基于Transformer和LSTM的药物相互作用预测  

Drug-Drug Interaction Prediction Based on Transformer and LSTM

作　　者：康雁[1] 徐玉龙 寇勇奇 谢思宇 杨学昆 李浩[1] KANG Yan;XU Yu-long;KOU Yong-qi;XIE Si-yu;YANG Xue-kun;LI Hao(School of Software,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学软件学院,昆明650504

出　　处：《计算机科学》2022年第S01期17-21,205,共6页Computer Science

摘　　要：药物相互作用的不良反应已经成为消化系统疾病、心血管疾病等发病率升高的重要原因之一,并且导致药物退出市场,因此准确预测药物相互作用受到了广泛关注。针对传统Encoder-Decoder模型无法捕捉药物子结构之间依赖的问题,提出了基于Transformer和LSTM的药物相互作用预测模型TransDDI(TransformerDDI)。TransDDI包括数据预处理模块、潜在特征抽取模块和映射模块3部分。数据预处理模块利用SPM算法从药物的SMILES格式输入中提取出表征药物的频繁子结构,形成药物特征向量,进而生成药物对的特征向量。潜在特征抽取模块利用Transformer充分挖掘特征向量中子结构之间蕴含的信息,突出不同子结构的不同重要作用,生成潜在特征向量。映射模块主要是将药物对的潜在特征向量和数据库中频繁子结构的向量形成字典表示,并且利用融合了LSTM的神经网络进行预测。在真实数据集BIOSNAP和DrugBank上,将所提模型与另外6种机器学习、深度学习方法进行实验比较。结果显示,TransDDI准确率更高,便于药物相互作用预测。The adverse reactions of drug-drug interactions have become one of the important reasons for the increase in the incidence of diseases such as digestive system diseases and cardiovascular diseases,and leads to the withdrawal of drugs from the market.Therefore,accurate prediction of drug interactions attracte widespread attention.Aiming at the problem that the traditional Encoder-Decoder model cannot capture the dependence between drug substructures,this paper proposes a TransDDI(TransformerDDI)based on Transformer and LSTM drug interaction prediction model.TransDDI includes three parts:data preprocessing module,latent feature extraction module and mapping module.The data preprocessing module uses the SPM algorithm to extract the frequent substructures that characterize the drug from the SMILES format input of the drug to form the drug feature vector,and then generate the feature vector of the drug pair.The latent feature extraction module uses Transformer to fully mine the information contained in the substructures of the feature vector,highlight the different important roles of different substructures,and generate potential feature vectors.The mapping module mainly forms a dictionary representation of the potential feature vector of the drug pair and the vector of the frequent substructure in the database,and uses the neural network fused with LSTM to make predictions.Onreal data sets BIOSNAP and DrugBank,the proposed method is compared with other 6 machine learning and deep learning methods by experiments.The results show that TransDDI has a higher accuracy rate and is convenient for drug interaction prediction.

关 键 词：药物相互作用 TRANSFORMER 长短期记忆网络 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]