基于图卷积神经网络的杀菌剂活性预测研究

Research on Predicting Fungicidal Activity Based on Graph Convolutional Network

摘要: 本研究提出了一种基于图神经网络(GCN)的杀菌剂活性预测模型。该模型通过直接解析分子SMILES字符串构建分子图结构，并利用GCN进行端到端的图表示学习。这种方法避免了传统方法中依赖手工特征工程的局限性，能够更全面且自动化地捕捉分子的结构信息。为评估模型性能，我们在相同杀菌剂活性数据集上，将GCN模型与基于分子描述符的支持向量机(SVM)、随机森林(RF)及深度神经网络(DNN)模型进行了系统对比。实验结果表明，GCN模型显著提高了真阳性率(TPR)，有效提高了筛选效率；同时，其在整体预测性能上全面超越了SVM、RF和DNN模型。本研究证实了图神经网络在直接从分子结构预测生物活性方面的强大能力，为高效发现新型杀菌剂候选分子提供了一种更具潜力的计算工具。

Abstract: This study proposes a graph neural network (GCN)-based model for predicting fungicide activity. The model directly parses molecular SMILES strings to construct molecular graph structures and utilizes GCN for end-to-end graph representation learning. This approach overcomes the limitations of traditional methods that rely on manual feature engineering, enabling a more comprehensive and automated capture of molecular structural information. To evaluate model performance, we conducted a systematic comparison between the GCN model and descriptor-based models—Support Vector Machine (SVM), Random Forest (RF), and Deep Neural Network (DNN)—on the same fungicide activity dataset. Experimental results demonstrate that the GCN model significantly improves the true positive rate (TPR), effectively enhancing screening efficiency. Concurrently, it surpassed the SVM, RF, and DNN models in overall predictive performance. This work validates the strong capability of graph neural networks in predicting bioactivity directly from molecular structures and provides a more promising computational tool for the efficient discovery of novel fungicide candidate molecules.