LncRNA-Disease Association Prediction Based on Integrated Application of Matrix Decomposition and Graph Contrastive Learning

Investigating the potential associations between long non-coding RNAs (lncRNAs) and diseases is crucial for advancing disease research and the development of therapeutic approaches. Nevertheless, the current lncRNA-disease association (LDA) data exhibits sparsity, hindering existing LDA prediction models from effectively capturing the features of lncRNAs and diseases. We introduce an LDA prediction model that integrates matrix decomposition and graph contrastive learning, MDGCLLDA, to address these challenges. Our approach accurately predicts lncRNA-disease associations (LDAs) by extracting features of lncRNAs, miRNAs, and embedded disease characteristics. We constructed a three-layer heterogeneous network encompassing lncRNAs, miRNAs, and diseases (LMDG), integrating their intricate interactions by examining similarities and associations. To capture comprehensive features of lncRNAs and diseases, we applied nonnegative matrix decomposition and singular value decomposition to the adjacency matrix of the heterogeneous network. Subsequently, an unsupervised embedding model enhanced local and global information exchange between nodes using a graph convolutional network encoder within graph contrastive learning. Finally, XGBoost was employed to predict LDA. The MDGCLLDA model demonstrated superior performance to the four benchmark models. Ablation experiments and case studies further validated the model’s reliability and effectiveness .