| Background:RNA interactions refer to the interaction between RNA molecules or between RNA molecules and the external environment.which include RNA-RNA interactions(RRIs),RNA-subcellular localization interactions(RLIs),RNA-disease interactions(RDIs)and so on.RNA interactions are involved in RNA modification,promoting or inhibiting protein translation,and other biological functions,and thus play an important role in a variety of pathological and physiological processes in life activities.Object:In recent years,RNA interactions have attracted widespread attention.Researchers have collected RNA interaction data that has been experimentally validated or predicted based on sequence and structural features of miRNA.As a result,RNA interaction-related databases have been constructed.However,existing text mining tools in the field of RNA interactions are limited to identifying named entities such as RNA,protein,or disease,and recognizing protein-protein interactions(PPIs)or RRIs.They still lack a bioinformatics computational method that can predict RRIs,RLIs,and RDIs simultaneously.Therefore,this study aims to develop an efficient and accurate text mining tool to predict these three types of RNA interactions from medical literature.Methods:Firstly,based on about 60,000 RNA interaction statements collected from databases such as RNAInter,RNALocate,and RNADisease,we built three positive RNA interaction corpora after manual screening and sentence standardization,including 23,803 RRIs sentences,1,338 RLIs sentences,and 16,873 RDIs sentences.Additionally,we randomly obtained an equal number of negative statements on PubMed which,together with the positive corpora,formed a benchmark dataset.Secondly,based on more than two million entities collected from seventeen RNA databases,three subcellular localization databases,and six disease databases,and processed by entity standardization and deduplication,we built three named entity libraries,namely RNA named entity library,subcellular localization named entity library,and disease named entity library containing 1,929,810,88,039,and 9,333 entities,respectively.Thirdly,PubMedBERT is a transfer learning model based on the BERT model specifically designed for the biomedical field.Based on the PubMedBERT model,we developed a text mining tool that integrates PubMedBERTBiLSTM-CRF for three types of entity recognition and TextCNN for multi-label classification around positive and negative corpora of RRIs,RLIs,and RDIs.To evaluate the performance of the tool,we used a five-fold cross-validation method.Results:In the named entity recognition and multi-label classification tasks on the benchmark dataset,our method demonstrated certain advantages in multiple performance evaluation indicators compared with other methods.Furthermore,on an independent dataset from BioBERT,the tool extracted most of the valid interaction information,demonstrating better accuracy and robustness than other algorithms.In summary,based on the current lack of effective methods for predicting RRIs,RLIs and RDIs,this study developed a high-performance and high-accuracy text mining tool.The tool can identify named entities such as RNA,diseases,and subcellular localization,and determine whether there are interactions between these entities and the type of interaction.In the future,this tool will provide certain assistance for RNA interaction group research,helping researchers focus on exploring the mysteries of RNA interaction-related functional mechanisms. |
