Adaptive dynamic K-nearest neighbors and context-aware similarity optimization for microbe-disease association prediction

Abstract

Microbes play a crucial role in disease occurrence, progression, and treatment. Traditional experimental methods are time-consuming, prompting researchers to turn to computational models. However, existing models often suffer from limited data adaptability and improper feature selection, making them prone to noise interference. To address these limitations, we propose ADKNN-KFGCN, a novel adaptive framework that integrates dynamic K-nearest neighbors, graph convolutional networks, and context-aware similarity optimization. The model constructs multi-source similarity networks by integrating various similarity measures between microbes and diseases, forming a comprehensive foundation for association inference. To better capture complex local patterns, it employs adaptive dynamic K-nearest neighbors to adjust the number of neighbors based on local structure, enhancing the accuracy of network construction. This is followed by context-aware similarity optimization, which filters out low-similarity nodes to suppress noise and emphasize the most informative connections. On this refined graph, graph convolutional networks are used to extract high-level representations, effectively capturing intricate topological relationships. These features are then fused through kernel-based strategies, combining multiple similarity sources via averaging and weighted integration to form a unified representation. Finally, Laplacian Regularized Least Squares leverages the global graph structure during prediction, improving generalization and ensuring robust performance. Experimental results show that ADKNN-KFGCN outperforms seven state-of-the-art methods, achieving an AUC of 0.9851±0.0025 and AUPR of 0.9587±0.0032 on the HMDAD dataset. Case studies on rheumatoid arthritis and inflammatory bowel disease further demonstrate its potential to uncover novel associations, provide insights into disease mechanisms, and support therapeutic target discovery.