CryoACE

CryoACE is an end-to-end deep learning framework that builds atomic protein models directly from cryo-electron microscopy (cryo-EM) density maps. Developed by researchers at ShanghaiTech University and released as an arXiv preprint on June 30, 2026, it automates the interpretation of a three-dimensional density map into a physically valid atomic structure — the final and often most difficult stage of the cryo-EM structure-determination pipeline. The work comes from the ShanghaiTech group behind earlier cryo-EM AI methods such as CryoGEM, DRACO, and CryoFormer.

Two challenges have long limited automated model building: enforcing the physicochemical validity of the reconstructed atoms, and coping with conformational heterogeneity, where a single map represents an ensemble of moving states rather than one rigid structure. CryoACE addresses both with an atom-centric design. Instead of running expensive three-dimensional convolutions over the full voxel grid, it samples density features directly at candidate atomic coordinates and iteratively recycles them to refine the model, keeping the representation anchored to physical atoms while making sequence-and-density fusion efficient.

Within the cryo-EM machine learning landscape, CryoACE is a pretrained structural-biology model in the same family as automated builders such as ModelAngelo, but it couples density interpretation with a Boltz-1-style sequence prior and a diffusion decoder, and it explicitly targets heterogeneous, dynamic structures rather than assuming a single static conformation.

#Key Features

• Atom-centric reconstruction: Samples the local density environment at candidate atomic positions via trilinear interpolation and iteratively recycles those features, replacing costly voxel convolutions for efficient multimodal fusion.
• Diffusion-based atom decoder: An atom-transformer decoder generates atomic coordinates through a conditional reverse diffusion process conditioned on the fused sequence and density representation.
• Sequence and density fusion: A Boltz-1-style sequence encoder (MSA module plus Pairformer) is combined with a 3D ResUNet density encoder through cross-attention blocks.
• Training-free resolution guidance: An inference-time guidance mechanism leverages predicted local-resolution and Q-score priors to resolve dynamic ambiguity without any per-map retraining.
• Heterogeneity handling: Reconstructs precise atomic graphs for both homogeneous structures and heterogeneous, multi-state assemblies.

#Technical Details

The architecture pairs a sequence encoder (MSA module and Pairformer) with a 3D ResUNet density encoder using three-dimensional rotary positional embeddings, fuses the two through cross-attention, and decodes coordinates with an atom-transformer diffusion decoder, alongside auxiliary local-resolution and Q-score prediction heads. CryoACE is trained once on a curated dataset of 10,915 high-quality density–structure–sequence triplets, distilled from 13,775 candidates, restricted to resolutions better than 4 Angstroms and to structures released before 2025. Training ran for 500 epochs on 32 NVIDIA H20 GPUs with the AdamW optimizer. At inference the weights are frozen, and only the iterative training-free guidance varies per input map. The model is evaluated on the held-out Cryo2StructData test set against ModelAngelo, CryoAtom, E3-CryoFold, Boltz-1, and CryoBoltz across completeness, backbone and all-atom accuracy, RMSD, Q-score, weighted cross-correlation, and MolProbity, with the authors reporting state-of-the-art results on these benchmarks.

#Applications

CryoACE targets structural biologists who need to convert cryo-EM density maps into atomic models, particularly for flexible complexes where a single map captures multiple conformational states. Because it does not require a pre-built static reference structure, it can recover dynamic ensembles directly: on the real heterogeneous datasets EMPIAR-10345 (αVβ8 integrin) and EMPIAR-10516 (SARS-CoV-2 spike protein), it reconstructs atomic-level dynamic structures without per-target retraining. A stated limitation is that the training set lacks nucleic-acid complexes, so DNA and RNA accuracy relies on pretrained Boltz priors rather than representations learned within the pipeline.

#Impact

CryoACE extends automated cryo-EM model building beyond the common assumption of a single rigid conformation, pushing atom-centric reconstruction toward the conformational heterogeneity that pervades real specimens. Its combination of a sequence prior, density encoder, and diffusion decoder demonstrates that experimentally observed density and evolutionary sequence signal can be fused efficiently at the level of individual atoms. The authors state that they will release the code, model weights, and training dataset. As an unreviewed preprint with no public repository located yet, its adoption and reproducibility remain open questions pending that release and peer review.