Research map / Image generation / Autoregressive and latent generation
CryoGEM: Physics-Informed Generative Cryo-Electron Microscopy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Generates annotated synthetic cryo-EM micrographs by combining a physics-based imaging simulation with mask-guided contrastive unpaired noise translation, improving particle picking and pose estimation.
Problem
Cryo-EM particle picking and pose estimation are limited by the lack of high-quality annotated training micrographs, and existing noise translation methods ignore physical imaging and random noise structure in densely packed particle images.
Approach
Simulates micrographs from a virtual specimen built from coarse cryoSPARC/cryoDRGN volumes using a weak phase object and CTF model, then translates them to real noise with a CUT-style contrastive loss whose positive/negative patches are chosen via a particle-background mask, plus Gaussian noise injection for non-deterministic output.
Key results
- CryoGEM achieves average FID 37.91 versus 60.84 for the next best (CUT) across five datasets.
- Fine-tuned Topaz reaches average AUPRC 0.674 versus 0.466 for original Topaz, and average reconstruction resolution 5.27 Å versus 6.71 Å.
- Pose estimation with CryoGEM data gives average Res(px) 4.29 versus 5.32 for the best baseline (CycleGAN).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train Topaz using CryoGEM-generated annotated images | Topaz · Proteasome, Ribosome, Integrin, PhageMS2, HumanBAF | AUPRC: 0.466 -> 0.674 (Avg.) | Helped |
| Train CryoFIRE using CryoGEM-generated images with ground-truth poses | CryoFIRE · Proteasome, Ribosome, Integrin, PhageMS2, HumanBAF | Res(px): 12.18 -> 4.29 (Avg.) | Helped |
| Use CryoGEM rather than the strongest listed baseline for image generation | CryoGEM · Proteasome, Ribosome, Integrin, PhageMS2, HumanBAF | FID: 60.84 -> 37.91 (Avg.) | Helped |
| Add zero-mean Gaussian noise to physical simulations (ablation) | CryoGEM noise translation · Ribosome | FID | Helped |
| Use particle-background mask-guided patch sampling (ablation) | CryoGEM contrastive learning · Ribosome | FID | Helped |
| Remove physical priors (ablation) | CryoGEM · Ribosome | FID | Hurt |
| Use noise in the encoder's third layer (ablation) | CryoGEM noise translation · Ribosome | FID | Mixed |
| Use noise in the encoder's fifth layer (ablation) | CryoGEM noise translation · Ribosome | FID | Mixed |
| Feed random noise as a second input channel (ablation) | CryoGEM noise translation · Ribosome | FID | Mixed |
| Scale random noise to SNR=0.01 (ablation) | CryoGEM noise translation · Ribosome | FID | Mixed |
| Scale random noise to SNR=1.0 (ablation) | CryoGEM noise translation · Ribosome | FID | Mixed |
| Remove Gaussian noise (ablation) | CryoGEM noise translation · Ribosome | FID | Hurt |
Surprising findings
- CycleDiffusion performs worst on FID (avg 415.42), worse than simple Gaussian and Poisson noise baselines.
- CycleGAN training is unstable, with its bijection assumption breaking down on cryo-EM data.
Limitations
- Relies on a coarse 3D volume as input, which may be hard to obtain for small or dynamic molecules.
- Trained per dataset; sacrifices generalization for a lightweight framework.
- Evaluated on only five datasets with per-dataset training.
Experimental setup
Datasets: EMPIAR-10025 (T20S Proteasome) EMPIAR-10028 (80S Ribosome) EMPIAR-10345 (Integrin) EMPIAR-10075 (PhageMS2) EMPIAR-10590 (HumanBAF) cryoPPP
Benchmarks: EMPIAR-10025 EMPIAR-10028 EMPIAR-10345 EMPIAR-10075 EMPIAR-10590
Metrics: FID AUPRC Res(Å) Res(px) Rot(rad)
Baselines: Poisson noise Gaussian noise Poisson-Gaussian noise CycleGAN CUT CycleDiffusion Topaz CryoFIRE
Builds on: CUT CycleGAN Topaz cryoDRGN cryoFIRE
Closest papers by meaning
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.90)
- CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (ICLR 2025 · similarity 0.89)
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.88)
- CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference (NeurIPS 2024 · similarity 0.88)
- Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data (NeurIPS 2025 · similarity 0.87)
- Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.