Automated atomic force microscopy analysis using convolutional and recurrent neural networks.

Area of Science:

• Biophysics
• Materials Science
• Cell Biology

Background:

• Atomic force microscopy (AFM) is crucial for characterizing cell and tissue biomechanics.
• Analyzing AFM force curves is challenging due to noise and contact point uncertainty.
• Existing methods lack speed, reproducibility, and quantitative accuracy.

Purpose of the Study:

• To develop a novel machine learning algorithm for processing AFM force curves.
• To improve the accuracy and reliability of contact point determination in AFM data.
• To enable high-throughput, precise biomechanical analysis of cells and tissues.

Main Methods:

• A convolutional bidirectional recurrent neural network (COBRA) was developed.
• The algorithm was trained on over 5000 curated AFM force curves from diverse cell types.
• COBRA was compared against classical and other machine learning techniques.

Main Results:

• COBRA demonstrated superior identification of low-quality AFM curves (AUC 0.92).
• The algorithm achieved minimal contact point error (28 ± 3 nm).
• Pointwise elastic modulus was determined with a mean absolute percentage error of 5.3% ± 0.7%.

Conclusions:

• COBRA reliably filters low-quality AFM force curves and determines the contact point.
• The method enhances precision and reproducibility in high-throughput AFM analyses.
• COBRA offers a significant advancement for quantitative biomechanical characterization.