Cryo-EM density-map-to-atomic-structure modeling that fuses protein language model embeddings with density voxels, then refines with AlphaFold3.
No providers recorded yet. Browse all providers
Cryo-electron microscopy resolves large protein assemblies at near-atomic detail, but converting the resulting density map into a chain-traced atomic model is still a bottleneck. Missing electron density and locally poor resolution break automated tracers, and multi-chain complexes compound the problem because every chain's sequence must be registered onto the right stretch of density.
Cryo2Struct2, developed by Nabin Giri and Jianlin Cheng at the University of Missouri, attacks this by treating the map and the sequence as two modalities of one problem. A shared transformer encoder reads the density map while embeddings from the ESM-2 3-billion-parameter protein language model are injected into its feature maps; two task-specific decoders then label every voxel with an atom type and an amino acid type. Those voxel probabilities become the states of a hidden Markov model, and a modified Viterbi algorithm threads each chain's sequence through the predicted alpha-carbon positions to build a backbone. The backbone is not the final answer: it is handed to AlphaFold 3 as a structural template.
That last step is what distinguishes Cryo2Struct2 from its predecessor, Cryo2Struct, and from the same group's MICA, which fuses AlphaFold3 predictions into the network's input. Cryo2Struct2 uses late fusion instead — the experimental evidence enters AlphaFold3 as a template, letting the predictor refine regions the density cannot constrain while staying anchored to the map where it can.
The network is a 3D SegFormer variant. Full density maps are resampled and cut into 32x32x32 sub-cubes, each a single-channel tensor of electron density values, and passed through four transformer stages using efficient self-attention with spatial-reduction ratios of 4, 2, 1, and 1; a 3D Mix-FFN block supplies positional cues in place of fixed encodings. The atom decoder emits four classes (alpha-carbon, N, C, and no atom) and the amino acid decoder 21. Training used Cryo2StructData — cryo-EM maps at 1.0-4.0 Å released through 27 March 2023, split into 6,652 training and 740 validation maps — with a weighted cross-entropy loss to counter voxel class imbalance, Adam at a learning rate of 1e-4, a batch size of 1,000, and four 80 GB GPUs.
Evaluation used 61 EMDB maps deposited in 2024, post-dating both the Cryo2StructData cutoff and AlphaFold3's own September 2021 training cutoff. The maps average 3.10 Å resolution (range 2.4-3.9 Å) and 1,230.8 residues (range 374-3,245). Scored with US-align across all chains and normalized by the experimental structure length, AlphaFold3 guided by Cryo2Struct2 templates averaged a TM-score of 0.32, against 0.28 for AlphaFold3 without templates and 0.21 for the Cryo2Struct2 backbones alone. Sequence identity followed the same ordering: 0.35, 0.25, and 0.12. These are whole-assembly scores on large multi-chain complexes, where absolute values are low for every method tested; the result is the consistent gain from adding experimental templates.
Cryo2Struct2 fits the model-building stage of a conventional cryo-EM workflow: a structural biologist supplies a reconstructed map and the complex's sequences and receives an atomic model to inspect and correct. It targets the cases that defeat density-only tracers — large multi-chain assemblies, flexible conformations, and maps whose local resolution degrades across the reconstruction. Because the refinement step runs AlphaFold3 in a Docker container, deployment inherits that model's access terms and hardware requirements alongside ChimeraX and the ESM stack.
The work is a bioRxiv preprint and has not been peer reviewed. Its practical contribution is a demonstration that experimental density and a general structure predictor reinforce each other in both directions: cryo-EM tracing supplies templates that measurably improve AlphaFold3, and AlphaFold3 repairs regions where the density is uninformative. The alignment machinery inherited from Cryo2Struct is reusable on its own, and the authors point toward extending the pipeline to complexes containing nucleic acids, ligands, and modified residues. Released code, checkpoints, and a labeled training corpus make the reported numbers reproducible.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.