Interpretable Deep Learning for Single-Molecule Nanopore Fingerprinting Using Physics-Guided Preprocessing.

Area of Science:

• Biotechnology
• Data Science
• Analytical Chemistry

Background:

• Nanopore sensing offers single-molecule analysis via ionic current pulses.
• Current methods rely on handcrafted features, potentially missing structural information.
• Need for robust and interpretable molecular fingerprinting in various applications.

Purpose of the Study:

• To develop an interpretable machine learning framework for analyzing raw nanopore sensing data.
• To improve accuracy and provide physically consistent attributions for molecular identification.
• To establish a practical design principle for custom filters in pulse-based sensing.

Main Methods:

• Developed a physics-guided time-frequency transform paired with a compact neural classifier.
• Operated directly on raw ionic current pulses, bypassing conventional feature extraction.
• Included baseline models: Support Vector Machines (SVMs) and a 1D classifier on raw pulses.

Main Results:

• Achieved high accuracy in distinguishing DNA nanostructures with overlapping standard features.
• Generated physically consistent feature attributions, highlighting discriminative signal motifs.
• Demonstrated the benefit of physics-guided preprocessing for improved reliability.

Conclusions:

• The interpretable machine learning framework enhances nanopore sensing accuracy and transparency.
• The approach is modular, lightweight, and applicable to diverse pulse-based sensing platforms.
• This work enables deployable sensing solutions for regulated and mission-critical environments.