CryoAtom

CryoAtom builds atomic models of proteins directly from cryogenic electron microscopy (cryo-EM) density maps. Deposition of cryo-EM-derived structures in the Protein Data Bank grew from 67 in 2012 to 5,788 in 2024, and interpretation — converting a 3D density map into an atomic model with correct sequence assignment — has become the rate-limiting step. Historically that meant months of manual fitting: developing the modeling and refinement protocol for a single 3.2 Å map of the yeast mitoribosomal large subunit took roughly six months. CryoAtom takes a map plus the amino-acid sequences of the target proteins and returns an all-atom model without a reference structure.

The method was developed at Shandong University with collaborators at Westlake University, the University of Münster, and the Max Planck Institute of Molecular Physiology. It was first posted as a preprint in November 2024 under the name CryFold and published in Nature Structural & Molecular Biology in 2025 as CryoAtom; the software and its repository now carry the CryoAtom name.

Its central insight is that a density map already supplies most of the spatial information that AlphaFold 2 has to infer from coevolution. Global all-against-all attention is therefore unnecessary, and the freed capacity can be spent on network depth instead. CryoAtom keeps AlphaFold2's encoder-decoder shape and recycling but swaps in geometry-aware local attention, improving on ModelAngelo, the prior state of the art.

#Key Features

• 3D rotary position embedding: Rotary position encoding, originally designed for 1D token sequences, is extended to node positions in 3D space so attention scores decay naturally with inter-residue distance. Removing it costs 26.4% amino-acid accuracy and destabilizes training.
• All-against-k local attention: Each node attends only to its k spatial neighbors rather than every other node, exploiting map-derived geometry and making a deeper network tractable.
• Per-residue confidence scores: A 0-100 score derived from the predicted FAPE loss tells users which parts of a model to trust, mirroring the confidence estimates familiar from structure prediction.
• Modeling in blurred regions: Across 177 test maps, CryoAtom places roughly 75,000 high-confidence residues absent from the corresponding deposited structures, including a 340-residue segment of the type III-E CRISPR Craspase complex that an independent AlphaFold 3 prediction corroborates at TM-score 0.86.
• Nucleic-acid extension: The maintained implementation, CryoAtom2, adds RNA and DNA modeling and can assign residue types without any input sequence, bundling ESM-2 and RNA-FM language-model weights.

#Technical Details

CryoAtom runs in two stages. A 3D U-Net predicts Cα positions from the density map, cropped into 64x64x64 patches and recombined; Cry-Net then generates all-atom coordinates, comprising an 18-block Cryformer encoder with non-shared weights and a 7-block Structure Module with shared weights, run over three recycles. ESM-2 sequence embeddings enter as attention keys and values. Post-processing — HMM-based sequence correction and pruning — follows ModelAngelo. Training used 6,422 EMDB maps released before April 2022 at better than 4 Å resolution, filtered by predicted-Cα precision to 5,731 map-model pairs.

On 177 held-out maps better than 4 Å, CryoAtom raises backbone recall to 97.17% from 94.37% and network-assigned amino-acid accuracy to 73.14% from 58.54% (87.09% versus 84.67% after HMM correction), with lower Cα RMSD (0.310 Å versus 0.334 Å), better MolProbity (3.56 versus 3.69) and better EMRinger (3.02 versus 2.90). On 104 maps at 4-7 Å the margin widens: completeness 36.8% versus 23.6%, backbone recall 43.0% versus 25.8%. A ~40,000-residue structure takes about 3.5 hours on a single A100, scaling linearly with length.

#Applications

CryoAtom targets the interpretation stage of single-particle cryo-EM, where a structural biologist has a reconstruction and needs an atomic model. It distinguishes paralogous sequences in noisy density, identifies uncharacterized components in native-source complexes whose composition is not fully known, isolates non-protein cofactors by compartmentalizing the map, and resolves subtle conformational differences between two reconstructions of the same assembly — a 104-protein complex was built in 5.6 hours, and a peripheral single-domain shift was detected by comparison. A Colab notebook covers users without a 14 GB GPU.

#Impact

By pairing map-derived spatial constraints with a structure-prediction architecture, CryoAtom demonstrates that model building benefits from the same advances that transformed folding, and its results are consistent enough across resolution that map quality, not training-set similarity, dominates performance: on a non-redundant subset filtered at 40% sequence identity, completeness rose rather than fell. The tradeoff is precision — because CryoAtom models flexible regions that deposited structures omit, those residues score as false positives, lowering backbone precision relative to ModelAngelo (75.5% versus 90.4% at stage 1). Low-resolution regions remain fundamentally hard, since side-chain density is unresolvable there; the authors point to conformational ensembles and molecular-dynamics flexible fitting as directions. Development continues under the CryoAtom2 name, which extends the approach to protein-nucleic acid complexes.