Gene bridge metabolite-disease association prediction based on GCN-DeepSeekMoE contrastive learning and its applications in oral diseases.

The variation of metabolite levels is strongly associated with the occurrence as well as development in diseases, especially oral diseases such as periodontitis and celiac disease, where metabolic dysregulation often mediates pathological processes by interacting with genetic factors. Thus, metabolite-disease association prediction, particularly for oral disease-related metabolic biomarkers, represents a crucial challenge in biomedical and dental research. Conventional approaches often fail to account for the regulatory influence of intermediate biological entities (e.g., genes) and face challenges related to inadequate long-range dependency modeling in complex oral-related biological networks. To address this, we propose a novel model, GCM-KAN, which achieves deep feature extraction through the dual-path collaborative contrastive learning of Graph Convolutional Network (GCN) and the DeepSeek Mixture-of-Experts (MoE) system, and conducts nonlinear decoding prediction based on an improved Kolmogorov-Arnold Network (KAN). Firstly, we integrated similarities to construct a tripartite heterogeneous network based on the metabolite-gene-disease multi-faceted hetero-relations. Secondly, we designed a gated residual GCN to capture local topological features while introducing MoE to dynamically select different experts to handle the global dependencies in heterogeneous graphs, alleviating the over-smoothing problem commonly seen in traditional message passing through conditional computation. Furthermore, cross-modal contrastive learning is employed to align the feature representations of the GCN and MoE paths. Then, multi-perspective features are integrated through cross-attention. Finally, the KAN decoder enhances nonlinear modeling capabilities, yielding the association probabilities of metabolites and diseases. Results from 5-fold cross-validation show that GCM-KAN achieves an AUC of 0.9880 and an AUPR of 0.9880, outperforming six existing state-of-the-art prediction methods, validating its superiority in complex biological networks. In-depth validations on representative oral diseases (celiac disease and periodontitis) further confirm that GCM-KAN can effectively identify potential metabolic biomarkers corresponding to oral diseases, providing reliable computational tools for oral disease mechanism research.