DRPnet

DRPnet automates particle picking in cryo-electron micrographs for single-particle analysis of macromolecular structures.

Key Features:

• Deep learning approach: Employs a double convolutional neural network (CNN) cascade to address low signal-to-noise ratios, variable particle sizes and shapes, diverse distributions, grayscale variations, and other cryoEM artifacts.
• Network architecture: Combines a fully convolutional regression network (FCRN) that maps input micrographs to a continuous distance map (serving as a probability density for particle centers) with a classification CNN that refines initial detections to reduce false positives.
• Transfer learning: The first CNN can be pretrained on a single cryoEM dataset and subsequently applied to different datasets without retraining.
• Performance metrics: Outperforms template-based autopicking methods such as RELION on recall, precision, and F-measure and produces particle sets that improve downstream 3D reconstructions as measured by number of 2D class averages, efficiency/angular coverage, Rosenthal-Henderson plots, and local/global resolution.
• Processing efficiency: Reduces overall processing time for generating initial particle datasets compared to manual picking workflows.
• Challenging dataset applicability: Performs well on low-contrast or clumped-particle datasets, supports higher-resolution 3D reconstructions with fewer particles or unknown symmetry, and improves angular orientation coverage.

Scientific Applications:

• Automated particle selection: Automated identification and localization of protein particles in cryo-electron micrographs for single-particle analysis.
• 3D reconstruction preparation: Generation of high-quality particle sets for 2D classification and 3D reconstruction to improve resolution and angular coverage.
• Challenging dataset analysis: Handling of low-contrast, clumped, or symmetry-ambiguous cryoEM datasets.
• Cross-dataset deployment: Application of pretrained models to accelerate particle picking across different cryoEM experiments.

Methodology:

Implements a double CNN cascade: a fully convolutional regression network (FCRN) that maps micrographs to a continuous distance map as a probability density for particle centers, followed by a classification CNN that refines detections; the first CNN supports transfer learning via pretraining on one dataset and application to others.