Research map / Probabilistic and scientific modeling / Molecules and proteins
CryoFM: A Flow-based Foundation Model for Cryo-EM Densities
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.
Trains CryoFM, a hierarchical flow-matching transformer on ~3.5k curated EMDB density maps, and uses flow posterior sampling to apply it as a training-free prior for cryo-EM denoising, missing-wedge restoration and ab initio modeling.
Problem
Cryo-EM and cryo-ET reconstruction are inverse problems whose quality depends on a good prior over biomolecular density maps, but existing data-driven tools are task-specific and not general foundation models.
Approach
Learn an unconditional flow-matching prior over high-quality sharpened EMDB density maps with a hierarchical 3D transformer (HDiT-based, neighborhood attention at the ends), then convert it to a posterior sampler for arbitrary forward operators via flow posterior sampling, without fine-tuning.
Key results
- Spectral noise denoising at 3.2 Å: CryoFM achieves FSC_AUC 0.95 with 0% fail rate vs DeepEMhancer 0.92 (0.06 FR) and EMReady 0.73.
- Anisotropic noise denoising (±15°): CryoFM FSC_AUC 0.81 vs DeepEMhancer 0.77 and spIsoNet 0.61, with zero fail rate.
- Missing wedge restoration FSC_AUC improves from 0.80 to 0.92 (missing region 0.76); CryoFM-S trained on 3479 maps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CryoFM instead of DeepEMhancer for spectral-noise restoration | CryoFM · Spectral noise denoising, 3.2 Å | FSC_AUC: 0.92\pm0.02 -> 0.95\pm0.01 | Helped |
| Use CryoFM instead of DeepEMhancer for spectral-noise restoration | CryoFM · Spectral noise denoising, 4.3 Å | FSC_0.5: 4.3\pm0.2 -> 4.4\pm0.2 | Hurt |
| Use CryoFM instead of DeepEMhancer for spectral-noise restoration | CryoFM · Spectral noise denoising, 6.1 Å | FSC_AUC: 0.62\pm0.05 -> 0.62\pm0.04 | No clear effect |
| Use CryoFM instead of DeepEMhancer for anisotropic-noise restoration | CryoFM · Anisotropic noise denoising, [-45°, +45°] | FSC_AUC: 0.80\pm0.03 -> 0.88\pm0.03 | Helped |
| Use CryoFM instead of DeepEMhancer for anisotropic-noise restoration | CryoFM · Anisotropic noise denoising, [-30°, +30°] | FSC_AUC: 0.79\pm0.05 -> 0.84\pm0.03 | Helped |
| Use CryoFM instead of DeepEMhancer for anisotropic-noise restoration | CryoFM · Anisotropic noise denoising, [-15°, +15°] | FSC_0.5: 3.4\pm0.2 -> 3.4\pm0.2 | No clear effect |
| Apply CryoFM flow posterior sampling to a missing-wedge observation | CryoFM · Missing wedge restoration | FSC_AUC (missing region): 0.0000 -> 0.76\pm0.06 | Helped |
| Add data normalization during training (ablation) | CryoFM · Three downstream tasks | FSC_AUC | Helped |
| Increase the number of posterior-sampling time steps (ablation) | CryoFM · Three density-restoration downstream tasks | Averaged ranking | Helped |
| Reduce the maximum posterior-sampling step size (ablation) | CryoFM · Three density-restoration downstream tasks | Averaged ranking | Hurt |
| Train the flow model in voxel space instead of latent space (ablation) | CryoFM · Unconditional density sampling | Fine-detail quality | Helped |
| Reduce patchifying and downsampling parameters (ablation) | CryoFM · Test set | Test loss | Helped |
Surprising findings
- Voxel-space flow models converged faster and better than latent flow models, which produced noisier high-frequency details.
- DeepEMhancer achieves the best FSC_0.5 at some noise levels despite lower FSC_AUC, and spIsoNet preserves signal but does not improve correlation with ground truth.
Limitations
- Applying to real-world noisy experimental densities remains challenging; evaluations use synthetic degradations.
- Does not address reconstruction from raw 2D particles.
- Training data limited to around 3.5k curated maps, with latent-space models underperforming possibly due to limited data.
Experimental setup
Datasets: EMDB
Benchmarks: EMDB test maps with synthetic spectral noise EMDB test maps with anisotropic noise EMDB test maps with missing wedge synthetic and EMPIAR datasets for ab initio modeling
Metrics: FSC_AUC FSC_0.5 fail rate
Baselines: DeepEMhancer EMReady spIsoNet cryoSPARC
Builds on: flow matching diffusion posterior sampling HDiT rectified flow neighborhood attention
Closest papers by meaning
- Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data (NeurIPS 2025 · similarity 0.92)
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.91)
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.91)
- Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery (NeurIPS 2025 · similarity 0.90)
- CryoDomain: Sequence-free Protein Domain Identification from Low-resolution Cryo-EM Density Maps (AAAI 2025 · similarity 0.90)
- Mixture of neural fields for heterogeneous reconstruction in cryo-EM (NeurIPS 2024 · similarity 0.90)
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.