A ML Framework for Genetic Sequence Identification Using 2D Electrical Conductance Probability Distributions from...

Area of Science:

• Biophysics
• Computational Biology
• Machine Learning

Background:

• Electrical characterization of single molecules offers a novel approach to genetic sequence identification.
• High noise levels in single-molecule electrical data challenge accurate detection and discrimination of single base-pair mismatches.

Purpose of the Study:

• To develop an improved computational method for accurate genetic sequence identification from noisy single-molecule electrical data.
• To enhance the detection of single base-pair mismatches in DNA sequences.

Main Methods:

• A hybrid machine learning architecture combining a convolutional neural network (CNN) with an ensemble learning method (XGBoost).
• Exploration of four input feature representations: 1D and 2D conductance probability distributions, with and without averaging over experimental parameters.
• Utilizing averaged conductance probability distributions as feature matrices derived from mixed datasets.

Main Results:

• The proposed architecture significantly enhances prediction accuracy for genetic sequence identification.
• 2D probability distributions improve classifier accuracy, with averaged conductance probability distributions showing the most substantial impact.
• An approximate 10% performance increase was observed across multiple DNA sequences, effectively distinguishing them from single base-pair mismatches.

Conclusions:

• The developed CNN-XGBoost architecture effectively addresses the challenge of noisy single-molecule electrical data for genetic identification.
• Averaged conductance probability distributions are a highly impactful feature representation for improving classification accuracy.
• The method demonstrates broad applicability for single-molecule identification tasks beyond DNA sequencing, based on conductance properties.