differentiable BTR

differentiable BTR reconstructs unknown AFM tip shapes from noisy atomic force microscopy (AFM) images to enable accurate recovery of biomolecular surface geometries.

Key Features:

• End-to-end differentiable BTR: Implements a fully differentiable blind tip reconstruction approach for AFM image deconvolution.
• Regularized loss function: Uses a loss that includes a regularization term to mitigate overfitting caused by image noise.
• Optimization by automatic differentiation: Optimizes tip shape parameters using automatic differentiation and backpropagation algorithms from deep learning frameworks.
• Robustness to noise: Demonstrated robustness on noisy pseudo-AFM images, including tests on myosin V motor domain data.
• Complex tip detection: Detects complex tip geometries such as double-tips and enables deconvolution of doubled molecular images.
• Benchmark reference: Contrasts with traditional BTR methods (Villarrubia 1997) based on mathematical morphology operators that perform poorly on noisy data.
• High-speed AFM applicability: Applicable to high-speed AFM datasets, including myosin V walking on actin filament recordings.

Scientific Applications:

• Surface geometry reconstruction: Recovers accurate biomolecular surface geometries from noisy AFM images for structural interpretation.
• High-speed AFM analysis: Applied to high-speed AFM data of myosin V walking on actin filament to produce geometries consistent with actomyosin structural models.
• Artifact detection and correction: Identifies tip-induced artifacts such as double-tips and deconvolves doubled molecular images.
• Benchmarking BTR methods: Provides a modern, noise-robust alternative for comparing against mathematical morphology–based BTR approaches.

Methodology:

End-to-end differentiable blind tip reconstruction using a regularized loss function with optimization via automatic differentiation and backpropagation.