Niu Yanqing, Zhang Wen
School of Mathematics and Statistics, South-central University for Nationalities, Wuhan, 430074, China.
School of Computer, Wuhan University, Wuhan, 430072, China.
Interdiscip Sci. 2017 Sep;9(3):434-444. doi: 10.1007/s12539-017-0236-5. Epub 2017 May 17.
Unexpected side effects of drugs are great concern in the drug development, and the identification of side effects is an important task. Recently, machine learning methods are proposed to predict the presence or absence of interested side effects for drugs, but it is difficult to make the accurate prediction for all of them.
In this paper, we transform side effect profiles of drugs as their quantitative scores, by summing up their side effects with weights. The quantitative scores may measure the dangers of drugs, and thus help to compare the risk of different drugs. Here, we attempt to predict quantitative scores of drugs, namely the quantitative prediction. Specifically, we explore a variety of drug-related features and evaluate their discriminative powers for the quantitative prediction. Then, we consider several feature combination strategies (direct combination, average scoring ensemble combination) to integrate three informative features: chemical substructures, targets, and treatment indications. Finally, the average scoring ensemble model which produces the better performances is used as the final quantitative prediction model.
Since weights for side effects are empirical values, we randomly generate different weights in the simulation experiments. The experimental results show that the quantitative method is robust to different weights, and produces satisfying results. Although other state-of-the-art methods cannot make the quantitative prediction directly, the prediction results can be transformed as the quantitative scores. By indirect comparison, the proposed method produces much better results than benchmark methods in the quantitative prediction. In conclusion, the proposed method is promising for the quantitative prediction of side effects, which may work cooperatively with existing state-of-the-art methods to reveal dangers of drugs.
药物的意外副作用是药物研发中备受关注的问题，副作用的识别是一项重要任务。最近，有人提出使用机器学习方法来预测药物是否存在感兴趣的副作用，但要对所有副作用进行准确预测却很困难。
在本文中，我们通过对药物的副作用进行加权求和，将药物的副作用概况转化为定量分数。这些定量分数可以衡量药物的危险性，从而有助于比较不同药物的风险。在此，我们尝试预测药物的定量分数，即定量预测。具体而言，我们探索了各种与药物相关的特征，并评估它们在定量预测中的判别能力。然后，我们考虑了几种特征组合策略（直接组合、平均评分集成组合），以整合三种信息特征：化学子结构、靶点和治疗适应症。最后，将表现更好的平均评分集成模型用作最终的定量预测模型。
由于副作用的权重是经验值，我们在模拟实验中随机生成不同的权重。实验结果表明，定量方法对不同权重具有鲁棒性，并产生了令人满意的结果。虽然其他现有方法不能直接进行定量预测，但预测结果可以转化为定量分数。通过间接比较，在定量预测中，本文提出的方法比基准方法产生的结果要好得多。总之，本文提出的方法在副作用定量预测方面很有前景，它可以与现有的先进方法协同工作，以揭示药物的危险性。
