structural-biology/ai produced the result/Communications Biology 2019 · v2
Neural network locates single protein particles in cryo-electron microscope images
Researchers built crYOLO, a convolutional neural network that scans cryo-electron microscopy images and marks the positions of individual protein particles, replacing a step normally done by hand or by conventional picking algorithms.
spectrum · one line per step, placed by what the step does · bright lines used AI
SPHIRE-crYOLO is a fast and accurate fully automated particle picker for cryo-EM
Communications Biology, 2019
doi:10.1038/s42003-019-0437-z · record aix-00008 v2 · checked 2026-10-07
- AI was for
- Detection
- Model family
- Convolutional neural network
- Checked by
- Benchmark
- Code
- available
The finding the paper is about came from the AI.
What this research was about
To work out the shape of a protein, researchers can freeze a watery sample of it and photograph it with an electron microscope. Each image, called a micrograph, shows many copies of the molecule lying at random angles in the ice. Software then combines thousands of these views into a single three-dimensional map. Before that can happen, someone has to find the copies: to mark where in each noisy image a particle sits, and to avoid marking ice, dirt or empty background. The images are faint and grainy by design, because a stronger electron beam would damage the sample, so telling a particle from noise is genuinely hard.
The researchers set out to hand this marking step to a trained network. They took an object-detection approach from computer vision, in which a network looks at a whole image once and predicts boxes around the things it finds, and applied it to whole micrographs.
Where AI came in
The AI is the particle picker itself. A network of 21 convolutional layers, trained from scratch, divides a micrograph into a grid and predicts for each cell whether a particle centre lies there, along with the position and size of a box around it. Training used micrographs in which a person had marked the particles by hand; only the particles were labelled, and every other position counted as background. The images were varied on the fly during training by blurring, flipping, adding noise and changing contrast. Separate networks were trained for individual samples, and a general network was trained on pooled micrographs from many datasets, including simulated images and images containing only contamination.
The trained networks were then run across full sets of micrographs to output particle coordinates in formats that existing structural-biology software reads. Those coordinates fed the conventional later steps, in which particles are sorted into two-dimensional averages and refined into three-dimensional density maps. The general network was also used on samples it had not been trained on, namely RNA polymerase and glutamate dehydrogenase. The picks were scored against hand-marked images held back from training, using measures of how many picks were correct, how many particles were missed and how well the boxes overlapped the manual ones, and compared with a published benchmark set.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
crYOLO applies the YOLO convolutional object-detection approach to whole cryo-EM micrographs to locate single particles, predicting for each grid cell whether it contains a particle centre and regressing the box position and size. Networks trained on 200-2500 particles per dataset were used to pick the TcdA1, NOMPC and Prx3 datasets, and the picked particles were carried through 2-D classification with ISAC and 3-D reconstruction in SPHIRE; picking ran at up to five micrographs per second on a single GPU and about one second per micrograph on a multi-core CPU. Performance was also measured on simulated TRPC4 micrographs at five noise levels and on the published KLH benchmark, using precision, recall, area under the precision-recall curve and intersection over union against manual labels. A general network trained on 840 micrographs from 45 datasets was used to pick RNA polymerase and glutamate dehydrogenase, which were not in its training set.
How AI was used
A convolutional neural network with 21 convolutional layers and 5 max-pooling layers for feature extraction, a passthrough connection between layers 13 and 21, a dropout layer and a final 1 x 1 convolutional detection layer was trained to detect particle bounding boxes in cryo-EM micrographs. Only positive particle boxes were labelled; all other positions were treated as negative, and the loss weighted localisation, box size and object/no-object confidence terms. Training used backpropagation with the ADAM optimiser on manually picked micrographs, with on-the-fly augmentation by Gaussian and average blurring, flipping, added Gaussian noise, pixel dropout and contrast normalisation. For small particles the input micrograph was divided into overlapping patches, each rescaled to the 1024 x 1024 network input and classified in a single batch. Separate networks were trained per dataset, and a general network was trained on a pooled set of manually picked, simulated and particle-free contamination micrographs; an Inception-ResNet feature-extraction variant was trained on the same pooled set for comparison. The trained networks were then run over full micrograph sets to output particle coordinates, which were passed to conventional 2-D classification and 3-D refinement software.
The shape of the work
Structural · the record, drawn
no AI
Manually label particles for training
Cleaning, filtering, normalising or labelling data already obtained.
To train crYOLO we manually selected particles for initial training datasets.where the paper describes this · verbatim
AI
Train dataset-specific crYOLO network
Fitting model parameters, including fine-tuning an existing model.
The network was trained using backpropagation with the stochastic optimization procedure ADAM.where the paper describes this · verbatim
AI
Pick particles across full datasets
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
For TcdA1, crYOLO was trained on 10 micrographs with 1100 particles and selected 10,854 particles from 98 micrographs.where the paper describes this · verbatim
no AI
Score picks against held-out manual labels
Testing outputs against ground truth.
The scores were calculated on 20% of the micrographs that were used for manual selection, but not for training.where the paper describes this · verbatim
AI
Train general network on pooled datasets
Fitting model parameters, including fine-tuning an existing model.
we trained the network with a combination of 840 micrographs from 45 datasetswhere the paper describes this · verbatim
AI
Pick previously unseen datasets with general network
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
Using this generalized crYOLO network, we automatically selected particles of RNA polymerase (EMPIAR 10190) and glutamate dehydrogenase (EMPIAR 10217).where the paper describes this · verbatim
no AI
2-D classification and 3-D reconstruction from picked particles
Testing outputs against ground truth.
we additionally calculated two-dimensional (2-D) classes using the iterative stable alignment and clustering approach (ISAC)where the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's subject is the particle-picking network itself; the reported particle sets, and the reconstructions built from them, come from the trained model's output
CrYOLO trains a deep CNN for automated particle selection.where the paper describes this · verbatim
A common benchmark dataset for particle picking is a published set of cryo-EM images of KLH.where the paper describes this · verbatim
crYOLO is available as a standalone program under http://sphire.mpg.de/where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of crYOLO (YOLO-based 21-layer CNN)Which version of the model was used is not stated.
- Version of crYOLO general modelWhich version of the model was used is not stated.
- Version of crYOLO with Inception-ResNet feature extractorWhich version of the model was used is not stated.
- What step 2 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00008, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error