Scholarlink

<p>Proteins serve as the central executors of life activities, performing diverse functions such as metabolic catalysis, genetic regulation, signal transduction, and cytoskeletal maintenance. However, microbial proteins face unique challenges: their rapid evolution leads to low sequence conservation, and structural diversity complicates functional inference. Experimental annotation lags far behind due to scalability limits—over 70% of microbial proteins in UniProt remain uncharacterized, compared to roughly 50% for model eukaryotes. Traditional homology-based tools (e.g., FASTA/BLAST) often fail on highly divergent microbial families, and existing machine-learning methods rarely account for microbial-specific signals such as horizontal gene transfer. To address this gap, this study presents the first publicly available dataset for microbial protein function annotation and introduces ProMoHGT, a novel model that extracts evolutionary and contextual sequence features using ESM-2, constructs three-dimensional spatial proximity graphs from AlphaFold2 predictions, and encodes residue-specific physicochemical properties. Its core heterogeneous Transformer architecture incorporates super-nodes and multi-head self-attention to integrate global topology with long-range dependencies, while graph contrastive learning adds regularization to enhance robustness and prevent overfitting. ProMoHGT outperforms state-of-the-art methods across all three Gene Ontology categories (MF, BP, CC) and in Enzyme Commission number prediction, with the smallest performance decay observed across varying homology scenarios, thereby validating its superior generalization capability. A case study on three representative microbial proteins (ArcA, CodY, and SPT16) further confirmed these advantages, where ProMoHGT most accurately recovered key experimentally validated functions such as DNA binding, transcription activation, chromatin remodeling, and metabolic regulation, achieving the highest F1 scores among all methods.</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

ProMoHGT: a heterogeneous graph transformer with graph contrastive learning for robust microbial protein function prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月17日

查看全文
  • Jianan Sui,
  • Xiaomiao Wang,
  • Yufeng Su,
  • Menglong Gao,
  • Huanjie Huang,
  • Hao Liu,
  • Jianfeng Zhang,
  • Yaohan Tang

摘要

Proteins serve as the central executors of life activities, performing diverse functions such as metabolic catalysis, genetic regulation, signal transduction, and cytoskeletal maintenance. However, microbial proteins face unique challenges: their rapid evolution leads to low sequence conservation, and structural diversity complicates functional inference. Experimental annotation lags far behind due to scalability limits—over 70% of microbial proteins in UniProt remain uncharacterized, compared to roughly 50% for model eukaryotes. Traditional homology-based tools (e.g., FASTA/BLAST) often fail on highly divergent microbial families, and existing machine-learning methods rarely account for microbial-specific signals such as horizontal gene transfer. To address this gap, this study presents the first publicly available dataset for microbial protein function annotation and introduces ProMoHGT, a novel model that extracts evolutionary and contextual sequence features using ESM-2, constructs three-dimensional spatial proximity graphs from AlphaFold2 predictions, and encodes residue-specific physicochemical properties. Its core heterogeneous Transformer architecture incorporates super-nodes and multi-head self-attention to integrate global topology with long-range dependencies, while graph contrastive learning adds regularization to enhance robustness and prevent overfitting. ProMoHGT outperforms state-of-the-art methods across all three Gene Ontology categories (MF, BP, CC) and in Enzyme Commission number prediction, with the smallest performance decay observed across varying homology scenarios, thereby validating its superior generalization capability. A case study on three representative microbial proteins (ArcA, CodY, and SPT16) further confirmed these advantages, where ProMoHGT most accurately recovered key experimentally validated functions such as DNA binding, transcription activation, chromatin remodeling, and metabolic regulation, achieving the highest F1 scores among all methods.