Algorithms For Identifying Potential Associations Between Non-coding RNA And Disease/Drug Sensitivity

Existing researches indicate that non-coding RNA are involved in the regulation of cellular processes such as proliferation,differentiation,metastasis,and apoptosis.Therefore,the investigation of predicting associations between ncRNAs and diseases/drug sensitivity will help researchers deeply understand the pathogenesis of complex diseases from the ncRNA level and offer new means for disease prevention,diagnosis,treatment and prognosis,and it is of great significance for drug discovery and precision medicine.In recent years,researchers have proposed many algorithms for predicting ncRNA-disease or ncRNA-drug sensitivity associations for further biological experimental verification.However,there is room for further performance improvement about the existing algorithms.It has become one of the hot topics how to applying deep learning to predict associatios between ncRNA and disease/drug sensitivity to obtain more accuracy prediction results.The main contributions of this dissertation are summarized as follows.(1)For addressing the problem of miRNA-disease association prediction,a prediction algorithm using deep collaborative called DCFMDA is proposed.The proposed algorithm combines the advantages of deep learning and matrix factorization.The proposed prediction model includes the non-negative matrix factorization sub-model based on deep neural network(NNMF)and the sub-model based on multi-layer perceptron(MLP).The MLP sub-model learns miRNA and disease similarity features respectively.The NNMF sub-model learns miRNA and disease association features respectively.The outputs of the two sub-models are merged to obtain prediction scores of miRNA-disease pairs.The experimental results show that compared with other existing methods,the proposed method DCFMDA can obtain significant improvement in ROC,AUC,PR and F1-score evaluation metrics,and the case study further verifies that the method DCFMDA can effectively predict the miRNAs associated with a given disease.(2)For addressing the problem of lncRNA-disease association prediction,a lncRNA-disease association prediction algorithm called g GATLDA based on graph?level graph attention network is proposed.The algorithm g GATLDA extracts the enclosing subgraphs of lncRNA-disease pairs according to lncRNA-disease association,lncRNA similarity,and disease similarity,and use them to train the prediction model based on multi-layer graph attention network,and predict lncRNA-disease association.In addition,an effect method to calculate the disease similarity based on gene-gene interaction network is proposed.The experimental results show that our method g GATLDA is in general better than other existing methods in terms of AUC,AUPR,F1-score,accuracy,and recall,and the case study further verifies that the algorithm g GATLDA can effectively predict the lncRNAs associated with a given disease.(3)For addressing the problem of drug sensitivity prediction model how to accurately apply in clinic,a lncRNA-drug sensitivity association prediction algorithm called SSTL-Lnc DR using semi-supervised transfer learning is proposed.The algorithm SSTL-Lnc DR consists of three sub-networks,namely the feature extractor,the prediction sub-network,and the adversavial learning sub-network.By transfer learning,the proposed algorithm can migrate the features learned from cell line to clinical prediction task to overcome rarely clinical patient sample data.By adversavial learning,the proposed method can further improve feature extractor for obtaining common feature space shared by cell line and cancer sample.By semi-supervised learning,the proposed method fully uses the labeled and unlabeled cell lines and cancer sample data such that the feature extractor can learn features that more inclined to the clinical target domain.The experimental results show that the proposed prediction algorithm based on is superior to other existing algorithms in AUC and AUPR evaluation metrics.The work achievement of this dissertation will provide reference for further research of miRNA and lncRNA relatitive prediction problems,and the obtained prediction results will provide data range reference for next biological experiment verification.