Reconstruction results

What CryoGEN / CryoWGEN actually achieve on real cryo-ET datasets — the headline results and metrics.

The other pages explain the methods themselves — casting missing-wedge restoration as a Bayesian inverse problem and solving it with statistical machine learning. This page takes the other view and collects what these methods actually produce: what they do on real cryo-ET data, and where they help. (Full figures and tables are in the papers [1][2].)

closerto the true structure — clearly beats IsoNet, DeepDeWedge

restoreswhat the missing wedge erased

noground-truth labels needed

fastertraining, markedly less time

Validated on real data

CryoGEN has been validated on real biological datasets, including immature HIV-1 virions and ribosomes. From each tomogram corrupted by noise and the missing wedge, it reconstructs a more isotropic, structurally more complete volume — and does so with no ground-truth labels.

More isotropic. Conventional pipelines leave structure stretched along the missing direction; CryoGEN’s energy prior (and, later, optimal transport) markedly reduces that anisotropy, restoring detail along the smeared vertical axis.
No recursive subtomogram averaging. The classical route repeatedly aligns and averages many copies to raise SNR and fill the wedge; CryoGEN folds this into a single self-supervised reconstruction, which is why training takes markedly less time.
More complete, more interpretable. Reconstructed volumes are more coherent across membranes and subunits, easing downstream picking and interpretation.
Quantified uncertainty (CryoWGEN). With entropic regularization added, CryoWGEN returns not one volume but a family of reconstructions — and the spread of that family is the missing-wedge uncertainty, made readable.

Key result figure

WBP reconstruction of apoferritin — Reconstruction of the apoferritin (PDB 6Z6U) low-density volume. WBP leaves missing-wedge artifacts (smearing and gaps along the vertical axis), IsoNet still shows inconsistencies, and CryoGEN gives a smoother, more complete structure. From CryoGEN (ICLR 2025), Figure 7.

IsoNet reconstruction of apoferritin — Reconstruction of the apoferritin (PDB 6Z6U) low-density volume. WBP leaves missing-wedge artifacts (smearing and gaps along the vertical axis), IsoNet still shows inconsistencies, and CryoGEN gives a smoother, more complete structure. From CryoGEN (ICLR 2025), Figure 7.

3D reconstruction comparison of Vipp1 stacked rings — 3D reconstruction of C13 Vipp1 stacked rings (EMDB:18424): (a) WBP → (b) IsoNet → (c) CryoGEN (ours) → (d) real structure. WBP and IsoNet leave artifacts; CryoGEN’s result is closest to the real structure. From CryoGEN (ICLR 2025), Figure 2.

Quantitative comparison

On synthetic data with known ground truth (sphere, prism, Vipp1 assembly), CryoGEN clearly beats the baselines on both PSNR and SSIM (higher is better):

Data state	Sphere PSNR	Sphere SSIM	Prism PSNR	Prism SSIM	Vipp1 PSNR	Vipp1 SSIM
Corrupted	21.12	0.8113	14.82	0.6931	26.68	0.8000
IsoNet	22.98	0.8770	19.11	0.8857	27.12	0.8191
DeepDeWedge	23.17	0.8824	21.10	0.9278	28.75	0.8758
CryoGEN	29.19	0.9706	32.69	0.9949	30.65	0.9199

Data from CryoGEN (ICLR 2025), Table 1.

Note

The figures and table on this page are from the published CryoGEN paper [1] (see the references at the end); CryoGEN-II is listed in reference [2]. To render your own .mrc volume into 3-D figures like these, see the software & visualization tutorial.

References

Yunfei Teng, Yuxuan Ren, Kai Chen, Xi Chen, Zhaoming Chen, Qiwei Ye. CryoGEN: Generative Energy-based Models for Cryogenic Electron Tomography Reconstruction. International Conference on Learning Representations (ICLR), 2025. openreview.net/forum?id=uOb7rij7sR
Yunfei Teng et al. CryoGEN-II: Cryogenic Electron Tomography Reconstruction via Generative Network. CVPR 2026 Workshops (VISION). openaccess.thecvf.com

← Cryo-ET Reconstruction