Cryogenic-electron tomography (cryo-ET) is an imaging technique that allows the observation of biomolecules at sub-nanometer resolution and that is extensively used for the visualization of macromolecules in their cellular environment. A major challenge while studying cryo-electron tomograms is the localization of macromolecules: in existing methods, for each macromolecule that is being studied, users must manually annotate it in the images to train the algorithms, which is a time-consuming and error-prone process. In a recent study, Gavin Rice and colleagues proposed TomoTwin, a framework for structural data mining of cryo-electron tomograms that alleviates this problem.
TomoTwin learns generalized three-dimensional (3D) shapes in tomograms using deep metric learning. The authors provided two workflows for identifying macromolecule location in embedding spaces: a reference-based workflow that identifies an example protein in a tomogram and maps it to the embedding space, and a clustering workflow that explores the macromolecular contents without prior knowledge. The system embeds tomograms based on macromolecular content similarity, allowing particles to be picked by identifying their associated regions, and consolidating them into a single centralized pick. The framework is based on a 3D convolutional neural network (CNN) consisting of five convolutional blocks followed by a head network. The 3D CNN can locate macromolecules from the training set and generalize to new ones, preserving fidelity and avoiding retraining for each protein. The proposed approach, unlike previous methods, avoids the tedious task of retraining the algorithm for each macromolecule to be observed.
This is a preview of subscription content, access via your institution