Research On Key Technologies Of Biomedical Entity-interaction Mining Based On Deep Model

In recent years,the rapid accumulation of biomedical big data has strongly promoted the vital transformation from hypothesis-driven to data-driven in biomedical research.Biomedical big data contains rich biomedical knowledge and includes many internal cor-relation rules between biomedical entities.It is an essential support for complex patho-logical analysis,epidemic disease prediction,new drug discovery,and clinical medica-tion decisions.Especially when public health emergencies happened,like the plague,Ebola virus,and COVID-19,biomedical knowledge can quickly respond to the needs of pathogen analysis and candidate drug screening,greatly accelerating the development process of new drugs,reducing the cost of research and development,and effectively controlling the expansion and deterioration of public health events.However,much med-ical information and knowledge are hidden in a large number of unstructured data carriers.How to quickly transform unstructured biomedical big data into semi-structured and struc-tured knowledge is a research hotspot in the biomedical field.As a practical and primary way to represent knowledge,the entity relation extraction is the core task in information extraction.The interaction information between biomedical entities can reveal the com-plex interaction mechanism between biomolecules,which plays a vital role in promoting the development of life science.In this paper,we take advantage of the deep learning tech-nology on automatically extracting relations between biomedical entities from biomedical literature and biological molecular network,providing decision-making suggestions for drug use and new drug research and development.The main research contents of this paper include the following aspects:Firstly,in view of the adverse drug reactions in the process of taking drugs,we pro-pose a dependency-based bi-directional long short-term memory network model for drug-drug extraction.According to the description of text sentences,our model can automati-cally recognize whether there is interaction between two candidate drugs in an instance,or identify the interaction types between them.In terms of the complex structure of the biomedical text,our model integrates the dependency relationship between words into the long short-term memory network.There are three channels.The first one is Linear chan-nel.The second one is DFS channel.The third one is BFS channel.They are constructed with three network layers,including embedding layer,LSTM layer and max pooling layer.There are two types of features in the embedding layer.one is distance-based feature and another is dependency-based feature.Lastly,the outputs of all channels are concatenated together.Then it is linked to the softmax layer for relation identification.Besides,the experimental results show that our model outperforms the baseline models and excels at balancing the Precision and Recall values.Secondly,considering the interpretability of the output,we propose a model based on attention mechanism for drug-drug extraction.Our model is able to compute the im-portance scores in the neural network,which can be utilized to measure the contributions of different words while the model makes decision.our model contains three layers: the Bi-LSTM layer,the Attention layer and the Dense layer.In the Bi-LSTM layer,our model user two long short-term memory network to learn the context information of words.In the Attention layer,the model uses a multi-head attention mechanism to learn the weight distribution of all features in different semantic spaces to improve the ability of the neu-ral network to enhance key features or suppress useless features.In the dense layer,we design a flag to capture global information across the whole sentence.The experimental results show that our model has achieved great performance in drug-drug interaction ex-traction.Besides,the weight distribution calculated by our model can effectively identify the critical features in the learning process,which is of great significance to explain the reliability of the results.Thirdly,in terms of drug target recognition and drug redirection,a heterogeneous graph containing chemical nodes,gene nodes,and pathway nodes is constructed.We present a chemical-gene interaction extraction model based on graph convolutional net-work.Our model uses a framework of encoder-decoder and formulates the chemical-gene interaction identification problem as a task of multi-relational link prediction.The encoder aggregates,transforms,and propagates neighborhood information over the graph with the graph convolutional network.The node embeddings are learned with two different per-spectives.The first perspective process the graph as a whole.The second one adopts a subgraph view that the node embeddings are learned by two steps.The first step is to learn the initial node embeddings with the binary association.Then the initial node embeddings are transferred to the multi-interaction subgraph for final node embeddings learning.In the subgraph view,we reconstruct the topology of the biological molecular network with the latent links based on the inference hypothesis.The experimental results show that our model perform better on the chemical-gene interaction prediction.Moreover,the model which adopts a subgraph perspective can show better performance and reduce the training time.