This webinar introduces TomoNet, a GUI-based software for cryo-electron tomography (CryoET) designed to streamline the data processing workflow and automate particle picking. The primary motivation is to manage the large datasets required for high-resolution structural determination and to provide efficient tools for identifying particles, particularly those in flexible or distorted lattice-like arrangements.
Two main automated picking methods are presented. The first, "Expand," is a geometry-based algorithm that iteratively identifies particles by expanding from a few user-defined seed points along a specified lattice configuration. This method is accurate but requires a repetitive structure and significant user input to define the lattice parameters. The second is a U-Net-based neural network picker that treats particle identification as a segmentation task. It requires a small set of manually picked examples for training and is more general, applicable to non-lattice structures, but as far as I can tell, offers less control over false positives and is likely less accurate. The talk also covers practical aspects like data preparation, parameter tuning for both methods, and result visualization.
TomoNet: A Workflow for Automated Particle Picking in CryoET
2:50 The CryoET Problem: Achieving high-resolution structures (< 5 Γ ) in CryoET requires averaging thousands to tens of thousands of particles (subtomograms). This creates significant challenges in data management and particle picking, which TomoNet aims to solve.
8:16 TomoNet's Goal: The software provides an integrated GUI to manage the entire CryoET pipeline (motion correction, reconstruction, picking) and offers two primary automated particle picking tools.
11:05 Two Picking Strategies: TomoNet includes two distinct picking algorithms:
Geometry-Based ("Expand"): Highly accurate for structures arranged in a 1D, 2D, or 3D lattice. It requires manual "seed" particles and precise lattice parameters from the user.
Neural Network-Based: More generalized and does not require a lattice. It learns to identify particles from a small set of user-provided examples but offers less control over the results.
13:31 The "Expand" Algorithm: This function starts with user-picked seed particles and searches for neighboring candidates based on a defined lattice geometry (e.g., vectors between particles). Candidates are accepted or rejected based on a cross-correlation score against a reference map. This process repeats iteratively until it reaches the lattice edge or a user-defined limit.
17:25 Preparing for "Expand": The user must provide (1) a few sparsely distributed seed particles with their coordinates and orientations, and (2) a parameter file defining the lattice vectors, search ranges, and a cross-correlation threshold for accepting new particles.
33:46 Neural Network Picker: This tool treats particle picking as a segmentation problem using a U-Net architecture.
Training: The user manually picks a small number of particles (~13 in the ribosome example) to create a training set of density patches and corresponding binary masks.
Prediction: The trained model is applied to the full tomogram to generate a probability map, which is then clustered to find particle centers. A "tolerance" parameter can be tuned to adjust the sensitivity, balancing true positives against false positives.
Curved Lattices: Handling significant curvature (e.g., in viral capsids) requires careful manual seeding in different views (top, side, bottom) and loosening the search parameters and cross-correlation thresholds. It's a manual tuning process.
Lattice Imperfections: The Expand tool can tolerate distortions and defects in the lattice. It simply stops picking when it encounters a broken region or a significant deviation from the defined geometry.
Training Data Centering: For the neural network, the manually picked training particles do not need to be perfectly centered. The algorithm is robust to minor, random offsets, as they average out during training.
Validating Results: To determine if gaps in a picked lattice are real defects or picking errors, the speaker recommends using TomoNet's "place back" function to visually overlay the picked particle locations onto the raw tomogram density.