A multi-view contrastive learning framework with information bottleneck for metabolite-disease association prediction
TL;DR
Experimental results demonstrate that MVCL-IB consistently outperforms competing methods across multiple evaluation metrics, and suggest that MVCL-IB provides an effective framework for prioritizing potential metabolite-disease associations.
Abstract
Metabolite-disease association prediction is important for understanding disease mechanisms and identifying biomarkers. However, most existing methods do not effectively integrate similarity-derived structural information with cross-type interaction information, which limits the quality of learned representations. To address this issue, we propose MVCL-IB, a multi-view contrastive learning framework with an information bottleneck for metabolite-disease association prediction. Specifically, a graph attention network is employed to encode similarity networks and capture view-specific structural patterns, while a heterogeneous graph transformer is used to model metabolite-disease interactions and learn cross-entity dependencies. Furthermore, a view-level graph information bottleneck is introduced to suppress redundant information within each view, and cross-view contrastive learning is incorporated to enhance representation consistency and complementarity across views. Experimental results on the HMDB-based dataset under five-fold cross-validation demonstrate that MVCL-IB consistently outperforms competing methods across multiple evaluation metrics. These results suggest that MVCL-IB provides an effective framework for prioritizing potential metabolite-disease associations.