DeepSignal: detecting DNA methylation state from Nanopore sequencing reads using deep-learning

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Oxford Nanopore sequencing offers direct DNA methylation detection without extra lab work.
• Accurate prediction of DNA methylation states from Nanopore reads requires novel computational methods.

Purpose of the Study:

• To develop DeepSignal, a deep learning method for DNA methylation state detection using Nanopore sequencing reads.
• To evaluate DeepSignal's performance against existing methods for accuracy, robustness, and coverage requirements.

Main Methods:

• Development of DeepSignal, a deep learning algorithm.
• Testing DeepSignal on Nanopore sequencing reads from Homo sapiens, Escherichia coli, and pUC19.
• Comparison of DeepSignal with Hidden Markov Model (HMM) based methods.

Main Results:

• DeepSignal achieves higher performance in detecting 6mA and 5mC methylation states at both read and genome levels.
• The method demonstrates robust performance across different methylation bases, motifs, and CpG types.
• DeepSignal requires significantly lower sequencing coverage (e.g., 90% accuracy for 5mC/6mA with 2× coverage).
• Achieves 90% correlation with bisulfite sequencing for DNA CpG methylation prediction at 20× coverage.
• Enables prediction of methylation states for an additional 5% of DNA CpGs compared to bisulfite sequencing.

Conclusions:

• DeepSignal is a robust and accurate method for DNA methylation state detection using Nanopore sequencing.
• The method significantly reduces the required sequencing coverage for reliable methylation analysis.
• DeepSignal enhances the capability of identifying previously undetectable DNA CpG methylation states.