Scholarlink

This study explores ways deep learning models enhance human DNA sequence classification by optimizing feature representation, architecture, and hyperparameters. The blended LSTM and CNN model are considerably better than typical machine learning classifiers, with an accuracy of 82.99%, by effectively embracing long-range relations and local structure in DNA information. The LSTM + CNN model works better consistently, sometimes even getting a 100% accuracy, while XGBoost (81.50%) and k-nearest neighbor (70.77%) also perform well. Deep learning algorithms such as DeepSea work at 76.59%, while others like DeepVariant (67.00%) and graph neural networks (30.71%) work poorly. Standard models such as logistic regression (45.31%) and naïve Bayes (17.80%) cannot handle the nonlinear nature of DNA. The study emphasizes the importance of preprocessing techniques, including one-hot encoding and DNA embeddings, with the mention that hyperparameter tuning would enhance model performance further. Future research should enhance encoding techniques, tune hyperparameters, and develop more advanced genomic classification models.

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

Improving DNA Sequence Classification Accuracy Using a Hybrid LSTM-CNN Deep Learning Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Elias Tabane,
  • Zenghui Wang,
  • Ernest Mnkandla

摘要

This study explores ways deep learning models enhance human DNA sequence classification by optimizing feature representation, architecture, and hyperparameters. The blended LSTM and CNN model are considerably better than typical machine learning classifiers, with an accuracy of 82.99%, by effectively embracing long-range relations and local structure in DNA information. The LSTM + CNN model works better consistently, sometimes even getting a 100% accuracy, while XGBoost (81.50%) and k-nearest neighbor (70.77%) also perform well. Deep learning algorithms such as DeepSea work at 76.59%, while others like DeepVariant (67.00%) and graph neural networks (30.71%) work poorly. Standard models such as logistic regression (45.31%) and naïve Bayes (17.80%) cannot handle the nonlinear nature of DNA. The study emphasizes the importance of preprocessing techniques, including one-hot encoding and DNA embeddings, with the mention that hyperparameter tuning would enhance model performance further. Future research should enhance encoding techniques, tune hyperparameters, and develop more advanced genomic classification models.