Reconstruction-Aware Cryo-EM Particle Picking

Particle picking, contamination removal, and 2D class selection as one pipeline, judged by the 3D map it delivers.

Riku Itsuji^1,2 Yuanhao Wang³ Xingjian Li³ Seonghui Min³ Hideo Saito^1,2 Min Xu³

¹Keio University, Yokohama, Japan ²Keio AI Research Center, Yokohama, Japan ³Carnegie Mellon University, Pittsburgh, PA, USA

arXiv Dataset Weights Code

Overview of the reconstruction-aware particle picking pipeline — **Reconstruction-aware particle picking.** Top: one round of the feedback loop. CryoTransformer picks candidates, a MicrographCleaner mask discards the candidates on contamination, and 2D class selection keeps the classes CryoSift scores as good. The surviving particles serve as pseudo-labels for fine-tuning the picker. Bottom: the checkpoint the loop delivers picks the full micrograph set, the same masking and 2D class selection purify the picks, and the surviving stack goes to 3D reconstruction.

Abstract

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. Code will be available upon acceptance.

Contributions

Reconstruction-aware formulation

We formulate particle picking, contamination removal, and 2D class selection as a single selection problem posed against downstream reconstruction, and evaluate it by reconstruction resolution rather than by 2D F1 against incomplete annotations.

A purification pipeline that outperforms existing pickers

We combine a high-recall picker with contamination masking and 2D class selection, and close the loop by re-training the picker on the purified particle stack. The pipeline gives the best reconstruction of any picker we compare.

The best F1 is not the best reconstruction

Across the same pickers and entries, 2D detection scores do not reliably order the reconstructions. Agreement with 2D annotations is therefore only a partial guide, and we judge a selection by the map it delivers.

Main Results

4.12 ÅEMPIAR-10081
(CryoTransformer: 5.02)

4.37 ÅEMPIAR-10093
(CryoTransformer: 6.78)

3.54 ÅEMPIAR-10345
(CryoTransformer: 7.11)

3.40 ÅEMPIAR-10532
(CryoTransformer: 4.03)

Method	10081	10093	10345	10532
crYOLO	4.25	4.55	3.56	3.54
Topaz	4.56	6.55	3.88	3.96
CryoSegNet	4.47	6.96	3.76	3.86
CryoTransformer	5.02	6.78	7.11	4.03
Ours	4.12	4.37	3.54	3.40

Local resolution maps of the reconstructions on EMPIAR-10081 and EMPIAR-10532 — **3D reconstructions of EMPIAR-10081 and 10532.** The best-of-three map of each picker and of our pipeline, colored by the local resolution estimated on it, blue better and red worse, on a scale shared within a row. Our map resolves the most fine-grained structure of the five.

Ablation Study

Ablation study. The stages between picking and reconstruction are toggled on the same CryoTransformer candidate set.
mask	select	fb	10081	10093	10345	10532
Resolution (GSFSC 0.143, Å, best of three seeds, lower is better)
–	–	–	5.02	6.78	7.11	4.03
✓	–	–	4.74	7.19	6.96	3.98
–	✓	–	4.15	4.61	3.55	3.47
✓	✓	–	4.08	4.52	3.59	3.43
✓	✓	✓	4.12	4.37	3.54	3.40
Particles retained, with the cumulative reduction from the raw picks
–	–	–	259,335	754,434	494,061	604,430
✓	–	–	244,924 (−5.6%)	753,440 (−0.1%)	484,649 (−1.9%)	576,473 (−4.6%)
✓	✓	–	126,181 (−51.3%)	301,589 (−60.0%)	29,214 (−94.1%)	178,722 (−70.4%)

Picks remaining after each stage of the pipeline on EMPIAR-10081 and 10532 — **What survives each stage of the pipeline.** The picks that remain after each stage on one micrograph of EMPIAR-10081 (top) and one of EMPIAR-10532 (bottom) at round 1, with the annotated particles repeated on the right.

The Best 2D F1 Is Not the Best Reconstruction

Scatter plot of macro F1 against GSFSC resolution for the four base pickers — Macro F1 of the four base pickers on the 300 annotated micrographs against the GSFSC 0.143 resolution of the main results, with marker shape the picker and color the dataset. The dashed line is a guide to the eye for reading orderings within one color, not distances across colors.

BibTeX

@inproceedings{itsuji2027reconstruction,
  title     = {Reconstruction-Aware Cryo-EM Particle Picking},
  author    = {Itsuji, Riku and Wang, Yuanhao and Li, Xingjian and
               Min, Seonghui and Saito, Hideo and Xu, Min},
  booktitle = {TBD},
  year      = {2027}
}