面向生物医学文献的药物重定位隐含知识发现方法研究

新药研发的周期漫长、耗资巨大，且成功率低。从生物医学文献中有效发现药物重定位隐含知识能够大幅度减少新药研发的周期和成本，对于提高新药研发的效率和水平具有重要的理论价值和现实意义。生物医学文献中语言表达的多样性、隐含性和复杂性给药物重定位隐含知识发现带来了极大挑战。准确性低和解释性不足是制约药物重定位隐含知识发现走向实际应用的瓶颈问题。本课题采用对抗学习和多任务学习策略，研究提升药物重定位实体识别模型泛化性能的方法；基于跨句句法关系和领域知识，探索建立适应于复杂语境的药物重定位跨句实体关系抽取模型；引入元路径理论和注意力机制，有效学习不同关联路径模式的作用，提高隐含知识发现的推理能力和可解释性。本课题预期将建立高效、准确、解释性强的药物重定位隐含知识发现方法，为新药研发有效提供实验线索和实验假设。

Drug discovery has grown to be time-consuming and costly, which has a low success rate. Drug repositioning knowledge discovery from biomedical literature can substantially reduce the cycle and cost of drug discovery, which will have profound guiding significance for our pharmaceutical companies to improve the efficiency and level of drug discovery. The diversity, implication and complexity of natural language pose huge challenges to drug repositioning knowledge discovery. The low accuracy and interpretation are two key problems that need to be solved urgently in the drug repositioning knowledge discovery, which constraint the practicality of drug repositioning knowledge discovery. Firstly, we adopt the adversarial learning and multi-task learning and study how to improve the generalization ability of named entity recognition model. Secondly, we will explore to extract the cross-sentence relations based on the vital cross-sentence syntactic feature and domain knowledge. Thirdly, we induce the meta-path theory and attention mechanism to learn the value of the different path pattern, which improve the inferential capability and interpretability of the knowledge discovery model. This research will build the efficient, accuracy and interpretable models for drug repositioning knowledge discovery, and provide the experimental clues and hypotheses for drug discovery.