DeepCryoRNA

DeepCryoRNA rebuilds an all-atom RNA three-dimensional structure directly from a cryo-EM density map and the molecule's known sequence. It was developed by Jun Li and Shi-Jie Chen at the University of Missouri, the group behind the Vfold family of RNA structure tools, and posted as a preprint in April 2025.

Cryo-EM has made large, non-crystallizable RNAs tractable, but the software that turns a density map into coordinates was built overwhelmingly for proteins. RNA is more flexible and conformationally variable, and the few automated tools that handle it — auto-DRRAFTER, DeepTracer-2.0, and CryoREAD — either require a secondary structure and expensive conformational sampling, or produce fragmented chains that do not follow the native path. Benchmarking them on every pure-RNA cryo-EM structure solved to better than 6.0 Å made the gap concrete: DeepTracer failed outright on 41 of 51 targets, and CryoREAD returned structures for all 51 but with RMSDs from 18.4 Å to 73.6 Å.

DeepCryoRNA attacks the problem with a segmentation network followed by a long, deliberate assembly stage. A 3D U-Net classifies each voxel of the map into one of 19 classes — background plus 18 RNA atom types — and post-processing then clusters those atoms into nucleotides, links nucleotides into short chains, threads short chains into long chains, and assigns the native sequence by modified global alignment before a final energy minimization. The authors are explicit that the gain over prior deep-learning methods comes mostly from this post-processing, not the network alone.

#Key Features

• Eighteen atom classes: The network predicts 12 backbone atoms (P, OP1, OP2, C5', O5', C4', O4', C3', O3', C2', O2', C1') and 6 base atoms — a broader target set than competing methods, making nucleotide assembly far more reliable.
• Purine/pyrimidine simplification: Rather than distinguish all four bases, the model classifies only purines versus pyrimidines, acknowledging that adenine-versus-guanine discrimination is not recoverable from medium-resolution density.
• Global sequence alignment: Chain identity and nucleotide indexing come from a modified global alignment of the predicted chain against the known sequence, rather than the fragment-level alignment CryoREAD uses.
• Dual patch sizes at inference: Predictions from 64³ and 128³ map patches are merged, reducing patch-boundary artifacts and improving the odds of recovering the correct chain path over either size alone.
• No secondary structure required: The pipeline runs automatically from map plus sequence, with no secondary structure input and no conformational sampling.

#Technical Details

The network is a 3D MultiResUNet — a U-Net variant that replaces plain skip connections with multi-scale MultiRes blocks and Res Paths — with five encoder/decoder levels whose filter counts run 32, 64, 128, 256, and 512, and a final softmax convolution over the 19 atom classes. It has 18,683,016 parameters and is implemented in TensorFlow. Maps are resampled to a 0.5 Å voxel size with ChimeraX and normalized before use.

Because pure-RNA cryo-EM structures are scarce, training used 131 non-redundant RNAs extracted from RNA-protein complexes deposited before February 2022 at better than 7 Å, each at least 50 nucleotides long, with map volume within 5 Å of the RNA cropped out of the parent map. Those sequences span 50 to 15,871 nucleotides and map resolutions 2.0 to 7.0 Å, trained on 64³ patches with augmentation. The test set is 51 pure RNAs of 101 to 720 nucleotides with maps from 2.3 to 5.7 Å; over those targets the combined-patch model reached average and median heavy-atom RMSDs of 12.1 Å and 5.0 Å, with 36 structures between 1.6 Å and 6.9 Å, against CryoREAD's 44.3 Å average and 42.4 Å median. Runtimes ranged from 4 minutes to 3.5 hours on one GTX 1080 Ti with 10 CPU cores.

#Applications

The intended user is a cryo-EM practitioner holding a protein-free RNA map and its sequence who wants a starting model without weeks of manual tracing. Accuracy tracks map resolution closely: better than 4.5 Å consistently yields near-native models, below that performance becomes variable. Even in weaker cases it rebuilds more than 60% of nucleotides, enough to seed manual rebuilding or refinement in tools such as SimRNA or Phenix.

#Impact

DeepCryoRNA addresses a real bottleneck — RNA is the modality where automated cryo-EM model building has lagged furthest behind protein — and shows a large margin over the two existing deep-learning options on identical inputs. The scope is narrow by design and openly stated: it handles pure RNA only, not DNA and not RNA-protein complexes, and heterogeneous density can leave loops or even helices unmodeled. The work remains an unreviewed preprint, and the benchmark is a single set of 51 structures. Source code, the trained weights, and the input/output directories for all 51 test cases are distributed on GitHub under GPL-3.0, so the fixed checkpoint can be applied to a new map without retraining.