FusionPHI: A phage-host interaction prediction network model based on attention-driven multi-modal feature fusion

Phage therapy has become an important strategy against the crisis of antibiotic resistance for its potential to specifically target pathogenic bacteria. However, the narrow host range of phage makes the screening of precise matches of clinical strains inefficient, while existing computational tools are difficult to capture the dynamics of phage-host interactions due to their reliance on single modal features (genome or proteins). In this paper, we propose FusionPHI, a phage-host interaction prediction model that adopts a fully connected neural network architecture and takes multi-modal features as input, including the k-mer statistics of genome sequences, the physicochemical properties of proteins, and the embedded representations of evolutionarily conserved gene motifs for a comprehensive understanding of phage-host interactions. Moreover, FusionPHI contains a dual-stage attention-drive feature fusion module that integrates a self-attention mechanism to optimize the correlation among the features within a single modality followed by a cross-attention mechanism to dynamically fuse genetic distribution patterns with the protein function information in global or local regions of sequences. The experiments of phage-host interaction prediction show that FusionPHI achieves 91% ROC AUC in cross-validation, demonstrating competitive performance compared to the evaluated baseline methods on our dataset, and ablation experiments further validate the necessity of multi-modal features and attention mechanism. The case study of E. coli infected by the M13K07 phage further validate the prediction ability of the proposed FusionPHI model.