CenterNet-based Methods For Pathway Entity Recognition And Relationship Determination

A biological pathway(Pathway)is a series of interactions between molecules in a cell that result in a change in certain cells or a certain product in the cell.Three relatively common types of biological pathways are involved in metabolism,regulation of gene expression,and signaling.Biological pathways play important roles at different levels,such as organ,cellular and genetic,to maintain the proper development of the organism.Therefore,the study of biological pathways has been the focus of the scientific community,and a large number of papers are published every year.These massive papers provide a great deal of information for researchers in related fields,especially the Pathway diagrams in the papers give the most direct information about Pathway.Therefore,it is of great practical importance to extract relevant information from these illustrations effectively.In this paper,the implementation of Pathway information extraction mainly relies on the object detection algorithm.Anchor-based object detection usually requires more hyperparameters for the anchor,and the training sample selection strategy has a great impact on the model,and the main representative algorithms are R-CNN series and SSD algorithms.Anchor-free object detection models are elegant and simple,and do not require pre-set size and scale of the anchor,but directly regress from the network and further generate boxes,so they can be relatively easy to extend to other vision tasks,of which the main representative algorithms are Corner Net and CenterNet algorithms.The first main work of this paper is to address the recognition of Pathway entities in the Pathway image dataset.For Pathway entity recognition,the detection model CenterNet-CA(CenterNet with Channel Attention)is proposed with the addition of an area loss to the loss function to further enhance the network detection performance.The final experiments show that the proposed model CenterNet-CA improves the detection metric m AP by 3.5 percentage points over the original CenterNet model in the Pathway dataset.The second main work of this paper is to propose a novel relationship determination model,Pathway-RDM(Pathway Relationship Determination Model),to extract the relationships among genes in Pathway.Since spatial information is very important in relationship determination,the relationship determination model introduces the convolutional module CBAM,which considers both channel attention and spatial attention,to better capture the spatial information of genes in the Pathway.In the relationship determination model,the CBAM module first processes the output features from the CenterNet-CA model,and the embedding vectors representing each location in the input features are obtained by further convolutional processing,then the embedding vectors of the corresponding entities are stacked with the detection results of the CenterNet-CA model,and finally the stacked vectors are fed into the fully connected network to complete the relationship determination.The final experimental results verify the feasibility of the model.To summarize,this paper mainly does two aspects.Firstly,based on the CenterNet model,by adding the channel attention mechanism,this paper proposes the CenterNet-CA model for Pathway image entity recognition,and adds area loss in the training process to further improve the detection effect.On top of this,a novel relationship determination model Pathway-RDM is proposed for the need of extracting activation and inhibition relationships between genes.The results of this work can help scholars in related fields to quickly obtain biological channel-related information,which is of good practical significance.