Research On CircRNA-Disease Association Prediction Based On Graph Relations

Circular RNAs(circRNAs)is closely related to a series of physiological and pathological processes,and it generally have low expression levels and relatively stable structures compared with messenger RNAs(m RNAs).Many research have shown that circRNAs can regulate biological process of malignant tumors in a variety of ways.Since it is costly and timeconsuming to infer the associations between circRNAs and diseases via traditional experiments,computational approaches are become a necessary research direction in this area.Although great progress is made by intelligent modeling methods,there are still many challenges,two of which are as follows:(1)existing intelligent methods need to be improved in feature learning,and more targeted feature learning skills need to be introduced.(2)In the process of feature extraction and construction,the encoding of circRNA-disease cooperative signals which are generally hidden in the circRNA-disease networks are ignored in the existing methods.This paper provides a comprehensive study and proposes two new strategies to address the two important challenges described above,an online service system for the association prediction was also developed for use.The following are the two main work of this paper to predict circRNA-disease associations.For the first challenge,we propose a multi-view dual attention embedding model with cooperative ensemble learning method called MDGF-MCEC for circRNA-disease association prediction.In the existing intelligent methods,single data of circRNA and disease are usually used as features,so it is important to introduce more comprehensive feature learning skills.In terms of the classification,the utilization of different views is not sufficient,so there is a need of more comprehensive classification methods.In MDGF-MCEC,the method first obtains different disease relation graphs and different circRNA relation graphs of multi-views with similarity computational algorithms.Then,the relation graphs are fed into graph convolution neural network for feature learning and embedding.In order to obtain high utility features,a dual-attention mechanism is introduced in this progress,which adjusts the weight at the channel level and the spatial level to balance the contribution of different features.Finally,based on the embedding features of disease and circRNA acquired by feature learning,a multi-view cooperative ensemble classifier was proposed to predict the associations of circRNA and disease by the use of multi-view features and ensemble learning skills.Our experimental study shows that the proposed new method MDGF-MCEC prediction performance is greatly improved compared with existing methods.For the second challenge,we propose a multi-layer cooperative attention neural graph collaborative filtering method MLNGCF,the method find a new way to emphasize the importance of mining circRNA-disease central network’s value and comprehensively explore the key collaborative signals between circRNAs and diseases by addressing issues that insufficient consideration of the interactions between circRNAs and diseases.For feature construction,multi-similarity information of circRNAs and diseases constructed from different views are fused to generate unified features of circRNAs and diseases by algorithm and deep auto-encoder.In the further feature learning,the collaborative messages between different network layers were propogated by establishing a central network about diseases and circRNAs based on the circRNA-disease adjacency matrix.This process injects collaborative signals between different layers into circRNA and disease features effectively,and the process can be divided into message propagation and message aggregation.Besides that,a multi-layer cooperative attention mechanism was designed to obtain the contribution values of different nodes in the central network to optimize feature construction process.Finally,a collaborative filtering is used to make association prediction by the use of generated circRNA and disease features with higher-order connectivity to predict unknown circRNA-disease associations.The experimental results shows that by making full use of the circRNA-disease central network,the proposed MLNGCF method achieves better results than other comparative methods in performance tests.