henhao
Sign in
Molecular BiologyScientific Reports
Peer-reviewed journal article

Cascaded-CNN: Deep Learning to Predict Protein Backbone Structure from High-Resolution Cryo-EM Density Maps

Dong SiiD, Spencer A. Moritz, Jonas PfabiD, Jie HouiD, Renzhi CaoiD, Liguo WangiD, Tianqi WuiD, Jianlin Jack ChengiD

Abstract

Cryo-electron microscopy (cryo-EM) has become a leading technology for determining protein structures. Recent advances in this field have allowed for atomic resolution. However, predicting the backbone trace of a protein has remained a challenge on all but the most pristine density maps (< 2.5[A] resolution). Here we introduce a deep learning model that uses a set of cascaded convolutional neural networks (CNNs) to predict C atoms along a proteins backbone structure. The cascaded-CNN (C-CNN) is a novel deep learning architecture comprised of multiple CNNs, each predicting a specific aspect of a proteins structure. This model predicts secondary structure elements (SSEs), backbone structure, and C atoms, combining the results of each to produce a complete prediction map. The cascaded-CNN is a semantic segmentation image classifier and was trained using thousands of simulated density maps. This method is largely automatic and only requires a recommended threshold value for each evaluated protein. A specialized tabu-search path walking algorithm was used to produce an initial backbone trace with C placements. A helix-refinement algorithm made further improvements to the -helix SSEs of the backbone trace. Finally, a novel quality assessment-based combinatorial algorithm was used to effectively map C traces to obtain full-atom protein structures. This method was tested on 50 experimental maps between 2.6[A] and 4.4[A] resolution. It outperformed several state-of-the-art prediction methods including RosettaES, MAINMAST, and a Phenix based method by producing the most complete prediction models, as measured by percentage of found C atoms. This method accurately predicted 88.5% (mean) of the C atoms within 3[A] of a proteins backbone structure surpassing the 66.8% mark achieved by the leading alternate method (Phenix based fully automatic method) on the same set of density maps. The C-CNN also achieved an average RMSD of 1.23[A] for all 50 experimental density maps which is similar to the Phenix based fully automatic method. The source code and demo of this research has been published at https://github.com/DrDongSi/Ca-Backbone-Prediction.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
108
FWCI
5.19
Field percentile
top 3%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (12 Mar 2019) · CC BY 4.0

Cascaded CNNs segmenting secondary structure, backbone and Cα voxels, plus tabu-search tracing, place 88.5% of Cα atoms within 3Å across 50 cryo-EM maps versus Phenix's 66.8%.

Tracing backbones and placing Cα atoms in cryo-EM density is reliable only on the cleanest maps (<2.5Å), and existing tools (a Phenix-based automatic pipeline, RosettaES, MAINMAST) leave models incomplete or demand heavy compute. This work trains semantic-segmentation 3D CNNs on simulated maps and adds tracing, helix refinement and sequence mapping to obtain more complete Cα models from 2.6Å to 4.4Å experimental maps.

Key results

  1. ShownOn experimental cryo-EM maps, the pipeline builds substantially more complete Cα backbone models than the Phenix-based fully automatic method, mainly because it traces low-confidence regions where Phenix places no atoms.

    Mean 88.5% of ground-truth Cα atoms matched within 3Å vs 66.8% for the Phenix method on the same 50 EMDB maps spanning 2.6Å to 4.4Å; the C-CNN exceeded Phenix at every resolution (Fig. 15). No statistical test is reported. · Abstract; C. Experimental Density Maps; G. Comparison of Deep-Learning C-NN and the…

  2. ShownThe placed Cα atoms are as accurately positioned as Phenix's, and false-positive placements stay low on better-resolved maps.

    Mean RMSD 1.23Å (C-CNN) vs 1.22Å (Phenix) over the 50 maps, rising with labeled resolution for both methods. The Cα error rate (false positives more than 3Å from any true Cα, divided by predicted Cα) was mostly below ∼5% for maps at 3.7Å or better and varied widely for poorer maps. · C. Experimental Density Maps; G. Comparison of Deep-Learning C-NN and the…; A. False Positive Cα Error Rate

  3. SuggestedIn a head-to-head test on one map, the C-CNN built the most complete model and had fewer false positives than Phenix. It also needed far less compute than RosettaES or MAINMAST.

    EMDB-5778 chain A: RosettaES was most accurate but found only 7.9% of Cα; MAINMAST had RMSD 2.20Å and the highest false-positive share; the C-CNN's RMSD was slightly higher than Phenix's. Runtime was about 20 minutes for ~1000 Cα (GTX 1070, single CPU core) vs 5 days on 20 CPUs (RosettaES), about 18 hours on 1 CPU (MAINMAST), several hours on 6 CPUs (Phenix). · E. Comparison of Prediction Methods; F. Computation Time of Prediction Models

  4. ShownA geometric α-helix refinement, which fits ideal helices along an estimated screw axis, slightly raises completeness overall and raises it sharply on some maps, without changing mean RMSD.

    Across experimental maps, refinement raised the share of Cα within 3Å by 1.5% while average RMSD held steady. The largest gain was up to ∼15% for EMDB-8637 (Table 3). RMSD improved less because the added helix Cα atoms carry more weight in the overall average. · I. Helix Refinement; D. Impact of Helix Refinement

  5. SuggestedA combinatorial sequence-mapping algorithm guided by quality assessment can assign sequence to long, well-traced Cα segments and produce full-atom models. Results on the shorter segments were mixed, with several scoring below a TM-score of 0.5.

    EMDB-5778: the longest segment (185 residues) scored TM-score 0.9478 and GDT-TS 0.8723; the second segment scored TM-score 0.3947 (incorrect topology) and the third 0.5978. EMDB-8410: the longest segment (234 residues) scored TM-score 0.8816; the other segments fell below 0.5. Only two targets were tested. · J. Mapping Protein Sequences onto Cα traces; H. Evaluating the Results of Mapping Protein…

Methods, data and reproducibility

Experimental maps are resampled to 1Å voxels, thresholded, normalized and split into overlapping 64×64×64 cubes. These pass through three chained dilated 3D CNNs (SSE, backbone, Cα) trained on simulated maps. Tabu-search path walking, graph pruning, helix refinement and Qprob-guided sequence mapping then yield Cα traces and full-atom models.

  • Simulated training density maps made with EMAN2 pdb2mrc from PDB entries that are at least 50% sequence-unique, simulated at varied resolutions and rotated for augmentation · n = over 7,000 PDB files; approximately 24,000 training maps
  • Simulated held-out test maps used to monitor network accuracy during training · n = 1,024 maps
  • Simulated 6Å pdb2mrc evaluation maps (described as comparable to ~3.5Å experimental maps) · n = seven maps · PDB 6eyw (example shown)
  • Experimental cryo-EM density maps from the EM Data Bank at 2.6Å to 4.4Å, cropped to about 5Å around the deposited model · n = 50 maps · EMDB-6272, EMDB-8410, EMDB-5778, EMDB-8637, EMDB-8642 (examples)
  • Map used for the four-method comparison (C-CNN, Phenix, MAINMAST, RosettaES), chain A at 3.28Å · EMDB-5778
  • Validation targets for sequence-to-trace mapping, scored with TM-score and GDT-TS · n = two targets · EMDB-5778; EMDB-8410
Methods & tools
EMAN2 pdb2mrc, TensorFlow (Python), UCSF Chimera (vop resample), Custom tabu-search path-walking algorithm, Pulchra, Scwrl, Qprob, TM-score / GDT-TS, SciPy optimize.minimize, Phenix-based fully automatic method (benchmark), RosettaES (benchmark), MAINMAST (benchmark)
Code
https://github.com/DrDongSi/Ca-Backbone-Prediction
Data & accessions
https://www.rcsb.org/https://www.emdataresource.org/index.html
Software & environment
Python with the TensorFlow library; UCSF Chimera vop resample for 1Å voxel resampling; SciPy minimize for helix fitting. No software versions or containers are given.
Assay conditions & compute
Nvidia GTX 1070 GPU: training ran 15 epochs in about 24 hours; C-CNN inference took about 15 seconds for a ~100×100×100 voxel map. Path walking took about 20 minutes for ~1000 Cα atoms on an Intel i7-8700K (6 core, 3.7GHz, 16GB RAM).
Reproducibility gaps
Each experimental map needs a density threshold chosen by eye, and evaluation masks density using the deposited model. Release of trained weights and of the simulated training set is not stated; only the source code and a demo are mentioned.
Limitations (6)
  • Each experimental map needs a manually chosen density threshold, picked by eye in Chimera so its SSEs resemble those in the simulated maps. The authors call this step subjective and propose automating it. · stated by the authors
  • Training used only simulated maps, which may cause overfitting to simulator features. The near-perfect simulated test results (mean RMSD 0.90Å, 98.7% within 3Å) used the same pdb2mrc script as training and contain no experimental noise. · stated by the authors
  • The RMSD metric, adopted from the Phenix method, pairs each predicted Cα with its nearest true Cα and allows skipped residues, which the authors note gives lower RMSD values than other methods. · stated by the authors
  • The Cα error rate becomes highly variable for maps worse than 3.7Å. Sequence mapping gave incorrect topologies (TM-score below 0.5) for some shorter segments, which the authors attribute to trace noise, mapping inaccuracy, or both. · stated by the authors
  • Experimental maps were masked using the deposited atomic model (voxels beyond about 5Å or 6Å zeroed), which is information unavailable in true de novo building. The text does not state that Phenix received identically masked maps, so the 88.5% vs 66.8% gap may be inflated. · noted in this summary
  • The comparison with RosettaES and MAINMAST rests on a single map (EMDB-5778), the Phenix comparison reports means without statistical testing, and the training-set size is reported inconsistently (approximately 24,000 vs 25,000 maps). · noted in this summary

Who should read:Read this if you build or benchmark automated Cα backbone tracing from 2.6Å to 4.4Å cryo-EM maps and want a cascaded 3D-CNN segmentation pipeline with tabu-search tracing and sequence mapping.

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
THISPfab '20He '21Li '25Terashi '23Mondal '25Giri '24
85% semantic matchBioinformatics
Jonas Pfab, Dong Si·bioRxiv·Feb 12, 2020· 3 citations