henhaobio
Sign in

NatureMolecular BiologyPeer-reviewed journal article

Automated model building and protein identification in cryo-EM maps

Kiarash JamaliiD, Lukas KälliD, Rui ZhangiD, Alan BrowniD, Dari KimaniusiD, Sjors H. W. ScheresiD

PDFDOIPreprintCode
In-depth full-text digest available · Jump to digest ↓

Abstract

Interpreting electron cryo-microscopy (cryo-EM) maps with atomic models requires high levels of expertise and labour-intensive manual intervention. We present ModelAngelo, a machine-learning approach for automated atomic model building in cryo-EM maps. By combining information from the cryo-EM map with information from protein sequence and structure in a single graph neural network, ModelAngelo builds atomic models for proteins that are of similar quality as those generated by human experts. For nucleotides, ModelAngelo builds backbones with similar accuracy as humans. By using its predicted amino acid probabilities for each residue in hidden Markov model sequence searches, ModelAngelo outperforms human experts in the identification of proteins with unknown sequences. ModelAngelo will thus remove bottlenecks and increase objectivity in cryo-EM structure determination.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 2 (17 Oct 2023) · CC BY 4.0

ModelAngelo combines convolutional and graph neural networks with protein language model embeddings and hidden Markov model searches to automate atomic model building and identify unknown proteins in cryo-EM maps.

Building atomic models into 2-4 A cryo-EM density maps and identifying unassigned subunits in endogenous complexes traditionally require slow, subjective manual intervention by human experts. Automating protein and nucleic acid backbone construction alongside probabilistic proteome searches removes a major structural biology bottleneck and identifies subunits missed by manual tracing.

Key results

  1. ShownModelAngelo builds atomic protein models in cryo-EM maps with accuracy and quality comparable to human-deposited structures across 177 held-out test structures.

    Across 410,585 residues in 177 test structures (<4 A resolution, <10% sequence identity to training data), ModelAngelo achieved backbone RMSDs <1.0 A down to Q-scores of 0.4, built 77% of all residues and 85% of residues with Q-scores >=0.4, and matched deposited models in Q-scores and Fourier shell correlation. · ModelAngelo builds protein models of comparable…; Figure legends

  2. ShownModelAngelo achieves substantially higher model completeness at lower local resolutions than existing automated cryo-EM builders.

    On a 27-structure benchmark subset spanning single-chain, homo-oligomeric, and hetero-oligomeric proteins, ModelAngelo built up to 80% of deposited residues at local resolutions down to 3.5-4 A, whereas DeepTracer reached ~80% completeness only at 2.5-3 A, outperforming Phenix, Demo-EM, and Buccaneer. · ModelAngelo outperforms alternative approaches; Figure legends; Tables

  3. ShownFor nucleic acids, ModelAngelo constructs accurate backbones superior to DeepTracer and CryoREAD, though purine/pyrimidine base assignment drops below 2.5 A.

    Across 11 ribosome structures (1.98-3.80 A) and a Scytonema hofmanni CRISPR-associated transpososome, nucleotide backbone RMSD ranged from <0.5 A in high-Q-score regions to 2 A in low-Q-score regions with >80% backbone completeness in well-resolved regions, while sequence-accurate completeness ranged from 20% to 80%. · ModelAngelo builds good nucleic acid backbones; Figure legends; Tables

  4. ShownUsing predicted 20-amino-acid probability distributions in HMMER3 searches without input sequences, ModelAngelo identifies unassigned endogenous protein chains.

    In the 16.7 MDa Porphyridium purpureum PBS-PSII-PSI-LHC supercomplex (PDB 7Y5E), sequence-free ModelAngelo identified all 6 unassigned chains (LPP1, CNT, PsbW, Psb34, LRH, LPS1; E-values 5.8e-66 to 6.4e-2) vs 2-3 found by findMySequence; in Chlamydomonas reinhardtii ciliary maps, it identified 4 radial spoke (RSP24-27) and 2 central apparatus (FAP92, FAP374) proteins. · Protein identification; ModelAngelo identifies protein chains that were…; Figure legends; Tables

Methods, data and reproducibility

A modified feature-pyramid convolutional neural network predicts protein C-alpha and nucleic acid phosphorus positions from 1 A-resampled cryo-EM maps to initialize a residue graph. An eight-layer graph neural network with cryo-EM, ESM1b sequence, and invariant point attention modules refines atomic coordinates and HMMER3 profile probabilities over three recycling rounds.

  • Cryo-EM map-model training pairs from EMDB/PDB deposited before April 1, 2022 with resolution <4 A · multiple species · n = 3715 map-model pairs
  • Held-out cryo-EM protein test set (<=30,000 residues, <10% sequence identity to training set) and benchmark subset · multiple species · n = 177 structures (410,585 residues); 27-structure subset · 8DWI
  • Ribosome cryo-EM structures (1.98-3.80 A) and Scytonema hofmanni V-K CAST transpososome for nucleic acids · multiple species · n = 11 ribosomes and 1 CAST transpososome · 8EA4
  • In situ PBS-PSII-PSI-LHC supercomplex (Porphyridium purpureum) and ciliary axoneme maps (Chlamydomonas reinhardtii) · plant (other) (P. purpureum and C. reinhardtii algae) · n = 1 supercomplex (16.7 MDa, 81 chains) and 3 ciliary maps · 7Y5E
Methods & tools
ModelAngelo, ESM1b, HMMER3, HHblits, Servalcat, ResMap, Phenix, Demo-EM, Buccaneer, DeepTracer, CryoREAD, findMySequence
Assay conditions & compute
Single A100 GPU: 2 min for smallest test structure (PDB ID: 8DWI, 54.7 kDa), 53 min for largest test structure (PDB ID: 7UMS, 1.85 MDa), and 23 h for the 16.7 MDa supercomplex (PDB ID: 7Y5E).
Limitations (3)
  • ModelAngelo omits nucleotide sequence embeddings and often cannot separate purine bases (A vs G) or pyrimidine bases (C vs U/T) when cryo-EM resolution is worse than 2.5 A, requiring manual base checking. · stated by the authors
  • Output models still require downstream stereochemical refinement (such as in Servalcat or Phenix) and manual inspection, particularly in map regions with local resolution worse than 3.5-4.0 A. · stated by the authors
  • ModelAngelo leaves cofactor densities (such as phycocyanobilin and phycoerythrobilin) unmodeled and does not perform rigid-body docking of known domains or loop building in poorly resolved density regions. · stated by the authors

Read this if you determine macromolecular structures by cryo-EM and want to automate atomic model building or identify unknown protein subunits from density maps.

Citations
685
FWCI
119.61
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSu '24Li '26Li '25Dai '23Mondal '25Li '23
78% semantic matchBioinformatics
Baoquan Su, Kun Huang, Zhenling Peng+2 more·bioRxiv·Nov 13, 2024· 8 citations