Deep Learning of Nanopore Sensing Signals Using a Bi-Path Network

Area of Science:

• Biophysics
• Nanotechnology
• Machine Learning

Background:

• Nanopore sensors detect analytes via electrical resistance changes, generating pulse-like signals.
• Current feature extraction methods rely on subjective, user-defined thresholds, limiting objectivity and performance in noisy data.
• Existing algorithms struggle with low signal-to-noise ratios, hindering accurate analysis of translocating molecules.

Purpose of the Study:

• To develop an objective, data-driven method for feature extraction from nanopore sensor current traces.
• To overcome limitations of traditional threshold-based algorithms in analyzing noisy pulse-like signals.
• To introduce a versatile deep learning architecture for analyzing biological molecule translocation events.

Main Methods:

• Implementation of a bi-path network (B-Net) deep learning architecture for pulse recognition and feature extraction.
• Training the B-Net on simulated and experimental datasets of DNA and protein translocation.
• Evaluation of B-Net performance against traditional threshold-based methods, particularly in low signal-to-noise conditions.

Main Results:

• The B-Net successfully extracts features and recognizes pulses without requiring predefined parameters.
• Results demonstrate small relative errors and stable trends in both simulated and experimental data.
• The B-Net effectively processes data with a signal-to-noise ratio of 1, a feat impossible for threshold-based algorithms.

Conclusions:

• The bi-path network (B-Net) offers an objective and robust solution for feature extraction in nanopore sensing.
• This deep learning approach significantly enhances the analysis of biological molecule translocation events, especially in challenging noisy environments.
• The B-Net architecture is generalizable and applicable to various pulse-like signal analysis tasks beyond nanopore sensing.