CryoAtom
v2.2.0Automated atomic model building for cryo-EM maps
Automated atomic model building for cryo-EM maps
CryoAtom builds atomic models of proteins, DNA, RNA and their complexes into cryo-EM density maps. It traces the map with a neural network, places residues and nucleotides, and writes an mmCIF model with a confidence value for every residue.
This page runs CryoAtom2 (version 2.2.0) by Su, Huang, Peng, Amunts and Yang. With sequences, CryoAtom aligns each built chain to the sequence you provide. Without sequences, it assigns residue types from the map alone and can identify the chains against a sequence database with HMMER, which is useful when the identity of a protein in the map is unknown.
A backbone file of Cα or P atoms replaces map tracing, so CryoAtom only builds and identifies the chains you give it. A mask map limits building to one region of the map.
Runs cost 13 credits per minute of measured runtime. The minimum reservation is 13 credits, not a minimum final charge. Completed runs are charged in proportion to elapsed time, rounded up to a whole credit, and unused reserved credits are returned. Set a spending limit before submission.
The quote before submission is an estimate based on the README example, the map EMD-33198 of PDB entry 7XHT with 656 residues, which took about 15 minutes on default settings. Runtime grows with the number of residues built, so a larger map costs more than the table shows.
| Cryo-Net rounds | Repeats per residue | No database | One database | Two databases |
|---|---|---|---|---|
| 1 | 3 | 81 | 94 | 107 |
| 3 | 1 | 81 | 94 | 107 |
| 3 (default) | 3 (default) | 191 | 204 | 217 |
| 5 | 3 | 301 | 314 | 327 |
| 3 | 5 | 301 | 314 | 327 |
| 5 | 5 | 484 | 497 | 510 |
With a backbone file CryoAtom runs one round, so the estimate falls to 81 credits, plus 13 for each sequence database. The thresholds and RU-Net box settings do not change the estimate.
A map is required. Sequences are optional: leave all three sequence inputs empty to build without sequences.
| Input | Accepted | Description |
|---|---|---|
Cryo-EM density map | .mrc, .map, up to 2,000 MiB | The reconstruction to build into. Decompress .map.gz files from EMDB before uploading. |
Protein sequences | .fasta, .fa, .faa, .mpfa, pasted FASTA, or a PDB ID fetched from RCSB, up to 50 MiB | Each distinct protein chain once, in one file. Every record needs a header line. |
RNA sequences | .fasta, .fa, .faa, .mpfa or pasted FASTA, up to 50 MiB | Each distinct RNA chain once. |
DNA sequences | .fasta, .fa, .faa, .mpfa or pasted FASTA, up to 50 MiB | Each distinct DNA strand once. |
Mask map | .mrc, .map, up to 2,000 MiB | Optional. A mask on the same grid as the map, to model one region only. Not used with a backbone file. |
Protein sequence database | .fasta, .fa, .faa, .fna, .mpfa, up to 2,000 MiB | Optional. A database such as a UniProt proteome, searched to identify the protein chains built in the map. |
Nucleic acid sequence database | .fasta, .fa, .fna, .mpfa, up to 2,000 MiB | Optional. A database of RNA or DNA sequences, searched to identify the nucleic acid chains. |
Backbone atoms | .pdb, .cif, up to 2,000 MiB | Optional. A model with at least the protein Cα or nucleic acid P atoms. CryoAtom builds and identifies these chains instead of tracing the map. |
| Parameter | Type | Default | Description |
|---|---|---|---|
Cryo-Net rounds | integer | 3 | Rounds of all-atom model building, at least 1. With a backbone file CryoAtom always runs one round. |
Repeats per residue | integer | 3 | Minimum number of times Cryo-Net processes each residue in one round, at least 1. |
Chain confidence threshold | number | 50 | With sequences, chains of the final model whose average confidence is at or below this value are removed. 0 to 100. Nucleic acid chains are always kept. |
Filter the raw model too | boolean | off | With sequences, also removes low-confidence chains from the raw model. |
| Parameter | Type | Default | Description |
|---|---|---|---|
Search confidence threshold | number | 25 | Only chains with an average confidence above this value are searched against the databases. 0 to 100. Without sequences it also trims the final model. |
HMMER E-value | number | 1 | Hits above this E-value are discarded. Without sequences CryoAtom multiplies it by 100, up to 10. |
| Parameter | Type | Default | Description |
|---|---|---|---|
Cα probability threshold | number | 0.6 | Probability above which a voxel can become a Cα atom. 0 to 1. Lower values place more protein residues. |
P probability threshold | number | 0.465 | Probability above which a voxel can become a nucleotide P atom. 0 to 1. Lower values place more nucleotides. |
Residue mask threshold | number | 0.3 | Threshold Cryo-Net uses to filter residues. 0 to 1. Set to 0 with a backbone file. |
RU-Net box stride (voxels) | integer | 50 | Step of the box that slides across the map, resampled to 1.51 Å voxels. |
RU-Net box size (voxels) | integer | 65 | Side length of the box cropped from the map. A map no larger than this box on any side is predicted in one pass. |
Raising Cryo-Net rounds or Repeats per residue takes longer and can improve models in regions of low local resolution.
The final model opens in the viewer. The Identified sequences table lists the best database hit for each searched chain, and every file is listed under Files.
| File | Description |
|---|---|
cryoatom/cryoatom.cif | The final model, with per-residue confidence from 0 to 100 in the B-factor column. |
cryoatom/cryoatom_raw.cif | The raw model, with every region that was built. |
cryoatom/new_hits_prot.xlsx, cryoatom/new_hits_na.xlsx | HMMER hits for protein and nucleic acid chains, with an all_hits and a best_hits sheet sorted by E-value. |
cryoatom/cryoatom_prot.fasta, cryoatom/cryoatom_na.fasta | Your input sequences followed by the sequences identified in the database. |
cryoatom/hmm_profiles/ | Without sequences, the HMMER profile and hit alignment of each chain. |
cryoatom.log, config.json, provenance.json | The run log, the settings CryoAtom ran with, and the source version and input checksums. |
The .xlsx files and the table appear only when a database was searched.
Color the final model by B-factor to see the confidence of each residue. High values mark residues CryoAtom placed and typed with certainty; low values usually sit in weak or poorly resolved density, at chain ends and in flexible loops. Treat low-confidence regions as a starting point for manual rebuilding.
The final model keeps only chains above Chain confidence threshold when sequences were given, or above Search confidence threshold without them. If a chain you expected is missing from the final model, look for it in cryoatom_raw.cif before lowering a threshold.
With sequences, each built chain is aligned to the sequence it matches best, so residue names and numbering follow your FASTA. Give each distinct chain once: copies of the same chain in the map are matched to the same record. A chain shorter than its sequence usually means part of the protein is disordered or absent from the map, not that the build failed.
Without sequences, residue types come from the density alone. They are reliable where side chains are resolved and uncertain elsewhere, which is why the database search uses an HMMER profile of the predicted residue probabilities rather than a single sequence.
| Column | Meaning |
|---|---|
database | protein or nucleic acid. |
query_name | The chain of the built model that was searched. |
target_name, accession, description | The database entry it matched. |
E-value | The number of hits this good expected by chance. Lower is better. |
score, bias | HMMER bit score and the part of it from low-complexity sequence. |
query_len | Length of the searched chain. |
A small E-value with a high score is a confident identification. Hits near the E-value limit, or with a bias close to the score, are weak and should be checked against the density. Short chains rarely give significant hits, because a few residues carry little information.
Validate the geometry of the model with MolProbity and refine it against the map before deposition. CryoAtom models are a first draft: real-space refinement and manual correction remain part of the workflow.
Automated atomic model building in cryo-EM maps
Assess docking quality between model and native structures
Validate protein structures with clashscore, Ramachandran, rotamer, and geometry checks.
Predict 3D structures for proteins, ligands, DNA, RNA, and multi-component complexes.
Predict controllable biomolecular structures for proteins, ligands, DNA, RNA, and complexes.
All-atom biomolecular complex prediction with Mol-1
All-atom biomolecular structure prediction
Open-source AlphaFold3-based structure prediction for proteins, ligands, DNA, and RNA
Open-source biomolecular structure prediction with Protenix
Calculate RNA density-of-states energy distributions.