| With the advancement of bioinformatics,researchers have discovered that the abnormal expression of miRNAs is closely associated with the development of diseases.Nevertheless,predicting miRNA-disease associations(MDAs)by traditional wet experiments requires significant resources.Therefore,designing efficient computational models to accurately identify potential MDAs has become one of the hotspots of current research.Although the existing models have achieved good performance,most of them suffer from the single priori information and simple fusion methods.In this thesis,four space-oriented prediction models are proposed to identify disease-associated miRNAs from the perspectives of enriching prior information,adopting different fusion methods,and using different heterogeneous network construction methods.Meaningfully,the proposal of prediction models offers novel solutions for early screening and clinical treatment of diseases.Details of the research are as follows:(1)Considering the problem of single prior information in traditional prediction models,an ensemble learning model based on locality-constrained linear coding is proposed to predict MDAs(LLCELF).Firstly,interaction profile(IP)similarity is calculated by localityconstrained linear coding(LLC).Then,the functional similarity,sequence similarity,and IP similarity on the miRNA side,as well as semantic similarity and IP similarity on the disease side are fed into the ensemble learning to enrich the prior information.In addition,the higherorder relationships between samples are described by hypergraph learning.(2)To reasonably fuse multi-source similarity information and solve the cold-start problem in existing graph models,the bipartite graph diffusion model with fusion network distance analysis is proposed to predict MDAs(GDFND).Firstly,miRNA functional similarity,disease semantic similarity,and Hamming similarity,Jaccard similarity,cosine similarity in miRNA(disease)space are calculated.Next,multi-kernel fusion is employed to integrate multi-source similarity features.In addition,network distance analysis improves the reliability of the fusion network by adjusting the distance between data nodes.Ultimately,bipartite graph diffusion solves the cold start problem and uncovers potential MDAs.(3)In response to the problems of the noise in existing fusion methods and the sparse association matrix,a link prediction model with centered kernel alignment is proposed to predict MDAs(CKALP).Firstly,Gaussian Interaction Profile(GIP)kernel similarity and qkernel similarity are introduced to increase the topological information and to take the neighborhood information into account.Secondly,multiple similarities are fused by centered kernel alignment-based multiple kernel learning(CKA-MKL).Next,pre-processing is employed to reduce the sparsity of the known association matrix.Finally,a large-scale heterogeneous network is constructed and unknown links are predicted using link prediction method.(4)To address the problems of large and sparse heterogeneous network,a space projection model based on block matrix is proposed to predict MDAs(BMPMDA).Among them,block matrix is constructed in both sides of the space to improve the comprehensiveness of the prediction and reduce the size of the problem.In addition,matrix completion(MC)compensates for missing terms and reduces sparsity by using the low-rank information in the block matrix.Then,linear neighborhood similarity(LNS)is introduced to reconstruct the similarity network.Finally,more potential MDAs are obtained by network consistency projection(NCP).The four bipartite space-oriented models proposed in this thesis are used to identify potential associations between miRNA and disease,and achieve stable prediction performance.Compared with the existing prediction models,experimental results show that better prediction effects can be achieved by fully integrating multi-source biological information and reasonably constructing heterogeneous network. |
PDF Full Text Request