Brownian motion data augmentation: a method to push neural network performance on nanopore sensors.

Area of Science:

• Biophysics
• Computational Biology
• Machine Learning

Background:

• Nanopores are sensitive sensors used in DNA/RNA sequencing, with potential for protein sequencing and biomarker identification.
• Solid-state nanopores exhibit instability and low signal-to-noise ratios, necessitating data-driven analysis.
• Restricted data acquisition poses a challenge for nanopore signal analysis.

Purpose of the Study:

• To address data scarcity in nanopore sensing through data augmentation.
• To improve the accuracy of nanopore signal classification tasks.
• To introduce a novel neural network architecture for enhanced nanopore data analysis.

Main Methods:

• Augmenting training data with simulated Brownian motion traces based on existing dynamic models.
• Applying the data augmentation method to a dataset of DNA nanopore reads with encoded barcodes.
• Developing and evaluating a new neural network, YupanaNet, incorporating skip connections and a soft attention mask.

Main Results:

• The Brownian motion data augmentation method noticeably increased the accuracy of the existing QuipuNet neural network.
• The novel YupanaNet achieved a higher classification accuracy of 95.8% compared to QuipuNet's 94.6%.
• YupanaNet demonstrated improved generalization due to data augmentation and novel architectural features.

Conclusions:

• Data augmentation using Brownian motion emulation effectively addresses data scarcity in nanopore sensing.
• The developed YupanaNet architecture offers superior performance for DNA nanopore read classification.
• This work enhances the potential of nanopore technology through improved data analysis techniques.