【摘要】 构建基于药物多特征融合的药物疾病关联预测模型，为药物知识发现提供新思路。借助药物的化学结构、药物-副作用关联、药物-靶标关联的3个特征，构建融合的药物综合相似度及基于MeSH的疾病语义相似度特征表示方法。利用图卷积神经网络模型抽取药物-疾病图数据特征信息，构建基于多特征融合的药物疾病关联预测模型(MFFGCN),进而实现未知的药物疾病关联发现。利用269种药物、598种疾病及其之间的18 416种关联关系，对药物疾病存在的未知关联进行预测，借助AUC、AUPR、准确率、灵敏度、召回率、F1等多个评价指标进行评价。结果表明，多特征融合的药物疾病关联预测方法的AUC指标为0.866 2,较单一特征的平均预测指标最大相对提升为2.48%,较4种代表性基线方法的指标最大相对提升为1.67%;AUPR指标为0.341 2,较单一特征预测结果最大相对提升为1.67%,较4种代表性基线方法提升27.49%。对预测结果中药物-疾病预测关联得分中排名前10的组合及阿霉素为例的单一药物预测组合进行文献研究验证、临床治疗验证，同样证明MFFGCN在未知的药物疾病关联预测上表现良好，能有效地发现药物的新适应症，为药物重定位提供方法借鉴和理论依据。更多还原

【Abstract】 We constructed a drug-disease association prediction model on the basis of drug multi feature fusion, which can provide theoretical foundation for drug knowledge discovery. Three similarities fused into drug comprehensive similarity by drug chemical structure, drug-side effect and drug-target multi-features. Disease similarity was calculated based on MeSH tree number. Next, GCN method was used to extract feature information of drug-disease graph data. Finally, MFFGCN was constructed for drug-disease association prediction. The association of drug diseases was predicted on the same data set, with the help of multiple evaluation indicators such as AUC, AUPR, accuracy, sensitivity, recall and F1, MFFGCN has better performance than the single feature association prediction method and 4 existing representative algorithms. The AUC index is 0.866 2, which is 2.48% higher than the average predicted AUC index of single feature and 1.67% higher than the baseline method. The AUPR index is 0.3412, which is 1.67% higher than the average predicted AUC index of single feature and 27.49% higher than the baseline method. MFFGCN has achieved good performance in the prediction of unknown drug disease association. This methods can find new indications of drugs, and also provide methodological reference and theoretical basis for drug relocation.更多还原