BiLSTM-5mC: A Bidirectional Long Short-Term Memory-Based Approach for Predicting 5-Methylcytosine Sites in...

Area of Science:

• Genomics
• Epigenetics
• Computational Biology

Background:

• Cancer proliferation is linked to altered DNA methylation patterns, including hypermethylation of tumor-suppressor gene promoters and decreased 5-methylcytosine (5mC) levels.
• Accurate identification of 5mC sites in promoters is vital for understanding its role in cancer and aging.
• Current experimental methods for detecting 5mC sites are often time-consuming and labor-intensive.

Purpose of the Study:

• To develop a deep learning-based computational approach for accurate identification of 5mC sites in genome-wide DNA promoters.
• To address the limitations of experimental techniques in detecting 5mC sites efficiently.

Main Methods:

• A deep learning model, BiLSTM-5mC, was proposed, utilizing a bidirectional long short-term memory (BiLSTM) network.
• Feature vectors were generated using one-hot encoding and nucleotide property and frequency (NPF) methods.
• A majority vote strategy integrated outputs from 22 submodels trained on balanced positive and negative 5mC sample subsets.

Main Results:

• The BiLSTM-5mC model demonstrated high accuracy in identifying 5mC sites in DNA promoters.
• Experimental results showed that BiLSTM-5mC outperformed existing methods when evaluated on an independent dataset.

Conclusions:

• BiLSTM-5mC offers an efficient and accurate computational solution for identifying 5mC sites in DNA promoters.
• This approach can significantly aid research into the role of DNA methylation in genetic diseases like cancer and aging.