Source-linked AI summary
Convolutional Neural Networks for Automated Annotation of Cellular Cryo-Electron Tomograms
Muyuan Chen, Wei Dai, Ying Sun, Darius Jonasch, Cynthia Y He, Michael F. Schmid, Wah Chiu, Steven J Ludtke
TL;DR
The method addresses the challenge of identifying molecular features in crowded cellular cryo-electron tomograms. It trains convolutional neural networks on manually annotated examples, processes tomograms efficiently, and supports particle extraction and subtomogram averaging, while performance remains constrained by training scope and detectable feature scale.
Problem
Identifying molecular features in crowded cellular tomograms requires considering surrounding pixels rather than each pixel independently.
Method
The method trains a four-layer convolutional neural network on manually annotated positive and negative tiles, then applies, normalizes, and merges network annotations across tomograms.
Results
The annotation outputs support extraction of discrete and continuous particles for classification and subtomogram averaging, while four CNNs each identified approximately 800–900 ribosomes in a test tomogram.
Takeaways & Limitations
The workflow reduces annotation effort by processing an entire tomogram slice simultaneously and enables downstream in-situ structural analysis through particle extraction and averaging.
Takeaways & Limitations
The network’s maximum detectable feature scale is 30 pixels, and performance is unpredictable for features not specifically represented in the training set.
Abstract
from arXiv · showhide
Cellular Electron Cryotomography (CryoET) offers the ability to look inside cells and observe macromolecules frozen in action. A primary challenge for this technique is identifying and extracting the molecular components within the crowded cellular environment. We introduce a method using neural networks to dramatically reduce the time and human effort required for subcellular annotation and feature extraction. Subsequent subtomogram classification and averaging yields in-situ structures of molecular components of interest.
Figure Captions:
The method trains lightweight CNNs on manually annotated 2D tomogram tiles, applies them slice-by-slice, and merges feature-specific outputs for annotation and subtomogram extraction.
- Training sets: Ten positive and roughly 100 negative 64x64-pixel tiles provide the training examples for each feature-specific CNN.Positive tiles are manually segmented; negative tiles require no annotation and can be selected rapidly.
- Neural network structure: The four-layer CNN uses convolution, max pooling, wider kernels, and a final output neuron to recognize features across multiple scales.The pooling layer enables receptive fields covering features up to 30x30 pixels.
- Applying trained CNNs: The trained network processes entire pre-filtered tomogram slices rather than separate tiles for every pixel, reducing computational complexity.Full-sized slices are filtered using the trained kernels and matrix operations combined with FFTs.
- Post processing and merging features: Outputs from independently trained CNNs are normalized and merged by assigning each voxel to the network with the highest value, with threshold 1 used for display and extraction.This produces a multi-component annotation while supporting particle identification for subtomogram averaging.
- Particle extraction and averaging: Discrete objects are extracted from connected annotated regions, while continuous features use randomly seeded points; EMAN2 classification and alignment support averaging.Z-axis projections help remove bad particles before 3D alignment and averaging.
- Reusability: A trained CNN can be reused across tomograms with the same voxel size and preprocessing, preferably within the same cell type and similar imaging conditions.The authors report robustness to reasonable differences in defocus or signal-to-noise ratio, but feature sensitivities are difficult to quantify.