Tan He, Ji Xiangmin, Xu Chen-Zhen, Zhao Xiaoyu, Hou Jie, Liu Mao, Ren Yan
School of Automation and Electrical Engineering, Inner Mongolia University of Science and Technology, Baotou 014010, China.
Key Laboratory of Synthetical Automation for Process Industries at Universities of Inner Mongolia Autonomous Region, Inner Mongolia University of Science and Technology, Baotou 014010, China.
ACS Omega. 2025 Aug 18;10(34):38753-38765. doi: 10.1021/acsomega.5c04006. eCollection 2025 Sep 2.
Identifying side effects is crucial for drug development and postmarket surveillance. Several computational methods based on graph neural networks (GNNs) have been developed, leveraging the topological structure and node attributes in graphs with promising results. However, existing heterogeneous-network-based approaches often fail to fully capture the complex structure and rich semantic information within these networks. Furthermore, the oversmoothing problem in GNNs remains a major challenge. In this study, we propose AGRL-DSE, a novel adaptive graph representation learning framework designed to enhance node-feature learning for predicting drug side effects. First, we construct a heterogeneous graph with intra- and interlayer connections to represent similarities and associations between drugs and side effects, capturing hidden topological relationships in heterogeneous contexts. Second, we integrate three GNN modules in AGRL-DSE, graph convolutional network (GCN), graph sample and aggregation (GraphSAGE), and graph attention network (GAT) at the graph, node, and edge levels, respectively, with the aim of capturing semantic information at different levels in graph data in a hierarchical manner, gradually extracting and enhancing the features of the graph. Additionally, we introduce an adaptive layer attention mechanism that dynamically assigns weights to each layer's features to achieve adaptive fusion of multilevel features, thereby automatically adjusting the contribution of each layer to the final embedding. Experimental results demonstrate that AGRL-DSE outperforms state-of-the-art predictive models in both hot- and cold-start scenarios, highlighting its superiority and generalizability. AGRL-DSE's ability to capture complex relationships and provide deeper insights into drug-side effect interactions could transform drug evaluation, monitoring, and prescription, leading to better health outcomes and more efficient drug development processes.
识别副作用对于药物研发和上市后监测至关重要。基于图神经网络（GNN）的几种计算方法已经被开发出来，利用图中的拓扑结构和节点属性，取得了不错的成果。然而，现有的基于异质网络的方法往往无法充分捕捉这些网络中的复杂结构和丰富语义信息。此外，GNN中的过平滑问题仍然是一个主要挑战。在本研究中，我们提出了AGRL-DSE，这是一种新颖的自适应图表示学习框架，旨在增强节点特征学习以预测药物副作用。首先，我们构建了一个具有层内和层间连接的异质图，以表示药物和副作用之间的相似性和关联性，捕捉异质环境中的隐藏拓扑关系。其次，我们在AGRL-DSE中分别在图层面、节点层面和边层面集成了三个GNN模块，即图卷积网络（GCN）、图采样与聚合（GraphSAGE）和图注意力网络（GAT），目的是以分层方式捕捉图数据中不同层面的语义信息，逐步提取并增强图的特征。此外，我们引入了一种自适应层注意力机制，动态地为各层特征分配权重，以实现多级特征的自适应融合，从而自动调整各层对最终嵌入的贡献。实验结果表明，AGRL-DSE在热启动和冷启动场景中均优于现有预测模型，凸显了其优越性和通用性。AGRL-DSE捕捉复杂关系并深入洞察药物-副作用相互作用的能力可以改变药物评估、监测和处方，带来更好的健康结果和更高效的药物研发过程。
