Classifying Force Spectroscopy of DNA Pulling Measurements Using Supervised and Unsupervised Machine Learning...

Area of Science:

• Biophysics
• Computational Biology
• Materials Science

Background:

• Dynamic force spectroscopy (DFS) generates large datasets requiring manual classification of force spectra.
• Automating the selection of successful force curves is crucial for efficient data analysis in DFS.
• Atomic force microscopy (AFM) is a key technique for DFS measurements on biomolecules.

Purpose of the Study:

• To investigate the application of machine learning algorithms for automated classification of AFM-based DFS measurements.
• To develop and evaluate supervised and unsupervised models for selecting successful force curves.
• To assess the performance of random forest and Gaussian mixture models in classifying force-distance curves.

Main Methods:

• Generated a dataset using photoswitch-modified DNA before and after UV light exposure.
• Extracted a feature set of six properties from force-distance curves.
• Trained supervised random forest models for binary and multiclass classification.
• Employed principal component analysis (PCA) and Gaussian mixture models (GMM) for unsupervised classification.

Main Results:

• Supervised random forest models achieved 94% accuracy for binary classification of successful pulls.
• Random forest models reached 90% accuracy for classifying force curves into five classes.
• The unsupervised GMM method achieved approximately 80% accuracy for binary classification.

Conclusions:

• Machine learning, particularly random forest algorithms, offers a highly accurate and automated solution for classifying DFS force curves.
• Automated classification significantly enhances the efficiency of data analysis in AFM-based dynamic force spectroscopy.
• Both supervised and unsupervised ML approaches demonstrate potential for improving DFS data processing workflows.