Protein structure heterogeneity analysis method and system based on generative model
Through the variational autoencoder and high-low frequency separation training strategy combined with physical imaging simulation methods, the problem of difficult analysis of rare conformations in Cryo-EM technology is solved, and efficient and accurate analysis of continuous conformations of biological macromolecules is achieved.
Patent Information
- Application Number
- CN202510522431.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-24
- Publication Date
- 2025-08-08
AI Technical Summary
The existing Cryo-EM technology has problems such as uneven data distribution, strong model dependence, and insufficient generalization ability when analyzing the rare conformation and continuous conformation changes of biological macromolecules, which makes it difficult to accurately identify and reconstruct rare conformations.
A dynamic conformation generation and screening framework was constructed using a variational autoencoder (VAE), combining high- and low-frequency information separation training strategies and physical imaging simulations, learning local details through high-frequency VAE and learning the overall structure through high-frequency VAE, generating density maps and screening the most reliable conformation with the global scoring mechanism.
The detection probability and analysis capabilities of rare conformations are improved, ensuring that the generated results are consistent with the original electron microscope data, overcoming the bias and rare conformation underestimation problems of existing methods, and providing an efficient and accurate structural analysis method.
Smart Images
Figure CN120452524A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of biological computing and structural analysis, and specifically relates to a method and system for analyzing protein structural heterogeneity based on a generative model. Background Art
[0002] The structure of biological macromolecules determines their function, and analyzing their three-dimensional conformation is crucial for understanding life processes and biochemical mechanisms. Cryo-electron microscopy (Cryo-EM) has developed rapidly in recent years. It has become one of the core technologies for analyzing the structure of biological macromolecules because of its advantages of not requiring crystallization, being applicable to macromolecular complexes, and being able to achieve atomic or near-atomic resolution. Cryo-EM captures the natural conformation of biological samples by vitrifying them in a very short time, avoiding the interference of conformational state transitions in solution. However, biological macromolecules are inherently dynamic and often exhibit continuous conformations. Movement is closely linked to function, and existing Cryo-EM data processing methods still face significant challenges in analyzing continuous conformational changes and rare conformations.
[0003] According to the principles of thermodynamic equilibrium, the conformational distribution of a sample during vitrification differs significantly from that in equilibrium. Stable conformations, due to their high prevalence, provide sufficient particle data for high-resolution reconstruction. However, rare conformations, due to their low abundance under equilibrium conditions, often face difficulties in obtaining sufficient data for reliable analysis. These rare conformations often play key roles in biological processes, making their structural elucidation of great scientific value. However, in traditional cryo-EM single-particle analysis, achieving ideal resolution for rare conformations is often difficult due to limited available image data and the difficulty of discrete clustering methods in fully capturing continuous conformational changes. To address this issue, several improved strategies have been proposed in recent years. For example, time-resolved cryo-electron microscopy (TREEM) precisely controls biochemical reactions before freezing, allowing the sample to vitrify at a moment when high-energy, quasi-stable conformations dominate, thereby increasing the observability of rare conformations. However, this method relies on a deep understanding of biochemical dynamics and requires highly precise microfluidic mixing and grid preparation, making the experimental operation relatively challenging. In addition, liquid cell electron microscopy (LCEM) increases the proportion of high-energy conformations by maintaining non-equilibrium conditions in a liquid environment, but due to the lack of cryo-fixation protection, the resolution is limited, making it difficult to achieve the atomic-level precision of Cryo-EM.
[0004] Given the inherent challenges of existing methods in resolving rare conformations, there is an urgent need to develop specialized image processing algorithms to reduce the uneven distribution of different conformations. In recent years, the Continuous Heterogeneity Reconstruction method has gradually emerged as an alternative. CryoDRGN directly infers three-dimensional conformational change trajectories from two-dimensional particle images, resolving the energy landscape of dynamic complexes without presetting the number of conformational classes. Its nonlinear dimensionality reduction properties are particularly adept at capturing transitional conformations between metastable states, significantly improving the reconstruction accuracy of rare conformations. In addition, deep learning-based post-processing techniques have been proposed to improve the resolution of rare conformations. EMReady utilizes a network architecture based on the three-dimensional Swin-Conv-UNet, which combines conventional residual convolution for local modeling, shifted windows for non-local modeling, and a multi-scale UNet to further enhance the advantages of local and non-local modeling. Through a neural network trained on large cryo-EM datasets, EMReady is able to predict and refine features in low-resolution input images, thereby improving the resolution accuracy of rare conformations.
[0005] Disadvantages of CryoDRGN:
[0006] CryoDRGN is based on data-driven modeling, but its analytical results often rely on the distribution of data rather than directly from the raw electron microscopy particle data. When the amount of data is uneven, the reconstruction algorithm may underestimate the existence of certain rare conformations, or even lead to deviations in the conformational transition pathway, resulting in non-physiological conformational transitions. In addition, rare conformations that are too low in the dataset may not be accurately identified, or even be automatically classified by the model as more common conformations, causing the reconstruction results to favor mainstream conformations while ignoring low-abundance biological states.
[0007] Disadvantages of EMReady:
[0008] While EMReady can optimize reconstruction results, its model training relies on existing protein model datasets, making it susceptible to "data bias," meaning the model cannot correctly predict conformations not found in the training set. This can lead to some newly discovered structures being merely "speculations" by the deep learning algorithm rather than actual biological conformations. Furthermore, the generalization ability of the EMReady model depends on the diversity of the training data. When rare conformations are not included in the training dataset, the model may fail to correctly identify and optimize these conformations, preferring to map them to known, common conformations, thus affecting the accuracy of the analysis. Summary of the Invention
[0009] In view of this, a method and system for analyzing protein structural heterogeneity based on generative models are proposed. A variational autoencoder (VAE) is used to construct a dynamic conformation generation and screening framework. Combined with physical imaging simulation and a global scoring mechanism, this method achieves unbiased characterization of the continuous dynamic changes of biological macromolecules and effectively compensates for the shortcomings of existing methods in terms of the lack of direct verification with raw electron microscopy data. The present invention uses a variational autoencoder (VAE) as the encoder and decoder of the protein density map. By using a strategy of separate training of high-frequency and low-frequency information, the model's ability to learn structural information at different scales of biological macromolecules is enhanced. The low-frequency VAE is responsible for learning the overall structure and morphology of the density map, ensuring that the generative model can accurately capture large-scale conformational changes in proteins. The high-frequency VAE is responsible for learning the local details and texture information of the density map to improve the resolution and accuracy of the generated density map.
[0010] Through this training process, the present invention constructs a latent space model and uniformly samples within the latent space to generate a set of derivative density atlases covering the continuous conformational distribution. Subsequently, the present invention projects the generated three-dimensional density maps using physical imaging simulations and calculates a similarity metric between each projection and the original cryo-EM particle image, combining Euler angle information from the actual particles. Based on this metric, the present invention constructs a comprehensive scoring function to globally rank all generated density maps, ultimately selecting the most credible conformations.
[0011] The present invention effectively improves the ability to analyze the continuous heterogeneity of cryo-EM data through variational autoencoders + physical imaging simulation + scoring mechanism, overcomes the shortcomings of existing methods in dealing with rare conformations and continuous conformational change paths, and ensures that the analysis results of the generated model can be directly verified with the original electron microscopy data, thereby providing an efficient, accurate and unbiased structural analysis method.
[0012] A protein structural heterogeneity analysis system based on a generative model, comprising a generative algorithm module and a model analysis module;
[0013] The generation algorithm module includes a preprocessing module, a low-frequency encoder module, a high-frequency encoder module, and a decoder module;
[0014] The model analysis module includes a sampling module, an indicator evaluation module, and a model evaluation module;
[0015] The data preprocessing module performs high-frequency and low-frequency separation processing on the input original protein density map to obtain a high-frequency density map and a low-frequency density map;
[0016] The high-frequency encoder module is used to receive the high-frequency density map, and is responsible for learning the local details and texture information of the density map to obtain the deep features of the high-frequency density map;
[0017] The low-frequency encoder module is used to receive the low-frequency density map, and is responsible for learning the overall structure and morphology of the density map to obtain the deep features of the low-frequency density map;
[0018] The decoder module receives the deep features from the high-frequency encoder module and the low-frequency encoder module, performs multiple random uniform samplings based on the deep features, generates image pairs consisting of high-frequency and low-frequency three-dimensional density maps, and splices each pair of images and outputs them to the indicator evaluation module;
[0019] The index evaluation module receives all generated density maps from the decoder module, and projects the generated density maps at the same angle as the Euler angle based on the Euler angle information of the original electron microscope data set, calculates the similarity index between each projection and the real particle image, and outputs the similarity index corresponding to all projections to the model evaluation module;
[0020] The model evaluation module receives all similarity indices obtained from the index evaluation module, performs weighted average on the similarity indices of the same density map to obtain a similarity score, and ultimately screens out the conformation with the highest similarity score for subsequent structural analysis.
[0021] Preferably, the data preprocessing module performs high-frequency and low-frequency separation processing on the input original protein density map to obtain a high-frequency density map and a low-frequency density map, including:
[0022] The density map is transformed into the frequency domain through Hartley transform. A spherical mask is used with the maximum frequency distance as the radius. The area with a frequency domain radius of 0-0.2 is defined as low frequency, and the area with a radius of 0.2-0.6 is defined as high frequency. A linear transition is used in 0.2-0.6, and the extremely high frequency greater than 0.6 is set to 0; then the inverse Hartley transform is used to convert the separated frequency domain into high and low frequency density maps.
[0023] Preferably, the network structure adopted by the high-frequency encoder module is a five-layer standard 3D convolutional encoder, each layer includes convolution, GELU activation and downsampling operations, and uses mean square error loss to constrain the reconstruction accuracy so that the model can accurately restore high-frequency details. At the same time, the KL divergence is introduced to align with the prior distribution of the latent space to enhance the generation ability and the continuity of the latent space expression; and the gradient loss weight is used to enhance the focus on local structure to ensure that key high-frequency features are not blurred by the hierarchical abstraction process of the deep learning model.
[0024] Preferably, in the high-frequency encoder module, a comprehensive loss function of the high-frequency encoder module is constructed using mean square error loss, KL divergence loss and gradient loss.
[0025] Preferably, in the high frequency encoder module, the gradient loss function is:
[0026]
[0027] Among them, x is the original three-dimensional density map, To generate a 3D density map, and They are the depth gradient, height gradient and width gradient of the original three-dimensional density map respectively; and They are the depth direction gradient, height direction gradient and width direction gradient for generating the three-dimensional density map.
[0028] Preferably, the low-frequency encoder module adopts a five-layer standard 3D convolution downsampling structure and adds global average pooling in the forward direction; and uses mean square error loss and KL divergence loss to construct a comprehensive loss function of the low-frequency encoder module.
[0029] Preferably, the decoder module adopts a five-layer three-dimensional deconvolution structure, gradually upsampling and restoring to the original density map size; the activation function uniformly adopts GELU, the last layer uses the Tanh activation function, and the output value is normalized to the [-1,1] interval and aligned with the normalized density map label.
[0030] Preferably, the method of splicing the sampling modules to reconstruct a complete protein density map is: performing Hartley transform on the generated density map, performing frequency domain splicing in the frequency domain according to the threshold method in preprocessing, and then performing inverse Hartley transform to obtain the final protein density map.
[0031] A method for continuous heterogeneity analysis using cryo-electron microscopy to generate a model, the method comprising:
[0032] The first step is to perform high-frequency and low-frequency separation on the input original protein density map to obtain high-frequency density map and low-frequency density map;
[0033] The second step is to learn the local details and texture information of the high-frequency density map and obtain the deep features of the high-frequency density map;
[0034] The third step is to learn the overall structure and morphology of the low-frequency density map and obtain the deep features of the low-frequency density map.
[0035] The fourth step is to perform multiple random uniform samplings based on the deep features to generate image pairs consisting of high-frequency and low-frequency three-dimensional density maps, and then stitch each pair of images together;
[0036] Step 5: Based on the Euler angle information of the original electron microscope data set, the generated density map is projected at the same angle as the Euler angle, and the similarity index between each projection and the real particle image is calculated;
[0037] In the sixth step, the similarity index of the same density map is weighted averaged to obtain the similarity score, so as to finally screen out the conformation with the highest similarity score for subsequent structural analysis.
[0038] The present invention has the following beneficial effects:
[0039] This paper employs a separate training strategy for high- and low-frequency information. The low-frequency VAE focuses on the overall morphology and conformational changes of the density map, while the high-frequency VAE enhances the reconstruction accuracy of local details. Through a feature fusion mechanism, the two methods are synergistically optimized to overcome the limitations of traditional methods in balancing resolution and dynamic representation, while simultaneously improving the global rationality and local authenticity of the generated density map.
[0040] The present invention utilizes a latent space model constructed using a variational autoencoder (VAE), combined with a uniform sampling strategy, to comprehensively cover the continuous conformational distribution of biomacromolecules, avoiding the underestimation of rare conformations due to uneven data distribution. Compared to existing technologies, the present invention eliminates the need for predefined conformational classifications, significantly improving the ability to analyze continuous dynamic changes. This ensures that the generated protein density map covers all possible conformational states, thereby increasing the probability of detecting rare conformations.
[0041] The present invention ensures strict consistency between the generated density map and the experimental data by aligning the generated density map with the Euler angle information of the original electron micrograph particles and calculating a similarity index. This mechanism overcomes the shortcomings of existing deep learning algorithms that rely on mathematical speculation, allowing the authenticity of the analyzed conformation to be directly verified from the original electron micrographs, avoiding the risk of generating "pseudo-conformations."
[0042] This method uses a global scoring function to comprehensively consider the physical plausibility of the generated density map, its compatibility with the original data, and the uniformity of the conformational distribution, effectively screening for high-confidence conformations. Compared to existing techniques, this method significantly improves its adaptability to rare conformations outside the training set by leveraging a dual approach of physical constraints and data-driven analysis, making it particularly suitable for the analysis of newly discovered structures. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a schematic diagram of the system composition of the present invention;
[0044] Figure 2 Schematic diagram of frequency domain separation in the preprocessing module;
[0045] Figure 3 Schematic diagram of the comparison results between the conformation with the highest credibility and the original conformation. DETAILED DESCRIPTION
[0046] The present invention is described in detail below with reference to the accompanying drawings and embodiments.
[0047] A cryo-electron microscopy continuous heterogeneity analysis system for generating models, including a generation algorithm module and a model analysis module;
[0048] The generation algorithm module includes a preprocessing module, a low-frequency encoder module, a high-frequency encoder module, and a decoder module;
[0049] The model analysis module includes an indicator evaluation module and a model evaluation module.
[0050] The generative algorithm module is the core of the system, responsible for processing input data and training the generative model to construct a potential space that conforms to the continuous conformational changes of biomacromolecules;
[0051] The data preprocessing module preprocesses the input raw protein density map and outputs the preprocessed density maps to the high-frequency encoder module and the low-frequency encoder module, respectively. Density maps of multiple proteins of the same species are first obtained from the EMDB for training, and raw electron microscopy 2D images of the same proteins are obtained from the EMPIAR public dataset for subsequent metric evaluation. Preprocessing includes model conversion, high- and low-frequency separation, and normalization. First, the input protein model is given a level threshold to control display detail for subsequent training. The density map is transformed into the frequency domain using a Hartley transform. A spherical mask is applied, using the maximum frequency distance as the radius. The region with a frequency domain radius of 0-0.2 is defined as low-frequency, and the region with a radius of 0.2-0.6 is defined as high-frequency. A linear transition is applied between 0.2 and 0.6 (i.e., the mask value for low frequencies decreases linearly from 0.2 to 0.6, while the mask value for high frequencies decreases inversely). Extremely high frequencies greater than 0.6 are set to zero due to excessive noise. The separated frequency domains are then converted back to the spatial domain and normalized. The output is fed into two high-frequency and low-frequency encoder modules respectively and trained separately to improve the accuracy of subsequent data generation and model training;
[0052] The high-frequency encoder module is used to receive the high-frequency density map and is responsible for learning the local details and texture information of the density map to improve the resolution and accuracy of the generated density map. The network structure of this module is a five-layer standard 3D convolutional encoder. Each layer includes convolution + GELU activation + downsampling operations. The reconstruction accuracy is constrained by the mean square error loss, so that the model can accurately restore high-frequency details. At the same time, the KL divergence is introduced to align with the prior distribution of the latent space to enhance the generation ability and the continuity of the latent space expression. The gradient loss weight is used to enhance the focus on local structure to ensure that key high-frequency features are not blurred by the hierarchical abstraction process of the deep learning model.
[0053] The low-frequency encoder module is used for the low-frequency density map and is responsible for learning the overall structure and morphology of the density map to ensure that the generated model can accurately capture the large-scale conformational changes of the protein. This module adopts a five-layer standard 3D convolutional downsampling structure, but adds global average pooling in the forward pass, so that the model can effectively extract the overall morphological information of the protein, reduce the number of parameters, and improve computational efficiency. It also uses MSE and KL divergence losses with different weights from the high-frequency encoder to enable it to maintain consistency with the overall morphology of different conformations, avoiding unreasonable distortion or deviation in the global morphology of the generated density map. Its training focus is more on reconstructing large-scale conformational trends.
[0054] The decoder module receives deep features from the high-frequency encoder module and the low-frequency encoder module, mapping the high-dimensional latent vector from the encoder back to the three-dimensional density map space. The module uses a five-layer three-dimensional deconvolution structure, which gradually upsamples and restores the original density map size. The activation function uniformly adopts GELU, and the final layer uses the Tanh activation function. The output value is normalized to the range [-1, 1] and aligned with the normalized density map label. This module uniformly samples in the latent space, generating a sufficient number of high- and low-frequency generated density map pairs as required. The two density maps are spliced together by reversely applying the frequency domain separation method used in preprocessing to obtain the final generated density map. This process does not rely on the original data input, but instead conducts conformational exploration based on the trained latent space to ensure sufficient coverage of all possible biological conformations. The goal is to generate a reasonable generated protein density map to support subsequent credibility assessment.
[0055] The index evaluation module receives all generated density maps from the decoder module, and based on the Euler angle information of the original electron microscope data set, projects the generated density maps at the same angle, calculates the similarity index between each projection and the real particle image, and summarizes all the calculation results and outputs them to the model evaluation module. The purpose of this step is to directly verify whether the generated density map truly conforms to the distribution of the Cryo-EM original particle data to ensure the credibility of the conformation generated by the model.
[0056] The model evaluation module receives all the score sets from the index evaluation module, performs weighted average on the score sets to sort the scores of all generated density maps, and finally obtains the conformation ranking with the highest credibility for subsequent structural analysis.
[0057] A method for analyzing protein structural heterogeneity based on a generative model, the method comprising:
[0058] In the first step, the data preprocessing module performs preprocessing on the input protein density map. The high-frequency and low-frequency protein density maps obtained after processing are output to the high-frequency encoder module and the low-frequency encoder module respectively.
[0059] In the second step, the high-frequency five-layer convolutional downsampling encoder module receives the high-frequency protein model and outputs the deep features extracted by the high-frequency encoder to the decoder module;
[0060] In the third step, the low-frequency five-layer convolutional downsampling encoder module receives the low-frequency protein model and outputs the deep features extracted by the low-frequency encoder to the decoder module;
[0061] In the fourth step, the decoder module receives the high-frequency and low-frequency deep features and maps them into the latent space. After uniform sampling, it generates image pairs of high-frequency and low-frequency three-dimensional density maps. Each pair of high-frequency and low-frequency three-dimensional density maps is spliced and output to the indicator evaluation module.
[0062] In the fifth step, the index evaluation module receives the spliced protein density map, projects each generated protein density map, calculates the similarity index between the generated density map projection and the same Euler angle particle image in the original electron microscope data, and obtains the similarity index data of each density map and outputs it to the model evaluation module.
[0063] In the sixth step, the model evaluation module receives all similarity indices obtained from the index evaluation module, performs weighted average on the similarity indices of the same sub-density map to obtain a similarity score, and finally selects the conformation with the highest similarity score for subsequent structural analysis.
[0064] Example:
[0065] Figure 1 This is the overall flow chart of the system. Figure 1 As shown, first the original protein density map enters the generation algorithm module. After the generation algorithm module, the data passes through the preprocessing module, the high-frequency encoder module, the low-frequency encoder module and the decoder module to obtain the latent space. Figure 1 On the left is the original protein density map.
[0066] like Figure 2As shown in FIG, in the preprocessing module, the density map is separated into high and low frequencies by a frequency domain radius of a certain threshold, where the left side is the high frequency domain and the right side is the low frequency domain.
[0067] like Figure 1 As shown in the figure, the generated density map is input into the model analysis module, and the density map of the conformation with the highest credibility is obtained through the index evaluation module and the model evaluation module respectively. Compared with the original protein density map, as shown in the figure, Figure 3 shown.
[0068] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A protein structural heterogeneity analysis system based on generative model, characterized by: Includes generation algorithm module and model analysis module; The generation algorithm module includes a preprocessing module, a low-frequency encoder module, a high-frequency encoder module, and a decoder module; The model analysis module includes a sampling module, an indicator evaluation module, and a model evaluation module; The data preprocessing module performs high-frequency and low-frequency separation processing on the input original protein density map to obtain a high-frequency density map and a low-frequency density map; The high-frequency encoder module is used to receive the high-frequency density map, and is responsible for learning the local details and texture information of the density map to obtain the deep features of the high-frequency density map; The low-frequency encoder module is used to receive the low-frequency density map, and is responsible for learning the overall structure and morphology of the density map to obtain the deep features of the low-frequency density map; The decoder module receives the deep features from the high-frequency encoder module and the low-frequency encoder module, performs multiple random uniform samplings based on the deep features, generates image pairs consisting of high-frequency and low-frequency three-dimensional density maps, and splices each pair of images and outputs them to the indicator evaluation module; The index evaluation module receives all generated density maps from the decoder module, and projects the generated density maps at the same angle as the Euler angle based on the Euler angle information of the original electron microscope data set, calculates the similarity index between each projection and the real particle image, and outputs the similarity index corresponding to all projections to the model evaluation module; The model evaluation module receives all similarity indices obtained from the index evaluation module, performs weighted average on the similarity indices of the same density map to obtain a similarity score, and ultimately screens out the conformation with the highest similarity score for subsequent structural analysis.
2. A cryo-electron microscopy continuous heterogeneity analysis system for generating a model according to claim 1, characterized in that: The data preprocessing module performs high-frequency and low-frequency separation processing on the input original protein density map to obtain a high-frequency density map and a low-frequency density map, including: The density map is transformed into the frequency domain through Hartley transform. A spherical mask is used with the maximum frequency distance as the radius. The area with a frequency domain radius of 0-0.2 is defined as low frequency, and the area with a radius of 0.2-0.6 is defined as high frequency. A linear transition is used in 0.2-0.6, and the extremely high frequency greater than 0.6 is set to 0; then the inverse Hartley transform is used to convert the separated frequency domain into high and low frequency density maps.
3. The system for continuous heterogeneity analysis using cryo-electron microscopy for generating a model according to claim 1, characterized in that: The network structure adopted by the high-frequency encoder module is a five-layer standard 3D convolutional encoder. Each layer includes convolution, GELU activation and downsampling operations. The reconstruction accuracy is constrained by the mean square error loss, so that the model can accurately restore high-frequency details. At the same time, the KL divergence is introduced to align with the prior distribution of the latent space to enhance the generation ability and the continuity of the latent space expression; and the gradient loss weight is used to enhance the focus on local structure to ensure that key high-frequency features are not blurred by the hierarchical abstraction process of the deep learning model.
4. The system for continuous heterogeneity analysis using cryo-electron microscopy for generating a model according to claim 3, wherein: In the high-frequency encoder module, a comprehensive loss function of the high-frequency encoder module is constructed using mean square error loss, KL divergence loss, and gradient loss.
5. The system for continuous heterogeneity analysis using cryo-electron microscopy for generating a model according to claim 4, characterized in that: In the high-frequency encoder module, the gradient loss function is: Among them, x is the original three-dimensional density map, To generate a 3D density map, and They are the depth gradient, height gradient and width gradient of the original three-dimensional density map respectively; and They are the depth direction gradient, height direction gradient and width direction gradient for generating the three-dimensional density map.
6. The system for continuous heterogeneity analysis using cryo-electron microscopy for generating a model according to claim 3, characterized in that: The low-frequency encoder module adopts a five-layer standard 3D convolutional downsampling structure and adds global average pooling in the forward pass; the comprehensive loss function of the low-frequency encoder module is constructed using mean square error loss and KL divergence loss.
7. The system for continuous heterogeneity analysis using cryo-electron microscopy for generating a model according to claim 1, characterized in that: The decoder module adopts a five-layer three-dimensional deconvolution structure, which gradually upsamples and restores to the original density map size; the activation function uniformly adopts GELU, and the last layer uses the Tanh activation function. The output value is normalized to the [-1, 1] interval and aligned with the normalized density map label.
8. The system for continuous heterogeneity analysis using cryo-electron microscopy for generating a model according to claim 1, characterized in that: The method for splicing the sampling modules to reconstruct a complete protein density map is as follows: performing Hartley transform on the generated density map, performing frequency domain splicing in the frequency domain according to the threshold method in the preprocessing, and then performing inverse Hartley transform to obtain the final protein density map.
9. A method for analyzing continuous heterogeneity in cryo-electron microscopy using a generative model, characterized in that: The steps of the method include: The first step is to perform high-frequency and low-frequency separation on the input original protein density map to obtain high-frequency density map and low-frequency density map; The second step is to learn the local details and texture information of the high-frequency density map and obtain the deep features of the high-frequency density map; The third step is to learn the overall structure and morphology of the low-frequency density map and obtain the deep features of the low-frequency density map. The fourth step is to perform multiple random uniform samplings based on the deep features to generate image pairs consisting of high-frequency and low-frequency three-dimensional density maps, and then stitch each pair of images together; Step 5: Based on the Euler angle information of the original electron microscope data set, the generated density map is projected at the same angle as the Euler angle, and the similarity index between each projection and the real particle image is calculated; In the sixth step, the similarity index of the same density map is weighted averaged to obtain the similarity score, so as to finally screen out the conformation with the highest similarity score for subsequent structural analysis.