Adverse drug-drug interaction (ADDI) is an important issue in drug developments and clinical applications, which causes a significant burden in the healthcare system and leads to serious morbidity and mortality in patients. Many methods are proposed for ADDI prediction due to the accumulation of healthcare data in a massive scale. However, these methods are insufficient in exploring the potential adverse mechanisms among drugs and incapable of revealing the leading factors of ADDIs. In this paper, we propose a Multi-Task Multi-Attribute (MTMA) learning model for ADDI prediction. In MTMA, two drug attributes, molecular structure and side effect, are adopted to model the adverse interactions among drugs and two interpretable tensors, adverse molecular structure-molecular structure interaction tensor and adverse side effect-side effect interaction tensor, are designed to uncover the adverse mechanisms among drugs. Meanwhile, we impose $l_{2,1}$-norm on the predicted attribute matrices to explore the leading molecular substructures and side effects for each specific ADDI. The optimization problem of MTMA is solved by an alternating algorithm based on the methods of low-rank tensor decomposition and stochastic gradient descent. Experiments on the real-world dataset demonstrate the considerable performance of MTMA when compared with nine baseline methods and its three variants.

药物不良相互作用（ADDI）是药物开发和临床应用中的重要问题，这给医疗保健系统造成了沉重负担，并导致患者严重的发病率和死亡率。由于大规模积累医疗保健数据，提出了许多用于ADDI预测的方法。但是，这些方法不足以探索药物之间的潜在不良作用机制，也无法揭示ADDI的主要因素。在本文中，我们提出了用于ADDI预测的多任务多属性（MTMA）学习模型。在MTMA中，采用两种药物属性（分子结构和副作用）来模拟药物与两个可解释的张量之间的不利相互作用，不良分子结构-分子结构相互作用张量和不良副作用-副作用相互作用张量旨在揭示药物之间的不良机理。同时，我们强加${升}_{2，\mathrm{1个}}$-对预测的属性矩阵进行规范，以探索每种特定ADDI的主要分子亚结构和副作用。基于低秩张量分解和随机梯度下降的交替算法，解决了MTMA的优化问题。与九种基线方法及其三种变体相比，真实数据集上的实验证明了MTMA的可观性能。