Scholarlink

<p> Purpose Classification of viral DNA sequences poses significant challenges in computational genomics, particularly in multi-class classification across diverse diseases. Existing studies often focus on binary or small-scale classifications, leaving a gap in addressing large-scale problems. Unbalanced datasets, where certain classes dominate due to limited data availability, further hinder model generalizability and accuracy. This article addresses these challenges and gaps in literature. We selected 17 distinct variants across six viruses: SARS-CoV-2 variants (Alpha, Delta, Omicron), Influenza subtypes (Alpha, Beta, Delta, Gamma), Hepatitis virus variants (B, C, D, E), Dengue Virus (Type1, Type 2, Type 3, Type 4), HIV, and Ebola. This study includes the largest number of distinct variants in a balanced dataset for classification reported in the literature. Methods Using a balanced dataset of 17,000 samples, the study ensures robust model performance while mitigating data imbalance biases. A novel Multi-Modal Dual Pathway Network is proposed, leveraging hierarchical channel-spatial attention and adaptive feature fusion to integrate complementary features. Genomic image processing techniques, including frequency chaos game representation (FCGR), Markov transition fields (MTF), and Gramian angular summation fields (GAF), enhance feature extraction. ResNet-50 pathways with convolutional block attention module (CBAM) refine feature maps for both scratch-trained and pre-trained pathways. Results The proposed model achieves a peak accuracy of 99.17% with MTF Order 2 fused with MTF Order 4. Other combinations, such as FCGR K=2 fused with MTF Order 4, also exceed 99.09% accuracy, demonstrating scalability and effectiveness. Conclusion This work sets a new benchmark for viral DNA classification, addressing data imbalance and advancing computational genomics through innovative methodologies.</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

A novel multi-modal dual pathway network with hierarchical channel-spatial attention and adaptive feature fusion for viral genomic variant classification

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Love Fadia,
  • Vatsal Shah,
  • Mohammad Hassanzadeh,
  • Jonathan Wu,
  • Majid Ahmadi

摘要

Purpose Classification of viral DNA sequences poses significant challenges in computational genomics, particularly in multi-class classification across diverse diseases. Existing studies often focus on binary or small-scale classifications, leaving a gap in addressing large-scale problems. Unbalanced datasets, where certain classes dominate due to limited data availability, further hinder model generalizability and accuracy. This article addresses these challenges and gaps in literature. We selected 17 distinct variants across six viruses: SARS-CoV-2 variants (Alpha, Delta, Omicron), Influenza subtypes (Alpha, Beta, Delta, Gamma), Hepatitis virus variants (B, C, D, E), Dengue Virus (Type1, Type 2, Type 3, Type 4), HIV, and Ebola. This study includes the largest number of distinct variants in a balanced dataset for classification reported in the literature. Methods Using a balanced dataset of 17,000 samples, the study ensures robust model performance while mitigating data imbalance biases. A novel Multi-Modal Dual Pathway Network is proposed, leveraging hierarchical channel-spatial attention and adaptive feature fusion to integrate complementary features. Genomic image processing techniques, including frequency chaos game representation (FCGR), Markov transition fields (MTF), and Gramian angular summation fields (GAF), enhance feature extraction. ResNet-50 pathways with convolutional block attention module (CBAM) refine feature maps for both scratch-trained and pre-trained pathways. Results The proposed model achieves a peak accuracy of 99.17% with MTF Order 2 fused with MTF Order 4. Other combinations, such as FCGR K=2 fused with MTF Order 4, also exceed 99.09% accuracy, demonstrating scalability and effectiveness. Conclusion This work sets a new benchmark for viral DNA classification, addressing data imbalance and advancing computational genomics through innovative methodologies.