面向药物-靶标相互作用预测的多核融合方法研究

Identification of drug-target interactions is an important process in drug discovery. The traditional biochemical experiments are expensive and time consuming. In recent years, relevant computational methods have emerged to solve the identification problem. Most existing computational methods used a single molecular substructure of drug and target sequence information to construct predictive model. Few models were constructed by multiple drugs and target information. Employing heterogeneous information, extracting effective feature, integrating various features via optimized weight are the difficulties of predicting drug-target information. This project intends to build a predictive model using a machine learning method based on multi-kernel fusion. This project will be carried out in three aspects: (1) extracting features from the molecular structure, side effects of drug, indications and drug-drug interaction networks, and constructing multiple drug kernels; (2) extracting relevant features from the sequence, structure, evolution, functions and their interaction networks of target protein and calculating the corresponding kernel matrix; (3) The multi-kernel learning algorithm is used to obtain the optimized kernel linear weight parameters, which could combine kernels in drug and target space, respectively. The fused kernels and semi-supervised algorithms will be used to construct predictive models to detect new drug-target interactions. This project aims to further find the key factors of drug-target interaction through the predictive model, which could provide reference for the research and drug discover. The research results will accelerate the development of new drugs.

识别药物-靶标相互作用是药物发现中的重要过程，传统生化实验方法是昂贵且耗时的。近年，已出现相关计算方法以解决识别问题，多数已有计算方法使用单一药物子结构和靶标序列信息构建模型，鲜有利用多种药物和靶标信息构建模型。引入多种异构信息，提取有效特征，优化融合各种特征为该方向的研究难点。本项目拟采用多核线性融合的机器学习方法来构建预测模型，并由三方面进行：（1）从药物分子结构、副作用、适应症及其相互作用网络等方面提取特征，并分别构建药物核矩阵；（2）从靶标蛋白序列、结构、进化、功能及其相互作用网络等方面提取相关特征，且计算相对应核矩阵；（3）利用多核学习算法在药物和靶标空间分别获得核线性权重参数，并利用融合核和半监督算法构建预测模型用于检测新药物-靶标相互作用。本项目拟通过预测模型进一步揭示影响药物-靶标相互作用的关键因素，为新药的研发提供参考，有助于药物的开发。

本项目引入多种异构信息，提取有效特征，采用多核线性融合的机器学习方法来构建预测模型，并由三方面进行。提取药物特征方面：项目提取了药物的二维分子结构信息、药物分子子结构指纹特征、已知的药物副作用信息和已知的药物靶标关联信息。提取靶标蛋白特征方面：提取了其字符串序列信息、蛋白质相互作用网络信息、蛋白质功能信息（基因本体，GO）和已知的靶标药物关联信息。多核融合方法方面：使用了快速核学习算法、基于希尔伯特-施密特独立性准则的多核学习算法、多视角标签传播算法和多核三重矩阵分解算法来分别融合药物空间的信息和靶标空间的信息。从多核学习的结果上看，药物的二维分子结构信息、已知的药物靶标关联信息和蛋白质相互作用网络信息对于药物-靶标相互作用的预测起到重要促进作用。对于生化实验而言，药物的结构和靶标蛋白互作网络也是检测药物-靶标相互作用的重要指标，所以项目所构建的预测模型对生化实验有很好的支撑作用。本项目的模型也对新药进行了测试，并预测新的药物-靶标相互作用，且大量预测出的药物-靶标对在新发布的数据库中得到了验证，从侧面证实了本项目研究方法的有效性。