and treatment strategies.