CM-PHI: combining multi-hop attention graph neural network with sequence semantic analysis to predict phage-host interaction

Prediction of phage-host interactions (PHI) plays a crucial role in combating multi-resistant bacterial infections. The adsorption process mediated by phage tail proteins and host receptor-binding proteins represents the initial step of phage life cycle. However, traditional wet-lab experimental methods for identifying PHI are time-consuming, laborious-intensive and expensive. To address these challenges, we propose a novel deep learning framework named CM-PHI, which leverages a combination of multi-hop attention graph neural network (MHAGNN) and gated convolutional networks (GCNN) to predict PHI. Initially, we constructed a heterogeneous microbial network that centred on the proteins involved in the adsorption mechanisms. Then, the MHAGNN model is used to capture topology-level features, while GCNN encoded sequence-level information. These features are subsequently fused using a self-attention mechanism and trained through a dual-input fusion network (DIF-Net). Compared with various existing prediction methods, CM-PHI achieves superior accuracy and robustness. Moreover, case studies and molecular docking analyses (based on AlphaFold 3) further indicated the robustness and interpretability of our model. This study highlights the broad potential of CM-PHI to contribute to advancements in microbial ecology and therapeutic applications.