KMGTMDA: KAN-driven multiscale graph neural network and context-enhanced prediction for human microbe-disease...

Microorganisms play an irreplaceable role in critical physiological processes such as human metabolic regulation, cancer inhibition, and pathogen antagonism, yet traditional culture-based microbial experiments are expensive and time-consuming. Consequently, there is an urgent need to develop efficient computational prediction models to prioritize disease-related microbes, aiming to discover potential associations and reveal their underlying biological mechanisms.

A novel computational framework named KMGTMDA is proposed. First, a dynamic adjacency matrix is employed to dynamically integrate first-order adjacency and higher-order topological information, utilizing an attribute matrix generation network to adaptively adjust node connection weights, and incorporating multi-scale graph features through dual-pathway graph convolution. Secondly, considering the contextual relationships of microbial and disease node features, a graph Transformer is employed to capture global dependencies through a multi-head attention mechanism. In this model, graph structural bias and neighborhood token extraction are introduced to capture local semantic patterns. Moreover, a feature retention strategy is incorporated to complement detailed interaction representations from the feature perspective. Finally, the features are input into a Kolmogorov-Arnold Network (KAN), which constructs nonlinear mappings through learnable basis functions and piecewise splines to generate association scores.

The model was evaluated through five-fold cross-validation, achieving AUC and AUPR values of 0.9779 and 0.9786, respectively. Additionally, case studies confirmed that KMGTMDA is an effective tool for screening potential microbes associated with periodontal disease and bacterial vaginosis.

This approach efficiently reveals potential associations between human microbes and diseases through multi-scale dynamic topological graph neural networks and feature context learning. It provides more effective guidance for interpreting disease pathogenesis and designing targeted treatment strategies, and offers new directions for further research on potential disease targets.