Research On Prediction Method Of Multi-category MiRNA-disease Association

Studies have shown that the dysfunction of miRNAs may be related to a variety of complex human diseases and the dysregulations of miRNAs cause complex diseases through various kinds of underlying mechanisms(such as genetics,epigenetics,etc.).However,human beings are still not clear about these mechanisms.It is necessary for us to conduct a deeper study on the underlying mechanism of the dysregulations of miRNAs to find the miRNA-disease pairs that are related and determine the type of the miRNA-disease association,which can better reveal the relationship between miRNA expression and function to help people further understand the pathogenesis of complex diseases.Compared with traditional biological wet experiments which is time-consuming and expensive,calculation-based methods to predict multi-type relationships between miRNAs and diseases are both time-saving and cost-effective that are highly desired for us.And it can better assist biomedical workers in screening purposefully,thus it has important research significance.This paper constructs a heterogeneous network of multi-type miRNA-disease association based on the latest HMDD v3.2 database,and presents a novel data-driven end-to-end learning-based framework of neural multi-category miRNA-disease association prediction(NMCMDA)for predicting multi-type miRNA-disease associations.The NMCMDA mainly includes two components:(1)Encoder operates directly on the miRNA-disease heterogeneous network and leverages graph neural network to learn miRNA and disease embedding in the latent space respectively.(2)Decoder yields miRNA-disease association scores under each association type with the learned latent representations from the encoder as input.Various kinds of encoders and decoders are proposed for NMCMDA.And we have verified through the experiments that the NMCMDA with the encoder of Relational Graph Convolutional Network and the Neural Multi-Relational decoder(NMR-RGCN)achieves the best prediction performance,because NMR-RGCN can better learn and capture the complex and nonlinear interactions between miRNAs and diseases.And then,we compared the method proposed in this paper with other baselines through experiments on three datasets.The experimental results show that the NMR-RGCN is significantly superior to other state-of-the-art methods in terms of Top-1 precision,Top-1 Recall,and Top-1 F1 Score.In order to further prove the effectiveness of the proposed method,this paper conducted case studies on two common complex diseases(breast cancer and lung cancer)based on the HMDD v2.0 database.Additionally,this paper also trained the model based on the HMDD v3.2 database and ranked the prediction scores of output to obtain the prediction results of the top-10 miRNA-disease-category associations.Finally,we verified seven prediction results via querying the literature.