Improving biosensor accuracy and speed using dynamic signal change and theory-guided deep learning.

Area of Science:

• Biosensing and Diagnostics
• Computational Biology
• Materials Science

Background:

• Biosensing faces challenges with false results and time delays, impacting reliability.
• Deep learning offers potential for enhanced biosensor performance, but explainability and domain consistency are critical.
• Integrating theoretical knowledge into deep learning models is key for trustworthy biosensing.

Purpose of the Study:

• To develop a deep learning framework that ensures prediction consistency with biosensing domain knowledge.
• To enable rapid and accurate quantification of analytes using biosensor dynamic responses.
• To improve biosensor performance metrics like accuracy, precision, and recall.

Main Methods:

• Implemented cost function supervision to align deep learning predictions with biosensor theory.
• Utilized data augmentation and a theory-guided recurrent neural network (TGRNN) classifier.
• Validated the methodology using cantilever biosensors for microRNA (let-7a) quantification.

Main Results:

• The TGRNN classifier improved F1 score, precision, and recall by an average of 13.8%.
• Achieved 98.5% average prediction accuracy, precision, and recall for analyte quantification and false result detection.
• Demonstrated high accuracy using initial transient or entire dynamic response data.

Conclusions:

• Cost function supervision enables explainable and domain-consistent deep learning for biosensing.
• The TGRNN methodology significantly enhances biosensor speed and accuracy, minimizing false results.
• Established new relationships between biosensor performance characteristics and design parameters for improved characterization.