Particle Picking

Modern pickers like crYOLO and Warp utilize You Only Look Once (YOLO) or U-Net architectures to perform object detection directly on micrographs. These CNNs learn hierarchical features—from edges and blobs to complex particle shapes—enabling them to distinguish true particles from carbon edges, ice contamination, and aggregates. Unlike traditional difference-of-Gaussian (DoG) pickers, CNN-based methods generalize across varied defocus values and ice thicknesses without manual parameter tuning.

• crYOLO: Implements a YOLO-based architecture trained on annotated micrographs, outputting bounding boxes with confidence scores.
• Warp: Uses a U-Net for real-time particle probability map generation during data collection.
• Training data: Typically requires only a few hundred manually picked particles per micrograph to fine-tune a pre-trained general model.

Reference-based pickers like Gautomatch and RELION's AutoPick use 2D class averages from an initial manual picking round as templates. These templates are cross-correlated with the micrograph, and peaks above a threshold are selected as particle coordinates. While computationally efficient, this approach suffers from template bias—only particles resembling the initial references are selected, potentially missing rare views or conformations.

• Gautomatch: GPU-accelerated, uses a fast local correlation algorithm with a Gaussian blob as an initial reference.
• RELION AutoPick: Integrates with the RELION pipeline, using user-provided 2D class averages for correlation-based picking.
• Limitation: Performance degrades significantly for heterogeneous samples or when preferred orientation artifacts dominate the initial templates.

Specialized pickers like e2helixboxer (EMAN2) and Filament Tracer (cryoSPARC) are designed for helical assemblies and filamentous structures such as actin, microtubules, and amyloid fibrils. These tools trace the filament path rather than picking discrete particles, then segment the traced path into overlapping segments for helical reconstruction.

• e2helixboxer: Manual or semi-automated helix boxing with user-defined helical parameters.
• cryoSPARC Filament Tracer: Uses a Dijkstra shortest-path algorithm on a CNN-generated probability map to trace filament centerlines.
• Output: Generates particle stacks with defined helical symmetry parameters for subsequent Iterative Helical Real Space Reconstruction (IHRSR).

Picking particles in cryo-electron tomography (cryo-ET) presents unique challenges due to the missing wedge artifact and crowded cellular environments. Tools like DeePiCt and crYOLO for tomograms apply 3D CNNs or 2D CNNs on tomogram slices to detect macromolecular complexes in situ. Template matching (e.g., PyTom) remains widely used, scanning a 3D reference through the tomogram, but deep learning methods are rapidly gaining traction for their speed and reduced false-positive rates.

• DeePiCt: A 3D U-Net trained on annotated tomograms for ribosome and proteasome detection.
• PyTom: GPU-accelerated 3D template matching with missing wedge compensation.
• Subtomogram averaging: Picked particles are extracted as 3D sub-volumes for subsequent alignment and averaging.

Real-time particle picking during data acquisition enables immediate feedback on sample quality and ice thickness. Warp and cryoSPARC Live integrate with microscope control software (e.g., SerialEM, EPU) to process micrographs as they are collected. This allows operators to monitor particle distribution, detect preferred orientation issues, and adjust collection parameters without waiting for offline processing.

• Warp: Performs real-time CTF estimation, particle picking, and 2D classification during collection.
• cryoSPARC Live: Streams data from the microscope, performing on-the-fly motion correction, CTF estimation, and blob picking followed by 2D classification.
• Edge case detection: Immediate identification of empty or contaminated grid squares saves valuable microscope time.

After picking, particle images are aligned and grouped by similarity to improve signal-to-noise ratio. This step reveals distinct orientation views and serves as a quality filter—junk particles that fail to produce coherent averages are discarded before expensive 3D reconstruction.

• Uses maximum-likelihood alignment to iteratively refine class centers
• Reveals preferred orientation artifacts early in the workflow
• Implemented in RELION and cryoSPARC with GPU acceleration

The CTF mathematically describes how the microscope's objective lens aberrations modulate image contrast as a function of spatial frequency. Accurate CTF estimation and correction are essential—without it, high-resolution features are inverted or lost entirely.

• Characterized by defocus and astigmatism parameters
• Thon rings in the power spectrum reveal CTF oscillations
• Per-micrograph estimation using tools like CTFFIND4 or Gctf

The computational process of determining a macromolecule's Coulomb potential density map from 2D projections. Modern algorithms like weighted back-projection and iterative refinement assign Euler angles to each particle and build a consensus volume.

• Requires accurate orientation assignment via projection matching
• Gold-standard refinement splits data into independent half-sets
• Resolution measured by Fourier Shell Correlation (FSC) at 0.143 criterion

A per-particle beam-induced motion correction algorithm in RELION that models radiation damage trajectories. It uses a Bayesian framework to reverse the cumulative blurring that degrades high-resolution signal during exposure.

• Operates on dose-fractionated movie frames
• Models both global and local motion trajectories
• Typically improves reported resolution by 0.1–0.3 Å

A computational classification method that sorts particles into structurally distinct 3D classes. Essential for resolving compositional heterogeneity (different subunit stoichiometries) and conformational heterogeneity (flexible domains) within a single sample.

• Uses maximum-likelihood 3D classification
• Prevents averaging of distinct states into a blurred consensus
• Input to advanced tools like 3DVA and cryoDRGN

An automated atomic model building program that uses a graph neural network to trace the protein backbone and assign amino acid sequences directly into cryo-EM density maps. Dramatically reduces the manual effort required for model completion.

• Integrates sequence information with density geometry
• Handles maps at resolutions better than ~3.5 Å
• Outputs a near-complete Cα trace with sidechain assignments