Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs (1803.08207v2)

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 {\AA} 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.

• The paper introduces a PU-learning framework using the GE-binomial criterion, which minimizes manual labeling by training on a few positive examples.
• It achieves significant performance gains, improving resolution by up to 0.15 Å and detecting up to 3.68 times more particles than manual annotations.
• Topaz's modular design, from micrograph preprocessing to CNN-based segmentation, offers wide applicability in cryo-EM and other imaging fields.

Positive-Unlabeled Convolutional Neural Networks for Particle Picking in Cryo-Electron Micrographs

The paper presents a novel approach for automated particle picking in cryo-electron microscopy (cryoEM) utilizing a positive-unlabeled (PU) learning framework, implemented in a system called Topaz. CryoEM is a valuable technique for elucidating high-resolution protein structures but poses challenges for automated particle picking due to variability in particle shapes and the low signal-to-noise ratio prevalent in the data. Current methods, such as Difference of Gaussians (DoG) and template-based techniques, often suffer from high false-positive rates and limitations in detecting non-spherical particles.

The authors propose a PU learning paradigm to circumvent the need for extensive manually labeled datasets, which are cumbersome and time-intensive to produce. PU learning leverages a small set of positively labeled particle images while treating the rest as unlabeled data. The core advancement in their approach is the development of the GE-binomial criterion, which enhances the PU learning framework by minimizing overfitting and improving classifier accuracy. This objective function is particularly effective in handling the noisy and complex backgrounds typical of cryoEM micrographs.

Topaz was evaluated on multiple publicly available cryoEM datasets, exhibiting significant improvements in the structural resolution of protein reconstructions without requiring post-processing. Specifically, employing only 1,000 labeled examples, Topaz achieved an increase in resolution by up to 0.15 Å over existing curated datasets, a notable achievement that underscores its practical efficacy. In datasets like EMPIAR-10025 and EMPIAR-10028, Topaz detected more particles than prevailing manual annotations, demonstrating its capability to identify high-quality particles that may be overlooked by manual methods. This was particularly evident where Topaz discovered 3.22, 1.72, and 3.68 times more particles in selected datasets, markedly enhancing the reconstruction quality.

The paper scrutinizes the performance of the Topaz pipeline, detailing its three integral components: micrograph preprocessing, PU learning-based classifier training, and micrograph segmentation. Preprocessing involves downsampling and normalization to reduce noise and standardize micrograph intensity, enhancing the efficacy of the convolutional neural networks (CNNs) employed for particle identification. The CNN classifiers are designed to discern between particle and non-particle regions using the limited labeled data facilitated by the GE-binomial objective function.

On a theoretical level, the implications of applying PU learning to particle picking in cryoEM are profound. It illustrates how constraints on expectations over unlabeled data can significantly reduce the requirement for negative labeled data, a prominent bottleneck in existing CNN-based methods. The introduction of autoencoder-based regularization further advances this framework by improving performance with minimal labeled data, mitigating overfitting, and refining detection precision.

The practical utility of Topaz extends beyond structural biology, offering potential applications in light microscopy and medical imaging where similar challenges with labeling and noise are encountered. Moreover, the modular and flexible design of Topaz, with source code available under an open-source license, paves the way for its integration into existing pipelines such as Appion and other cryoEM software suites, facilitating broader adoption and continued development.

In conclusion, the integration of PU learning in the Topaz pipeline represents a significant contribution to cryoEM analysis. By leveraging a minimal number of labeled particles, this methodology effectively optimizes structural resolution and provides the cryoEM community with a robust tool for accelerating the pathway from image acquisition to high-resolution structural determination. Future developments could explore expanding the applicability of GE-based PU learning to other domains and enhancing particle detection algorithms through integration with advanced neural network architectures.