Source-linked AI summary
Reconstruction-Aware Cryo-EM Particle Picking
Riku Itsuji, Yuanhao Wang, Xingjian Li, Seonghui Min, Hideo Saito, Min Xu
TL;DR
The paper addresses the gap between independently optimized particle-picking stages and downstream reconstruction quality. It combines permissive picking, contamination masking, continuous 2D class selection, and feedback fine-tuning, achieving the best reconstruction among compared pickers while showing that 2D F1 does not reliably rank reconstructions.
Problem
Particle picking, contamination removal, and 2D class selection are typically trained and evaluated separately, even though their combination determines the reconstructed map.
Method
The pipeline combines high-recall CryoTransformer picking, MicrographCleaner contamination masking, continuous CryoSift class scoring, and fine-tuning on surviving particles.
Results
The pipeline gives the best reconstruction resolution among compared pickers on all four EMPIAR entries and improves over CryoTransformer by 0.63 to 3.57 ˚A.
Takeaways & Limitations
Particle selection is better treated as a reconstruction-aware pipeline judged by the map it delivers rather than by 2D F1 alone.
Takeaways & Limitations
The subtractive feedback loop can repeatedly discard candidates and remain below the base checkpoint’s F1, while evaluation 2D scores are not held out because annotated micrographs also train each round.
Abstract
from arXiv · showhide
Cryo-electron microscopy (cryo-EM) determines the structures of proteins and macromolecular assemblies at near-atomic resolution, and the final 3D reconstruction depends on extracting a clean particle stack from noisy micrographs. This extraction decomposes into three sub-tasks, namely particle picking, contamination removal, and 2D class selection. Each of them, however, is trained and evaluated in isolation, and none is optimized for the reconstruction. We instead integrate the three sub-tasks into a single pipeline posed against downstream reconstruction quality. We instantiate the pipeline with a state-of-the-art component for each sub-task, CryoTransformer picking permissively, MicrographCleaner masking contamination, and CryoSift selecting 2D classes by a continuous quality score, and close the loop with a fine-tuning step that returns the surviving particles to the picker. The pipeline achieves a better 3D resolution than every picker we compare. We also show that the best 2D F1 is not the best resolution, so particle selection is better treated as one reconstruction-aware pipeline judged by the map it delivers.
1. Introduction
Cryo-EM particle extraction is usually split into picking, contamination removal, and 2D class selection, although separately optimized stages do not necessarily yield good 3D reconstructions. The paper instead joins these stages in a reconstruction-aware pipeline and evaluates the delivered map.
- Low electron dose reduces particle–background contrast, while ice crystals and carbon edges create high-contrast false positives.
- Current practice separates particle picking, contamination removal, and 2D class selection into distinct sub-tasks.
- Separate optimization can miss rare views because a picker’s false-positive trade-off is judged before later purification stages remove outliers.
- Surviving particles can be mapped back to micrographs and reused as pseudo-labels, adapting the picker to target data without manual annotation.
- The pipeline combines CryoTransformer, MicrographCleaner, and continuous CryoSift class scores, then fine-tunes the picker on the purified set.
- 0.63 to 3.57 ˚A improvement over CryoTransformer across four entries yielded the best reconstruction among compared pickers.
2. Related Work
Prior work spans classical, CNN, transformer, and segmentation-based pickers, with feedback loops that retrain on selected 2D classes. This paper frames picking and purification as one reconstruction-scored system whose stages discard contamination and update a high-recall picker.
- Particle picking: CryoTransformer achieves high recall but incurs false positives and duplicate picks, while other modern approaches use CNN, transformer, or segmentation formulations.
- Feedback loops: Existing feedback loops retrain pickers on selected classes, but heuristic or binary decisions cannot express nuanced class-quality differences.
- Reconstruction-aware formulation: The paper defines a picker fθ and selection operator σ as one system, scoring the retained stack by reconstruction resolution and particle count.
- Purification and feedback: Contamination masking and 2D class selection only discard candidates, so the pipeline is designed for a high-recall picker and updates it using survivors.
- Purification and feedback: MicrographCleaner removes candidates at contaminated centers before CryoSift scores grouped 2D classes on a continuous 1.0-to-5.0 scale.
- Purification and feedback: Surviving particles are mapped back to micrograph coordinates and used as pseudo-labels for fine-tuning the picker each round.
4. Experimental Setup
The experiments use four CryoTransformer test-set entries, with annotated subsets supporting 2D metrics and full depositions supporting reconstruction. The primary outcome is gold-standard FSC resolution on reconstructed maps, alongside particle counts and auxiliary detection scores.
- Datasets: Four EMPIAR test entries provide the evaluation data, with CryoPPP annotations on 300 micrographs per entry and full unannotated depositions.
- Pipeline: The feedback loop uses CryoTransformer candidates, contamination masking, 2D class selection, and pseudo-label fine-tuning before full-set reconstruction.
- Datasets: Full depositions are used for reconstruction because 300 micrographs are insufficient for a stable reconstruction.
- Metrics: Gold-standard FSC between independently refined half-maps at the 0.143 criterion is the primary reconstruction-resolution measure.
- Metrics: Reconstructions use three random seeds, reporting the best GSFSC map, particle counts, and local-resolution maps as a qualitative check.
- Metrics: Precision, recall, and F1 against CryoPPP annotations are reported as auxiliary 2D detection metrics.
- Implementation: All conditions use the same CryoSPARC v4.7 reconstruction job chain.
5. Results
Across the ablations, 2D selection provides the largest resolution improvement, while contamination masking and feedback fine-tuning add smaller, dataset-dependent changes. The pipeline’s reconstruction outcomes therefore depend on how each stage affects the retained particle set and final map.
- Main results: CryoTransformer alone gives the worst reconstruction, whereas the complete pipeline achieves the best resolution on all four entries, improving by 0.63 to 3.57 ˚A.The pipeline’s maps also resolve the most fine-grained structure among the five compared systems.
- Ablation study: Contamination masking helps on three of four entries but costs 0.41 ˚A on EMPIAR-10093.On EMPIAR-10532, 16.6% of mask removals are annotated particles.
- Ablation study: 2D selection gives the largest improvement of any stage and improves every entry by 0.56 to 3.56 ˚A when applied to raw picks.Adding masking beforehand improves three of four entries by at most 0.09 ˚A.
- Ablation study: Masking and 2D selection together remove more than 50% of raw picks on every entry, while masking alone removes at most 5.6%.Almost all removals are attributable to 2D selection, the stage that also moves the resolution.
- Ablation study: Feedback re-picking improves three of four entries by at most 0.15 ˚A and costs 0.04 ˚A on the fourth.Round 1 is reported because later rounds remain within 0.013 macro F1 on most entries, although feedback can increase recall, precision, or degrade the picker depending on the dataset.
6. Discussion
The discussion shows that 2D F1 rankings can diverge from reconstruction rankings, while dataset-specific selection and contamination failures limit the pipeline’s reliability. The subtractive feedback loop also offers limited improvement when it can only discard candidates.
- 2D F1 versus reconstruction: On three of four entries, low-F1 picker rankings differ from reconstruction rankings, nearly reversing on EMPIAR-10532.F1 and resolution rankings agree only on EMPIAR-10081, where pickers reach their highest F1.
- 2D F1 versus reconstruction: The picker with lower F1 wins after masking and 2D selection because high-recall candidates retain particles discarded by CryoSegNet.After identical selection, the CryoTransformer set leads on all four entries.
- Failure cases: 44 of 50 classes and 88.6% of particles are discarded on EMPIAR-10345 because CryoSift’s absolute threshold meets a shifted score distribution.Later rounds discard more than 90% of picks and remain below the base checkpoint’s F1.
- Failure cases: Four EMPIAR-10532 micrographs lose more than half their matched particles because the contamination mask covers the particle field.The released MicrographCleaner weights did not include EMPIAR-10532 in their training data.
- Limitations of the subtractive design: The feedback loop does not substantially improve resolution because both purification stages only discard candidates and teach the picker to omit them.The discussion identifies additive collection of under-represented views as a future direction.
7. Conclusion
The paper frames picking, contamination removal, and 2D class selection as one reconstruction-directed selection problem. Its pipeline achieves the best resolution among compared pickers on all four entries, while 2D rankings do not reproduce reconstruction rankings.
- Conclusion: The method jointly poses particle picking, contamination removal, and 2D class selection against reconstruction rather than incomplete-annotation agreement.The instantiated pipeline uses CryoTransformer, MicrographCleaner, CryoSift, and picker fine-tuning.
- Conclusion: The pipeline gives the best resolution of any compared picker on all four EMPIAR entries.This conclusion is paired with the finding that 2D picker rankings do not reproduce reconstruction rankings.
- Conclusion: Particle selection is better treated as a reconstruction-aware pipeline that purifies a high-recall picker than as optimization of picker 2D F1 alone.The conclusion follows from the mismatch between 2D rankings and reconstruction rankings.
Supplementary Material
The supplementary material follows the main paper’s order and provides implementation details, failure-case visualization, reconstruction evidence, and base-picker detection scores.
- Supplementary Material: Section S1 diagrams the pipeline, Sections S2–S4 detail its components, and Section S5 shows the EMPIAR-10345 selection failure.The material follows the order of the main paper.
- Supplementary Material: Section S6 collects FSC curves and viewing-direction distributions behind the main tables, while Section S7 tabulates base-picker 2D detection scores.Section S8 lists additional supplementary material content.
S1. Details of the pipeline
The pipeline uses permissive candidate picking, contamination masking, sequential 2D class selection, and pseudo-label fine-tuning before full-set reconstruction. Supplementary details specify the operating point and training stages.
- Pipeline stages: CryoTransformer picks candidates, MicrographCleaner removes contamination, and CryoSPARC extracts the survivors in both pipeline blocks.The upper block runs one feedback-loop round; the lower block processes the full micrograph set with the resulting checkpoint.
- 2D class selection: The 2D selection stage classifies 50 averages at a time, repeatedly shrinking the discard pool while initially retained green classes skip intermediate cycles.The blue classes proceed to the next classification, and green classes rejoin at the final classification.
- Feedback and reconstruction: The feedback round produces pseudo-labels from kept classes without reconstruction, while the full-set stage performs ab-initio reconstruction and homogeneous refinement.Three random seeds are run, and the best refined map supplies the reported GSFSC resolution.
- Picker operating point: CryoTransformer emits 600 scored queries per micrograph, retains the top 75%, and removes duplicates using non-maximum suppression at overlap threshold 0.7.These settings retain the original implementation’s operating point.
- Training stages: Head repair trains only the classification head on frozen-detector features, whereas fine-tuning trains every picker weight on each round’s teacher set.The no-object class weights are set to zero after head repair, and fine-tuning uses the last epoch’s weights.
S3. Details of contamination masking
MicrographCleaner assembles window-level contamination predictions into a smooth micrograph mask, while the subsequent 2D selection uses class-level grouping that can retain mixed particle and non-particle classes.
- Contamination masking: The masking implementation is illustrated by comparing the released assembly with triangular blending on the same micrograph and network output.The released mask leaves vertical unmasked stripes, whereas the triangular mask has a smooth boundary.
- Contamination masking: The released MicrographCleaner assembly averages overlapping windows uniformly, creating mask steps that require later seam correction.The correction can overwrite values from a border to the image edge when a sharp border transition is detected.
- Contamination masking: Triangular blending weights overlapping-window predictions from one at each center to zero at each border, preventing seams without repair.The replacement uses half-size overlap and predicts each micrograph in a single orientation instead of rotation averaging.
- 2D class selection: The 2D selection stage scores class averages and divides them into set-aside, discarded, and repeatedly re-classified groups.Classes scoring 2.5 or higher can be set aside, with the best 70% of those candidates held out from later cycles.
S5. The 2D selection failure on EMPIAR-10345
EMPIAR-10345 loses most candidates during contamination masking and early 2D selection, and the resulting reconstructions show that apparent directional coverage need not represent genuine views.
- 2D selection failure: 94.1% of EMPIAR-10345 candidates are removed by contamination masking followed by 2D class selection.The first cycle discards 44 of 50 classes, compared with 23 of 50 on EMPIAR-10081, and later cycles cannot recover discarded classes.
- Figure setup: The figure compares full-set reconstruction rows with the CryoPPP reconstruction from 300 annotated micrographs per dataset.The annotated reconstruction is shown below the dotted line and is not one of the main table rows.
- Reconstruction diagnostics: The EMPIAR-10345 baseline and +mask FSC curves nearly coincide, while separation returns after selection.The coincidence is consistent with CryoTransformer’s precision of 0.187 on this entry and little remaining signal beyond solvent noise.
- Reconstruction diagnostics: The baseline and +mask rows on EMPIAR-10345 deliver 7.11 and 6.96 ˚A while populating nearly every direction.After 2D selection, rows deliver 3.54–3.59 ˚A from fewer directions, and the remaining bias is attributed to discarding too many particles.
- Figure setup: The directional-distribution panels use red for the most populated direction and blue for the least, with colors scaled within each panel.The distributions are estimated pose assignments rather than ground-truth orientations.
- Contamination masking: Triangular blending produces a smooth contamination-mask boundary rather than the straight-edged vertical stripes of the released assembly.The comparison uses the same input micrograph and network for both post-processing methods.
S7. 2D detection scores of the base pickers
The base-picker F1 scores are measured against CryoPPP annotations, but training overlap and incomplete alignment between F1 and reconstruction quality complicate their interpretation.
- Evaluation caveats: The CryoPPP scores are not uniformly independent of training data because EMPIAR-10081 is part of crYOLO’s general-model training data.Topaz’s released general-model training data is undocumented, so overlap cannot be ruled out; affected values are grayed and excluded from ranking.
- Selection context: The first 2D-selection cycle discards 23 of 50 classes on EMPIAR-10081 and 44 of 50 on EMPIAR-10345.A discarded class is permanently removed from subsequent processing.
- Picker comparison: CryoTransformer has the highest recall on every entry among pickers with no known training overlap, supporting its use as the high-recall pipeline base.Purification stages can discard candidates, making recall especially relevant to the initial picker choice.
S8. Limitations of the evaluation
The evaluation is bounded by non-held-out annotation usage and an entry-specific pixel-size issue, while supplementary figures document reconstruction and viewing-direction diagnostics under the experimental setup.
- Evaluation limitations: The 2D scores against CryoPPP annotations are not held-out because 50 of 300 annotated micrographs train each feedback round.This limitation applies to the reported annotation-based evaluation.
- Evaluation limitations: EMPIAR-10345 resolutions use the CryoPPP-declared pixel size and are about half the physical figure, so comparisons are valid only within that entry.The scope boundary affects cross-entry interpretation of the resolution values.
- Reconstruction diagnostics: Figure S6 shows one reconstruction row per dataset and refinement, with the CryoPPP reconstruction below the dotted line as a reference.The reference uses 300 annotated micrographs, whereas the other rows use each dataset’s full micrograph set.
- Viewing-direction diagnostics: Figure S7 encodes azimuth and elevation distributions, with red as the most populated direction and blue as the least within each panel.Colors cannot be compared across panels because CryoSPARC scales each panel to its own counts.
- Compute setup: Experiments use one node with RTX A5000 GPUs, 128 physical CPU cores, and 2 TB of RAM, with each pipeline stage occupying one GPU.A fine-tuning round uses 50 micrographs and takes just under two hours for 50 epochs.