DeepSelectNet: deep neural network based selective sequencing for oxford nanopore sequencing.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Nanopore sequencing enables selective sequencing, rejecting unwanted reads for applications like species classification.
• Current selective sequencing methods for species identification are immature, with accuracy varying widely (77-97%).

Purpose of the Study:

• To introduce DeepSelectNet, a novel deep learning method for accurate species classification directly from nanopore sequencing current signals.
• To improve the accuracy and efficiency of selective sequencing for species identification.

Main Methods:

• Developed DeepSelectNet, a deep learning model utilizing advanced data preprocessing and neural network architecture.
• Trained and evaluated DeepSelectNet on five diverse datasets for species classification tasks.

Main Results:

• DeepSelectNet achieved high accuracy (91-99%, average 95%) across five datasets.
• Demonstrated a significant accuracy increase (up to 12%) compared to the deep learning predecessor, SquiggleNet.
• Exhibited strong precision and recall (>89% average) and outperformed SquiggleNet in execution performance by 13%.

Conclusions:

• DeepSelectNet substantially enhances accuracy for real-time species classification via selective nanopore sequencing.
• The method surpasses traditional base alignment and existing deep learning approaches.
• Source code is publicly available for further research and application.