PaperlayerView exact revision on arXiv

Source-linked AI summary

A deep convolutional neural network approach to single-particle recognition in cryo-electron microscopy

Yanan Zhu, Qi Ouyang, Youdong Mao

arXiv:1605.05543v2physics.data-ancs.CV

TL;DR

Particle extraction is a practical bottleneck in cryo-EM, and template-based picking can depend on low-resolution references. DeepEM uses a deep CNN for template-free particle recognition, with improved precision-recall performance and greater tolerance of lower-SNR micrographs than template-matching methods.

  • Problem

    Cryo-EM reconstruction requires many particle images, making extraction laborious, while template-based picking can introduce reference-dependent bias.

    Source: Background · p. 3

  • Method

    DeepEM applies a recursively trainable deep CNN to automate particle picking, selection, and verification without templates.

    Source: Discussion · p. 14, Conclusion · p. 15

  • Results

    DeepEM shows improved precision-recall performance and increased tolerance to much lower SNRs than template-matching methods.

    Source: Conclusion · p. 15

  • Takeaways & Limitations

    The approach enables automated particle picking, selection, and verification from noisy micrographs without templates.

    Source: Conclusion · p. 15

    A dataset with particles of significantly different features may make previously trained hyper-parameters suboptimal.

    Source: Discussion · p. 14

Abstract

from arXiv · show

Background: Single-particle cryo-electron microscopy (cryo-EM) has become a popular tool for structural determination of biological macromolecular complexes. High-resolution cryo-EM reconstruction often requires hundreds of thousands of single-particle images. Particle extraction from experimental micrographs thus can be laborious and presents a major practical bottleneck in cryo-EM structural determination. Existing computational methods of particle picking often use low-resolution templates as inputs for particle matching, making it possible to cause reference-dependent bias. It is critical to develop a highly efficient template-free method to automatically recognize particle images from cryo-EM micrographs. Results: We developed a deep learning-based algorithmic framework, DeepEM, for single-particle recognition from noisy cryo-EM micrographs, enabling automated particle picking, selection and verification in an integrated fashion. The kernel of DeepEM is built upon a convolutional neural network (CNN) of eight layers, which can be recursively trained to be highly "knowledgeable". Our approach exhibits improved performance and high precision when tested on the standard KLH dataset. Application of DeepEM to several challenging experimental cryo-EM datasets demonstrates its capability in avoiding selection of un-wanted particles and non-particles even when true particles contain fewer features. Conclusions: The DeepEM method derived from a deep CNN allows automated particle extraction from raw cryo-EM micrographs in the absence of templates, which demonstrated improved performance, objectivity and accuracy. Application of this novel approach is expected to free the labor involved in single-particle verification, thus promoting the efficiency of cryo-EM data processing.

Background

Particle recognition is a practical bottleneck because cryo-EM needs many images and micrographs are noisy, while existing automated approaches can struggle to distinguish usable particles from contaminants. DeepEM applies an eight-layer CNN to template-free particle recognition, aiming to improve accuracy across varied micrograph conditions.

Background

  • Particle extraction is a bottleneck for cryo-EM reconstruction, which requires large numbers of single-particle images from noisy, low-contrast micrographs.Limited electron dose is used to reduce radiation damage.Source: Background · p. 3
  • Template-matching methods depend on local cross-correlation sensitive to noise, while edge- and feature-based methods lose performance as micrograph contrast falls.A prior three-layer pyramidal network had one hidden layer, considered insufficient for extracting rich single-particle features.Source: Background · p. 3
  • Automated methods may fail to distinguish good particles from overlaps, aggregates, noise fluctuations, ice contamination, and carbon areas, leaving further classification or manual selection necessary.These difficulties add work after initial particle picking.Source: Background · p. 3
  • DeepEM uses an eight-layer CNN for template-free particle recognition, targeting good-particle detection across varied situations with improved accuracy over existing methods.The described architecture includes an input layer, three convolutional layers, three subsampling layers, and an output layer.Source: Background · p. 4

Methods

DeepEM uses a convolutional neural network with convolution, subsampling, and fully connected layers. Its architecture extracts feature maps through local operations and progressively reduces their spatial size.

Methods

  • Convolution applies learnable kernels to local inputs across the image, producing feature maps through shared connectivity weights.The convolutional layer's shared weights and bias combine inputs from feature maps in the previous layer.Source: Design of the DeepEM model · pp. 4–5
  • Subsampling after each convolutional layer reduces representation size, parameter count, and computation, and is intended to help control over-fitting.The network computes a subsampling average after each convolutional layer.Source: Design of the DeepEM model · p. 5
  • The DeepEM architecture alternates three convolutional layers with three subsampling layers, followed by a fully connected output layer that predicts the input image's class.The convolutional layers are the first, third, and fifth layers; the subsampling layers are the second, fourth, and sixth.Source: Design of the DeepEM model · p. 5

Training of the DeepEM model

DeepEM is trained on manually assembled particle and non-particle images, then can be retrained with representative examples added after independent testing. This iterative process addresses the difficulty of assembling balanced data across heterogeneous objects and particle views.

Training of the DeepEM model

  • Training uses positive particle images and negative non-particle images, with network outputs of 1 and 0, respectively.Particle images are rotated by 90°, 180°, and 270° to augment the training data fourfold.Source: Training of the DeepEM model · pp. 5–6
  • The model minimizes squared-error loss between each training image's target and its output-layer response using error back-propagation and gradient-based updates.The learning rate is reported as set to 1 in this study.Source: Training of the DeepEM model · p. 6
  • Heterogeneous objects and different projections of the same molecule make it difficult to assemble balanced representative training data, risking missed particles or mistaken detections.The described objects include impurities, ice, carbon, overlaps, and local aggregates.Source: Training of the DeepEM model · p. 6
  • After testing on independent micrographs, representative particle images can be added to the original training set and the network retrained.The workflow is presented as a way to improve differentiation between positive and negative particles.Source: Training of the DeepEM model · pp. 6–7

Particle recognition and selection in the DeepEM model

DeepEM scans micrographs with image boxes, averages CNN scores across four orientations, and applies score and pixel-variation criteria to select particles. The score threshold is chosen using F2, which weights recall more heavily than precision.

Particle recognition and selection in the DeepEM model

  • Each boxed region from a testing micrograph is scored as four separate CNN inputs—the unrotated image and its 90°, 180°, and 270° rotations—and their output values are averaged as the box raster-scans the whole micrograph.Source: Particle recognition and selection in… · p. 7
  • The threshold is set at the highest F2 score, which weights recall higher than precision, and candidate images below the CNN-score threshold are rejected.F-measure combines precision and recall for binary classification.Source: Particle recognition and selection in… · pp. 7–8
  • A narrow pixel-intensity standard-deviation range further filters candidates to remove particles in carbon or contaminant areas.Such areas tend to have higher or lower pixel standard deviations than clean amorphous ice.Source: Particle recognition and selection in… · p. 8

DeepEM algorithm workflow

DeepEM trains a CNN through iterative testing and dataset or hyperparameter updates, then scans micrographs with rotated image patches to select particles. Performance is assessed by varying the selection threshold and examining precision–recall trade-offs.

DeepEM algorithm workflow

  • Training assigns positive and negative examples outputs of 1 and 0, then updates CNN weights by backpropagation and tests learning error on independent data.The training set or hyperparameters can be adjusted and training repeated until the testing error falls below a cutoff.Source: DeepEM algorithm workflow · p. 8
  • DeepEM scans micrographs with overlapping square patches, averages CNN scores across four rotations, and selects local-maximum candidates above a threshold.Selected images are filtered by standard deviation, then their coordinates are written to box files.Source: DeepEM algorithm workflow · p. 9

Performance evaluation

  • Precision measures the fraction of selected images that are true particles, while recall measures the fraction of all true particles selected.Changing the score threshold trades precision against recall, so the curve evaluates their joint performance across thresholds.Source: Performance evaluation · pp. 9–10

DeepEM training on the KLH dataset

The KLH training set used manually selected particles and randomly chosen non-particle images, with rotations augmenting the examples. Independent testing informed iterative hyperparameter and training-data updates.

DeepEM training on the KLH dataset

  • The training set used 800 manually selected particle images as positive examples and the same number of randomly selected non-particle images as negative examples.Source: DeepEM training on the KLH dataset · p. 10
  • Four 90-degree rotations augmented each training image, producing four times as many positive and negative examples.The passage describes rotation-based augmentation of both classes.Source: DeepEM training on the KLH dataset · p. 10
  • Positive and negative particle images excluded from prior training tested an intermediately trained CNN; its test-set accuracy or error fed back into hyperparameter tuning and training-dataset updates.Source: DeepEM training on the KLH dataset · p. 10

Application to experimental cryo-EM data

On KLH, DeepEM reached about 90% precision and recall together and outperformed template-matching methods in precision–recall comparisons. Increasing training examples improved performance up to a point.

Application to experimental cryo-EM data

Experiments on the KLH dataset

  • About 90% precision and recall were reached together on KLH, and DeepEM outperformed RELION and TMACS in precision–recall comparisons.The test used a manually selected set from 32 micrographs that did not include training particles; precision was higher while recall remained high.Source: Experiments on the KLH dataset · pp. 11–12
  • Increasing KLH training particles from 100 to 400 improved the precision–recall curve, with little further improvement at larger sizes; the best result used 800 positives.The tested range was 100 to 1,200 training particles.Source: Experiments on the KLH dataset · p. 12

Experiments on challenging cryo-EM datasets

DeepEM recognized particles in challenging cryo-EM datasets, including difficult inflammasome cases and 19S micrographs, while avoiding non-particles. Runtime varied by dataset and image size.

Experiments on challenging cryo-EM datasets

  • On 19S micrographs, precision and recall both reached about 80%, with nearly all true particles selected while avoiding non-particles in aggregates and carbon-film areas.Picked particles were approximately as well centered as manually boxed particles.Source: Experiments on challenging cryo-EM… · p. 12
  • In two difficult inflammasome cases, DeepEM performed well at particle recognition despite excessive particle overlap, ice, and lower-SNR side views, while avoiding selection of overlapped particles and non-particles.Source: Experiments on challenging cryo-EM… · p. 13

Computational efficiency

  • For the proteasome datasets, processing took about 790, 560, and 1,160 seconds per micrograph, with binned micrographs measuring 1,855 by 1,919 pixels.Source: Computational efficiency · p. 13

Discussion

DeepEM selects true particles without non-particles in one step and is reported to perform as well as a human worker. Its performance depends on training and image features, while new datasets may require additional hyperparameter tuning.

Discussion

  • DeepEM automates particle extraction from raw cryo-EM micrographs, with the stated goal of improving data-processing efficiency.The method is presented as a deep-CNN approach to single-particle recognition.Source: Discussion · p. 14
  • New datasets with significantly different features can make trained hyperparameters suboptimal, requiring user-led empirical screening that may be laborious.The authors note that automated systematic hyperparameter tuning could be developed in the future.Source: Discussion · p. 14
  • DeepEM was more effective on feature-rich inflammasome images; side views performed worse than top views and had fewer low-frequency features.The authors also examined model noise tolerance using simulated datasets.Source: Discussion · p. 14
  • At an SNR of 0.01, DeepEM could still recognize particle images well after proper training.The result was reported for simulated datasets.Source: Discussion · p. 14
  • DeepEM selects true particles without non-particles in one recognition step and performs as well as a human worker.Previously, comparable performance required combining automated picking, unsupervised classification, and manual curation.Source: Discussion · p. 14
  • Iteratively updating or optimizing the training dataset can further improve DeepEM’s particle-recognition performance.The authors contrast this with conventional recognition algorithms whose performance was bound by their mathematics and control parameters.Source: Discussion · p. 14

Conclusion

DeepEM is presented as a template-free tool for extracting particles from noisy micrographs, with improved precision-recall performance and tolerance to lower SNRs than template-matching methods.

Conclusion

  • DeepEM supports automated particle picking, selection, and verification, with improved precision-recall performance and greater tolerance to lower SNRs than template-matching methods.The authors expect this integrated approach to reduce labor in single-particle verification and promote cryo-EM data processing efficiency.Source: Conclusion · p. 15

Abbreviations

  • The paper defines CNN as convolutional neural network, Cryo-EM as cryo-electron microscopy, SNR as signal-to-noise ratio, AI as artificial intelligence, and KLH as Keyhole Limpet Hemocyanin.Source: Abbreviations · p. 15

Funding

  • The work reports support from the Thousand Talents Plan of China, the National Natural Science Foundation of China, and the Intel Parallel Computing Center program.It also reports NSF award no. 1541959 supporting the Center for Nanoscale Systems facilities used in part for the cryo-EM experiments.Source: Funding · p. 16

Availability of data and material

Authors’ contributions

  • The manuscript assigns experiment design to YZ, QO, and YM; experiments to YZ; and data analysis to YZ and YM.It credits QO and YM with reagents, materials, and analysis tools, and ZY and YM with writing the paper.Source: Authors’ contributions · p. 16

Consent for publication

These passages describe DeepEM’s CNN architecture and workflow, then present figures and a table covering KLH and 19S results, training-set size, challenging examples, SNR, and dataset hyperparameters.

Loading 1605.05543v2…