A dual-view fusion framework using residual gated graph convolution and Mamba for metabolite-disease association prediction with explorations on salivary metabolites.

Metabolites are important indicators of physiological and pathological states, and their alterations are closely associated with disease prediction. Therefore, prediction of metabolite-disease associations is valuable for disease research and biomarker discovery. Here, we propose DFFRGM, a dual-view fusion framework that integrates residual gated graph convolution and Mamba-based multi-hop dependency modeling for metabolite-disease association prediction. The framework jointly learns from homogeneous similarity networks and heterogeneous association networks, and then fuses the two views through bidirectional cross-attention. On two datasets, DFFRGM achieved AUC values of 98.54 and 98.89%, together with AUPR values of 98.61 and 98.91%. To further examine its applicability in a biofluid-specific scenario, we additionally constructed a salivary metabolite-disease dataset from the HMDB salivary metabolite resource. Validation on this dataset further supported the applicability of DFFRGM in saliva-oriented metabolite-disease association prediction.