A protein cryo-em structure quality evaluation method based on multi-feature fusion

Abstract

The application discloses a protein cryo-EM structure quality evaluation method based on multi-feature fusion, belongs to the field of bioinformatics, and comprises the following steps: firstly, constructing a simulated density map data set and a real experiment density map data set, and adopting a transfer training strategy of simulated density map pre-training and real experiment density map fine-tuning; preprocessing the density map and the protein structure to generate perturbed structure samples of different quality levels; extracting protein structure features, and constructing two types of residue-level density-structure correlation features, respectively representing the direct correlation of single-residue simulated density and experimental density, and the weighted correlation obtained by distributing the experimental density according to the simulated density contribution proportion in the density overlap area; after multi-scale feature fusion, inputting the deep learning evaluation network, and outputting the residue-level and overall-level quality scores. The application can improve the accuracy of protein structure model quality evaluation obtained by cryo-EM modeling, and has strong robustness.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics, structural biology detection, and computer applications, and in particular relates to a method for assessing the structural quality of proteins using cryo-electron microscopy based on multi-feature fusion. Background Technology

[0002] In recent years, cryo-electron microscopy has become an important tool for resolving the structures of biomacromolecule complexes. However, the protein atomic structures constructed from cryo-electron microscopy density maps can still be affected by factors such as density map resolution, local mass fluctuations, and noise interference, leading to problems such as local conformational biases, residue misalignments, or overall structural inaccuracies in the models. These errors not only affect the reliability of the structural models but also further restrict subsequent model screening, structural correction, and refinement optimization. Therefore, it is necessary to conduct quality assessments on protein structural models obtained through cryo-electron microscopy to determine model reliability and identify potential error regions.

[0003] Existing methods for assessing protein structure quality can be broadly categorized into two types: physical statistical methods and deep learning methods. Traditional physical statistical methods, such as VoroMQA, typically assess structural quality by describing atomic contact relationships and combining them with statistical potential functions. While possessing relatively clear physical meaning, their ability to characterize complex nonlinear features is relatively limited. With the development of artificial intelligence, deep learning methods have been gradually applied to the field of protein model quality assessment, with representative methods including DeepAccNet, QATEN, and EnQA-MSA. These methods significantly improve the accuracy and robustness of assessments by fusing sequence evolution information, three-dimensional geometric features, and some prior physical knowledge through neural networks. However, existing methods still primarily rely on the geometric and topological consistency of the structural model itself, failing to adequately utilize the experimental constraints contained in cryo-electron microscopy density maps. Therefore, in the presence of experimental noise, local resolution inhomogeneity, or protein conformational heterogeneity, relying solely on the structural features often fails to accurately determine the deviation of the model from real experimental data. Therefore, there is an urgent need to construct a protein structure quality assessment method that can integrate sequence information, geometric structural features, physical prior knowledge, and cryo-electron microscopy density map constraints, so as to more accurately evaluate the rationality of the structural model and its consistency with experimental data. Summary of the Invention

[0004] To overcome the shortcomings of existing protein structure quality assessment methods, which mainly rely on the geometric and topological information of the structural model itself and cannot directly utilize the original constraint information in the density map of cryo-electron microscopy experiments, this invention proposes a protein cryo-electron microscopy structure quality assessment method based on multi-feature fusion. Compared with traditional methods, this invention can be applied to the quality assessment of protein structures resolved by cryo-electron microscopy, providing an objective and reliable basis for the screening and refinement of protein structures based on cryo-electron microscopy.

[0005] The technical solution adopted by this invention to solve its technical problem is: A method for assessing the structural quality of proteins using cryo-electron microscopy based on multi-feature fusion, the method comprising the following steps: Step 1) Obtain the structure of the protein to be evaluated and its corresponding cryo-electron microscopy density map, and perform preprocessing and spatial alignment on the protein structure and the cryo-electron microscopy density map to obtain input data suitable for quality assessment; Step 2) Extract protein structural features and construct residue-level density-structure correlation features based on the protein structure and the corresponding cryo-electron microscopy experimental density map; wherein, the correlation features include at least a first correlation feature characterizing the direct correlation between the simulated density map and the experimental density map of a single residue, and a second correlation feature characterizing the correlation in the case of density overlap. Step 3) Input the above features into the trained quality assessment model for joint modeling, output the residue-level quality score of the protein structure to be evaluated, and obtain the model-level overall quality score based on the residue-level quality score, thereby realizing the quality assessment of protein structure.

[0006] Furthermore, in step 2), the residue-level density-structure correlation characteristics are obtained through the following process: 2.1) Based on residues Using the basic unit, simulated density values ​​of the residue hierarchy are generated on a density map voxel grid based on the atomic coordinates of the residue. In the local voxel region corresponding to this residue Internal extraction experimental density value ,in For voxel mesh points; 2.2) In local voxel regions Within, calculate residue-level simulated density values. Compared with experimental density value The correlation between them yields the first correlation feature. The first correlation feature is used to characterize the direct correlation between the simulated density map and the experimental density map of a single residue; 2.3) For cases where multiple residues overlap in density within the same voxel region, according to the residue... At voxel points The proportion of the simulated density value at a given location to the total simulated density values ​​of all residues, relative to the experimental density value. A weighted allocation is performed to obtain the residues. Weighted experimental density ; 2.4) In local voxel regions Inside, calculation With weighted experimental density value The correlation between them yields the second correlation feature. .

[0007] Furthermore, in step 3), the process of constructing the training dataset for the quality assessment model is as follows: screening protein structure samples; for samples with corresponding cryo-electron microscopy experimental density maps, obtaining local density regions corresponding to the spatial range of the target protein chain, and constructing an experimental density map dataset; for samples without corresponding experimental density maps, generating simulated cryo-electron microscopy density maps based on the protein structure, and constructing a simulated density map dataset; subsequently, removing redundancy, preprocessing, and dividing the datasets, and generating decoy conformation samples of different quality levels through structure prediction methods and / or random local perturbation strategies, and then spatially registering the decoy conformations with the original structures.

[0008] Furthermore, in step 2), the protein structural features include one-dimensional residue-level features, two-dimensional residue pair features, and three-dimensional local voxel features; wherein, one-dimensional residue-level features are used to characterize the sequence, physicochemical properties, and local conformation information of a single residue, two-dimensional residue pair features are used to characterize the geometric relationship, relative orientation, and interaction information between residue pairs, and three-dimensional local voxel features are used to characterize the local spatial environment information truncated with the residue as the center.

[0009] In step 3), the quality assessment model includes a convolutional module, a backbone residual network, and two prediction branches. The convolutional module is used to extract and fuse three-dimensional local voxel features, one-dimensional residue-level features, and two-dimensional residue pair features. First, the three-dimensional convolutional block of the convolutional module extracts three-dimensional local voxel features, which are then concatenated with one-dimensional residue-level features and input into the one-dimensional convolutional block of the convolutional module for preliminary feature mapping. Subsequently, the obtained residue-level representation is expanded along two dimensions to form a two-dimensional feature matrix, which is then fused with two-dimensional residue pair features and input into the two-dimensional convolutional block of the convolutional module to complete cross-dimensional information integration. The backbone residual network is used to perform multi-level spatial topology modeling on the fused two-dimensional features. It includes a feature mapping layer and multiple residual blocks. Each residual block contains multiple expanded residual units with different expansion rates to simultaneously model local atomic environments and long-range spatial constraints. The two prediction branches receive the shared features output by the backbone residual network in parallel, and output distance error distribution maps and contact probability mask maps, respectively, and further derive the residue-level quality score and overall quality score of the protein structure to be evaluated based on these.

[0010] In step 3), the quality assessment model is trained using a transfer training strategy based on pre-training on simulated density maps and fine-tuning on real cryo-electron microscopy experimental density maps. During pre-training on the simulated density map dataset, a joint optimization objective consisting of residue-level quality score loss, residue pair distance error distribution prediction loss, and residue contact relationship prediction loss is used to optimize the model and obtain pre-trained model parameters. During fine-tuning training on the real cryo-electron microscopy experimental density map dataset, the pre-trained model parameters are transferred to the quality assessment model. While retaining the residue pair distance error distribution prediction loss and residue contact relationship prediction loss, the residue-level quality score loss is adjusted to a weighted loss with an asymmetric penalty term and an error margin constraint term. Simultaneously, a layer-by-layer unfreezing training strategy is used to gradually expand the range of trainable parameters, ultimately obtaining the quality assessment model.

[0011] The technical concept of this invention is as follows: First, a training data system combining simulated density maps and cryo-electron microscopy experimental density maps is constructed, and protein structure, density maps, and perturbed conformations are uniformly preprocessed and spatially aligned. Then, based on the multi-scale structural features of proteins, two residue-level density-structure correlation features are further constructed. One feature characterizes the direct fitting relationship between the simulated density map and the experimental density map for a single residue, while the other characterizes the relative consistency relationship after allocating the actual density according to the contribution ratio of the simulated density within the overlapping region of multiple residue densities. Next, the aforementioned density constraint features and structural features are jointly input into a deep learning network for joint learning, enabling the model to simultaneously utilize prior structural information and cryo-electron microscopy experimental density information to discriminate local conformational deviations. Subsequently, a transfer training strategy based on pre-training with simulated density maps and fine-tuning with actual cryo-electron microscopy experimental density maps is adopted to improve the model's adaptability to noise, resolution variations, and distribution shifts in the cryo-electron microscopy experimental density maps. Finally, residue-level quality scores and model-level overall quality scores are output, achieving a quantitative assessment of the local error regions and overall reliability of the protein structure model.

[0012] The beneficial effects of this invention are as follows: First, it introduces cryo-electron microscopy density map information into protein structure quality assessment and characterizes it as residue-level density-structure correlation features, so that the assessment no longer relies solely on the information of the structural model itself, thereby improving the ability to identify local conformational deviations. Second, it constructs two types of residue-level density-structure correlation features, which can more effectively distinguish complex situations such as inconsistencies between local structure and experimental density, residue position deviations, and conformational shifts. Third, it adopts a transfer training strategy of pre-training with simulated density maps and fine-tuning with real cryo-electron microscopy density maps, which improves the model's adaptability to noise interference and differences in data distribution. Finally, it realizes the joint modeling of structural features, experimental density constraints, and quality score output, outputting residue-level quality scores and overall quality scores, providing a basis for model screening, error localization, and subsequent refinement. Attached Figure Description

[0013] Figure 1 This is a basic flowchart of a protein cryo-electron microscopy structure quality assessment method based on multi-feature fusion.

[0014] Figure 2 This is the structure of the A chain in the protein complex 7vq3 to be evaluated.

[0015] Figure 3 This is the local cryo-electron microscopy density map (EMD-32084) of the A chain in the protein complex 7vq3 to be evaluated.

[0016] Figure 4 This is the true quality assessment result at the A chain residue level of the protein complex 7vq3.

[0017] Figure 5 This is the quality assessment result predicted by the A chain residue-level model of the protein complex 7vq3.

[0018] Figure 6 This is a schematic diagram of a protein cryo-electron microscopy structural quality assessment method based on multi-feature fusion. Detailed Implementation

[0019] The present invention will now be further described with reference to the accompanying drawings.

[0020] Reference Figures 1-6 A method for assessing the structural quality of proteins using cryo-electron microscopy based on multi-feature fusion includes the following steps: Step 1) Construct a training data system consisting of a simulated density map dataset and a real cryo-electron microscopy experimental density map dataset. Perform uniform normalization preprocessing on the density maps and spatially align the protein structures with the corresponding cryo-electron microscopy density maps. In this embodiment, protein structure samples for constructing simulated cryo-electron microscopy density maps were obtained from the DeepAccNet public dataset (https: / / github.com / hiranumn / DeepAccNet), and single-stranded proteins with a length of less than 400 were selected. For protein chains with corresponding cryo-electron microscopy density maps, the corresponding density maps within a resolution range of 3–6 Å were downloaded from the EMDB database (https: / / www.ebi.ac.uk / emdb / ). Only the local density region corresponding to the spatial range of the protein chain was extracted as the cryo-electron microscopy density map for that chain, and samples with a cross-correlation coefficient between the protein structure and the density map below 0.6 were removed to construct the experimental density map dataset. For protein chains without corresponding cryo-electron microscopy density maps, simulated cryo-electron microscopy density maps were generated by sampling within a resolution range of 3–6 Å using the Chimera tool to construct the simulated density map dataset. Subsequently, the cd-hit tool (https: / / github.com / weizhongli / cdhit / releases) was used to remove redundancy from the selected samples based on a 40% sequence similarity threshold. The final experimental dataset contained 7684 protein structures, while the simulated dataset contained 1778 protein structures. Finally, the two datasets were randomly divided into training, validation, and test sets at ratios of 80%, 10%, and 10%, respectively, for use in subsequent phased training.

[0021] The density map voxel values ​​are z-score normalized, with voxel values ​​less than 0 set to 0, and an upper limit determined based on the 98th percentile of the positive density voxel distribution. Voxel values ​​exceeding this upper limit are then set to 1. Finally, the density values ​​are linearly normalized to the [0,1] interval.

[0022] The decoy conformation samples of different quality levels are generated by the structure prediction algorithm and the random local perturbation strategy, and the decoy conformation is spatially registered with the original structure to ensure that the decoy conformation and the density map are accurately matched in a unified reference coordinate system. In this embodiment, ColabFold (https: / / gitcode.com / gh_mirrors / co / ColabFold) prediction and Rosetta (https: / / rosettacommons.org / software / getting-started / ) local perturbation generate diverse perturbation structure samples. The TM-score index (https: / / www.aideepmed.com / TM-score / ) is used to spatially register the perturbation conformation with the original structure to ensure that the perturbation structure maintains pose matching with the density map in a unified reference coordinate system.

[0023] Step 2) Extract structural features of the protein, including one-dimensional residue-level features, two-dimensional residue pair features, and three-dimensional local voxel features. Construct residue-level density-structure correlation features based on the protein structure and the corresponding experimental density map. The residue-level density-structure correlation features include at least a first correlation feature characterizing the direct correlation between the simulated density map and the experimental density map of a single residue. And a second correlation feature characterizing the weighted correlation between simulated density maps of single residues and the allocated experimental density maps under density overlap conditions. ; In this embodiment, for each residue, its main chain atoms are... , and A local reference coordinate system is constructed, and the atomic coordinates within the residue neighborhood are represented in voxel format within this system. Two-dimensional residue pair features are used to characterize the association information between residue pairs, including the geometric distance between residues, relative orientation, and residue pair energy terms calculated using molecular dynamics software, thus characterizing the topological constraints of the structure at the residue pair scale. One-dimensional residue-level features are used to characterize the local structural state of individual residues, including amino acid type and its physicochemical properties, main chain dihedral angles (…). , The calculated residue energy terms and secondary structure information determined through pattern recognition are used. Density map-structure correlation features are obtained by calculating the cross-correlation coefficients between simulated and experimental density maps of single residues. For each residue... First, based on its atomic coordinate information, the corresponding residue-level simulated density values ​​are constructed. ,in y This represents the voxel grid points in the density map. The simulated density values ​​are obtained by superimposing the Gaussian functions of the contributions of each atom in the residue: ; in Represents residues i No. k The scattering intensity coefficient corresponding to each atom, This represents the three-dimensional spatial coordinates of the atom. The width parameter of the Gaussian function. σ From density map resolution R Decision, that is Based on this, Used to measure residues i Correlation between simulated density and experimental density: ; in, Represents residues Simulated density value at voxel point y This represents the density value at the corresponding voxel point in the true density map. Simulated density representing residue 𝑖 The average value across all voxel points involved in the calculation. Experimental density The residue This corresponds to the average value within the local voxel region. To further address the issue of spatial overlap caused by different residue density contributions, the values ​​of each residue at the voxel point are considered. y The proportion of the simulated density map contribution to the real density map Perform a weighted allocation to obtain the residues. True density value : ; in This indicates that all residues are at voxel points. y The sum of the simulated density values ​​at each location, j For residue indexing. Defined as the simulated density value of residues Its corresponding weighted true density value Correlation coefficient between them: ; in, Indicates allocation to residues Weighted true density The average value within the corresponding local voxel region.

[0024] 3) The protein structural features obtained in step 2) are fused with the residue-level density-structure correlation features and input into a deep learning evaluation network for joint modeling. The residue-level density-structure correlation features are concatenated with one-dimensional residue-level structural features to form a comprehensive residue-level representation, which is then combined with a local spatial representation extracted from three-dimensional local voxel features for feature mapping. Subsequently, the comprehensive residue-level representation is extended to two dimensions and fused with two-dimensional residue pair structural features to obtain a two-dimensional fused feature for quality assessment. The deep learning evaluation network outputs a residue-level quality score based on the two-dimensional fused feature and further obtains a model-level overall quality score. The deep learning evaluation network in this embodiment consists of a convolutional module, a backbone residual network, and two prediction branches, used to achieve feature integration and inference from local atomic environments to global topological constraints. In the convolutional module, a 3D convolutional block composed of four 3D convolutional layers and one pooling layer is first used to extract 3D local voxel features, which are then flattened and correlated with indices including sequence physicochemical properties, main chain dihedral angles, and density correlation. , The one-dimensional residue-level features are concatenated and then input into a 1D convolutional block consisting of two one-dimensional convolutional layers for initial feature mapping. Subsequently, the residue-level representations output by the 1D convolutional block are copied and expanded along both row and column dimensions to construct a... A two-dimensional matrix of dimension, where The residue length of the protein chain is calculated and fused with two-dimensional features containing residue pair distances, relative orientations, and energy terms. The fused two-dimensional features are then input into a 2D convolutional block consisting of a two-dimensional convolutional layer and a normalization layer to complete cross-dimensional information integration. The backbone residual network consists of a feature mapping layer and five residual blocks. Each residual block contains four expanded residual units with expansion rates of 1, 2, 4, and 8. Each expanded residual unit employs a bottleneck structure, including features for channel compression. Two-dimensional convolutional layers are used to extract spatial topological relationships. Dilated 2D convolutional layers and methods for restoring channel dimensions Two-dimensional convolutional layers are used, with normalization operations and ELU nonlinear activation functions applied before each convolutional layer. Input and output features are fused through skip connections, enabling the network to model long-range spatial constraints and complex local interactions of proteins while preserving details of low-level atomic distributions. Shared features from the backbone residual network are input in parallel to two prediction branches: a distance error prediction branch and a contact map prediction branch. These prediction branches employ a lightweight structure, consisting of a residual block and two expansion units with an expansion rate of 1, and do not include normalization layers. The final outputs are a distance error distribution map discretely divided into 15 intervals, and a mask map representing the contact probability of residues within a 15 Å range, based on which residue-level and overall LDDT (Local Distance Difference Test) scores are further derived.

[0025] During the training phase, a transfer training strategy was adopted. First, pre-training was performed on a simulated density map dataset to learn general representation capabilities, and then the training was transferred to a real cryo-electron microscopy experimental density map dataset for fine-tuning to enhance adaptability to real noise, resolution variations and data distribution differences. The process of the transfer training strategy in this embodiment is as follows: 3.1) The deep learning evaluation network is pre-trained on a simulated density map dataset. Protein structural features and residue-level density-structure correlation features are jointly input into the deep learning evaluation network, and residue-level quality score prediction, residue pair distance error distribution prediction, and residue contact relationship prediction are used as pre-training optimization objectives; wherein, the residue-level quality score loss term... Loss term for prediction of distance error distribution of residue pairs Loss term predicted by residue contact relationship The joint loss function is used to optimize the pre-training optimization objective, thereby obtaining initial model parameters with general structure-density representation capabilities; 3.2) After completing the pre-training on the simulated density map dataset, the initial model parameters are transferred as initialization parameters to the real cryo-electron microscopy experimental density map dataset for fine-tuning training; during this fine-tuning stage, the features are retained. and Two auxiliary loss terms are used, and the residue-level quality score loss term used in the pre-training phase is input into the deep learning evaluation network; simultaneously, the residue-level quality score loss term used in the pre-training phase is also included. Replace with residue-level weighted quality score loss with asymmetric penalty and error margin constraints. To suppress the model's systematic overestimation of low-quality residue regions and reduce extreme prediction errors; the residue-level weighted quality score loss The method is constructed by introducing asymmetric penalties for low-quality residue regions and additional constraints for portions exceeding a preset error margin, based on the basic residue-level quality score loss. 3.3) During the fine-tuning of the real cryo-electron microscopy experimental density map dataset, a layer-by-layer thawing training strategy was adopted to gradually expand the range of trainable parameters, so that the model could gradually adapt to the noise interference, resolution changes and data distribution differences in the real cryo-electron microscopy experimental density map while retaining the general structural representation ability obtained from the simulated density map pre-training. In this embodiment, the model is first pre-trained on a simulated density map dataset. A joint loss function consisting of a core loss term and an auxiliary loss term is used to constrain the model's intermediate geometric representation learning and final quality score prediction. ; in, Cross-entropy loss is used to predict the distance error distribution of residue pairs; A binary cross-entropy loss with logits is used to predict residue contact relationships; The mean squared error between the predicted and actual values ​​is used for calculation. The weights of the three loss terms are set as follows: =10、 =1 and =0.25. After pre-training on the simulated density map dataset, the initial model parameters were transferred to the cryo-electron microscopy experimental density map dataset for fine-tuning. During this fine-tuning phase, the parameters were retained... Two auxiliary loss terms, and the residue-level quality score loss term from the pre-training phase. Replace with residue-level weighted quality score loss with asymmetric penalty and error margin constraints. To suppress the model's systematic overestimation of low-quality residue regions and reduce extreme prediction errors; the loss function expression for the fine-tuning stage is: ; in, This represents the residue-level weighted quality score loss after introducing the asymmetric penalty term and the error margin constraint term.

[0026] Finally, during the fine-tuning of the cryo-electron microscopy density map dataset, a layer-by-layer thawing training strategy was adopted to gradually expand the range of trainable parameters. This allowed the model to gradually adapt to noise interference, resolution variations, and data distribution differences in the cryo-electron microscopy density map while retaining the general structural representation capabilities obtained from the simulated density map pre-training. Ultimately, a protein structure quality assessment model suitable for cryo-electron microscopy density maps was obtained.

[0027] During inference in this embodiment, the structure of the protein to be evaluated and its corresponding cryo-electron microscopy density map are input into the trained network, and residue-level quality scores and overall quality scores are output. In this embodiment, chain A of the protein 7vq3 to be predicted is predicted at the residue level according to the steps described above. After using this method, refer to... Figure 5 The model predicts a residue-level LDDT score of 0.7715; (Refer to...) Figure 4 The actual residue-level LDDT score is 0.7779.

[0028] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for assessing the structural quality of proteins using cryo-electron microscopy based on multi-feature fusion, characterized in that, The method includes the following steps: Step 1) Obtain the structure of the protein to be evaluated and its corresponding cryo-electron microscopy density map, and perform preprocessing and spatial alignment on the protein structure and the cryo-electron microscopy density map to obtain input data suitable for quality assessment; Step 2) Extract protein structural features and construct residue-level density-structure correlation features based on the protein structure and the corresponding cryo-electron microscopy experimental density map; wherein, the correlation features include a first correlation feature characterizing the direct correlation between the simulated density map and the experimental density map of a single residue, and a second correlation feature characterizing the correlation in the case of density overlap. Step 3) Input the above features into the trained quality assessment model for joint modeling, output the residue-level quality score of the protein structure to be evaluated, and obtain the overall model-level quality score based on the residue-level quality score, thereby realizing the quality assessment of protein structure.

2. The protein cryo-electron microscopy structural quality assessment method based on multi-feature fusion as described in claim 1, characterized in that, In step 2), the residue-level density-structure correlation characteristics are obtained through the following process: 2.1) Based on residues Using the basic unit, simulated density values ​​of the residue hierarchy are generated on a density map voxel grid based on the atomic coordinates of the residue. In the local voxel region corresponding to this residue Internal extraction experimental density value ,in For voxel mesh points; 2.2) In local voxel regions Within, calculate residue-level simulated density values. Compared with experimental density value The correlation between them yields the first correlation feature. The first correlation feature is used to characterize the direct correlation between the simulated density map and the experimental density map of a single residue; 2.3) For cases where multiple residues overlap in density within the same voxel region, according to the residue... At voxel points The proportion of the simulated density value at a given location to the total simulated density values ​​of all residues, relative to the experimental density value. A weighted allocation is performed to obtain the residues. Weighted experimental density ; 2.4) In local voxel regions Inside, calculation With weighted experimental density value The correlation between them yields the second correlation feature. .

3. A protein cryo-electron microscopy structural quality assessment method based on multi-feature fusion as described in claim 1 or 2, characterized in that, In step 3), the construction process of the training dataset for the quality assessment model is as follows: screening protein structure samples; for samples with corresponding cryo-electron microscopy experimental density maps, obtaining local density regions corresponding to the spatial range of the target protein chain, and constructing an experimental density map dataset; for samples without corresponding experimental density maps, generating simulated cryo-electron microscopy density maps based on the protein structure, and constructing a simulated density map dataset; subsequently, removing redundancy, preprocessing, and dividing the samples into datasets, and generating decoy conformation samples of different quality levels through structure prediction methods and / or random local perturbation strategies, and then spatially registering the decoy conformations with the original structures.

4. The protein cryo-electron microscopy structural quality assessment method based on multi-feature fusion as described in claim 1, characterized in that, In step 2), the protein structural features include one-dimensional residue-level features, two-dimensional residue pair features, and three-dimensional local voxel features. The one-dimensional residue-level features are used to characterize the sequence, physicochemical properties, and local conformation information of a single residue. The two-dimensional residue pair features are used to characterize the geometric relationship, relative orientation, and interaction information between residue pairs. The three-dimensional local voxel features are used to characterize the local spatial environment information truncated around the residue.

5. A protein cryo-electron microscopy structural quality assessment method based on multi-feature fusion as described in claim 1 or 2, characterized in that, In step 3), the quality assessment model includes a convolutional module, a backbone residual network, and two prediction branches. The convolutional module is used to extract and fuse three-dimensional local voxel features, one-dimensional residue-level features, and two-dimensional residue pair features. First, the three-dimensional convolutional block of the convolutional module extracts three-dimensional local voxel features, which are then concatenated with one-dimensional residue-level features and input into the one-dimensional convolutional block of the convolutional module for preliminary feature mapping. Subsequently, the obtained residue-level representation is expanded along two dimensions to form a two-dimensional feature matrix, which is then fused with two-dimensional residue pair features and input into the two-dimensional convolutional block of the convolutional module to complete cross-dimensional information integration. The backbone residual network is used to perform multi-level spatial topology modeling on the fused two-dimensional features. It includes a feature mapping layer and multiple residual blocks. Each residual block contains multiple expanded residual units with different expansion rates to simultaneously model local atomic environments and long-range spatial constraints. The two prediction branches receive the shared features output by the backbone residual network in parallel, and output distance error distribution maps and contact probability mask maps, respectively, and further derive the residue-level quality score and overall quality score of the protein structure to be evaluated based on these.

6. The protein cryo-electron microscopy structural quality assessment method based on multi-feature fusion as described in claim 5, characterized in that, The quality assessment model is trained using a transfer training strategy based on pre-training on simulated density maps and fine-tuning on real cryo-electron microscopy experimental density maps. During pre-training on the simulated density map dataset, a joint optimization objective consisting of residue-level quality score loss, residue pair distance error distribution prediction loss, and residue contact relationship prediction loss is used to optimize the model and obtain pre-trained model parameters. During fine-tuning training on the real cryo-electron microscopy experimental density map dataset, the pre-trained model parameters are transferred to the quality assessment model. While retaining the residue pair distance error distribution prediction loss and residue contact relationship prediction loss, the residue-level quality score loss is adjusted to a weighted loss with an asymmetric penalty term and an error margin constraint term. Simultaneously, a layer-by-layer unfreezing training strategy is used to gradually expand the range of trainable parameters, ultimately obtaining the quality assessment model.

Images