End-to-end differentiable blind tip reconstruction for noisy atomic force microscopy images

Area of Science:

• Biophysics
• Microscopy
• Computational Biology

Background:

• Understanding biomolecular dynamics is crucial for function. High-speed atomic force microscopy (HS-AFM) visualizes biomolecules under near-physiological conditions.
• AFM imaging is affected by tip shape, distorting molecular surface data. Accurate tip shape determination is essential for reliable image analysis.
• The existing blind tip reconstruction (BTR) algorithm is sensitive to noise, limiting its application in real-world AFM data.

Purpose of the Study:

• To develop a noise-robust method for determining AFM tip shapes.
• To improve the accuracy of reconstructing biomolecular surface geometry from AFM images.
• To provide a general post-processing tool for AFM image analysis.

Main Methods:

• Introduced an end-to-end differentiable blind tip reconstruction (BTR) method utilizing machine learning.
• Incorporated a regularization term in the loss function to mitigate noise overfitting.
• Optimized tip shape using automatic differentiation and backpropagation within deep learning frameworks.

Main Results:

• The proposed differentiable BTR method demonstrated robustness against noise in simulated AFM images of myosin V.
• The method successfully identified double-tip shapes and deconvolved molecular images.
• Application to real HS-AFM data of myosin V on actin yielded accurate actomyosin surface geometry, consistent with existing structural models.

Conclusions:

• The end-to-end differentiable BTR method offers a significant advancement in AFM image analysis, particularly for noisy datasets.
• This approach enhances the reconstruction of biomolecular surfaces, aiding the study of structural dynamics.
• The developed method serves as a versatile post-processing tool for various AFM imaging applications.