Contrastive Hypergraph Networks with Adaptive Negative Sampling for Accurate Discovery and Risk Calibration of Low-Frequency Drug-Drug Interactions
Md. Jisan Mashrafi, Hasan Ahamed Alif, Muhammad Jasim Uddin, Sebagat Selim, Urmi Haldar, Meher Tabassum
Abstract
Drug-drug interactions (DDIs) pose significant challenges in clinical pharmacology, with rare interactions often causing severe adverse events despite being underreported in traditional screening methods. Existing computational approaches struggle with class imbalance and fail to capture complex interaction patterns effectively. This work presents a novel contrastive hypergraph network framework that integrates three key innovations: adaptive negative sampling for mitigating class imbalance, hypergraph convolutional learning for modeling higher-order relationships, and temperature scaling for calibrated risk prediction. Our Siamese-style architecture employs contrastive loss optimization to learn discriminative drug pair embeddings, while similarity-based negative sampling generates semantically meaningful non-DDI training examples. Hypergraph convolution captures complex multi-drug relationships beyond traditional pairwise approaches. The ChCh-Miner benchmark dataset demonstrates that it outperforms state-of-the-art algorithms like SPARSE and DeepDDI, achieving a ROC-AUC of 0.8655 and an F1-score of 0.7239. Our framework provides a scalable solution for detecting rare DDI, offering calibrated probability estimates for clinical decision support and personalized medicine.