Enhancing lncRNA-Disease Association Prediction Through Collaborative Representation of Multi-Source Heterogeneous Features

Research has shown that exploring the relationship between long non-coding RNAs (lncRNAs) and diseases can contribute to understanding disease mechanisms and aid in developing new therapeutic strategies. However, existing methods for predicting lncRNA-disease associations still have limitations. They tend to rely on incomplete utilization of the extensive information present in heterogeneous networks, an overreliance on local structures, and a lack of consideration for global topological information. In order to address these issues, this paper puts forward a multimodal fusion model based on graph Transformer, matrix decomposition and automatic metapath generation. Firstly, a network of Transformer encoders is employed to encode the nonlinear features of nodes and to capture the complex interactions between different types of nodes. Secondly, a matrix decomposition method is employed to learn the linear embedding vectors of diseases and lncRNAs, thereby representing their potential features. In addition, we use a dynamic metapath generation algorithm, which adaptively generates and selects optimal metapaths based on the network structure, to capture the global topological characteristics of the nodes. Ultimately, the diverse node representations are consolidated and conveyed to the multilayer perceptron (MLP) for comprehensive learning, thereby deriving the final prediction scores. In 5-fold cross-validation experiments, our approach outperforms existing methods across multiple performance metrics for both datasets. Furthermore, case studies substantiate the efficacy of our method.