MoEPH: an adaptive fusion-based LLM for predicting phage-host interactions in health informatics.

Abstract

Phage-host interaction prediction plays a crucial role in the development of phage therapy, particularly in combating antimicrobial resistance (AMR). Current in silico models often suffer from limited generalizability and low interpretability. To address these gaps, we introduce MoEPH (Mixture-of-Experts for Phage-Host prediction), a novel framework that integrates transformer-based protein embeddings (ProtBERT and ProT5) with domain-specific statistical descriptors. Our model dynamically combines features using a gated fusion mechanism, ensuring robust and adaptive prediction. We evaluate MoEPH on three publicly available phage-host interaction databases: Dataset 1 (101 host strains, 129 phages), Dataset 2 (38 host strains, 176 phages), and Dataset 3 (combined). Experimental results demonstrate that MoEPH outperforms existing methods, achieving an accuracy of 99.6% on balanced datasets and a 31% improvement on highly imbalanced data. The model provides a transparent, dynamic and knowledge-driven fusion solution for phage-host prediction, contributing to more effective phage therapy recommendations. Future work will focus on incorporating structural protein features and exploring alternative neural backbones for further enhancement.