Journal: Acta Crystallographica Section F: Structural Biology Communications

Article Title: CryoSift : an accessible and automated CNN-driven tool for cryo-EM 2D class selection

doi: 10.1107/S2053230X25008866

Figure Lengend Snippet: Architecture, training and benchmarking of CryoSift . ( a ) GUI of tkteach for 2D class labeling and basic architecture of the deep convolutional neural network of CryoSift . RELION or cryoSPARC 2D projections of 31 × 31 to 210 × 210 px input. Three data features from the mass estimator and three metadata features (FRC resolution, class distribution and pixel size) are also fed into the model, resulting in a predicted quality score. Example output of AP2 averages with grade-based labels (red). ( b ) 2D averages with predicted quality scores are grouped by protein, sampled across the full range of class-quality scores. ( c ) Mean-square error loss over epochs of training and validation with features. The inset shows the prediction error between true and predicted score as a confusion matrix (density on a log scale). ( d ) Overview of the CNN layers (details are given in Supplementary Fig. S2).

Article Snippet: AP2 Tgn38 , EMPIAR-11605 , EMD-24712 * , CST , EMPIAR-10718 , EMD-21567 , GS-GN , EMPIAR-11139 , EMD-14587 *.

Techniques: Labeling, Biomarker Discovery