CryoNet.Refine: A One-step Diffusion Model for Rapid Refinement of Structural Models with Cryo-EM Density Map Restraints

Fuyao Huang; Xiaozhu Yu; Kui Xu; Qiangfeng Cliff Zhang

q-bio.BMFeb 25, 2026

CryoNet.Refine: A One-step Diffusion Model for Rapid Refinement of Structural Models with Cryo-EM Density Map Restraints

Authors: Fuyao Huang, Xiaozhu Yu, Kui Xu, Qiangfeng Cliff Zhang

Organizations: State Key Laboratory of Membrane Biology–Membrane Structure and Artificial Intelligence Biology Branch, Hangzhou, China · State Key Laboratory of Membrane Biology, Beijing Tsinghua Institute for Frontier Interdisciplinary Innovation, Beijing, China · State Key Laboratory of Membrane Biology, Beijing Frontier Research Center of Biological Structures, Tsinghua-Peking Joint Center for Life Sciences, School of Life Sciences, Tsinghua University, Beijing, China

Abstract

High-resolution structure determination by cryo-electron microscopy (cryo-EM) requires the accurate fitting of an atomic model into an experimental density map. Traditional refinement pipelines such as Phenix.real_space_refine and Rosetta are computationally expensive, demand extensive manual tuning, and present a significant bottleneck for researchers. We present CryoNet.Refine, an end-to-end deep learning framework that automates and accelerates molecular structure refinement. Our approach utilizes a one-step diffusion model that integrates a density-aware loss function with robust stereochemical restraints, enabling rapid optimization of a structure against experimental data. CryoNet.Refine provides a unified and versatile solution capable of refining protein complexes as well as DNA/RNA-protein complexes. In benchmarks against Phenix.real_space_refine, CryoNet.Refine consistently achieves substantial improvements in both model-map correlation and overall geometric quality metrics. By offering a scalable, automated, and powerful alternative, CryoNet.Refine aims to serve as an essential tool for next-generation cryo-EM structure refinement. Web server: https://cryonet.ai/refine; Source code: https://github.com/kuixu/cryonet.refine.

Figures & tables

Figure 1: (a) Workflow of CryoNet.Refine . Atomic models are colored by model–map correlation coefficients (CC), with blue indicating high CC and red indicating low CC. (b) The input atomic model, refined atomic model within cryo-EM density map.

Figure 2: (a) Overview of the CryoNet.Refine framework. The CryoNet.Refine Network consists of five modules: Atom encoder, Sequence embedder, Pairformer module, Diffusion module, and Density generator. The input atomic model is first processed by the encoders, and the resulting features are fed into a one-step diffusion module to generate an initial refined atomic model. Subsequently, the Density generator creates a synthetic density map, which is used to compute density loss against the input density map and geometry loss based on geometry restraints. These losses are then backpropagated to optimize the diffusion module, while the atomic model is further refined through multiple recycle steps until convergence. (b) Model-map correlation coefficient trajectory over 234 recycling CryoNet.Refine iterations on the structure of the human concentrative nucleoside transporter CNT3( ( Bank, ) ). The input density map is EMD-0775, the input atomic model is predicted by AlphaFold3 .

Category	Metrics	AlphaFold3	Phenix.real_space_refine	CryoNet.Refine
Model-map Correlation Coefficients	$\rm CC_{mask}\uparrow$	0.38	0.54	0.59
	$\rm CC_{box}\uparrow$	0.41	0.53	0.57
	$\rm CC_{mc}\uparrow$	0.40	0.55	0.60
	$\rm CC_{sc}\uparrow$	0.39	0.55	0.58
	$\rm CC_{peaks}\uparrow$	0.27	0.40	0.45
	$\rm CC_{volume}\uparrow$	0.42	0.55	0.60

Table 1: Performance on benchmark of 110 protein complexes(↑: higher is better, ↓: lower is better).

Figure 3: Model–map correlation coefficients on protein complex benchmark.

Figure 4: Model geometric metrics. (Color gradient: blue for better, red for worse.)

Figure 5: The input atomic models from AlphaFold3 , the refined atomic model from Phenix.real_space_refine and CryoNet.Refine on the Medicago truncatula HISN5 protein (PDB- 7oj5 ; EMD-22692) and the human concentrative nucleoside transporter CNT3 (PDB- 6ksw , EMD-0775) complex. Inserts in the right panel show that the main-chains and side-chains generated from CryoNet.Refine model align well with the density map.

Category	Metrics	AlphaFold3	Phenix.real_space_refine	CryoNet.Refine
Model–map Correlation Coefficients	$\rm CC_{mask}\uparrow$	0.40	0.57	0.65
	$\rm CC_{box}\uparrow$	0.49	0.61	0.67
	$\rm CC_{mc}\uparrow$	0.45	0.62	0.61
	$\rm CC_{sc}\uparrow$	0.42	0.58	0.67
	$\rm CC_{peaks}\uparrow$	0.35	0.51	0.60
	$\rm CC_{volume}\uparrow$	0.48	0.61	0.69

Table 2: Performance on DNA/RNA–protein complexes.

Figure 6: Model–map correlation coefficients on DNA/RNA-protein complex.

Figure 7: The input atomic models from AlphaFold3 , the refined atomic models from Phenix.real_space_refine and CryoNet.Refine are shown on the Cas12i(E894A)–crRNA–dsDNA (PDB- 6w5c , EMD-21541) and the active-state Dot1 bound to the H4K16ac nucleosome (PDB- 7k6q , EMD-22692) complex. Inserts in the right panel show that the main-chains and side-chains generated by CryoNet.Refine align well with the density map.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Dataset statistics of density map resolution, number of chains, and number of residues.

Figure 10: Data Patterns and Visualizations of the Density Generator Output

Loss Functions	Weight	Value
Density loss ( $\mathcal{L}_{\text{den}}$ )	$\gamma_{\text{den}}$	20.0
Ramachandran loss ( $\mathcal{L}_{\text{rama}}$ )	$\gamma_{\text{rama}}$	500
Rotation loss ( $\mathcal{L}_{\text{rot}}$ )	$\gamma_{\text{rot}}$	500
Angle loss ( $\mathcal{L}_{\text{angle}}$ )	$\gamma_{\text{angle}}$	2
C β loss ( $\mathcal{L}_{\text{C}_{\beta}}$ )	$\gamma_{\text{C}_{\beta}}$	50
Violation loss ( $\mathcal{L}_{\text{viol}}$ )	$\gamma_{\text{viol}}$	1000

Appendix

Table 3: Loss function weights

Figure 11: Visualization of Ramachandran distributions

Figure 12: Model-map Correlation Coefficients (CC) analysis of protein complexes across different recycling numbers (We provide a trajectory video over the entire 234 refinement recyclings in supplementary material).

Figure 13: Model-map Correlation Coefficients (CC) analysis of DNA/RNA–protein complex across different diffusion sampling steps.

Category	Metrics	CryoNet.Refine (One-step)	Classical Diffusion (200 steps)
Model-map Correlation Coefficients	$\rm CC_{mask}\uparrow$	0.65	0.30
	$\rm CC_{box}\uparrow$	0.58	0.35
	$\rm CC_{mc}\uparrow$	0.66	0.31
	$\rm CC_{sc}\uparrow$	0.63	0.31
	$\rm CC_{peaks}\uparrow$	0.47	0.24
	$\rm CC_{volume}\uparrow$	0.65	0.34

Appendix

Table 4: Quantitative comparison between CryoNet.Refine (One-step) and Classical Diffusion (200 steps) across 27 cases. The multi-step approach fails to preserve structural biology constraints when applied to refinement tasks.

Category	Metrics	$\gamma_{\mathrm{den}}=0$	$\gamma_{\mathrm{rama}}=0$	$\gamma_{\mathrm{rot}}=0$	CryoNet.Refine
Model-map Correlation Coefficients	$\rm CC_{mask}\uparrow$	$0.41$	0.65	0.64	0.65
	$\rm CC_{box}\uparrow$	$0.42$	0.58	$0.57$	0.58
	$\rm CC_{mc}\uparrow$	$0.42$	0.66	0.65	0.66
	$\rm CC_{sc}\uparrow$	$0.41$	0.64	0.64	0.63
	$\rm CC_{peaks}\uparrow$	$0.24$	0.48	0.47	$0.47$
	$\rm CC_{volume}\uparrow$	$0.43$	0.65	0.64	0.65

Appendix

Table 5: Ablation study: how loss functions influence model-map correlation coefficients and model geometric metrics

Category	Metrics	AlphaFold3	Phenix. real_space_refine	Numerical Optimization	CryoNet. Refine
Model-map Correlation Coefficients	$\rm CC_{mask}\uparrow$	0.44	0.61	0.46	0.65
	$\rm CC_{box}\uparrow$	0.44	0.55	0.47	0.58
	$\rm CC_{mc}\uparrow$	0.45	0.62	0.46	0.66
	$\rm CC_{sc}\uparrow$	0.44	0.61	0.45	0.64
	$\rm CC_{peaks}\uparrow$	0.31	0.44	0.32	0.47
	$\rm CC_{volume}\uparrow$	0.47	0.61	0.48	0.65

Appendix

Table 6: Comparison between CryoNet.Refine and Direct Numerical Optimization (SGD) on a subset of 27 protein complexes. The numerical baseline optimizes geometry well but fails to fit the density map.

Component	Time (ms)	Percentage (%)
Data Loading & Preprocessing	150	$\sim$ 6%
Model Forward Pass (One-step)	200	$\sim$ 7%
Loss Calculation (Density + Geometry)	940	$\sim$ 35%
Backpropagation & Optimization	1420	$\sim$ 52%
Total per Recycle	2710	100%

Appendix

Table 7: Runtime profile of a single refinement recycle (averaged over total recycle).

Cryo-Electron Microscopy

Cryo-electron microscopy reconstructs biological structures from images of frozen specimens, from cellular volumes to atomic-resolution molecular models.

See all 27 papers

CryoACE: An Atom-centric Framework for Accurate and Automated Model Building in Cryo-EM
Jun 30, 2026Minzhang Li, Mingrui Li, Weichen Qin +5Protein Structure Prediction Cryo-Electron Microscopy
School of Information Science and Technology, ShanghaiTech University, Shanghai, China · Cellverse Co, Ltd., Shanghai, China.
Enhancing Cryo-EM Density Map Segmentation in Phenix for Improved Atomic Model Building
May 6, 2026Chenwei ZhangAI-Assisted Scientific Research Cryo-Electron Microscopy
Department of Computer Science The University of British Columbia Vancouver, Canada
Fold'EM: Direct atomic structure inference from Cryo-EM particles
Oct 1, 2026Advaith Maddipatla, Märt-Erik Mäeots, Marco Pegoraro +5Cryo-Electron Microscopy Electron Microscopy
Institute of Science and Technology Austria · University of Zurich · Frankfurt Institute for Advanced Studies +2

Jun 30, 2026J/K move · Enter open · S save

CryoACE: An Atom-centric Framework for Accurate and Automated Model Building in Cryo-EM

Minzhang Li, Mingrui Li, Weichen Qin, Qihe Chen+4

School of Information Science and Technology, ShanghaiTech University, Shanghai, China · Cellverse Co, Ltd., Shanghai, China.

Protein Structure Prediction Cryo-Electron Microscopy

Open paper AskPDF arXiv

Protein automodeling from cryo-EM density maps faces unique challenges in enforcing physicochemical validity and managing conformational heterogeneity. Current solvers are often limited to static predictions or require computationally intensive heuristic searches. We present CryoACE, an end-to-end framework that reconstructs precise atomic graphs for both homogeneous and heterogeneous structures. Our method features two key innovations: an atom-centric reconstruction paradigm, where density features are sampled directly at atomic coordinates and iteratively recycled to refine structures, replacing expensive voxel convolutions for efficient multimodal fusion; and a training-free guidance mechanism that leverages predicted local resolution priors to resolve dynamic ambiguity. Validated on a newly constructed high-quality dataset, CryoACE significantly outperforms existing baselines on static benchmarks and, for the first time, unveils atomic-level dynamic conformations on complex real-world datasets like EMPIAR-10345 without relying on pre-built static structures.

Abstract

Cryo-Electron Microscopy

Explore similar work

CryoACE: An Atom-centric Framework for Accurate and Automated Model Building in Cryo-EM

Enhancing Cryo-EM Density Map Segmentation in Phenix for Improved Atomic Model Building

Fold'EM: Direct atomic structure inference from Cryo-EM particles