Integrating Pharmacogenomics and Network Topology for Machine Learning Prediction of HLA-Associated Severe Cutaneous Adverse Drug Reactions.

Adverse drug reactions (ADRs) remain a major clinical challenge and a leading cause of morbidity and mortality worldwide. Among them, severe cutaneous adverse drug reactions (SCARs), including Stevens-Johnson syndrome (SJS) and toxic epidermal necrolysis (TEN), represent life-threatening immune-mediated hypersensitivity responses strongly associated with specific human leukocyte antigen (HLA) alleles. Despite well-established pharmacogenetic associations, current diagnostic strategies remain largely retrospective and lack predictive capability for novel drug-HLA risk pairs. Here, we present an integrative network-informed machine learning framework for predicting HLA-associated SCAR risk by combining pharmacogenomic features, drug chemical structure, and topological descriptors derived from drug-drug and drug-symptom interaction networks. An Extreme Gradient Boosting (XGBoost) classifier trained on integrated HLA allele and drug features, labeled using curated HLA-SCAR associations, achieved an accuracy of 0.860 ± 0.005, an F1-score of 0.689 ± 0.010, with an area under the receiver operating characteristic curve (AUROC) of 0.922 ± 0.003 and an area under the precision-recall curve (AUPRC) of 0.768 ± 0.007. Notably, several predicted positive associations absent from the training data corresponded to biologically plausible and literature-supported cases, including carbamazepine-HLA-B*15:11, supporting the model's ability to generalize beyond known associations. Molecular docking provides structural evidence for the predicted associations, highlighting allele-specific binding patterns underlying these results. Overall, our results demonstrate that network-informed machine learning provides a proactive and integrative approach to SCAR risk prediction and may support early risk stratification and personalized drug safety assessment in precision medicine.