Denoising Autoencoder Trained on Simulation-Derived Structures for Noise Reduction in Chromatin Scanning Transmission Electron Microscopy

Abstract

Scanning transmission electron microscopy tomography with ChromEM staining (ChromSTEM), has allowed for the three-dimensional study of genome organization. By leveraging convolutional neural networks and molecular dynamics simulations, we have developed a denoising autoencoder (DAE) capable of postprocessing experimental ChromSTEM images to provide nucleosome-level resolution. Our DAE is trained on synthetic images generated from simulations of the chromatin fiber using the 1-cylinder per nucleosome (1CPN) model of chromatin. We find that our DAE is capable of removing noise commonly found in high-angle annular dark field (HAADF) STEM experiments and is able to learn structural features driven by the physics of chromatin folding. The DAE outperforms other well-known denoising algorithms without degradation of structural features and permits the resolution of α-tetrahedron tetranucleosome motifs that induce local chromatin compaction and mediate DNA accessibility. Notably, we find no evidence for the 30 nm fiber, which has been suggested to serve as the higher-order structure of the chromatin fiber. This approach provides high-resolution STEM images that allow for the resolution of single nucleosomes and organized domains within chromatin dense regions comprising of folding motifs that modulate the accessibility of DNA to external biological machinery.

Convolutional Machine Learning Method for Accelerating Nonequilibrium Green’s Function Simulations in Nanosheet Transistor

“…Our approach can be described as a similar method used in face age recognition or the removal of the noise from images [18], [19]. If our hypothesis is correct, an Auto-Encoder (AE) [11] NN architecture, and more specifically a sub-type known as a Denoising Auto-Encoder (DAE) [12]- [14], can be used to learn the systematic changes in the charge and the potential. Hence, such DAE NN will be able to predict main device characteristics, such as charge and potential distributions, but it can also be adapted to predict other parameters.…”

LCPOM: Precise Reconstruction of Polarized Optical Microscopy Images of Liquid Crystals

“…Machine learning methods are revolutionizing microscopy analysis in the physical sciences, finding applications in denoising images, labeling features, and revealing the underlying physics of complex systems. − Reports of their applications to microscopic images of LC systems to date mainly rely on the feature changes in the brightness profile while ignoring the information from the colors. , We anticipate that the LCPOM package will take machine learning efforts one step further by aiding in generating training data with color information.…”