This talk by Dr. Fei Wong introduces TomoNet, a GUI-based software package designed to streamline the cryo-electron tomography (cryo-ET) workflow, with a specific focus on automated particle picking. The software addresses challenges in managing large datasets and integrating multiple analysis tools. TomoNet features two main particle picking methods. The first, "Expand," is a geometry-based tool optimized for accuracy in picking particles arranged in 1D, 2D, or 3D lattices. It operates by starting from user-defined seed particles and iteratively searching for neighbors based on the lattice geometry. The second method is a more generalized neural network-based picker (using a U-Net architecture) that treats picking as a segmentation task, requiring minimal training data and no lattice assumption. The talk presents successful applications of these tools on various biological samples, including bacterial S-layers, viral capsids, and ribosomes, demonstrating their effectiveness in achieving high-resolution structures.
TomoNet: An Integrated Solution for Automated Particle Picking in Cryo-ET
0:02:23 - The Challenge in Cryo-ET: The speaker explains that achieving high-resolution structures via cryo-ET requires averaging thousands of particles (subtomograms), which involves collecting and processing hundreds of tomograms. This creates major challenges in data management and workflow efficiency.
0:08:16 - TomoNet as a Solution: TomoNet is a GUI-based platform designed to create an integrated and seamless pipeline. It combines existing tools (like IMOD and Relion) and new picking algorithms into a single interface, simplifying tomogram management and analysis.
0:11:08 - Two Automated Picking Tools: The talk focuses on TomoNet's two main automated particle picking tools:
Geometry-based ("Expand"): Designed for particles in a lattice. It is highly accurate but requires the sample to have a repeating structure and needs more user input to define the lattice.
Neural Network-based: A more general tool that works for isolated or irregularly distributed particles. It is faster to set up but can be less accurate and harder to control for false positives.
0:13:31 - How "Expand" Works: This geometry-based method starts with a few user-selected "seed" particles. It then iteratively searches for neighboring particles based on user-defined vectors that describe the lattice geometry. A cross-correlation score is used to validate and accept new particles, stopping when it reaches the edge of a lattice or a defect.
0:27:22 - Applications of "Expand": The speaker shows several examples where "Expand" successfully picked particles from 1D (microtubules), 2D (viral capsids, S-layers), and 3D (paracrystalline rods) lattices, leading to high-resolution structures. It is shown to be robust to significant curvature and defects in the lattice.
0:33:46 - How Neural Network Picking Works: This method uses a U-Net architecture to treat particle picking as an image segmentation problem. The user manually picks a small number of particles (~13 in the ribosome example) to create a training set. The trained model then identifies and extracts the coordinates of all similar particles in the tomograms.
0:40:02 - Applications of Neural Network Picking: This approach is very fast for samples with distinct features, such as ribosomes. In one example, training and prediction for an entire tomogram took only about one minute each. It has also been used successfully for more complex targets like viral glycoproteins.
0:42:26 - Availability: TomoNet is an open-source tool available on GitHub, with tutorials and documentation for users.