Knowledge Mining And Knowledge Discovering For Biomedical Text And Graph

With the rapid growth of bioinformatic research,a large amount of data and knowl-edge has been accumulated.It is effective to organize the various biomedical knowledge by constructing knowledge graph.Researchers can facilitate computation approaches on knowledge graph for information retrieval,data mining and knowledge discovering,which supports the theory of biology,pathology and pharmacology.The construction and application of biology knowledge graph needs several key steps,including knowl-edge retrieval,knowledge representation,knowledge integration and knowledge discov-ering.This article focuses on the two typical problems of these steps,which are the information retrieval of biomedical literature and the link prediction of gene-disease as-sociation network.We proposed specified machine learning model for these problems.The main works of this article include:(1)We proposed a novel combination strat-egy based on hybrid neural networks for biomedical event extraction from literature.A large amount of knowledge in unstructured formation is recorded in massive liter-ature,and event is an effective structure to describe these knowledge.To extract the information,this article utilized a hybrid neural networks for event extraction to elimi-nate the dependence on manual feature engineering,and used a combination strategy as post-process to alleviate error accumulation.The experiment results,conducted on the BioNLP shared tasks,showed that our model had achieved good performance.(2)We proposed a graph convolutional networks to predict gene-disease associations based on knowledge graph.There are many existing datasets that contain large number of associ-ations between genes and diseases,and latent knowledge can be mined from the knowl-edge graph integrated by these datasets.This article used graph convolutional networks to predict the unknown gene-disease associations.We described an adjacency matrix dropout technology and a new cluster loss function,which were applied to enhance the generalization ability.The experiments on the DisGeNet dataset indicated that the pro-posed method had achieved state-of-the-art performance among existing methods.(3)To handle the insufficient of annotated datasets for the literature mining and associa-tion prediction,we introduced a semi-supervised learning model based on self-training.Most of the supervised learning models in biomedical domain are hindered by the insuf-ficient annotated samples.This article applied the self-training method on biomedical text mining and gene-disease association prediction,which made use of annotated data and massive unannotated data.We selected samples from unannotated dataset by their indicator of reliability to augment training set.The contrast experiments demonstrated that self-training brought positive effect to original models.