Topaz

Topaz detects particles in cryo-electron microscopy (cryo-EM) micrographs using convolutional neural networks (CNNs) and a positive-unlabeled learning framework to improve particle identification for protein structure determination.

Key Features:

• Positive-Unlabeled Learning Framework: Employs positive-unlabeled learning to train models with sparsely labeled particles without requiring labeled negative examples, reducing the need for extensive manual labeling.
• Convolutional Neural Networks (CNNs): Uses CNNs for automated particle detection in cryo-EM micrographs.
• High Accuracy and Low False-Positive Rates: Retrieves a substantially higher number of real particles compared to conventional methods while maintaining low false-positive rates.
• Detection of Challenging Particle Shapes: Excels at detecting small, non-globular, and asymmetric proteins that are difficult for traditional algorithms.
• Robust Performance Across Datasets: Demonstrates robustness on both difficult and conventional cryo-EM datasets.

Scientific Applications:

• Particle Picking for Single-Particle Analysis: Produces particle coordinates for downstream protein structure determination from cryo-EM data.
• Improving Particle Sets for Structure Determination: Increases yield of true particles while limiting false positives to improve quality of datasets used for structural analysis.
• Analysis of Small or Asymmetric Complexes: Enables detection of small, non-globular, or asymmetric protein complexes that challenge traditional pickers.
• Reduction of Manual Annotation Effort: Lowers the amount of manual particle labeling required by leveraging positive-unlabeled training.

Methodology:

Topaz trains convolutional neural networks using a general-purpose positive-unlabeled learning method applied to cryo-EM micrographs, and its performance has been demonstrated on both difficult and conventional datasets.