Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs

Tristan Bepler; Andrew Morin; Julia Brasch; Lawrence Shapiro; Alex J. Noble; Bonnie Berger

Source-linked AI summary

Positive-unlabeled convolutional neural networks for particle picking in cryo-electron micrographs

Tristan Bepler, Andrew Morin, Julia Brasch, Lawrence Shapiro, Alex J. Noble, Bonnie Berger

arXiv:1803.08207v2Mar 22, 2018q-bio.QMcs.CVstat.ML

TL;DR

Particle picking is a time-consuming bottleneck, particularly for small or non-globular particles. Topaz uses convolutional neural networks and positive-unlabeled learning to pick particles, outperforming other PU approaches and improving structure resolution without post-processing.

Problem
Particle picking is a major bottleneck, often taking weeks or months for small or non-globular particles.
Method
Topaz uses convolutional neural networks and positive-unlabeled learning to train a particle-picking pipeline from few labeled particles and remaining unlabeled data.
Results
Topaz's generalized-expectation approach outperforms other PU techniques for particle detection and improves structure resolution without time-consuming post-processing.
Takeaways & Limitations
Topaz enables particle picking for unusually shaped particles with low signal-to-noise.
Takeaways & Limitations
Particle detection remains challenging in micrographs with low signal-to-noise ratio and structured backgrounds.

Abstract

from arXiv · show

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.

1 Computational and Systems Biology, MIT, Cambridge, MA, USA … Introduction

Topaz frames cryoEM particle picking as positive-unlabeled learning, using few labeled positive regions together with unlabeled micrograph regions to train convolutional neural networks. With 1,000 labeled examples, Topaz improved reconstruction resolution by up to 0.15 Å without ad hoc post-processing and performed well on challenging non-globular particles.

3 Department of Mathematics, MIT, Cambridge, MA, USA: The authors’ affiliations include Computational and Systems Biology, Computer Science and Artificial Intelligence Laboratory, and Mathematics at MIT, plus the National Resource for Automated Molecular Microscopy and the Zuckerman Mind Brain Institute.The supplied affiliation passages also identify the New York Structural Biology Center.
Introduction: Weeks or even months of manual particle picking remain necessary for small or non-globular particles because of variable shapes and structured noise.Accurate particle identification is a major bottleneck in cryoEM workflows.
Introduction: DoG methods are limited to spherical, compact particles of known size, while template derivation from known structures can produce many false positives or fail for atypical particles.These limitations motivate learned approaches for more diverse particle morphologies.
Introduction: Topaz reframes particle picking as positive-unlabeled learning, training a classifier with a small number of labeled positive regions and remaining unlabeled micrograph regions.This approach addresses the limitations of workflows requiring extensive positive and negative labeling.
Introduction: A novel generalized expectation criteria constrains PU learning to mitigate overfitting and train high-accuracy particle classifiers with very few labeled data points.Autoencoder-based regularization can further reduce the labeled data required for high performance.
Introduction: Topaz predictions were fed directly into alignment and reconstruction without any ad hoc post-processing, while maintaining a remarkably low false positive rate at a relaxed probability threshold.The relaxed threshold enabled retrieval of large particle sets.
Introduction: The GE-based PU approach enabled particle picking with minimal labeled examples, including on a challenging dataset containing stick-like particles with low signal-to-noise.The study compared GE-based PU learning with other recent PU learning methods.

Results · 1. The Topaz Pipeline

Topaz is an end-to-end particle-picking pipeline that combines micrograph preprocessing, positive-unlabeled CNN training, and coordinate extraction. Its PU framework leverages many unlabeled regions alongside few labeled particles to train accurate classifiers for challenging cryoEM datasets.

1. The Topaz Pipeline: Topaz comprises three steps: micrograph preprocessing, PU-learning neural-network classifier training, and particle-coordinate extraction by non-maximum suppression.The implementation is freely available for academic use.
Micrograph preprocessing: Micrographs are downsampled to reduce pixel-level noise and improve training and particle-extraction efficiency.They are then normalized with a per-micrograph scaled two-component Gaussian mixture model.
Micrograph preprocessing: The Gaussian mixture normalization captures large-scale intensity differences between micrographs and pixel-value bimodality from dark grid sections.This corrects variation caused by micrograph-specific intensity differences and dark grid regions.
Classifier training from positive and unlabeled data: Topaz frames particle picking as PU learning, using few labeled particles and many unlabeled micrograph regions to discriminate particles from non-particles.Leveraging unlabeled data is identified as the core advancement of Topaz.
Classifier training from positive and unlabeled data: CNN classifiers use minibatched stochastic gradient descent with the GE-binomial objective, which models minibatch sampling statistics and regularizes predictions on unlabeled data.An optional autoencoder module supports training with very few positive examples.
Classifier training from positive and unlabeled data: The approach helps overcome overfitting associated with recent neural-network PU methods and enables particle picking in challenging cryoEM datasets.The paper states that these capabilities are demonstrated in experiments.
Micrograph segmentation and particle extraction: Topaz convolves the trained classifier over each micrograph to produce per-pixel probabilities, then applies non-maximum suppression to select high-scoring particle centers.The output is a list of predicted coordinates and associated model scores for each micrograph.

2. Topaz improves structure resolutions with no postprocessing

Topaz’s raw particle picks produced equal or better reconstructions than curated or published sets across three cryoEM datasets, despite using only 1,000 positives and no postprocessing. It retrieved substantially more particles, improved resolution on two datasets, matched the published resolution on aldolase, and yielded nearly identical structures from newly picked particles alone.

Evaluation: Topaz selected particle sets by optimizing reconstruction resolution across decreasing probability thresholds, using raw picks without postprocessing.The reconstructions used identical cryoSPARC homogeneous-refinement settings for Topaz and published particle sets.
Particle yield: 3.22, 1.72, and 3.68 times more particles were found for EMPIAR-10025, EMPIAR-10028, and NYSBC-aldo, respectively.Topaz was trained with 1,000 positives, while comparisons used curated particle sets.
Resolution gains: ~0.15 Å and ~0.05 Å resolution improvements were achieved for EMPIAR-10025 and EMPIAR-10028, respectively, despite no Topaz particle postprocessing.All predicted particle coordinates were fed directly into reconstruction.
Resolution gains: 3.0 Å was achieved for EMPIAR-10028, improving resolution by 0.1 Å over the best structure reported in the EM map challenge.This result came from Topaz trained using 1,000 initial examples.
Resolution gains: 2.63 Å by FSC0.143 was achieved for NYSBC-aldo by both Topaz and published particle sets, although Topaz found many more particles.The published set’s ~200k particles were sufficient to reach the data’s resolution limit with standard reconstruction methods.
Validation: Reconstructions using only Topaz’s newly picked particles produced nearly identical structures after removing the entire published particle set.This directly tested whether the additional particles found by Topaz were good particles.

3. Topaz particle predictions are well-ranked and contain few false positives · 4. Our GE criteria based PU learning framework outperforms other PU learning methods

Topaz ranks particles effectively: resolution improves as good particles are added, then declines when relaxed thresholds admit false positives, while noisy or unusual classes appear only at low thresholds. GE-criteria PU learning methods outperform NNPU and PN, with GE-binomial and GE-KL usually similar but differing by dataset and label count.

3. Topaz particle predictions are well-ranked and contain few false positives: Topaz reconstruction resolution increases as more good particles are included, then drops when low thresholds admit too many false positives.The final-threshold resolution dip is demonstrated for EMPIAR-10025.
3. Topaz particle predictions are well-ranked and contain few false positives: Topaz particles quickly match published-particle resolution for the proteasome and ribosome, while aldolase requires more particles before matching and exceeding the curated set.The aldolase difference may reflect insufficient side views until the probability threshold is lowered, unlike the curated set enriched for those views.
3. Topaz particle predictions are well-ranked and contain few false positives: False-positive fractions increase as probability thresholds decrease, but the number of false positives remains remarkably low even at relaxed thresholds.Particles are classified into ten classes and manually assessed through 2D class averages.
3. Topaz particle predictions are well-ranked and contain few false positives: Noisy or unusual particle classes appear only at low thresholds; in T20S, particles near contaminating gold are selected only as the threshold decreases.Gold particles appear as dark spots in the micrographs.
4. Our GE criteria based PU learning framework outperforms other PU learning methods: The GE approaches constrain classifiers using unlabeled data by matching classifier expectations to the positive-class prior through GE-KL or GE-binomial criteria.GE-binomial uses cross entropy between classifier posteriors and a Binomial prior over positives per minibatch.
4. Our GE criteria based PU learning framework outperforms other PU learning methods: GE-binomial and GE-KL generally perform similarly, but GE-binomial significantly outperforms GE-KL on EMPIAR-10096 with 10 positive examples.At larger numbers of positives on EMPIAR-10096, both methods are statistically equivalent.
4. Our GE criteria based PU learning framework outperforms other PU learning methods: On Shapiro-lab, GE-binomial significantly outperforms GE-KL at 1,000 labeled examples, whereas GE-KL performs better within the 50-250 labeled-example range.Both differences are reported as significant with p<0.05.

Discussion

Topaz uses positive-unlabeled learning to enable accurate cryoEM particle picking from few labeled positives, including unusually shaped, low-signal particles. Its GE-based approach, optional autoencoder, and efficient workflow reduce manual effort while supporting improved structure resolution without post-processing.

Contribution: Topaz is a particle picking pipeline using positive-unlabeled learning when only a small number of labeled positive examples are available.The method is designed for particle picking with incomplete positive labels.
Results: The GE-criteria approach outperforms other PU learning techniques for particle detection, while an autoencoder further improves performance when n < 500.The autoencoder benefit is observed in the regime of very few labeled data points.
Results: Topaz enables particle picking for unusually shaped particles with low signal-to-noise and improves structure resolution without time-consuming ad-hoc post-processing.These capabilities address difficult datasets and reduce manual processing in cryoEM structure determination.
Implementation: By reporting particle probabilities alongside coordinates, Topaz supports selecting particle sets of varying size through a probability threshold.Particles can be added iteratively by lowering the threshold until reconstruction resolution stops improving.
Impact: Topaz requires fewer labeled particles, performs independently of particle shape, and reduces manual effort for high-resolution structures.It opens automated picking to proteins that were previously difficult to locate computationally.
Impact: Training takes a few hours on a single GPU, predictions for hundreds of micrographs take only minutes, and trained models can be reused for new imaging runs of the same particle.The pipeline is computationally efficient and supports repeated application to the same particle.

Methods … 4. PU learning with generalized expectation criteria

The methods assemble and preprocess multiple cryoEM datasets, define positive-unlabeled learning baselines, and introduce GE-binomial to train neural classifiers by matching predicted and known positive proportions.

1. Dataset description: Published particle sets and aligned, summed micrographs were retrieved from EMPIAR for EMPIAR-10025, EMPIAR-10028, and EMPIAR-10096.Additional hand-labeled or curated particle sets and micrographs came from the Shapiro lab and New York Structural Biology Center.
1. Dataset description: Each dataset was downsampled to the resolution specified in table 1, normalized, and split into training and test sets at the micrograph level.The numbers of micrographs and labeled particles in each split were reported in table 1.
2. Micrograph normalization: Images were normalized with a per-image scaled two-component Gaussian mixture model whose parameters were estimated by expectation-maximization for each dataset.Pixels were then divided by the image scaling factor and standardized to the dominant mixture component.
3. PU learning baselines: PU learning uses labeled positive regions P and unlabeled regions U, with η denoting the fraction of positive examples within U.The classifier is trained to discriminate positive and negative regions from P and U.
3. PU learning baselines: When η is small, treating unlabeled examples as negatives can be effective, but the resulting objective can overfit because of poor classification-objective specification.The passage also describes unbiased and non-negative PU estimators proposed to reduce overfitting.
4. PU learning with generalized expectation criteria: The proposed alternative minimizes labeled-data classification loss while constraining the classifier’s expected output on U to equal the known positive fraction η.GE-KL imposes this constraint softly through KL divergence between the classifier expectation and η.
4. PU learning with generalized expectation criteria: GE-KL sample-based gradients are biased, so the authors propose GE-binomial, which minimizes the difference between minibatch positive-count and binomial distributions parameterized by η.The GE-binomial criterion is defined as cross entropy between these two distributions.
4. PU learning with generalized expectation criteria: Because exact computation of the minibatch positive-count distribution is slow, the method substitutes a Gaussian approximation using its mean and variance.The resulting objective combines labeled classification loss with the GE-binomial cross-entropy term.

5. Autoencoder-based classifier regularization · 6. Classifier and autoencoder architectures and hyperparameters

Topaz regularizes its classifier by adding a decoder that reconstructs input images from encoder features, penalizing expected reconstruction error in the GE-binomial objective. The implementation uses a three-layer strided convolutional classifier and a DCGAN-like transpose-convolutional decoder.

5. Autoencoder-based classifier regularization: The classifier is split into an encoder f containing all layers except the final linear layer and a linear classifier c, forming the full network g.
5. Autoencoder-based classifier regularization: A deconvolutional decoder d takes encoder features and returns a reconstruction of the input image.
5. Autoencoder-based classifier regularization: The objective adds a weighted term penalizing expected reconstruction error over all images in dataset D.
5. Autoencoder-based classifier regularization: This augmented objective is the full GE-binomial objective function with the autoencoder component used in Topaz.
6. Classifier and autoencoder architectures and hyperparameters: The classifier is a three-layer strided convolutional neural network using batch normalization and PReLU activations.
6. Classifier and autoencoder architectures and hyperparameters: Its layers comprise 32 conv7x7 filters, 64 conv5x5 filters, 128 conv5x5 filters, and a final fully connected layer with one output.
6. Classifier and autoencoder architectures and hyperparameters: The decoder uses repeated transpose convolutions with batch normalization and leaky ReLU activations, expanding spatial dimensions while reducing feature dimensions until the original image size.
6. Classifier and autoencoder architectures and hyperparameters: Specifically, z is mapped through transpose conv4x4 layers with 128-d, 64-d, and 32-d features, then transpose conv3x3 1-d to produce X'.

7. PU learning benchmarking · 8. Classifier evaluation · 9. Non-maximum suppression algorithm for extracting particle coordinates

The study benchmarks PU classifiers using sampled positives and unlabeled particles, evaluates them with average precision, and extracts particle coordinates through greedy non-maximum suppression. The suppression radius is set from particle size but may be reduced for closely packed or irregular particles.

7. PU learning benchmarking: Classifiers are benchmarked by randomly sampling varying numbers of particles as positives and treating all remaining particles as unlabeled.Each positive-count condition uses 10 independent particle samples.
7. PU learning benchmarking: Training uses cross-entropy loss for labeled particles, with GE-binomial’s criteria weight set to 1 and GE-KL’s set to 10.The GE-KL setting follows the cited recommendation, while GE-binomial uses a weight of 1.
8. Classifier evaluation: Average precision evaluates classifiers by measuring how well micrograph regions are ranked by predicted particle probability.Average precision corresponds to the area under the precision-recall curve.
8. Classifier evaluation: Average precision is calculated across ranked regions using precision at retrieved elements multiplied by the change in recall.Precision is the fraction of correct predictions, while recall is the fraction of labeled particles retrieved among the top predictions.
9. Non-maximum suppression algorithm for extracting particle coordinates: Non-maximum suppression greedily selects coordinates and predicted particle probabilities starting from the highest-scoring region.The procedure prevents nearby pixels from generating additional particle candidates.
9. Non-maximum suppression algorithm for extracting particle coordinates: When a coordinate is selected, all pixels within a user-defined radius are excluded from consideration as particle candidates.This radius controls the spatial exclusion applied during coordinate extraction.
9. Non-maximum suppression algorithm for extracting particle coordinates: The default suppression radius is half the particle’s major-axis length, although smaller radii may work better for closely packed, irregularly shaped particles.The alternative is motivated by the limitation that the default radius can be too restrictive for such particles.

10. Micrograph pre-processing … 13. 3D structure analysis

The study preprocesses public and newly collected cryoEM micrographs, reconstructs particle sets with cryoSPARC, generates standardized 2D class averages, and evaluates final maps using Chimera and 3DFSC.

10. Micrograph pre-processing: Public EMPIAR micrographs used aligned and summed images and released CTF estimates, while EMPIAR-1002826 frames were aligned and summed without dose compensation.MotionCor2 was used for frame alignment, and public whole-micrograph CTF estimates were retained for EMPIAR-1002826.
11. 3D reconstruction procedure: Each particle set was reconstructed in cryoSPARC by single-class ab initio modeling followed by homogeneous refinement with dataset-specific symmetry.Symmetry was D7 for the T20S proteasome, C1 for the 80S ribosome, and D2 for aldolase; aldolase ab initio determination used C2 symmetry.
11. 3D reconstruction procedure: Reconstructions across decreasing Topaz score thresholds were independently calculated from particles predicted at or above each threshold.This procedure evaluated reconstruction quality for particle sets selected by thresholded Topaz scores.
12. 2D class averages: 2D class averages were generated in cryoSPARC using the “2D Classification” option with 10 classes for every particle set.All other classification settings remained at their defaults.
13. 3D structure analysis: Final 3D reconstructions were assessed visually in UCSF Chimera and quantitatively with 3DFSC for resolution range and anisotropy.Published or previous reconstructions, with fitted PDB structures when available, were aligned to newly processed maps for visual comparison.

Code availability statement … Supplemental Figures

The paper provides public code, deposited datasets, reconstruction and particle-picking results, supplemental examples and analyses, and disclosures of competing interests.

Code availability statement: Topaz source code is publicly available on GitHub and licensed under the GNU General Public License v3.0.The repository is available at https://github.com/tbepler/topaz.
Data availability statement: Single-particle maps and masks for XXXX were deposited in EMDB under accession EMD-XXXX, while the NYSBC-aldo dataset was deposited in EMPIAR under EMPIAR-XXXX.The deposited materials include half maps, full sharpened maps, and masks for XXXX.
Competing financial interests: The authors declare no competing financial interests.
Tables: The dataset summary table lists cryoEM datasets and classifier-training hyperparameters, with each dataset downsampled and split into train and test sets at the whole-micrograph level.
Figures: Figure 1 outlines the single-particle cryoEM workflow from sample preparation and micrograph collection through particle picking, extraction, classification, and 3D alignment.Topaz focuses on the particle-picking stage.
Figures: Topaz trains CNNs with positive and unlabeled regions, convolves classifiers over micrographs, and extracts particle coordinates using non-maximum suppression.The pipeline illustration uses labeled particles as positive regions and all other regions as unlabeled.
Figures: Topaz picked 3.22x more particles for the T20S proteasome, resulting in a 0.16 Å increase in resolution, while additional particles produced nearly identical maps.The reconstruction comparison includes published particles, Topaz particles, and Topaz particles with published particles removed.
Supplemental Figures: Supplemental analyses show Topaz avoiding ice chunks and other artifacts while identifying particles missing from ground-truth or manually labeled sets, and report objective-function sensitivity, 2D classes, and precision-recall/F1 curves.The Shapiro-lab dataset is described as an extremely challenging low-signal-to-noise dataset involving a ~75kDa stick-lick protein; GE-binomial sensitivity uses 100 and 1000 labeled particles across 10 runs.