MVSL-DSF: Multiview Subspace Representation Learning and Cross-Modal Feature Dynamic Aggregation for Enhanced Drug...

Drug side effects increase morbidity and mortality in the relevant medical fields. Assessing the frequency of drug side effects is crucial for drug development and risk-effect analysis. Most current research approaches focus on modeling heterogeneous network graphs of drug-side effect frequency relationships using a single view of the data based on graph neural network approaches. However, the scarcity of drug-side effect frequency interaction networks restricts the effective utilization of similar information between drugs and side effects, and existing methods cannot fully utilize integrated multivariate data features from various pharmacological networks, resulting in a limited performance. In this paper, we propose a novel multiview subspace representation learning aggregation method (MVSL-DSF), which maps three types of views: drug, side effect, and drug-side effect interaction views, into a shared low-dimensional space. This mapping preserves the unique features of each view while capturing their shared intrinsic structure, enabling better representation, differentiation, and consistent complementarity of features within this space. We treat multiview consistency as alignment representation and complementarity as information augmentation, achieving their joint optimization through Canberra-distance-driven cross-modal attention: high distances trigger low weights to suppress redundant features, while low distances assign high weights to preserve key features. This approach sparsely aggregates features while eliminating redundant information, thereby enhancing the model's discriminative capability. In baseline experiments, compared with the existing state-of-the-art methods, MVSL-DSF optimized the RMSE and MAE to 0.297 and 0.167, respectively. Multiview modeling enables the precise quantification of drug side effect frequency from multisource data, supporting risk assessment and clinical decisions to improve medication safety. Finally, we constructed a frequency-severity matrix to predict drug risk levels and enhance pharmacovigilance assessment.