基于图神经网络的抗癌药物协同预测

Anti-Cancer Drug Synergy Prediction Based on Graph Neural Network

摘要: 药物组合作为个性化癌症治疗中的重要手段，因其低毒性和较小不良反应而受到广泛关注。然而，由于药物组合的指数性增长使得体外筛选药物协同组合变得耗时且费力。基于计算的方法逐渐被引入用于预测药物协同组合，以提高效率。然而，这些方法未能充分利用药物与细胞系之间复杂的关系。因此，本研究提出了一种名为GNNSynergy的基于图神经网络的抗癌药物协同预测方法。GNNSynergy利用Loewe评分构建了药物组合与细胞系之间的相互作用超图和单个药物与细胞系之间的相互作用图，其中药物和细胞系均作为图中的实体。然后，我们采用超图神经网络和图卷积神经网络分别学习两种图中的节点特征。最后，通过多层感知机和线性相关系数计算并融合两类网络的药物协同评分。GNNSynergy模型在基线实验中表现优于现有的先进模型，在新药与新细胞系实验中展现出了强大的泛化能力。

Abstract: Drug combination has gained wide attention in personalized cancer treatment due to its low toxicity and minimal adverse effects. However, the exponential growth of possible drug combinations makes it time-consuming and labor-intensive to screen for synergistic drug combinations in vitro. Computational approaches have been gradually introduced to predict drug synergy, aiming to improve efficiency. However, these methods have not fully exploited the complex relationship between drugs and cell lines. Therefore, this study proposes a graph neural network-based anticancer drug synergy prediction method called GNNSynergy. GNNSynergy utilizes Loewe scores to construct interaction hypergraphs between drug combinations and cell lines, as well as interaction graphs between individual drugs and cell lines, where drugs and cell lines are treated as entities in the graphs. Subsequently, we employ a hypergraph neural network and a graph convolutional neural network to learn node features in the two types of graphs, respectively. Finally, drug synergy scores from the two networks are calculated and fused using a multi-layer perceptron and linear correlation coefficient. The GNNSynergy model outperforms existing advanced models in baseline experiments and exhibits strong generalization ability in experiments involving new drugs and cell lines.