Optimizing transformer-based prediction of human microbe–disease associations through integrated loss strategies

Microorganisms play an important role in many complex diseases, influencing their onset, progression, and potential treatment outcomes. Exploring the associations between microbes and human diseases can deepen our understanding of disease mechanisms and assist in improving diagnosis and therapy. However, traditional biological experiments used to uncover such relationships often demand substantial time and resources. In response to these limitations, computational methods have gained traction as more practical tools for predicting microbe-disease associations. Despite their growing use, many of these models still face challenges in terms of accuracy, stability, and adaptability to noisy or sparse data. To overcome the aforementioned limitations, we propose a novel predictive framework, HyperGraph Neural Network with Transformer for Microbe-Disease Associations (HGNNTMDA), designed to infer potential associations between human microbes and diseases. The framework begins by integrating microbe-disease association data with similarity-based features to construct node representations. Two graph construction strategies are employed: a K-nearest neighbor (KNN)-based adjacency matrix to build a standard graph, and a K-means clustering approach that groups similar nodes into clusters, which serve as hyperedges to define the incidence matrix of a hypergraph. Separate hypergraph neural networks (HGNNs) are then applied to microbe and disease graphs to extract structured node-level features. An attention mechanism (AM) is subsequently introduced to emphasize informative signals, followed by a Transformer module to capture contextual dependencies and enhance global feature representation. A fully connected layer then projects these features into a unified space, where association scores between microbes and diseases are computed. For model optimization, we propose a hybrid loss strategy combining contrastive loss and Huber loss. The contrastive loss aids in learning discriminative embeddings, while the Huber loss enhances robustness against outliers and improves predictive stability. The effectiveness of HGNNTMDA is validated on two benchmark datasets-HMDAD and Disbiome-using five-fold cross-validation (5CV). Our model achieves an AUC of 0.9976 on HMDAD and 0.9423 on Disbiome, outperforming six existing state-of-the-art methods. Further case studies confirm its practical value in discovering novel microbe-disease associations.