Scholarlink

Background <p>Polysaccharide depolymerases derived from phages play a pivotal role in the degradation of bacterial surface polysaccharides, offering a promising strategy to address antibiotic-resistant bacterial strains. However, the identification of these enzymes along with their associated host capsular serotypes poses significant challenges. This study introduces DposFinder, a transformer-based deep learning framework designed to predict depolymerases alongside their host capsular serotypes, thereby facilitating large-scale discovery and application of these enzymes.</p> Methods <p>We developed DposFinder as an interpretable model based on transformer architecture, utilizing the pre-trained protein language model ESM-2. The model’s attention mechanism was harnessed to pinpoint β-helix domains of depolymerases and predict their corresponding host capsular serotypes. DposFinder’s performance was evaluated through extensive benchmarking and experimental validation involving sequenced phages of <i>Klebsiella pneumoniae</i>.</p> Results <p>DposFinder achieved state-of-the-art performance, with an area under the ROC curve (AUC) of 0.991 on an independent test dataset, outperforming existing depolymerase prediction tools. Experimental validation confirmed the predictions of six novel depolymerases, which exhibited less than 50% sequence identity with known depolymerases. Furthermore, DposFinder successfully predicted the host capsular serotypes for the identified depolymerases. The utility of the model was further demonstrated through improved plasmid electroporation efficiency in hypermucoviscous <i>K. pneumoniae</i> isolates and the facilitation of host-serotype-labeled phage library construction. A publicly accessible database was also established, containing over 100,000 putative depolymerases derived from more than 440,000 phage genomic sequences.</p> Conclusions <p>DposFinder is a novel framework capable of predicting both depolymerases and their associated host capsular serotypes. Its interpretative capabilities and scalability render it a powerful tool for expediting the discovery and application of phage-derived depolymerases in biomedical research and phage therapy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DposFinder: an interpretable transformer model for predicting phage-derived polysaccharide depolymerases and their host capsular serotypes

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月20日

查看全文
  • Yanxiang Shen,
  • Heyuan Lun,
  • Yumeng Zhang,
  • Zhikang Wang,
  • Cui Tai,
  • Xiaohua Chen,
  • Jiangning Song,
  • Ping He,
  • Hong-Yu Ou

摘要

Background

Polysaccharide depolymerases derived from phages play a pivotal role in the degradation of bacterial surface polysaccharides, offering a promising strategy to address antibiotic-resistant bacterial strains. However, the identification of these enzymes along with their associated host capsular serotypes poses significant challenges. This study introduces DposFinder, a transformer-based deep learning framework designed to predict depolymerases alongside their host capsular serotypes, thereby facilitating large-scale discovery and application of these enzymes.

Methods

We developed DposFinder as an interpretable model based on transformer architecture, utilizing the pre-trained protein language model ESM-2. The model’s attention mechanism was harnessed to pinpoint β-helix domains of depolymerases and predict their corresponding host capsular serotypes. DposFinder’s performance was evaluated through extensive benchmarking and experimental validation involving sequenced phages of Klebsiella pneumoniae.

Results

DposFinder achieved state-of-the-art performance, with an area under the ROC curve (AUC) of 0.991 on an independent test dataset, outperforming existing depolymerase prediction tools. Experimental validation confirmed the predictions of six novel depolymerases, which exhibited less than 50% sequence identity with known depolymerases. Furthermore, DposFinder successfully predicted the host capsular serotypes for the identified depolymerases. The utility of the model was further demonstrated through improved plasmid electroporation efficiency in hypermucoviscous K. pneumoniae isolates and the facilitation of host-serotype-labeled phage library construction. A publicly accessible database was also established, containing over 100,000 putative depolymerases derived from more than 440,000 phage genomic sequences.

Conclusions

DposFinder is a novel framework capable of predicting both depolymerases and their associated host capsular serotypes. Its interpretative capabilities and scalability render it a powerful tool for expediting the discovery and application of phage-derived depolymerases in biomedical research and phage therapy.