Decoding potential lncRNA and disease associations through graph representation learning and gradient boosting with histogram

Long noncoding RNAs (lncRNAs) are important regulators and promising targets for complex diseases. They have manifested dense relationships with various diseases. Although laboratory techniques have validated many lncRNA-disease associations (LDAs), they are costly, laborious, and time-consuming. This study introduces LDA-GMCB, an LDA inference model, by leveraging graph embedding learning, multi-head self-attention mechanism (MSA) with convolutional neural network (CNN), low-rank singular value decomposition (SVD), and histogram-based gradient boosting (HGBoost). For all lncRNAs and diseases, LDA-GMCB first deciphers their nonlinear features by incorporating graph embedding learning and MSA with CNN, then captures their linear features through low-rank SVD, and finally infers their relationships based on HGBoost. LDA-GMCB was compared with four baselines (i.e., SDLDA, LDNFSGB, IPCARF and LDA-VGHB) under 5-fold cross validation and two cold start scenarios, and four popular classifiers (i.e., multi-layer perceptron, SVM, random forest, and XGBoost). Additionally, LDA-GMCB implemented ablation study. The outcomes demonstrated that LDA-GMCB greatly surpassed the above models and gained significant improvement on two public databases (i.e., lncRNADisease and MNDR) under most conditions. Moreover, LDA-GMCB was further applied to infer potential lncRNAs for Alzheimer's disease and Parkinson's disease. It identified that DGCR5 and HIF1A could link with the two diseases, respectively. We hope that LDA-GMCB help infer potential lncRNAs for various complex diseases. LDA-GMCB is freely available at https://github.com/smiling199/LDA-GMCB .