【作者】 张旭东；

【导师】 宋弢；

【作者基本信息】 中国石油大学(华东) ， 软件工程（专业学位）， 2023， 硕士

【摘要】 研发一款新药物的成本高达数亿至数十亿美元,平均研发周期在10年以上。药物重定位和联合用药作为突破传统药物研发流程的新方法,可以大幅度缩短药物研发周期。药物重定位的关键步骤是对药物-靶标相互作用进行预测,联合用药的核心技术是药物-药物相互作用预测。近年来,基于深度学习方法的药物-靶标相互作用预测和药物-药物相互作用预测方法表现突出,但仍然存在许多问题。因此,本文围绕基于深度学习方法的药物-靶标相互作用预测和药物-药物相互作用预测进行研究。为解决以往深度学习方法在小样本药物数据集上表现不佳的问题,提出了基于子结构和相似性编码的多尺度特征融合模型来预测药物-靶标相互作用,该模型包括全局相似性特征提取通道、局部子结构特征通道以及相互作用预测模块。实验表明,该模型在多个公开数据集,特别是在小样本数据集上表现优秀。在案例研究部分,将训练好的模型用于药物重定位,并对预测结果使用分子对接和文献挖掘进行验证,证明了新预测的药物-靶标相互作用具有一定实际意义和应用前景。为解决以往深度学习方法不能捕获药物和蛋白质序列的双向依赖关系以及模型难以训练的问题,提出了基于桥联的多尺度特征融合模型来预测药物-靶标相互作用。该模型包括药物编码器、蛋白质编码器以及相互作用预测模块,每个编码器包括使用桥联的BiLSTM和Transformer两个子模块。实验表明,该方法优于其它先进方法。在案例研究部分,使用训练好的模型为蛋白碳酸酐酶Ⅱ推荐了10种可能与其发生相互作用的候选药物,并通过分子对接和文献挖掘进行了验证。为解决以往深度学习方法仅使用药物序列结构信息或药物原子信息,忽略原子空间结构信息的问题,提出了基于专注分子图空间结构的轻量化自注意力模型来预测药物-药物相互作用。该模型包括基于权重共享的原子特征嵌入、基于权重共享的边特征嵌入、堆叠的稀疏自注意力编码器以及解码器。实验表明,该模型取得了优异的预测表现。在案例研究部分,使用训练好的模型预测了阿利吉仑、赛乐西帕和沃拉帕沙三种药物之间新的相互作用,并通过文献挖掘验证了模型的预测。更多还原

【Abstract】 The cost of developing a new drug is hundreds of millions to billions of dollars,and the average development period is more than 10 years.Drug repositioning and drug combination,as new approaches to break through the traditional drug development pipeline,can significantly accelerate the drug development period.The key step of drug repositioning is the drug-target interactions prediction,and the core technology of drug combination is the drug-drug interactions prediction.In recent years,drug-target interaction prediction and drug-drug interaction prediction methods based on deep learning methods have performed outstandingly.However,there are still many problems.This thesis focuses on drug-target interaction prediction and drug-drug interaction prediction based on deep learning.To address the poor performance of previous deep learning methods on small sample datasets,a multi-scale feature fusion model based on sub-structure and similarity encoding is proposed to predict drug-target interactions,which includes a global similarity feature extraction channel,a local sub-structure feature channel,and an interaction prediction module.Experiments show that the model performs well on several publicly available datasets,especially on small sample datasets.In the case study section,the trained model is used for drug repositioning and the prediction results are validated using molecular docking and literature mining,demonstrating that the newly predicted drug-target interactions have some practical significance and application prospects.To address the problems that previous deep learning methods cannot capture the bi-directional dependence of drug and protein sequences and that the models are difficult to train,a multi-scale feature fusion model based on bridging BiLSTM and Transformer is proposed to predict drug-target interactions.The model includes a drug encoder,a protein encoder,and an interaction prediction module,each of which includes two sub-modules of BiLSTM and Transformer using bridging.Experiments show that this method outperforms other state-of-the-art methods.In the case study section,the trained model is used to recommend 10 drug candidates for protein carbonic anhydrase Ⅱ that may interact with it,and is validated by molecular docking and literature mining.To solve the problem that previous deep learning methods only use drug sequence structure information or drug atom information and ignore atomic space structure information,a lightweight self-attention model focusing on molecular graph space structure is proposed to predict drug-drug interactions.The model includes a weight-sharing based atomic feature embedding,a weight-sharing based edge feature embedding,a stacked sparse self-attention encoder,and a decoder.Experiments show that the model achieves excellent prediction performance.In the case study section,the trained model is used to predict new interactions between three drugs,Aliskiren,Selexipag,and Vorapaxar,and the predictions of the model are validated by literature mining.更多还原