[Paper Review] Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs

Tristan Bepler, Andrew Morin+4 more|PubMed|Mar 22, 2018

Advanced Electron Microscopy Techniques and Applications2 references179 citations

TL;DR

This paper introduces Topaz, a particle picking pipeline for cryo-EM that uses positive-unlabeled (PU) learning to train CNNs with few labeled particles and many unlabeled ones, achieving higher reconstruction quality with minimal labeling.

ABSTRACT

Cryo-electron microscopy (cryoEM) is an increasingly popular method for protein structure determination. However, identifying a sufficient number of particles for analysis (often >100,000) can take months of manual effort. Current computational approaches are limited by high false positive rates and require significant ad-hoc post-processing, especially for unusually shaped particles. To address this shortcoming, we develop Topaz, an efficient and accurate particle picking pipeline using neural networks trained with few labeled particles by newly leveraging the remaining unlabeled particles through the framework of positive-unlabeled (PU) learning. Remarkably, despite using minimal labeled particles, Topaz allows us to improve reconstruction resolution by up to 0.15 Å over published particles on three public cryoEM datasets without any post-processing. Furthermore, we show that our novel generalized-expectation criteria approach to PU learning outperforms existing general PU learning approaches when applied to particle detection, especially for challenging datasets of non-globular proteins. We expect Topaz to be an essential component of cryoEM analysis.

Motivation & Objective

• Motivate efficient particle picking in cryo-EM to reduce manual annotation effort.
• Propose a learning framework that leverages unlabeled particles via positive-unlabeled learning.
• Develop a CNN-based particle picker (Topaz) trained with limited labeled data.
• Demonstrate reconstruction quality gains on public cryo-EM datasets without post-processing.
• Show superiority of a generalized-expectation PU criterion for challenging datasets.

Proposed method

• Develop Topaz, a CNN-based particle picker for cryo-EM.
• Train the network using Positive-Unlabeled (PU) learning to exploit unlabeled particles.
• Introduce a generalized-expectation criterion for PU learning.
• Apply the method to three public cryo-EM datasets and compare reconstruction outcomes.
• Evaluate improvements in reconstruction resolution without post-processing.

Experimental results

Research questions

• RQ1 Can PU learning enable accurate particle picking with very few labeled examples in cryo-EM?
• RQ2 Does leveraging unlabeled particles improve picking accuracy and downstream reconstruction quality?
• RQ3 How does the generalized-expectation PU criterion perform relative to existing PU methods on non-globular particles?
• RQ4 What are the practical reconstruction accuracy gains achieved by Topaz on public datasets?

Key findings

• Topaz achieves improved reconstruction resolution by up to 0.15 Å over published particles on three public datasets.
• PU learning enables high-quality particle detection with minimal labeled data.
• The generalized-expectation PU criterion outperforms existing PU approaches for challenging, non-globular particles.
• Topaz reduces or eliminates the need for ad-hoc post-processing in particle picking.
• The method is presented as an essential component for cryo-EM analysis.