Adaptive Regularized Multiattribute Fuzzy Distance Learning for Predicting Adverse Drug–Drug Interaction

Adverse drug–drug interaction (ADDI) causes harmful injuries and accidental deaths in patients, posing as a significant life-threatening issue in public health. Early prediction of ADDIs has become an increasingly concerning task for the safety of pharmacotherapy during clinical treatments. In this article, we propose an adaptive regularized multiattribute fuzzy distance (MAFD) learning model for ADDI prediction. Unlike the existing works that only focus on whether an adverse interaction occurs or not for a specific drug pair and do not consider their implicit medication risks, MAFD employs fuzzy distance learning by designing a fuzzy membership matrix to model the adverse distance with a fuzziness level for exploring the medication risks of adverse drug pairs. Meanwhile, for each attribute, we develop two projection matrices to respectively map its original feature and adverse interaction spaces into a common space for eliminating noisy information and capturing their compact and informative representations. Besides, adaptive regularization is explicitly designed to investigate the underlying characteristics of different attributes in ADDI modeling and neighborhood structure preservation is seamlessly integrated to benefit the prediction results. The optimization problem is solved by an iterative algorithm based on the alternating direction method of multipliers with detailed convergence proofs. Experiments on real-world dataset demonstrate the effectiveness of MAFD when compared with ten baselines and its five variants.