Deep learning-assisted analysis of single molecule dynamics from liquid-phase electron microscopy

Abstract

We apply U-Net and UNet++ to analyze single-molecule movies obtained from liquid-phase electron microscopy. Neural networks allow full automation, and high throughput analysis of these low signal-to-noise ratio images, while achieving higher segmentation accuracy, and avoiding subjective errors as compared to the conventional threshold methods. The analysis enables the quantification of transient dynamics in chemical systems and the capture of rare intermediate states by resolving local conformational changes within a single molecule.