Heterogeneous Information Network-based Learning For MiRNA-disease Association Prediction Study

Micro RNAs(miRNAs)are a class of small non-coding RNA molecules encoded by endogenous genes.They are approximately 22-24 nucleotides in length and have been shown to regulate post-transcriptional gene expression,thereby influencing physiological processes such as cell growth,differentiation,metabolism,and immune responses.Some studies have also revealed the involvement of miRNAs in various biological processes related to human diseases.Predicting potential miRNA-disease associations is beneficial for understanding human diseases,including disease prevention,diagnosis,and drug development.Traditional biological approaches have achieved certain success in predicting miRNA-disease associations,but they suffer from issues such as high cost and low efficiency.With the advancement of bioinformatics,several computational models have been proposed for miRNA-disease association prediction.However,some challenges remain.Firstly,many methods rely on similarity networks between miRNAs and diseases,neglecting the complex relationships among miRNAs,diseases,and genes.Genes play a crucial role in connecting miRNAs and diseases,as gene dysfunction can lead to diseases.Secondly,some computational approaches based on heterogeneous information networks require the manual definition of meta-paths,which is subjective and dependent on specific domain knowledge.To address these challenges,this study proposes two computational models for miRNAdisease association prediction,utilizing a heterogeneous information network comprising three types of nodes(miRNAs,diseases,and genes)and their corresponding relationships,as well as the biological features of miRNAs and diseases.The HMM-MDA model is developed based on multi-channel meta-path graph learning in the heterogeneous information network.By assigning attention weights to each relationship in the network,the model scores multi-hop connections between different relationships,ultimately constructing a trainable meta-path graph.Multiple channels are generated using a multi-channel mechanism.A simplified graph convolution is deployed on the generated meta-path graph to integrate the biological features of miRNAs and diseases with the embedded features obtained from the meta-paths.The fused features from all channels are then utilized for matrix completion to predict miRNA-disease associations.The model automatically extracts useful meta-paths through a learnable attention mechanism instead of relying on predefined domain knowledge.The HGA-MDA model is proposed based on generative adversarial learning in the heterogeneous information network.It employs a generative embedding approach,where a relationship-aware generator learns the real node distribution and samples latent nodes from a continuous distribution,while a relationship-aware discriminator distinguishes between true and fake pairs of nodes.Through iterative training of the generator and discriminator,the model learns node embedding features from the heterogeneous information network without relying on meta-paths.Finally,the biological features of miRNAs and diseases are fused with the embedded features of nodes to obtain the final feature representations,and a random forest classifier is employed to predict potential miRNA-disease associations.The performance of the proposed models is evaluated using five-fold cross-validation.Experimental results demonstrate that the HMM-MDA model achieves an average AUC of0.9302,and the HGA-MDA model achieves an average AUC of 0.9313,outperforming the compared methods in the same dataset.Furthermore,the two models are applied to predict miRNA associations related to lymphoma and gastric cancer to further validate their performance.