An ensemble pipeline, PhageHost, for phage tail fiber discovery and accurate Klebsiella pneumoniae host prediction using protein language models

作者：Jiang Wu, Chao Wang, Yuan Zeng, Yiyao Song, Shengyao Xu, Jingyi Nie, Tingxuan Wang, Juncai Ma, Jie Feng, Linhuan Wu · 发表于：Cell Reports · 年份：2026 · DOI：10.1016/j.celrep.2026.117275 · 被引用次数：1 · 研究领域：Bacteriophages and microbial interactions、vaccines and immunoinformatics approaches、Machine Learning in Bioinformatics

The high specificity of phages toward their hosts holds great promise for phage therapy while posing a challenge to computational prediction approaches. To this end, we introduce PhageHost (P&H), an ensemble pipeline for large-scale discovery of phage tail fibers and strain-level host prediction in Klebsiella pneumoniae using protein language models (PLMs). The pipeline begins with TailSeek, a PLM for tail fiber detection from phage and prophage genomes. Building on TailSeek predictions, we developed HostBuster, a deep learning framework that integrates tail fiber features with host-specific information to predict the lytic potential of phage-K. pneumoniae pairs. In silico and experimental validations confirm that P&H achieves high sensitivity and precision in both tail fiber identification and host prediction. Moreover, our framework demonstrates strong generalizability, enabling large-scale mining of tail fibers from prophage data across diverse bacterial taxa. P&H significantly accelerates high-throughput phage screening, offering a scalable tool for clinical phage therapy applications.