Progressive Orthogonal Multimodal Similarity Learning for Metabolite-Disease Association Prediction

The identification of potential associations between metabolites and diseases is crucial for understanding the onset and progression of diseases. Although numerous methods have been developed to predict metabolite-disease associations (MDAs), few methods fully mine the complementary information between different similarity relation or learn higher-order relationships between metabolites or diseases. To address these challenges, we propose a method for MDAs prediction with Progressive Orthogonal Multimodal Similarity Learning (POMSL). POMSL first constructs two hypergraphs within single-modal similarity with KNN and K-means to capture higher-order complex relationships. Then, hierarchical contrastive learning is applied to enhance the consistency of multi-similarity features by performing contrastive learning for inter-similarity and intra-similarity views. Next, a progressive orthogonal multimodal similarity integration strategy is developed to ensure the effective fusion of cross-similarity complementary information and to enhance the divergence of multi-similarity features. Finally, a bilinear decoder is used to predict MDAs. Extensive experimental results demonstrate that POMSL performs excellently in MDA prediction and, more importantly, validate the effectiveness of hierarchical contrastive learning and progressive orthogonal multimodal similarity learning. The source code and dataset are available at https://github.com/1521250466LYP/POMSL.