Development of a deep learning-based method to identify “good” regions of a cryo-electron microscopy grid

摘要: Recent advances in cryo-electron microscopy (cryo-EM) have enabled protein structure determination at atomic resolutions. Cryo-EM specimens are prepared by rapidly freezing a solution on metal grid coated with holey carbon film; this results the formation of an ice film each hole. The thickness is critical factor for high-resolution determination; that too thick degrades contrast image while thin excludes from hole or denatures protein. Therefore, trained researchers need to manually select “good” regions appropriate thicknesses imaging. To reduce time spent such tasks, we developed deep learning program consisting “detector” and “classifier” identify good low-magnification EM images. In our method, holes detected via detector, classified as either bad classifier. detector more than 95% regardless type samples. classifier was different types samples because varies between sample types. accuracies classifiers were 93.8% soluble (β-galactosidase) 95.3% membrane (bovine heart cytochrome c oxidase). addition, found training data set containing ~ 2100 images 300 sufficient obtain accuracy, higher 90%. We expect throughput cryo-EM collection step will be greatly improved using method.