Research map / Probabilistic and scientific modeling / Molecules and proteins
CryoDomain: Sequence-free Protein Domain Identification from Low-resolution Cryo-EM Density Maps
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.
Proposes CryoDomain, a dual-tower network that aligns low-resolution cryo-EM density map embeddings with atomic-structure domain embeddings so protein domains can be identified by retrieval without sequences.
Problem
Identifying protein domains from low-resolution (>4 Å) cryo-EM density maps without prior sequence knowledge; existing sequence- and structure-search methods degrade sharply at low resolution.
Approach
Cross-modal alignment via a dual-tower network (DensityTower and AtomTower) that maps density maps into the embedding space of atomic domain structures, with a retrieval database (DateDB) of atomic-structure embeddings, expanding recognition units from amino acids to domains.
Key results
- On the SCOPe test set, CryoDomain reached mAP 0.47 versus 0.01 for cryoID and ModelAngelo; on CATH, mAP 0.15 versus 0.01 (cryoID) and 0.09 (ModelAngelo).
- CryoDomain achieved AUROC over 0.75 on both CATH and SCOPe test sets.
- Top-1 accuracy at 4-5 Å on CATH was 7.87% for CryoDomain versus 49.44% for ModelAngelo, which had a sharp drop at 5-10 Å (0.00%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CryoDomain instead of cryoID or ModelAngelo | Protein domain identification · CATH low-resolution maps (4∼10 ˚A) | mAP: 0.01, 0.09 -> 0.15 | Helped |
| Use CryoDomain instead of cryoID or ModelAngelo | Protein domain identification · SCOPe low-resolution maps (4∼10 ˚A) | mAP: 0.01 -> 0.47 | Helped |
| Use CryoDomain instead of ModelAngelo | Protein domain identification · CATH 4∼5 ˚A | Top-5 accuracy: 50.56 (45/89) -> 56.18 (50/89) | Helped |
| Use CryoDomain instead of cryoID or ModelAngelo | Protein domain identification · CATH 5∼10 ˚A | Top-5 accuracy: 0.00 (0/59) -> 18.64 (11/59) | Helped |
| Use CryoDomain instead of ModelAngelo | Protein domain identification · SCOPe 5∼10 ˚A | Top-5 accuracy: 0.00 (0/127) -> 74.02 (94/127) | Helped |
| Add the Residual U-Net branch (ablation) | CryoDomain · CATH | Top-5 accuracy: 31.76 (47✓) -> 41.22 (61✓) | Helped |
| Add the Swin-Conv U-Net branch (ablation) | CryoDomain · CATH | Top-5 accuracy: 37.16 (55✓) -> 41.22 (61✓) | Helped |
| Add self-supervised training (ablation) | CryoDomain · CATH | Top-5 accuracy: 7.43 (11✓) -> 41.22 (61✓) | Helped |
| Add cross-modal alignment training (ablation) | CryoDomain · CATH | Top-5 accuracy: 3.38 (5✓) -> 41.22 (61✓) | Helped |
| Add self-supervised training (ablation) | CryoDomain · SCOPe | Top-5 accuracy: 0.00 (0✓) -> 73.64 (95✓) | Helped |
| Add cross-modal alignment training (ablation) | CryoDomain · SCOPe | Top-5 accuracy: 8.53 (11✓) -> 73.64 (95✓) | Helped |
| Add the Residual U-Net branch (ablation) | CryoDomain · SCOPe | mAP: 0.34 -> 0.47 | Helped |
Surprising findings
- At 4-5 Å resolution on CATH, ModelAngelo had higher Top-1 accuracy than CryoDomain, which the authors attribute to CryoDomain's higher false positive rate.
- ModelAngelo collapsed to 11 of 34 successful cases in Top-5 at 5 Å after low-pass filtering.
Limitations
- Higher false positive rate than ModelAngelo at higher resolution (4-5 Å), and lower Top-1 accuracy there.
- Performance declines gradually at 9-10 Å resolution.
- Test sets are small (159 CATH and 129 SCOPe maps) and restricted to domain types with at least ten pairs.
Experimental setup
Datasets: EMDB PDB CATH SCOPe KluDo predictions
Benchmarks: CATH domain test set (159 maps, low-resolution >4 Å) SCOPe domain test set (129 maps)
Metrics: Top-1/5/10 accuracy mAP AUROC MSE between embeddings
Baselines: cryoID ModelAngelo DeepTracer-ID
Builds on: AlphaFold2 CLIP Faiss Swin-Conv U-Net
Closest papers by meaning
- Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data (NeurIPS 2025 · similarity 0.90)
- CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (ICLR 2025 · similarity 0.90)
- cryoSPHERE: Single-Particle HEterogeneous REconstruction from cryo EM (ICLR 2025 · similarity 0.89)
- CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM (NeurIPS 2024 · similarity 0.88)
- Mixture of neural fields for heterogeneous reconstruction in cryo-EM (NeurIPS 2024 · similarity 0.87)
- CryoSPIN: Improving Ab-Initio Cryo-EM Reconstruction with Semi-Amortized Pose Inference (NeurIPS 2024 · similarity 0.86)
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.