Research On Circrna-Drug Sensitivity Association Prediction Algorithm Based On Graph Neural Network

Identifying circRNA-drug susceptibility associations(CDAs)is beneficial for drug discovery as a growing body of research indicates that circRNAs play an important role in tumourigenesis and therapeutic resistance and that their expression also affects cellular sensitivity to drugs,thereby significantly influencing drug efficacy.Validating these associations through traditional biological experiments is time-consuming and costly.Therefore,designing effective algorithms to predict unknown circRNA-drug sensitivity associations is an important and urgent task.Graph neural networks and multi-kernel learning-related algorithms have shown excellent performance in association prediction problems.However,how to extract effective feature information and insufficient sources of sample information are challenges in the current circRNA-drug sensitivity association prediction problem.To address these problems,this thesis proposes two algorithms for predicting potential circRNA-drug sensitivity associations based on graph neural networks,as follows:(1)Prediction algorithm MNGACDA for learning multimodal networks based on graph auto-encoder and attention mechanism.the algorithm first uses multiple information sources from circRNA and drug to construct a multi-modal network,and then uses node-level attention graph auto-encoder to obtain low-dimensional embeddings of circRNA and drug in the multimodal network.In addition convolutional neural networks were used to integrate the embedding representations of each layer and finally,an inner product decoder was used to predict the association scores between circRNA and drug sensitivity based on the embedding representations of circRNA and drug.The code for the MNGACDA implementation is publicly available at https://github.com/youngbo9i/MNGACDA.(2)Prediction algorithm MIMKCDA via fusion of multiorder neighbourhood infor mation and multi-kernel learning.the algorithm first integrates known circRNA-drug se nsitivity associations,circRNA similarity,and drug similarity into a heterogeneous netw ork,and then uses improved graph convolutional neural networks to fuse the multiord er neighbourhood information of the circRNA-drug heterogeneous network to obtain th e circRNA and drug embedding representations,next calculate the GIP(Gaussian inter action profile)kernel similarity matrices for each layer of circRNA and drug embeddin g,and fuse these kernel similarity matrices using multichannel attention,and finally,u se Dual Laplace regularized least squares to predict potential circRNA-drug sensitivity associations.The code for the implementation of MIMKCDA is publicly available at h ttps://github.com/youngbo9i/MIMKCDA.Finally,the algorithm performance is evaluated via comparative experiments and case studies.The results of the comparative experiments validate that the two algorithms proposed in this thesis are more effective compared to other advanced prediction algorithms.Also,the case studies show that the proposed algorithms are able to predict unknown circRNA-drug sensitivity associations.The results of both the comparative experiments and the case studies demonstrated the effectiveness of the proposed algorithm...