药物-疾病关系预测:一种推荐系统模型

Mining drug-disease relationships: a recommendation system

目的药物重定位是指发掘已有药物新的治疗作用,然而具有潜在治疗作用的药物-疾病往往隐藏在数以百万计的关系对中。该研究基于医疗大数据分析,预测具有潜在治疗关系的药物-疾病关系对。方法将社交网络中推荐系统模型应用于药物重定位研究,并假设具有相似化学结构的药物可能具有相似的适应症。从开源数据库收集已知药物-疾病的治疗关系、副作用关系以及药物和疾病特征描述符,计算得到药物-药物的相似度和疾病-疾病相似度,再构建推荐模型将上述信息融合,并预测具有潜在治疗关系的药物-疾病,最终得到预测关系对的排序列表。结果列表排名前500的关系对中,有12.8%得到临床实验支持或综述报道,20%得到模式生物实验或细胞实验支持。结论相比于已有分类模型和随机抽样结果,本模型可明显提高具有潜在治疗作用药物-疾病的富集程度。

Aim Drug repositioning is to find new indications for existing drugs,however,potential drug-disease relationships are often hidden in millions of unknown relationship. With the analyzing of medical big data,we predict the potential drug-disease relationships. Methods Based on the assumption that similar drugs tend to have similar indications,we applied a recommendation-based strategy to drug repositioning. First,we collected the information of known drug-disease therapeutic effect,side effect,drug characters and disease characters; second,we calculated the drug-drug similarity measurements and diseasedisease similarity measurements; last,we used a collaborative filtering( CF) method to predict unknown drug-disease relationships based on the known drug-disease relationships by integrating the similarity measurements,and built a ranking list of prediction results. Results The experiments demonstrated that among the TOP 500 of the list,12. 8% were supported by clinical experiments or review,and 20% were supported by model organism or cell experiments. Conclusion Compared to the classification model and random sampling results,the CF model can effectively reduce the false positives.