A Graph Transformer-Based Method for Predicting LncRNA-Disease Associations Using Matrix Factorization and Automatic Meta-Path Generation

LncRNAs are crucial regulators of gene expression that exert their influence on diverse cellular processes. Exploring the potential connections between lncRNAs and various pathological conditions holds significant promise for unraveling the intricate mechanisms that underlie disease onset and progression. Due to traditional biological experimentation for probing lncRNA-disease associations is often hampered by substantial financial constraints and prolonged timelines. Consequently, computational modeling and bioinformatics methodologies have emerged as efficient and cost-effective alternatives. This paper proposes a lncRNA-disease association prediction model that integrates graph transformers, matrix factorization, and automatic meta-path generation. First, a transformer encoder network encodes the nonlinear node features, capturing their deep semantic representations. Simultaneously, matrix factorization learns linear embedding vectors for diseases and lncRNAs, representing their latent features. Moreover, an automatic meta-path generation algorithm dynamically updates the graph structure and reconstructs node representations to capture their topological characteristics. Finally, these diverse node representations are merged and fed into a multilayer perceptron for comprehensive learning, yielding the final prediction scores. Under 5-fold cross-validation experiments, our approach outperforms other existing methods in several performance metrics on the dataset. Additionally, case studies further provide further evidence of the effectiveness of our approach.