Allergenic protein classification with cross-modal fusion of physicochemical and structural features

Accurate identification of allergenic proteins is critical for allergic mechanism research, clinical allergy diagnosis and targeted immunotherapy. Current prediction methods suffer from three critical drawbacks: insufficient modeling of residue spatial structures, shallow cross-modal feature fusion without bidirectional interaction, and poor generalization ability on low-homology allergens with rare conformations. To tackle these challenges, this paper constructs a dual-branch multimodal fusion model combining Bidirectional Long Short-Term Memory (Bi-LSTM) and Graph Attention Network (GAT). The Bi-LSTM branch captures sequential contextual correlations of protein physicochemical features, while the GAT branch mines spatial interaction relationships between amino acid residues from protein graph structures. A four-head bidirectional cross-modal attention module is designed to realize deep dynamic information interaction between two modalities and fully exploit their complementary information. Furthermore, a self-supervised node feature reconstruction auxiliary task is integrated into the training process, paired with label-smoothed Focal Loss to form a dual-loss optimization framework, which enhances the discriminability of extracted features and improves model robustness on rare allergen samples. Experiments conducted on a balanced allergen dataset containing 10,000 protein sequences demonstrate that our model achieves state-of-the-art performance, with Accuracy and F1-Score both reaching 0.9980, outperforming 14 mainstream baseline models across all six evaluation metrics. Ablation experiments and t-SNE feature visualization further verify the synergistic effectiveness of each core module in the proposed architecture. The multimodal fusion framework proposed in this work provides an effective computational tool for large-scale high-precision allergen protein screening and offers a reference paradigm for other protein property prediction tasks in bioinformatics.