MACGA: Multi-scale Adaptive Convolution with Graph Attention for LncRNA-Disease Association Prediction

Abstract

Accurate prediction of lncRNA-disease associations (LDAs) is crucial for understanding complex disease mechanisms and advancing precision medicine. Existing computational methods based on graph neural networks (GNNs) often suffer from over-smoothing issues, limiting their predictive accuracy. To address this challenge, we propose MACGA, a novel neural network framework integrating Multi-scale Adaptive Convolution with Graph Attention networks. MACGA effectively captures both local neighborhood information and global high-order structural features by combining PCA-based dimensionality reduction, local node feature extraction via graph attention mechanisms, and multi-scale feature aggregation through adaptive convolution. Extensive experiments using five-fold cross-validation demonstrated that MACGA significantly outperformed four representative state-of-the-art methods (AUC=0.9463, AUPR=0.4119). Case studies on prostate and colon cancers further validated the reliability of MACGA, with top-ranked lncRNA predictions well-supported by existing experimental literature. Moreover, we extended the proposed framework to predict disease-drug associations via lncRNA-bridging strategy, highlighting its potential for drug repositioning. Overall, MACGA provides a powerful computational tool for accurately predicting lncRNA-disease associations and exploring novel therapeutic opportunities. The source code and dataset used in this paper are available at: https://github.com/huifeideyu518/MACGA.