PRiVDIS: A Pathway-Based Risk-Informed Variant Dosage Integration Score Framework for Machine Learning-Based Discovery of Disease-Specific Molecular Pathways

초록

Myasthenia Gravis (MG) and Guillain-Barre Syndrome (GBS) are both autoimmune neuromuscular diseases. Although genetic associations have been identified for each disease, no studies have estimated discriminative power of molecular pathways against healthy controls. In this study, we propose a novel machine learning framework that incorporating Pathway-based Riskinformed Variant Dosage Integration Score (PRiVDIS) to identify disease-specific molecular pathways with interpretation provided by SHapley Additive exPlanations (SHAP). Nonsynonymous exonic variants from the UK Biobank dataset were analyzed. Logistic regression was performed for MG versus healthy controls and GBS versus healthy controls. For each disease, pathway enrichment analysis was conducted using the gene list derived from variants with p-values < 0.1. Then, the risk direction for each variant was determined based on whether the odds ratio exceeded 1. PRiVDIS was calculated by averaging risk-informed variant dosages within each pathway. Together with age and sex, the PRiVDIS dataset was subjected to dimensionality reduction and machine learning model for disease prediction. Both variantlevel and pathway-level importance were assessed using SHAP within the same analysis pipeline, by adjusting the scope of the SHAP model. The best-performing soft voting ensemble model with SHAP interpretation identified the five most contributing molecular pathways including ATP hydrolysis activity and Extracellular matrix organization in MG and Cilium movement involved in cell motility and Centriole in GBS. Additionally, the five most significant interactions between top SHAP-contributing variants in each pathway are identified through epistasis analysis for both MG and GBS. Furthermore, important variants for each disease were subjected to comparative analysis with MG and GBS patients to identify genetic differences between them. This integrative approach enables accurate classification and suggests disease-specific molecular mechanisms, offering insights into pathway-level biomarkers and potentially informing precision medicine approaches in autoimmune neurology.