Scholarlink

Reliable prediction of protein secondary structure (PSS) is crucial for advancing structure-based drug discovery, particularly in addressing the challenges posed by antimicrobial resistance in various pathogens. Despite the development of various computational frameworks, they often depend on homology-based sequence alignments or incur high computational costs, limiting their applicability to less-characterized pathogens and resource-constrained settings. Here, we present FCTFold, a fusion deep learning model that integrates multiscale CNNs (kernel sizes 3, 5, and 7) with a windowed transformer encoder (window size, w = 16) to predict the PSS directly from one-hot encoded sequences, reducing complexity from O(n2) to O(n·w). Trained on high-resolution data from the Protein Data Bank (≤2.5 Å), FCTFold achieves an overall Q3 accuracy >89.9%, outperforms the PSIPRED, GOR, RNN, LSTM, BiLSTM, and DeepCNF models, with pathogen-specific average F1-scores: M. tuberculosis (88.6%), E. coli (89.1%), S. aureus (91.9%), and P. aeruginosa (87.5%). With efficient short- and long-range extraction of significant patterns and low computational overhead, FCTFold provides a scalable solution for PSS prediction and is a crucial step toward understanding protein function, designing therapeutics, and combating antimicrobial resistance.

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

FCTFold: A Fused CNN-Transformer Framework for Protein Fold Prediction in AMR Pathogens

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mousumi Banerjee,
  • Anushree Chakraborty,
  • Sandip Paul,
  • Rachana Banerjee,
  • Kausik Basak

摘要

Reliable prediction of protein secondary structure (PSS) is crucial for advancing structure-based drug discovery, particularly in addressing the challenges posed by antimicrobial resistance in various pathogens. Despite the development of various computational frameworks, they often depend on homology-based sequence alignments or incur high computational costs, limiting their applicability to less-characterized pathogens and resource-constrained settings. Here, we present FCTFold, a fusion deep learning model that integrates multiscale CNNs (kernel sizes 3, 5, and 7) with a windowed transformer encoder (window size, w = 16) to predict the PSS directly from one-hot encoded sequences, reducing complexity from O(n2) to O(n·w). Trained on high-resolution data from the Protein Data Bank (≤2.5 Å), FCTFold achieves an overall Q3 accuracy >89.9%, outperforms the PSIPRED, GOR, RNN, LSTM, BiLSTM, and DeepCNF models, with pathogen-specific average F1-scores: M. tuberculosis (88.6%), E. coli (89.1%), S. aureus (91.9%), and P. aeruginosa (87.5%). With efficient short- and long-range extraction of significant patterns and low computational overhead, FCTFold provides a scalable solution for PSS prediction and is a crucial step toward understanding protein function, designing therapeutics, and combating antimicrobial resistance.