pocketz.doi.bio/2026-01-10_085303_nature_s41594-025-01723-1

EMProt improves structure determination from cryo-EM maps | Nature Structural & Molecular Biology

URL: https://www.nature.com/articles/s41594-025-01723-1

Saved: 2026-01-10T16:53:08.556Z

Technical Report Published: 08 December 2025 EMProt improves structure determination from cryo-EM maps Tao Li, Ji Chen, Hao Li, Hong Cao & Sheng-You Huang 

Nature Structural & Molecular Biology (2025)Cite this article

2533 Accesses

1 Citations

3 Altmetric

Metrics details

Abstract

Cryo-electron microscopy (cryo-EM) has become the mainstream technique for macromolecular structure determination. However, because of intrinsic resolution heterogeneity, accurate modeling of all-atom structure from cryo-EM maps remains challenging even for maps at near-atomic resolution. Addressing the challenge, we present EMProt, a fully automated method for accurate protein structure determination from cryo-EM maps by efficiently integrating map information and structure prediction with a three-track attention network. EMProt is extensively evaluated on a diverse test set of 177 experimental cryo-EM maps with up to 54 chains in a case at <4-Å resolution, and compared to state-of-the-art methods including DeepMainmast, ModelAngelo, phenix.dockandrebuild and AlphaFold3. It is shown that EMProt greatly outperforms the existing methods in recovering the protein structure and building the complete structure. In addition, the built models by EMrot exhibit a high accuracy in model-to-map fit and structure validations.

You have full access to this article via your institution.

Download PDF Similar content being viewed by others Model building of protein complexes from intermediate-resolution cryo-EM maps with deep learning-guided automatic assembly Article Open access 13 July 2022 CryoAtom improves model building for cryo-EM Article 14 November 2025 Cryo-EM model validation recommendations based on outcomes of the 2019 EMDataResource challenge Article Open access 04 February 2021 Main

Proteins have important roles in many biological processes by themselves or interacting with other molecules. Therefore, determining the atomic structures of proteins in three dimensions is crucial for understanding molecular life processes. In the past decade, cryo-electron microscopy (cryo-EM) has greatly advanced the process to determine the structures of biological macromolecules, achieving the resolvability of individual atoms1,2,3. As a result, the EM Data Bank (EMDB)4 has seen an exponential increase in the number of deposited cryo-EM maps. Various methods have been developed for image processing and model building of cryo-EM maps5,6,7,8,9. Many macromolecular complex structures have been resolved and deposited to the Protein Data Bank (PDB)10. Nevertheless, model building of cryo-EM maps remains a labor-intensive and challenging process and requires extensive human expertise.

For years, a number of methods have been proposed to reduce the obstacles for accurately modeling structures from cryo-EM maps. Current cryo-EM model building methods can be broadly categorized into two types: de novo modeling and template-based fitting. De novo modeling methods, such as phenix.maptomodel11, MAINMAST12, DeepMM13, C-CNN14, A2-Net15, DeepTracer16 and Cryo2Struct17, often start by detecting critical atom positions from map and threading them to chains. ModelAngelo18, as the state-of-the-art de novo modeling tool, has received notable attention. While these methods demonstrate the feasibility of modeling a structure directly from raw density maps without additional knowledge, they often fail in density regions with poor local quality, even for maps at near-atomic resolution.

For template-based fitting approaches, researchers often fit a template structure into the density map as an initial placement and then refine the structure manually for multiple rounds. With the advancement of structure prediction algorithms such as AlphaFold2 (AF2) and AF3 (refs. 19,20), fitting highly accurate predicted structures into cryo-EM maps is frequently adopted in the field. Many automated tools such as gmfit21, Situs22,23, EMBuild24, DEMO-EM25,26 and phenix.dockandrebuild (PHENIX)27 have been developed to minimize the need for manual intervention. These automated tools are capable of modeling a reliable structure if the predicted structure is sufficiently accurate to be fitted into the map. However, they often fail if the predicted global fold deviates substantially from the actual ground-truth structure.

The difference between two modeling categories is sometimes blurred as several methods attempt to integrate the information from both maps and templates. CR-I-TASSER28 uses the Cα atom position predicted from density map to reselect a suitable template; however, it is primarily designed for segmented single-chain maps and does not use structure predictions by AF. DeepMainmast29, an enhanced version of MAINMAST, shows the potential of automated AF2-guided de novo modeling. However, the structure-to-map fitting in DeepMainmast is not exhaustive and the integration of structure and map is also kind of empirical. As such, its practical applicability may be limited by time cost and modeling accuracy.

With the exponential increase in released cryo-EM maps at near-atomic resolution, there is a pressing need for automated and accurate modeling for the complete structure from a map with weak local densities. Meeting this need, we introduce EMProt, an automated modeling approach for accurate protein structure determination from cryo-EM maps by effectively integrating de novo modeling and structure prediction using a three-track attention network. EMProt first uses deep leaning to predict the amino acid types and the positions of Cα atoms from raw density maps. Then, a three-track attention network is used to build the all-atom structure by integrating the one-dimensional (1D) and two-dimensional (2D) density features of predicted Cα positions and the three-dimensional (3D) backbone frame into the reasoning of 3D atomic coordinates. Moreover, EMProt is also able to integrate predicted structures through a combination of structure alignment to de novo built model and fast Fourier transform (FFT)-based structure-to-map fitting. EMProt is evaluated on a benchmark set of 177 experimental maps with resolutions better than 4 Å and compared to state-of-the-art methods, including DeepMainmast, ModelAngelo, PHENIX and AF3. Overall, EMProt achieves the best performance among the five approaches in terms of both recovering the protein structure and building the complete model.

Results The workflow of EMProt

The overall workflow of EMProt is illustrated in Fig. 1a. EMProt first threads an initial protein model using only information from the density map. To achieve this, two separately trained Swin-Conv-UNet (SCUNet) networks are used simultaneously; one predicts the voxel-wise probabilities of N, Cα and C atoms, while the other predicts the amino acid type probability. Afterward, the Cα positions are derived from the probability map using a mean-shift algorithm. At near-atomic resolution, density maps often lack the precision needed to accurately detect all backbone atoms, let alone side-chain atoms. In addition, the predicted Cα positions often deviate from their true locations. Therefore, to accurately build the all-atom structure, we design a three-track attention network19,30 to refine Cα positions and predict backbone frames and torsion angles from raw Cα atoms and the density map (Fig. 1c,d). The three-track attention network is trained by noising the PDB atoms and having the network to recover the true positions. As for feature generation, with initial Cα positions, we extract the densities around each atom as a 1D representation and the densities between pairwise atoms as a 2D representation (schematic illustration in Supplementary Fig. 1). The Cα positions and randomly initialized rotation matrices are used as 3D backbone frames. We use self-attention and invariant point attention to update the 1D representation with attention biases from the 2D representation and 3D backbone frames. We use axial attention to update the 2D representation with attention biases from the 3D backbone frames. The trained network is able to model all-atom positions given only Cα atoms by predicting backbone frames and torsion angles.

Fig. 1: Overview of the EMProt framework.

a, The workflow of EMProt. The input is the density map and sequence(s). Left: in the de novo pipeline, EMProt goes through Cα prediction, backbone frame prediction, chain tracing and sequence assignment. Right: in the integrative pipeline, structures are predicted using input sequences and fitted to the map at the fragment level; the output structure is assembled on the basis of all fragments. Finally, a postrefinement step can be performed to further improve the model-to-map fit and structure quality of the built model. b, The all-atom prediction and refinement module. Node and edge features are initialized using Cα atoms and the density map. After the node, edge and backbone frames communicate through the three-track attention network, a final layer is used to predict torsion angles and all-atom positions are built accordingly. c, The three-track attention module. The biased self-attention is used to update node features and the biased axial attention is used to update edge features. The invariant point attention is used to fuse node, edge features and backbone frame information. d, The structure fitting and assembling module. The AF2 models are integrated by structure alignment and semiflexible fitting. Finally, we select nonclash fragments on a graph built on all fragments.

Full size image

After the backbone frames and torsion angles are predicted, we thread individual residues into continuous chains under constraints from peptide bond length. With the threaded chains, we assign the amino acid type probability for each path and then perform a hidden Markov model (HMM) search against the provided sequences31,32. To achieve this, the initial amino acid probability for each residue is determined by finding the nearest voxel in the predicted amino acid type map. These probabilities are converted into the HMM profiles as position-specific scores using HMMER, which are then used to search the input sequences and identify the matching regions, yielding the final residue sequence. With the amino acid identities assigned for each path, the side-chain atoms are constructed according to the predicted torsion angles.

After de novo modeling, EMProt may further integrate predicted structures by matching them with the density map (Fig. 1b). We use AF2-predicted structures in this study. As the predicted model may have deviations from the true structure, EMProt first splits predicted chain models into fragments on the basis of structural contacts. Then, EMProt aligns the AF2 fragments with the de novo built fragments using TMalign33. The alignment establishes the residue correspondences between de novo built fragments and AF2 structure fragments, enabling EMProt to (1) identify the regions in de novo fragments with incorrect amino acid identities and (2) detect missing regions in the de novo model. If an AF2 fragment aligns well with a de novo fragment (for example, with a root-mean-square deviation (r.m.s.d.) < 1.5 Å), EMProt refines the amino acid identities in the de novo model and fills in missing residues using the AF2-derived structure. Next, EMProt performs model-to-map fitting, enabling local substructural movements to improve the fit into the main-chain probability map through a semiflexible refinement process. The corrected de novo fragments and the fitted AF2 structures are combined into a fragment library.

As multiple fragments may occupy adjacent positions, structural clashes can occur. To resolve this, EMProt applies a fragment-assembling algorithm that selects well-matched, nonclashing fragments to build the final model. A graph-based optimization algorithm is used, where each fragment is treated as a node and edges are added between nonclashing fragments. The algorithm solves a constraint optimization problem on this graph, retaining only the fragments that fit the density map while avoiding clashes. Finally, the assembled complex structure can be refined by a postrefinement program such as phenix.realspacerefine34, which is recommended for improving the local geometry and model-to-map fit. In this study, all the models were refined by phenix.realspacerefine during the evaluations, unless otherwise specified.

Performance in recovering the protein structure

EMProt is first validated on a test set of 177 nonredundant experimental maps and compared to state-of-the-art methods that use different modeling strategies, including ModelAngelo (de novo modeling), PHENIX (template-based modeling) and DeepMainmast (a hybrid of de novo and template-based modeling). Following previous studies11,16,17, three metrics, residue coverage, sequence match and sequence recall, are used to evaluate the performance of different methods in recovering the protein structure (Supplementary Table 1). The corresponding values are calculated using phenix.chain_comparison with a distance cutoff of 3.0 Å and Cα atoms as the representatives11.

Figure 2 shows the bar plots and head-to-head comparisons of the corresponding results by EMProt and the three other methods. As shown in Fig. 2a,d, EMProt outperformed the other methods in all metrics. Specifically, for residue coverage, EMProt achieved the highest average value of 91.20%, compared to 75.13% for DeepMainmast, 77.81% for ModelAngelo and 79.25% for PHENIX. Among the 177 maps, EMProt obtained a higher coverage than DeepMainmast for 130 cases, ModelAngelo for 163 cases and PHENIX for 131 cases. In addition, EMProt performed better than the other methods in both sequence match and sequence recall, achieving the high values of 95.87% and 87.79%, respectively, compared to 83.80% and 68.72% for DeepMainmast, 94.98% and 75.08% for ModelAngelo and 92.35% and 74.85% for PHENIX (Fig. 2b,c,e,f). In detail, EMProt had a better sequence match for 123, 111 and 83 cases and a better sequence recall for 136, 159 and 127 cases compared to DeepMainmast, ModelAngelo and PHENIX, respectively. The sequence match metric only considers the built parts, whereas the sequence recall accounts for all residues in the PDB structure. In other words, both EMProt and ModelAngelo could correctly assign nearly all amino acid types in the built model but EMProt correctly built many more residues than ModelAngelo.

Fig. 2: Comparison of EMProt, DeepMainmast, ModelAngelo and PHENIX in recovering the PDB structures on the test set of n = 177 maps.

a–c, Bar plots of residue coverage (a), sequence match (b) and sequence recall (c). The bars indicate the means and the error bars indicate the 95% confidence interval. Seq., sequence. d–f, Head-to-head comparison of residue coverage (d), sequence match (e) and sequence recall (f). g, Three examples of the high-quality models built by EMProt for three targets.

Source data

Full size image

As map resolution directly affects cryo-EM structure modeling, we next investigated the impact of map resolution on modeling performance. Sequence recall was chosen here to represent the overall performance of a method. For a comprehensive comparison, we report the performances of both the de novo modeling mode and the full pipeline of EMProt (Supplementary Fig. 2). It is shown that the sequence recall decreased as resolution decreased in both EMProt running modes, as expected. Overall, EMProt showed a consistently good sequence recall for maps with 2.5-Å resolution or better and its performances varied greatly for lower-resolution maps. However, when the AF2 structure was incorporated, the performance decline for the full pipeline was noticeably slower than that for the de novo mode, suggesting the necessity of integrating AF2 models into model building of cryo-EM maps.

Figure 2g lists three EMProt-built models with >99% sequence recall. The targets were EMD-27431 (PDB 8DH7), EMD-27639 (PDB 8DPN) and EMD-27661 (PDB 8DQV) from left to right, with resolutions of 2.99 Å, 2.49 Å and 1.52 Å, respectively. It can be seen from the figure that the built models for the three targets showed a perfect overlap with the PDB structures even in loop regions. It can also be noted that, in addition to the three examples shown here, there were a total of 25 built models with a sequence recall of >99% and 100 models with a sequence recall of >90%. Most cases had a resolution better than 3.0 Å, a relatively high resolution for EMProt to recover most residues from the map.

Examples in recovering the protein structure

Several examples are given in Fig. 3 to show the extent to which the models built by EMProt, DeepMainmast, ModelAngelo and PHENIX recover a protein structure. Figure 3a shows the built models for target EMD-15220 (PDB 8A7D) at 3.1-Å resolution. The built models are colored in blue and the PDB structures are colored in green. For this target, EMProt achieved a sequence recall of 91.06%, which was much higher than 75.85% for DeepMainmast, 73.04% for ModelAngelo and 76.68% for PHENIX. When compared to the PDB structure, the EMProt model showed good consistency with the reference PDB structure, whereas DeepMainmast and ModelAngelo failed to recover the residues in two regions, as indicated in the dashed rectangles, and PHENIX failed in one of these regions.

Fig. 3: Comparison of the built models by EMProt, DeepMainmast, ModelAngelo and PHENIX in terms of recovering the PDB structures.

a, Target EMD-15220 (PDB 8A7D) at 3.1-Å resolution. b, Target EMD-27253 (PDB 8D8O) at 3.4-Å resolution. The black dashed outlines indicate those regions where EMprot correctly builds the structures while the other methods fail.

Full size image

Figure 3b shows another comparison of the built models using the four methods on target EMD-27253 (PDB 8D8O) at 3.4-Å resolution. The sequence recall values were 86.35% for EMProt, 65.47% for DeepMainmast, 55.07% for ModelAngelo and 49.33% for PHENIX. Once again, EMProt demonstrated superior performance by achieving the highest sequence recall, while the other three methods exhibited different degrees of structural deficiency and failed to build the structure in the regions, as indicated by the dashed ellipses. Specifically, for ModelAngelo and PHENIX, a very large area of the structure was missing, resulting a low sequence match of <60%.

Performance in building the complete structure

We further evaluate the ability of EMProt and the other methods in building the complete structure using two global metrics: template modeling score (TM-score) and r.m.s.d. The metrics were calculated using USalign35 in multimer mode. Here, we intentionally exclude reporting the two metrics for ModelAngelo, as they are not suitable for the nonordered structural fragments built by ModelAngelo. In addition, as protein multimer structure prediction is greatly improved in AF3, we also list its performance for comparison on this test set. The AF3 structures are predicted by the AF3 server and the structure with the highest pTM score is selected among five predictions. The detailed results for TM-score and r.m.s.d. are listed in Supplementary Table 2.

Figure 4 shows the bar plots and head-to-head comparisons of the corresponding results using EMProt and the three other methods (DeepMainmast, PHENIX and AF3). It can be seen from the figure that EMProt outperformed the three other methods in terms of both TM-score and r.m.s.d. For TM-score (Fig. 4a,c), EMProt achieved the highest average value of 0.90, compared to 0.69 for DeepMainmast, 0.82 for PHENIX and 0.84 for AF3. Among the 177 maps, EMProt obtained a higher TM-score than DeepMainmast for 132 cases, PHENIX for 117 cases and AF3 for 102 cases. Interestingly, AF3 performed well on this test set and even better than PHENIX. Such a phenomenon can be understood because this test set was built before AF3 was released and some items were included in the training set of AF3.

Fig. 4: Comparison of EMProt, DeepMainmast, PHENIX and AF3 in building the complete structure on the test set of n = 177 maps.

a,b, Bar plots of TM-score (a) and r.m.s.d. (b). The bars indicate the means and error bars indicate the 95% confidence interval. c,d, Head-to-head comparison of TM-score (c) and r.m.s.d. (d). e, Three examples of high-quality models built by EMProt for three targets.

Source data

Full size image

For r.m.s.d. (Fig. 4b,d), EMProt also showed an improvement over the other approaches. The average r.m.s.d. was 1.42 Å for EMProt, compared to 3.57 Å for DeepMainmast, 1.91 Å for PHENIX and 3.42 Å for AF3. Specifically, EMProt obtained a lower r.m.s.d. than DeepMainmast for 140 cases, PHENIX for 101 cases and AF3 for 142 cases. Because of a lack of experimental map information, AF3 failed to build the models in atomic precision; conversely, PHENIX outperformed AF3 in r.m.s.d. The r.m.s.d. and TM-score calculation was based on matching query residues with target PDB residues globally and only the matched residue contributed to the r.m.s.d. EMProt achieved the highest TM-score while maintaining the lowest r.m.s.d., suggesting that it could obtain the correct fold at atomic-level accuracy.

Figure 4e show three EMProt high-quality models for targets EMD-16183 (PDB 8BQE), EMD-26885 (PDB 7UZE) and EMD-26801 (PDB 7UUR). All three models had a high TM-score of >0.99 and a low r.m.s.d. of <1.0 Å. Overall, there were 128 EMProt models with a TM-score > 0.90 and 37 cases with a TM-score > 0.99 in the test set.

Examples in building the complete structure

Several examples are displayed in Fig. 5 to illustrate how the models built by EMProt, DeepMainmast, PHENIX and AF3 are globally similar to the PDB structure. The built models are colored by chain names. Figure 5a shows the example for target EMD-15690 (PDB 8AW3). For chain 2, colored in green, all map-to-model methods could build the structure correctly. The major challenge for this target was building chain 3. In this regard, EMProt successfully built both chains and, thus, achieved a high TM-score of 0.888 for the built model, whereas DeepMainmast and PHENIX failed to build the correct structure for chain 3 on this target and AF3 could not find the correct global fold, resulting in a lower TM-score.

Fig. 5: Comparison of the models built by EMProt, DeepMainmast, PHENIX and AF3 in terms of building the complete structures.

a, Target EMD-15690 (PDB 8AW3) at 3.6-Å resolution. b, Target EMD-27875 (PDB 8E40) at 3.6-Å resolution. In a,b, nucleic acids in the PDB model and the fragments built into the nucleic acid regions are not shown. c,d, Comparison of EMProt and AF3 on targets EMD-15686 (PDB 8AVX) at 3.5-Å resolution (c) and EMD-27022 (PDB 8CVZ) at 3.5-Å resolution (d).

Full size image

Figure 5b shows another example of the models built by different methods for target EMD-27875 (PDB 8E40). This was a small target with only three chains and a total of 636 residues. Although this target had near-atomic resolution of 3.57 Å, its density map had severe heterogeneity. As such, ModelAngelo only yielded a low sequence recall of 29.56% for this target. The major challenge when using EMProt, DeepMainmast and PHENIX for this target was to recover chain C in the model. Although EMProt achieved the best TM-score of 0.930, one of the modeled chains (colored in red) still had a small deviation from the PDB model. DeepMainmast and PHENIX both failed to recover a complete model and performed moderately on this target. Although AF3 predicted individual chains (A, B and C) well with high TM-scores of 0.924, 0.953 and 0.893, respectively, the entire three-chain complex was wrongly assembled with a low TM-score of 0.580, indicating the necessity of incorporating the experimental map for accurate structure assembly.

We further compared EMProt and AF3 on two additional targets. Figure 5c shows the result for target EMD-15686 (PDB 8AVX), a homologous dimer complex with a total of 961 residues. On this target, EMProt achieved a TM-score of 0.971 and an r.m.s.d. of 1.94 Å, much better than the 0.471 and 5.29 Å for AF3. With the inclusion of experimental density map information, the complex was readily solved by EMProt with relatively high quality. However, the contacting interface was wrongly predicted by AF3, resulting in a globally inverted conformation for chain A, although the monomer TM-score was still good (0.813). Figure 5d shows the example of another target, EMD-27022 (PDB 8CVZ), including ten chains and 3,070 residues. The TM-score and r.m.s.d. were 0.946 and 1.13 Å for EMProt, much better than the 0.665 and 7.66 Å for AF3. Further examination of the AF3 structure revealed that, although the monomer prediction showed high accuracy, there were still challenges in predicting protein multimer structures. These results suggest that, without considering the experimental map information, methods such as AF3 still face problems in modeling the correct structural assembly, especially for large complexes.

Model-to-map fit validation

In addition to recovering the structure from a map, another critical criterion for evaluating the accuracy of a built model is the consistency between the model and the map. The model-to-map fit assessment examines how well the atomic model explains the experimental data and represents another gold-standard assessment for structure modeling from cryo-EM maps. Here, four metrics, CCmask, CCbox, CCpeaks and CCvolume, which are calculated using the PHENIX tool phenix.mapmodelcc36, were used to evaluate the consistency of the built model with the map.

For a comprehensive comparison, we divided the methods into two groups (de novo modeling and integrated modeling) during the evaluations. For de novo modeling, both the refined and the unrefined models were evaluated (Supplementary Table 3). Overall, EMProt performed slightly better than ModelAngelo in de novo modeling (Table 1). Specifically, for the unrefined models, although EMProt yielded slightly lower CCmask and CCvolume values than ModelAngelo (0.700 and 0.704 versus 0.706 and 0.706). EMProt was better than ModelAngelo in terms of CCbox and CCpeaks (0.632 and 0.556 versus 0.627 and 0.550). After the model refinement, EMProt performed better than ModelAngelo in all four metrics and obtained CCmask, CCbox, CCpeaks and CCvolume values of 0.732, 0.656, 0.589 and 0.732, respectively, compared to 0.730, 0.647, 0.578 and 0.727 for ModelAngelo. Overall, the performance gaps between both methods for both model geometry and model-to-map fit were marginal, indicating their comparable performance.

Table 1 Model-to-map fit validation for the built models by EMProt, ModelAnglelo, DeepMainmast and PHENIX on the test set of 177 cryo-EM maps Full size table

When comparing the full pipelines of EMProt and other methods, EMProt was consistently better than the other methods in all four metrics (Supplementary Table 3). Specifically, EMProt achieved average CCmask, CCbox, CCpeaks and CCvolume values of 0.677, 0.647, 0.578 and 0.692, respectively, compared to 0.597, 0.565, 0.462 and 0.613 for DeepMainmast and 0.638, 0.592, 0.493 and 0.652 for PHENIX (Table 1). It can be noted that both EMProt and DeepMainmast adopt an integrated modeling strategy; however, the model-to-map fit of EMProt was much better than that of DeepMainmast, indicating a more effective structure information integration of EMProt than DeepMainmast. In addition, the much better performance of EMProt than PHENIX also indicates that fitting predicted models into cryo-EM maps is not trivial and simply docking and rebuilding predicted models into densities will not solve the modeling problem. As a reference, Table 1 also lists the results of original AF3-predicted models. The worst correlations between the AF3 models and the density maps again suggest the challenge and necessity of integrated model building for cryo-EM maps.

Supplementary Fig. 3 shows an example of the EMProt-built model at atomic detail for the high-resolution target PDB 8DQV (1.52 Å). EMProt achieved a sequence recall of 99.70%, a high TM-score of 0.997 and an excellent r.m.s.d. of 0.28 Å on this target. Correspondingly, the EMProt model had a high CCmask of 0.749, CCbox of 0.689, CCpeaks of 0.685 and CCvolume of 0.769 in model-to-map correlation. In addition to the overall atomic details shown in Supplementary Fig. 3, we also examined those residues around the ligands. As shown in Supplementary Fig. 3a,c, the presence of the ligands did not hinder accurate modeling of either backbones or side chains and the EMProt-built model overlapped well with the corresponding PDB structure. We further checked whether the atoms fit the density well. As shown in Supplementary Fig. 3b, the presented fragment fit well in the density, demonstrating the ability of EMProt in recovering atomic details from cryo-EM maps.

Structure validation of the build models

We next conducted all-atom structure validation of the built models. Structure validation ensures that the built model has reasonable geometry characteristics and stereochemistry, such as bond lengths, bond angles, torsion angles and atom clash. Specifically, we used seven metrics to validate the build models, including Clashscore, MolProbity score, Ramachandran favored and outlier, rotamer outlier, r.m.s. (bonds) and r.m.s. (angles), which were calculated using the PHENIX tool phenix.molprobity6,37. Along with the model-to-map fit metrics, these complementary evaluations provide a comprehensive measure of the model reliability and how close the models are ready to be deposited to the PDB.

Similarly, we first compared the performances between EMProt and ModelAngelo in de novo modeling (Supplementary Table 4). Overall, EMProt exhibited comparable performance to ModelAngelo (Table 2). Specifically, for the unrefined models, although ModelAngelo was better than EMProt in terms of Clashscore, MolProbity score and rotamer outlier (66.51, 3.44 and 8.99% versus 69.36, 3.52 and 11.87%), EMProt was better than ModelAngelo in terms of Ramachandran favored, Ramachandran outlier, r.m.s. (bonds) and r.m.s. (angles) (92.90%, 2.12%, 0.0270 and 2.30 versus 91.80%, 2.78%, 0.0632 and 5.39). It is, thus, speculated that EMProt is better at modeling protein backbones, as measured by Ramachandran favored and outlier, whereas ModelAngelo is better at modeling side chains, as measured by rotamer outlier. Moreover, EMProt is better at modeling standard amino acid conformations, as measured by r.m.s. (bonds) and r.m.s. (angles), whereas ModelAngelo is better at modeling overall atom compatibility, as measured by Clashscore and MolProbity score. For the refined models, the performance differences between EMProt and ModelAngelo became smaller and more comparable (Table 2).

Table 2 Structure validation for the built models by EMProt, ModelAnglelo, DeepMainmast and PHENIX on the test set of 177 cryo-EM maps Full size table

We also evaluated the full pipeline of EMProt, DeepMainmast and PHENIX using the same metrics (Supplementary Table 4). As shown in Table 2, EMProt achieved an overall better performance than the other methods. Specifically, EMProt achieved a better performance than the other methods in six of seven metrics, yielding values of 13.44, 2.01, 95.92%, 0.26%, 0.0063 and 0.80 for Clashscore, MolProbity score, Ramachandran favored, Ramachandran outlier, r.m.s. (bonds) and r.m.s. (angles), respectively (Table 2). EMProt only performed worse than DeepMainmast in rotamer outlier but still gave a good value of 1.97%, compared to 1.70% for DeepMainmast and 2.10% for PHENIX. These results suggest the high accuracy of the models built by the full EMProt pipeline in terms of structure quality.

Quality assessment by local confidence scores

To assess the quality of the built model in a reference-structure-free manner, we propose a local confidence score for each built residue. Specifically, we chose Cα as the representative atom of a residue and defined the local confidence score as the probability of the nearest voxel in the Cα probability map. The local confidence score solely reflects the accuracy of Cα atom placement, without considering the amino acid identity. We further defined the displacement of a built residue as the distance between its Cα atom and the nearest Cα atom in the reference PDB structure. The confidence scores were grouped into bins of 0.05. As shown in Fig. 6a, the local confidence exhibited a good correlation with local displacement, suggesting that the local confidence score could effectively estimate the modeling accuracy of the built model at the residue level.

Fig. 6: Quality assessment of the built models by local confidence scores.

a, The residue displacements versus their local confidence scores for n = 371,706 Cα atoms, representing the corresponding residues on the test set of n = 177 maps. Confidences are merged into bins of 0.05. The dots indicate the medians and the error bars indicate the first and third quartiles. b, An example of local confidence score as a displacement measurement, where the PDB model is colored in green and the EMProt model is colored in a red–white–blue gradient. The black dashed outlines indicate those modeled regions with low confidence scores. c, Bar plots of residue displacements for different global resolution bins (n = 15, 47, 69 and 46 maps for <2.5 Å, 2.5–3.0 Å, 3.0–3.5 Å and 3.5–4.0 Å). d, Bar plots of residue displacements for different secondary-structure category for n = 177 maps. In c,d, the bars indicate the means and error bars indicate the 95% confidence interval.

Source data

Full size image

One representative example, EMD-33807 (PDB 7YG4), is shown in Fig. 6b. The PDB structure (colored in green) is superposed onto the EMProt-built model. In the top panel, the built model is colored according to residue displacement, ranging from blue for 0 Å to red for 2 Å. The bottom panel displays the same model but the model is colored by local confidence score, from blue for 0.90 to red for 0.10. As shown in the figure, most residues are in blue in the built model, indicating relatively high modeling confidence and small displacement. However, some residues, particularly in coil and loop regions, exhibit larger deviations from the PDB structure, as indicated by dashed rectangles. Therefore, the proposed local confidence score is consistent with the residue displacement and may be used to effectively estimate the modeling accuracy in a reference-structure-free manner. However, a low confidence score does not necessarily indicate incorrect or highly deviated models and may require a final round of expert inspection or manual pruning.

We further investigate the relationship between residue displacements and global resolutions in the EMProt-built models. Global resolutions are grouped into four bins: <2.5 Å, 2.5–3.0 Å, 3.0–3.5 Å and 3.5–4.0 Å. The bar plots of the results are shown in Fig. 6c. It can be seen from the figure that EMProt achieved a good accuracy of <1.0 Å for all ranges of resolutions. Specifically, the average displacements for different resolution ranges were 0.20 Å, 0.42 Å, 0.59 Å and 0.79 Å, respectively. The displacements increased as the resolution worsened, as expected.

We also analyzed the residue displacements in the built models for different secondary-structure categories. The built residues were split into ‘helix’, ‘sheet’ and ‘coil’ using secondary-structure algorithm DSSP38. The bar plots of the results are shown in Fig. 6d. Again, EMProt achieved a good accuracy of <1.0 Å for different structure categories. The average displacement for helix (0.53 Å) was worse than for sheet (0.49 Å) but both were better than for coil (0.68 Å). This phenomenon can be understood because helices and sheets exhibit more specific and stable spatial arrangements, whereas coils are more flexible. Additionally, coils are more prone to erroneous modeling because of their lower resolution in experimental maps compared to helices and sheets. Given the observation, it is recommended that users conduct more manual inspection and modeling on coils more than helices and sheets.

Running time

The full pipeline of EMProt includes two stages: de novo model building and integrated modeling with AF2 structures. We examined the running times of EMProt at different stages on the test set of 177 cryo-EM maps (Supplementary Table 5). Here, deep learning was performed on one Nvidia A100 GPU (40 GB) and the other computations were run on a single core of an Intel(R) Xeon(R) Gold 6240 CPU at 2.60 GHz. It was revealed that, for the de novo modeling protocol, the running times of EMProt showed a linear relationship with the number of residues in the target (Supplementary Fig. 4a). For both the integrated modeling mode and the full pipeline, the running times showed a similar trend (Supplementary Fig. 4b). Specifically, for the targets with fewer than 4,000 residues, the running times increased linearly with the size of the target to be built, whereas the running times varied greatly for the larger targets, with running times ranging from a few hundred seconds to 25,000 s (Supplementary Fig. 4). On average, EMProt could typically finish a job within 5 min for building the de novo model and 30 min for running the full pipeline (Supplementary Table 5).

Discussion

We propose an integrative modeling framework for accurate modeling of the complete protein structure from cryo-EM density maps, referred to as EMProt. EMProt accurately builds the complete structure from cryo-EM maps through the integration of multiple well-designed submodules, including 3D recognition networks for main-chain atom prediction and residue type prediction, three-track attention-based structure refinement networks for all-atom construction, a fragment-threading module for initial model building and a fragment-fitting module for resolving missing residues that are hard to interpret from raw density maps.

EMProt was extensively evaluated on a benchmark set of 177 experimental cryo-EM maps and compared to state-of-the-art methods, including DeepMainmast, ModelAngelo, PHENIX and AF3. Overall, EMProt achieved the highest sequence recall in recovering the protein structure when compared to DeepMainmast, ModelAngelo and PHENIX and yielded the best TM-score and r.m.s.d. in building the complete structure when compared to DeepMainmast, PHENIX and AF3. A key factor contributing to EMProt’s success lies in its effective integration of density-based de novo modeling with structure prediction. As the inclusion of predicted models in cryo-EM structure determination becomes increasingly mainstream, EMProt offers a practical and efficient solution for protein structure modeling from cryo-EM maps.

According to a comprehensive evaluation and comparison to related methods, EMProt exhibits improved performance over state-of-the-art methods. On one hand, EMProt can de novo model protein structures from cryo-EM density maps, achieving a performance comparable to the leading de novo method, ModelAngelo. On the other hand, by incorporating AF2 structures, EMProt greatly improves the completeness of the built models while maintaining overall good model geometry and model–map agreement. Nevertheless, the resulting models are not sufficiently perfect to omit subsequent manual processing steps. Therefore, it is recommended that users refine the built models with a third-party program such as phenix.realspacerefine before further analysis. In summary, EMProt marks a major step forward in cryo-EM structure modeling. Future improvements could further enhance model quality, particularly for difficult-to-resolve regions, paving the way for even more reliable and high-throughput protein structure determination.

Methods Data collection

The test set used in this study consisted of 177 nonredundant cryo-EM maps at <4-Å resolution sourced from the ModelAngelo paper. The test set was constructed by excluding the targets that (1) had >30,000 residues; (2) were viruses with I-group symmetry; and (3) contained insertion codes or other irregularities.

To construct a training set for the neural networks, we first queried the EMDB for a complete list of maps with resolution < 4 Å and their corresponding PDB structures. To remove redundancy between the test and training sets, we adopted MMseqs2 (ref. 39) to search for redundant chains in the PDB structures, using a sequence identity cutoff of 30%. If any chain in a candidate PDB structure was similar to a chain in the test set, the associated map and PDB structure were removed. The remaining maps and PDB structures were further filtered on the basis of three criteria. Firstly, the maps with nonorthogonal axes were excluded. Secondly, the maps reconstructed by methods other than single-particle analysis were discarded. Thirdly, the PDB structures with unassigned amino acid types or containing only backbone atoms were removed. Next, we calculated the cross-correlations between the experimental maps and the structure-simulated maps using Chimera40. The map–structure pairs with cross-correlations below a threshold of 0.60, indicating severe misfits, were filtered out. The remaining maps were then manually inspected for quality control. Subsequently, we applied MMseqs2 to cluster the chains with a sequence identity cutoff of 30%. To further reduce the redundancy, we developed an in-house greedy algorithm to cluster the corresponding PDB items on the basis of MMseqs2 clustering results. Specifically, if two chains from different PDB items were identified as similar, only one of the PDB items was retained. Finally, a total of 802 maps and their corresponding PDB structures were retained as the training set (Supplementary Table 6).

Main-chain atom and residue type prediction from density map The SCUNet network architecture

We used an SCUNet41 architecture to extract atom probabilities and amino acid identities from density maps. The network consists of three encoder blocks, one transition block and three decoder blocks, which were based on Swin-Conv (SC)41,42 modules, with skip connections linking the encoders and decoders. Each Swin-Conv block contains a Swin transformer (SwinT) block parallelized with a residual convolutional (RConv) block, which is sandwiched between two 1 × 1 convolutions. SwinT uses a window size of 3. For downsampling, the network applies a 3D convolution layer with a kernel size and stride of 2. For upsampling, it uses a 3D transposed convolution layer with the same kernel size and stride. The input to the network consists of density chunks of size 48 × 48 × 48 with a grid interval of 1.0 Å. The output dimensions are the same as the input. Both atom probability and amino acid identity predictions use the same network architecture, differing only in the number of output channels in the final layer.

Network training

To train the main-chain atom probability predictor, the N, Cα and C atoms were extracted from the structure and the training labels were generated on the basis of atom positions and map resolution. For each Cα atom, the density value at grid point x was calculated as follows43,44:

$$D({\bf{x}})=\max \left{{e}^{-k\parallel {\bf{x}}-{\bf{a}}{\parallel }_{2}^{2}},\forall {\bf{a}}\in A\right}$$ (1)

where a denotes the coordinate of the Cα atom in structure A and the kernel k is defined according to the map resolution R as follows45:

$$k={\left(\frac{\pi }{1.2+0.8R}\right)}^{2}$$

It should be noted that a local resolution might be used to get a more precise representation of local density quality. However, accurate local resolution requires the calculation between two half-maps, which may not be achievable because the authors do not always deposit the half-maps into the EMDB. Thus, we used a global resolution in this study for simplicity.

The N and C atom probability maps were generated in the same manner and concatenated with the Cα probability maps. The loss for atom probability prediction was a combination of smooth L1 loss and structural similarity index measure (SSIM)46,47 loss, as follows:

$${\mathcal{L}}=\,\text{smooth}\,{L_1}\,(\;{y}^{{\rm{pred}}},{y}^{{\rm{GT}}})+\,\text{SSIM}\,(\;{y}^{{\rm{pred}}},{y}^{{\rm{GT}}})$$ (2)

To train the amino acid type segmentation network, each voxel was labeled with the nearest amino acid type [0, 20) within 3 Å. Voxels that did not correspond to any amino acid within this distance were labeled as background. The loss function was a 20-class cross-entropy loss, with label-20 predictions ignored, as only predictions near the backbone were of interest:

$${\mathcal{L}}=\,\text{cross-entropy}\,(\;{y}^{{\rm{pred}}},{y}^{{\rm{GT}}})$$ (3)

To process the training data, the grid size of the map was standardized to 1.0 Å using trilinear interpolation. The density values of each experimental map were clipped to the range between ≥0.0% and the ≤99.999% of the density values and then normalized to [0, 1]. The experimental maps and labeled maps were partitioned into overlapping chunks of size 60 × 60 × 60 with a stride of 30. For computational efficiency, those experimental chunks with a maximum density value less than 0.0 were excluded from training.

During training, data augmentation was applied to prevent overfitting. The input box was randomly cropped to 48 × 48 × 48 from the 60 × 60 × 60 chunk and a random 90∘ rotation was applied. The network was implemented using PyTorch 2.1. The Adam optimizer was used to update the network weights. We trained the atom probability predictor and the amino acid type predictor separately using the same hyperparameters. The initial learning rate was set to 5 × 10−4, with an L2 regularization coefficient (weight_decay) of 1 × 10−4. During training, the learning rate was halved if the average training loss did not decrease for four consecutive epochs, with a minimum learning rate capped at 5 × 10−6. Each network was trained with a batch size of 96 on two Nvidia A100 GPUs (40 GB). The models with the lowest validation loss were selected as the final models.

All-atom construction in density map Three-track attention network architecture

After getting the Cα atom probability, the Cα atoms were derived from the probability map by detecting local maxima using a mean-shift algorithm. We then used a three-track attention network to predict backbone frames (R, t) and torsion angles τ for each Cα atom. The network consisted of 16 main blocks, iteratively processing 1D, 2D and 3D information.

The densities around each Cα (a cubic with shape of 233 Å3) were extracted as initial 1D features. The densities between pairwise Cα atoms (a cuboid with shape of 17 × 3 × 3 Å3) were extracted as initial 2D features. A schematic illustration for the feature extraction process is shown in Supplementary Fig. 2. For information flowing from 1D, 2D and 3D to 1D, we used a self-attention and invariant point attention module to update 1D features, with attention biases from 2D features and 3D frames. For information flowing from 1D to 2D, we used the outer product to update the 2D features. For information flowing from 2D and 3D to 2D, we used axial attention to update the 2D features, with attention biases from 3D frames. The backbone update module predicted the new backbone frames for each Cα atom, giving the updated 1D features. After going through 16 blocks, a final layer was used to predict torsion angles. Having the backbone frames and torsion angles, all-atom positions could be determined by transforming ideal residue atom groups19.

Network training

The neural network was trained using a customized noising–denoising schedule. For each residue, its Cα position tr was randomly noised according to the following formula:

$${{\bf{t}}}^{r}={{\bf{x}}}{{\rm{C}}{\upalpha}}^{r}={{\bf{x}}}{{\rm{C}}{\upalpha}}+{\mathcal{N}}\left(0,\frac{1}{\sqrt{3}}\right)$$ (4)

where ({\mathcal{N}}(0,\frac{1}{\sqrt{3}})) is a random Gaussian noise applied to the Cα position. The rotation matrix Rr of the backbone frame is noised using the following formulas:

$${{\bf{x}}}{\rm{N}}^{r}={{\bf{x}}}{{\rm{C}}{{\upalpha}} }^{r}+{\rm{rand}}_{\rm{like}}({{\bf{x}}}{{\rm{C}}{{\upalpha}} }^{r})$$ (5) $${{\bf{x}}}{\rm{C}}^{r}={{\bf{x}}}{{\rm{C}}{\upalpha}}^{r}+{\rm{rand}}_{\rm{like}}({{\bf{x}}}{{\rm{C}}{\upalpha}}^{r})$$ (6) $${R}^{r}={\rm{rot}}_{\rm{from}}\,\text{_}3\text{_}\,{\rm{points}}({{\bf{x}}}{\rm{N}}^{r},{{\bf{x}}}{{\rm{C}}{{\upalpha}} }^{r},{{\bf{x}}}_{\rm{C}}^{r})$$ (7)

The training procedure consisted of the following steps: (1) selecting a PDB structure; (2) randomly choosing a residue; (3) extracting its k neighboring residues (including itself); (4) applying noise to each residue; and (5) using the neural network to predict the true positions. For training, we used an effective batch size of 8 and set the number of neighbors to k = 128. The neural network parameters were updated using AdamW optimizer. The initial learning rate was set to 1 × 10−4. We stopped the gradient of rotation between consecutive blocks to stabilize the training.

The loss function consisted of three learning objectives. The first loss was the r.m.s.d. loss, which was defined on the Cα atoms, backbone N, Cα and C atoms and all atoms. The Cα and backbone r.m.s.d. losses were calculated in every intermediate layer of the network, with the lth layer having a weight of γ16−l (γ was set to 0.80 in this study). The Cα r.m.s.d. loss was defined as

$${{\mathcal{L}}}{{\rm{C}}{{\upalpha}} }=\sqrt{\frac{| | {{\bf{x}}}{{\rm{C}}{{\upalpha}} }^{{\rm{pred}}}-{{\bf{x}}}{{\rm{C}}{{\upalpha}} }^{{\rm{true}}}| {| }^{2}}{{{N}}{{\rm{C}}{{\upalpha}} }}}$$ (8)

and the backbone and all-atom losses were similarly defined. The total r.m.s.d. loss was defined as

$${{\mathcal{L}}}{{\rm{r.m.s.d.}}}={{\mathcal{L}}}{{\rm{C}}{{\upalpha}} }+{{\mathcal{L}}}{{\rm{backbone}}}+{{\mathcal{L}}}{{\rm{all}}\text{-}{\rm{atom}}}$$ (9)

The second loss was the torsion angle loss, which was defined as

$${{\mathcal{L}}}{{\rm{torsion}}}=\min\left (| {\tau }^{{\rm{pred}}}-{\tau }^{{\rm{true}}}| ,| {\tau }^{{\rm{pred}}}-{\tau }{{\rm{alt}}}^{{\rm{true}}}|\right )+0.02\times | | | {\tau }^{{\rm{pred}}}| {| }_{2}-1|$$ (10)

The third loss was the structure violation loss ({{\mathcal{L}}}_{{\rm{violation}}}). The structure violation loss penalized the structure irregularity (for example, atom clashes and irregular bond lengths or angles). The total loss for network training was the sum of all three losses:

$${\mathcal{L}}={{\mathcal{L}}}{{\rm{r.m.s.d.}}}+{{\mathcal{L}}}{{\rm{torsion}}}+{{\mathcal{L}}}_{{\rm{violation}}}$$ (11) Backbone tracing

As the predicted all-atom residues are unordered, it is needed to connect residues to chains. A peptide bond ({d}{{\rm{C}}{i}-{\rm{N}}{i+1}}) is defined as the distance between the C atom in the last residue and N atom in the next residue. Given the fact that ({d}{{\rm{C}}{i}-{\rm{N}}{i+1}}) only varies in small ranges, if we start from one residue, we can iteratively find its nearest next residue under the C − N distance restraint (d < 2.1) and the residues are threaded to fragments.

Sequence assignment

For each residue in a fragment, we located the nearest voxel to its Cα atom in the predicted amino acid type map and assigned the predicted logits to the residue. For each fragment, the amino acid logits were used to construct an HMM profile. A sequence search was then performed against the given sequences using HMMER. After the HMM search, residues were assigned with the corresponding amino acid types according to the alignment results.

Integration of predicted structures Structure prediction

For each chain, the full-length sequence in the PDB SEQRES record was used as the input to AF2 (version 2.3.0). The max-template-date was set to January 5, 2022, a time point before all the test models were released. We truncated the predicted model to have the same start and end residue using a simple sequence alignment. It should be noted that 22 chains in eight PDB structures failed in the AF2 prediction because their sequences were too long; thus, we used AF3 to predict their structures instead.

Chain partition

The predicted protein models, despite high accuracy, may have large deviation in interfragment level. To find the best match with the map, splitting of chains into subcomponents is necessary. Following UniDoc48, chains were split into domain level and fragment level on the basis of structure contacts. The Cα distances were converted to contact map pm,n, which was then used to calculate the interdomain interaction score Sinter(Di, Dk) and intradomain interaction score Sintra(Di). The definition of these scores can be found in the UniDoc study. In a top-down splitting stage, the structure was first cut into fragments to minimize Sinter iteratively. In a bottom-up merging stage, the fragments were merged to maximize the score S(i, k) iteratively.

$${{S}}(i,k)={{{S}}}{{\rm{inter}}}({D}{i},{D}{k})-\min \left{{{{S}}}{{\rm{intra}}}({D}{i}),{{{S}}}{{\rm{intra}}}({D}_{k})\right}$$ (12)

With such a splitting–merging strategy, a chain can be split to fragments (before merging) and domains (after merging). A chain contains ≥1 domain and a domain contains ≥1 fragment.

Matching template by structure alignment

To combine the information from the template structure (AF2 models in this study) and the density map, one way is to align the template to the de novo chains directly built from the map. The chains built from the map provide highly accurate backbone information but the predicted AF models may have large movement between domains49. Therefore, the first step is to split each template structure into substructure and perform the matching at substructure level. We split the template into fragments and the matching was performed on fragments.

Each substructure was aligned to the built chains using TMalign. If the r.m.s.d. between the aligned residues was <1.5 Å, the substructure was considered as being matched. The coordinates of the matched substructure were then transformed using the output transformation matrix from TMalign. The matched substructures also helped to correct the wrong sequence alignment and fix missing residues on the built chains. If the aligned residues between the substructure and the chain had a sequence identity of >80% (highly matched in sequence), then the sequence of the matched residues was updated to the corresponding sequence in the substructure and the missing residues were fixed using AF2 residues.

Matching template by structure fitting

In addition to aligning the template structure to the built chains, EMProt also directly fits the template structure to the density map. Especially when local resolution is poorer, de novo modeling often fails to give a chain model but template fitting fills the coverage gap. Rather than fitting to the raw map, we fitted the template to the predicted main-chain map as it is less noised.

EMProt uses an exhaustive search in rotational and translation space to maximize the matching score between structure and main-chain map:

$$S(i,j,k)=-\theta \mathop{\sum }\limits{l=1}^{M}\mathop{\sum }\limits{m=1}^{M}\mathop{\sum }\limits{n=1}^{M}\left({A}{l,m,n}\times {B}_{l+i,m+j,n+k}\right)$$ (13)

where grid A was generated by mapping the N, Cα and C atoms of the structure into a 3D grid and grid B denotes the main-chain map. Grid A was sampled with a grid interval of 3.0 Å and grid B was downsampled from the main-chain map to have the same grid interval. An angle interval of 18∘ was used to evenly discretize the Euler space, resulting in a total of 2,540 evenly distributed Euler angles. The searching process was accelerated by FFT and multithread programming. After clustering the sampled poses with an r.m.s.d. cutoff of 5 Å, each structure was placed at M different positions.

When fitting structure to the map, a semiflexible refinement was performed to improve local substructure conformation. We used a local SIMPLEX optimization algorithm to find the best match of the substructures. For a template structure, starting from a selected substructure, the positions of all substructures were optimized one after the other through a breadth-first search on the graph defined by substructure adjacency. Taking each substructure as a seed, the refinement output n different structure models. The results of substructure refinement for one structure consisted of M × n refined models plus M rigidly fitted models.

We adopted two strategies for fitting structures to the map. First, we fitted the structure of the chains, where the substructures are the domains. Second, we fitted the structure of the domain, where the substructures are the fragments. For both strategies, all chains and domains were split into fragments for finial assembly.

Structure assembling from fragments

After the matching step by structure alignment and fitting, the fitted AF2 fragments and the fixed built chains formed a structure library. These fragments were gathered for a final stage of assembly. For each fragment, a match score was defined as

$$s(i)=\mathop{\sum }\limits{k=1}^{{N}{{\rm{C}}{{\upalpha}} }^{i}}{D}^{{\prime} }\left({{\bf{x}}}_{{\rm{C}}{{\upalpha}} }^{i,k}\right)$$ (14)

where ({D}^{{\prime} }({\bf{x}})) is the predicted atom probability map (density values normalized between 0.0% and 99.999%). For fragments i and k, if the distance between a Cα atom in fragment i and another Cα atom in fragment k was <2.0 Å, the two atoms were considered to be clashing and the clash percentage of fragment i was defined as

$${p}{{\rm{clash}}}(i;k)=\frac{{{N}}{{\rm{clashed}}_{\rm{C}}{{\upalpha}} (i;k)}^{i}}{{{N}}_{{\rm{C}}\upalpha }^{i}}$$ (15)

where ({N}{{\rm{clashed}}_{\rm{C}}{{\upalpha}} (i;k)}^{i}) is the number of clashed Cα atoms and ({N}{{\rm{C}}{\upalpha}}^{i}) is the total Cα atoms in fragment i and pclash(k; i) is defined in the same way. Two fragments i and k were considered to be clashing if pclash(i; k) > pcutoff or pclash(k; i) > pcutoff. The problem of assembling the fragments was then converted to a constrained optimization problem, formulated as maximizing

$$s=\mathop{\sum }\limits{i=1}^{\rm{N}{{\rm{fragment}}}}s(i)x(i)$$ (16)

subject to

$$x(i)\in {0,1},\,\,i=1,2,\ldots ,{N}{{\rm{fragment}}}$$ (17) $$\begin{array}{l}x(i)+x(k) \le 1,\,\,{\rm{if}}\,{p}{{\rm{clash}}}(i;k) > {p}{{\rm{cutoff}}}\,{\rm{or}}\,{p}{{\rm{clash}}}(k;i) > {p}{{\rm{cutoff}}},\,\,\i,k=1,2,\ldots ,{N}{{\rm{fragment}}}\end{array}$$ (18)

We set pcutoff to 0.10 in this study, allowing a mild clash between fragments. Generally, if two fragments were clashing, we expected to keep the larger one. The assembled fragments were input to the all-atom refinement network again for a final round of all-atom refinement.

Postrefinement step

After the structure is assembled, users can refine the built model to improve local structure and model-to-map fit through a third-party program, which is highly recommended in real applications. For a fair evaluation and comparison, in this study, the models built by all methods, including EMProt, ModelAngelo, DeepMainmast and PHENIX, were refined using phenix.realspacerefine with default parameters, unless otherwise specified.

Evaluation of the built models

Two types of evaluations were used to assess the accuracy of a built model using the PDB structure as the reference. One involved recovering the protein structure to measure how much of the PDB structure was recovered by the built model regardless of the order and connectivity of built fragments. The other involved building the complete structure to measure how similar the built model was to the PDB structure in a global way.

To recover the protein structure, three metrics (residue coverage, sequence match and sequence recall) were used to measure the accuracy of the built protein model. The metrics were calculated using a distance cutoff of 3 Å and Cα as the representative atom for each residue using phenix.chain_comparion. Here, the residue coverage measured the fraction of residues in the built model that were within a specified distance cutoff from the map, irrespective of residue type, relative to the PDB structure. Sequence match indicated the fraction of the residues in the built model that had the same residue type as those in the PDB structure. The sequence recall was defined as the product of residue coverage and sequence match, reflecting the proportion of residues in the built model that were correctly placed in terms of both backbone position and residue type when compared to the PDB structure.

To build the complete structure, two metrics, TM-score and r.m.s.d., were used to measure the global similarity between the built model and the PDB structure. The two metrics were calculated using the USalign program.

In addition, we used four metrics (CCmask, CCbox, CCpeaks and CCvolume) calculated by phenix.mapmodelcc to evaluate model–map consistency. It should be noted that CC values vary with a model’s B factors. As the B factors presented by different methods were in different numerical ranges, we reset all model B factors to 0.0 for simplicity. For structure validation of the build models, we used seven metrics (Clashscore, MolProbity score, Ramachandran favored and outlier, rotamer outlier, r.m.s. (bonds) and r.m.s. (angles)) calculated by phenix.molprobity.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All published datasets used in this paper were taken from the EMDB and PDB (accession codes specified in the figure captions and Supplementary Tables). All raw data of the evaluation results are provided in the article and Supplementary Information. The list of items in the test set is available in Supplementary Table 1. The list of training items is available in Supplementary Table 6. Source data are provided with this paper.

Code availability

The EMProt package is freely available online for academic or noncommercial users (http://huanglab.phys.hust.edu.cn/EMProt or https://github.com/huang-laboratory/EMProt). A link to a demo that runs on Google Colab and requires no installation on the local device is available on the GitHub repository.

References

Nogales, E. The development of cryo-EM into a mainstream structural biology technique. Nat. Methods. 13, 24–27 (2016).

Article   CAS   PubMed   PubMed Central   Google Scholar

Frank, J. Advances in the field of single-particle cryo-electron microscopy over the last decade. Nat. Protoc. 12, 209–212 (2017).

Article   CAS   PubMed   PubMed Central   Google Scholar

Cheng, Y. Single-particle cryo-EM—how did it get here and where will it go. Science 361, 876–880 (2018).

Article   CAS   PubMed   PubMed Central   Google Scholar

Lawson, C. L. et al. EMDataBank unified data resource for 3DEM. Nucleic Acids Res. 44, D396–D403 (2016).

Article   CAS   PubMed   Google Scholar

Zivanov, J. et al. New tools for automated high-resolution cryo-EM structure determination in RELION-3. eLife 7, e42166 (2018).

Article   PubMed   PubMed Central   Google Scholar

Liebschner, D. et al. Macromolecular structure determination using X-rays, neutrons and electrons: recent developments in PHENIX. Acta Crystallogr. D Struct. Biol. 75, 861–877 (2019).

Article   CAS   PubMed   PubMed Central   Google Scholar

De la Rosa-Trevín, J. M. et al. Scipion: a software framework toward integration, reproducibility and validation in 3D electron microscopy. J. Struct. Biol. 195, 93–99 (2016).

Article   PubMed   Google Scholar

Punjani, A., Rubinstein, J. L., Fleet, D. J. & Brubaker, M. A. cryoSPARC: algorithms for rapid unsupervised cryo-EM structure determination. Nat. Methods. 14, 290–296 (2017).

Article   CAS   PubMed   Google Scholar

Vilas, J. L., Carazo, J. M. & Sorzano, C. O. S. Emerging themes in cryoEM—single particle analysis image processing. Chem. Rev. 122, 13915–13951 (2022).

Article   CAS   PubMed   PubMed Central   Google Scholar

Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000).

Article   CAS   PubMed   PubMed Central   Google Scholar

Terwilliger, T. C., Adams, P. D., Afonine, P. V. & Sobolev, O. V. A fully automatic method yielding initial models from high-resolution cryo-electron microscopy maps. Nat. Methods. 15, 905–908 (2018).

Article   CAS   PubMed   PubMed Central   Google Scholar

Terashi, G. & Kihara, D. De novo main-chain modeling for EM maps using MAINMAST. Nat. Commun. 9, 1618 (2018).

Article   PubMed   PubMed Central   Google Scholar

He, J. & Huang, S.-Y. Full-length de novo protein structure determination from cryo-EM maps using deep learning. Bioinformatics 37, 3480–3490 (2021).

Article   CAS   PubMed   Google Scholar

Si, D. et al. Deep learning to predict protein backbone structure from high-resolution cryo-em density maps. Sci. Rep. 10, 4282 (2020).

Article   CAS   PubMed   PubMed Central   Google Scholar

Xu, K., Wang, Z., Shi, J., Li, H. & Zhang, Q. C. A2-net: molecular structure estimation from cryo-EM density volumes. In Proc. 33rd AAAI Conference on Artificial Intelligence (eds van Hentenryck, P. & Zhou, Z.-H.) (AAAI, 2019).

Pfab, J., Phan, N. M. & Si, D. DeepTracer for fast de novo cryo-EM protein structure modeling and special studies on CoV-related complexes. Proc. Natl Acad. Sci. USA 118, e2017525118 (2021).

Article   CAS   PubMed   Google Scholar

Giri, N. & Cheng, J. De novo atomic protein structure modeling for cryoEM density maps using 3D transformer and HMM. Nat. Commun. 15, 5511 (2024).

Article   CAS   PubMed   PubMed Central   Google Scholar

Jamali, K. et al. Automated model building and protein identification in cryo-EM maps. Nature 628, 450–457 (2024).

Article   CAS   PubMed   PubMed Central   Google Scholar

Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).

Article   CAS   PubMed   PubMed Central   Google Scholar

Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).

Article   CAS   PubMed   PubMed Central   Google Scholar

Kawabata, T. Multiple subunit fitting into a low-resolution density map of a macromolecular complex using a Gaussian mixture model. Biophys. J. 95, 4643–4658 (2008).

Article   CAS   PubMed   PubMed Central   Google Scholar

Wriggers, W. Conventions and workflows for using Situs. Acta Crystallogr. D Biol. Crystallogr. 68, 344–351 (2012).

Article   CAS   PubMed   PubMed Central   Google Scholar

Kovacs, J. A., Galkin, V. E. & Wriggers, W. Accurate flexible refinement of atomic models against medium-resolution cryo-EM maps using damped dynamics. BMC Struct. Biol. 18, 12 (2018).

Article   PubMed   PubMed Central   Google Scholar

He, J., Lin, P., Chen, J., Cao, H. & Huang, S.-Y. Model building of protein complexes from intermediate-resolution cryo-EM maps with deep learning-guided automatic assembly. Nat. Commun. 13, 4066 (2022).

Article   CAS   PubMed   PubMed Central   Google Scholar

Zhou, X. et al. Progressive assembly of multi-domain protein structures from cryo-EM density maps. Nat. Comput. Sci. 2, 265–275 (2022).

Article   PubMed   PubMed Central   Google Scholar

Zhang, Z. et al. DEMO-EM2: assembling protein complex structures from cryo-EM maps through intertwined chain and domain fitting. Brief. Bioinform. 25, bbae113 (2024).

Article   CAS   PubMed   PubMed Central   Google Scholar

Terwilliger, T. C. et al. Improved AlphaFold modeling with implicit experimental information. Nat. Methods. 19, 1376–1382 (2022).

Article   CAS   PubMed   PubMed Central   Google Scholar

Zhang, X., Zhang, B., Freddolino, P. L. & Zhang, Y. CR-I-TASSER: assemble protein structures from cryo-EM density maps using deep convolutional neural networks. Nat. Methods. 19, 195–204 (2022).

Article   CAS   PubMed   PubMed Central   Google Scholar

Terashi, G., Wang, X., Prasad, D., Nakamura, T. & Kihara, D. DeepMainmast: integrated protocol of protein structure modeling for cryo-EM with deep learning and structure prediction. Nat. Methods. 21, 122–131 (2024).

Article   CAS   PubMed   Google Scholar

Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021).

Article   CAS   PubMed   PubMed Central   Google Scholar

Krogh, A., Brown, M., Mian, I. S., Sjölander, K. & Haussler, D. Hidden Markov models in computational biology. Applications to protein modeling. J. Mol. Biol. 235, 1501–1531 (1994).

Article   CAS   PubMed   Google Scholar

Larralde, M. & Zeller, G. PyHMMER: a Python library binding to HMMER for efficient sequence analysis. Bioinformatics 39, btad214 (2023).

Article   CAS   PubMed   PubMed Central   Google Scholar

Zhang, Y. & Skolnick, J. TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Res. 33, 2302–2309 (2005).

Article   CAS   PubMed   PubMed Central   Google Scholar

Afonine, P. V. et al. Real-space refinement in PHENIX for cryo-EM and crystallography. Acta Crystallogr. D Struct. Biol. 74, 531–544 (2018).

Article   CAS   PubMed   PubMed Central   Google Scholar

Zhang, C., Shine, M., Pyle, A. M. & Zhang, Y. US-align: universal structure alignments of proteins, nucleic acids, and macromolecular complexes. Nat. Methods 19, 1109–1115 (2022).

Article   CAS   PubMed   Google Scholar

Afonine, P. V. et al. New tools for the analysis and validation of cryo-EM maps and atomic models. Acta Crystallogr. D Struct. Biol. 74, 814–840 (2018).

Article   CAS   PubMed   PubMed Central   Google Scholar

Chen, V. B. et al. MolProbity: all-atom structure validation for macromolecular crystallography. Acta Crystallogr. D Biol. Crystallogr. 6, 12–21 (2010).

Article   Google Scholar

Kabsch, W. & Sander, C. Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 22, 2577–2637 (1983).

Article   CAS   PubMed   Google Scholar

Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol. 35, 1026–1028 (2017).

Article   CAS   PubMed   Google Scholar

Pettersen, E. F. et al. UCSF Chimera—a visualization system for exploratory research and analysis. J. Comput. Chem. 25, 1605–1612 (2004).

Article   CAS   PubMed   Google Scholar

Zhang, K. et al. Practical blind image denoising via Swin-Conv-UNet and data synthesis. Mach. Intell. Res. 20, 822–836 (2023).

Article   Google Scholar

Liu, Z. et al. Swin transformer: hierarchical vision transformer using shifted windows. In Proc. of the IEEE/CVF International Conference on Computer Vision (ed. O’Conner, L.) (IEEE, 2021).

Li, T. et al. All-atom RNA structure determination from cryo-EM maps. Nat. Biotechnol. 43, 97–105 (2025).

Article   PubMed   Google Scholar

Li, T., Cao, H., He, J. & Huang, S.-Y. Automated detection and de novo structure modeling of nucleic acids from cryo-EM maps. Nat. Commun. 15, 9367 (2024).

Article   CAS   PubMed   PubMed Central   Google Scholar

DiMaio, F., Tyka, M. D., Baker, M. L., Chiu, W. & Baker, D. Refinement of protein structures into low-resolution density maps using Rosetta. J. Mol. Biol. 392, 181–190 (2009).

Article   CAS   PubMed   PubMed Central   Google Scholar

He, J., Li, T. & Huang, S.-Y. Improvement of cryo-EM maps by simultaneous local and non-local deep learning. Nat. Commun. 14, 3217 (2023).

Article   CAS   PubMed   PubMed Central   Google Scholar

Wang, Z., Bovik, A. C., Sheikh, H. R. & Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process. 13, 600–612 (2004).

Article   PubMed   Google Scholar

Zhu, K., Su, H., Peng, Z. & Yang, J. A unified approach to protein domain parsing with inter-residue distance matrix. Bioinformatics 39, btad070 (2023).

Article   CAS   PubMed   PubMed Central   Google Scholar

Terwilliger, T. C. et al. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. Nat. Methods. 21, 110–116 (2024).

Article   CAS   PubMed   Google Scholar

Download references

Acknowledgements

This work was supported by the National Natural Science Foundation of China (32161133002, 32430020 and 62072199 to S.-Y.H.), the startup grant of Huazhong University of Science and Technology (to S.-Y.H.) and the Postdoctoral Fellowship Program of the China Postdoctoral Science Fund (GZB20250617 to T.L.). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Author information Authors and Affiliations

School of Physics and Key Laboratory of Molecular Biophysics of MOE, Huazhong University of Science and Technology, Wuhan, China

Tao Li, Ji Chen, Hao Li, Hong Cao & Sheng-You Huang

Contributions

S.-Y.H. conceptualized and supervised the project. T.L., J.C., H.L. and H.C. designed and performed the experiments. S.-Y.H. and T.L analyzed the data. T.L. and S.-Y.H. wrote the paper. All authors reviewed and approved the final version of the paper.

Corresponding author

Correspondence to Sheng-You Huang.

Ethics declarations Competing interests

The authors declare no competing interests.

Peer review Peer review information

Nature Structural & Molecular Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available. Primary Handling Editor: Sara Osman, in collaboration with the Nature Structural & Molecular Biology team.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information Supplementary Information

Supplementary Figs. 1–4.

Reporting Summary Peer Review File Supplementary Tables

Supplementary Tables 1–6.

Source data Source Data Fig. 2

Source data.

Source Data Fig. 4

Source data.

Source Data Fig. 6

Source data.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article Cite this article

Li, T., Chen, J., Li, H. et al. EMProt improves structure determination from cryo-EM maps. Nat Struct Mol Biol (2025). https://doi.org/10.1038/s41594-025-01723-1

Download citation

Received 16 July 2025

Accepted 24 October 2025

Published 08 December 2025

Version of record 08 December 2025

DOI https://doi.org/10.1038/s41594-025-01723-1

Share this article

Anyone you share the following link with will be able to read this content:

Get shareable link

Provided by the Springer Nature SharedIt content-sharing initiative

Subjects Cryoelectron microscopy Molecular modelling