AFP-GFuse: an antifungal peptide identification model with structural information fusion via multi-graph neural...

Antifungal peptides (AFPs) are natural defense molecules that inhibit fungal pathogens, protecting against external fungal invasion. Their mechanism of action can effectively combat fungal resistance, offering broad-spectrum efficacy, high safety, and other advantages. However, traditional laboratory methods for identifying AFPs are inefficient and expensive. Consequently, with the development of artificial intelligence, computational models for identifying and predicting AFPs have emerged. But existing methods often rely on datasets compiled from literature and inadequately consider AFP representation, such as ignoring spatial features. Furthermore, single Graph neural Networks (GNNs) can suffer from feature bias in capturing features. To this end, this study constructed a state-of-the-art and comprehensive dataset and developed a deep learning model, AFP-GFuse, that integrates sequence and structural information and three complementary GNNs. A hierarchical cross-attention mechanism is designed to dynamically align and fuse multi-graph feature representations. Experiments demonstrate that AFP-GFuse outperforms state-of-the-art models in predicting AFPs, achieving an accuracy of 0.9140. Ablation experiments further validate the effectiveness of structural representation. Furthermore, comparisons with three individual GNNs demonstrate that integrating a cross-attention mechanism effectively complements their representational limitations and improves overall model performance. To facilitate broader application, we also provide open data 8.4and an online service, AFP-GFuse, publicly available at http://www.bioai-lab.com/AFP-GFuse .