Cryoelectron microscope density map reconstruction method based on multi-modal deep learning framework

Abstract

The invention relates to a cryoelectron microscope density map reconstruction method based on a multi-modal deep learning framework, which comprises the following steps: obtaining a two-dimensional projection image from a cryoelectron microscope experiment, screening a particle image, and carrying out preprocessing including denoising, normalization and pose and CTF estimation; then dispersing the initial density field into a three-dimensional grid, uniformly sampling 3D coordinates, rotating and translating the coordinates in combination with pose information obtained by preprocessing, and simulating experimental conditions; thirdly, carrying out Fourier pose coding on the 3D coordinates, mapping the 3D coordinates to a high-dimensional space, combining the 3D coordinates with latent variables extracted by an image encoder, inputting the 3D coordinates into an MLP, generating an initial low-resolution 3D density map, carrying out U-Net refining to enhance local details, projecting the 3D density map into a 2D image through a differentiable model, and calculating loss with an original particle image to optimize MLP parameters; a density map with a more accurate structure is generated; and finally, the image quality under different resolutions is analyzed and evaluated through FSC, and the model performance is verified. According to the invention, the reconstruction efficiency and precision are improved.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics, intelligent optimization, and computer applications, and in particular to a method for reconstructing cryo-electron microscopy density maps based on a multimodal deep learning framework. Background Art

[0002] Proteins are the core functional carriers of life activities, and the analysis of their three-dimensional spatial structure is the cornerstone for revealing molecular mechanisms and physiological and pathological processes. The breakthrough development of cryo-electron microscopy (Cryo-EM) technology has made it possible to analyze the atomic-level structure of biological macromolecular complexes under near-physiological conditions, and has become an important technology in structural biology research. This technology rapidly freezes samples to maintain the native conformation of molecules, and combines transmission electron microscopy imaging with three-dimensional reconstruction algorithms to reconstruct a three-dimensional density map from massive two-dimensional projection data. Then, methods such as de novo modeling or isologous structure approximation are used to model the atomic spatial structure from the density map. Therefore, the quality of the density map reconstruction affects the accuracy of the subsequent atomic structure modeling.

[0003] The inherent conformational heterogeneity of biomacromolecules is a core challenge affecting reconstruction accuracy. Conformational heterogeneity refers to the dynamic structural diversity exhibited by the same protein molecule in different physiological states or microenvironments. This heterogeneity not only underlies the structural plasticity of protein function but is also a key factor influencing pathological aggregation. In response to this characteristic, the key to heterogeneous reconstruction using cryo-electron microscopy lies in resolving multiple conformational ensembles from a non-uniformly distributed 2D projected particle population.

[0004] Traditional methods rely on the Fourier slice theorem to reconstruct three-dimensional density maps through filtered back projection or algebraic iterative algorithms (such as Relion's Bayesian optimization). However, such methods are limited by the single conformational assumption and have difficulty handling continuous conformational changes or dynamic equilibrium systems. Early improvement strategies used discrete classification models (such as 3D classification and multi-body refinement) to guide particle grouping by introducing an initial structure, but prior model deviations can easily lead to artifact generation. Recently developed linear subspace models (such as CryoSPARC's 3D variability analysis) analyze conformational distributions through a manifold learning framework, but still face bottlenecks of high computational complexity and insufficient artifact suppression when dealing with nonlinear conformational trajectories.

[0005] The introduction of deep learning technology has opened up a new paradigm for heterogeneity reconstruction. For example, CryoDRGN uses variational autoencoders (VAE) to construct a latent variable space, mapping particle images to continuous conformational distributions to achieve unsupervised heterogeneity analysis; 3DFlex establishes a dynamic model of conformational flexible regions by fusing particle-specific motion parameters into the latent space. It is worth noting that CryoNeFEN innovatively adopts the principle of neural radiation field (NeRF) to directly optimize the mapping function from three-dimensional coordinates to density in the spatial domain, effectively overcoming the shortcomings of traditional Fourier domain methods in terms of insufficient visual angle coverage and loss of high-frequency information. However, existing deep learning methods still have significant gaps in signal-to-noise sensitivity and computing resource requirements, and the ability to physically interpret dynamic conformational trajectories needs to be improved.

[0006] Therefore, the existing cryo-EM density map reconstruction methods have defects in the accuracy of heterogeneous conformation reconstruction and need to be improved. Summary of the Invention

[0007] In order to overcome the shortcomings of existing cryo-electron microscopy density map reconstruction methods in terms of the accuracy of heterogeneous conformation reconstruction, the present invention provides a cryo-electron microscopy density map reconstruction method based on a multimodal deep learning framework to improve reconstruction efficiency and accuracy.

[0008] The technical solution adopted by the present invention to solve its technical problem is:

[0009] A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework, the method comprising the following steps:

[0010] 1) Input cryo-EM experimental particle images;

[0011] 2) Obtaining preprocessing preparation information and performing preprocessing;

[0012] 3) Discretize the initial density field into a regular three-dimensional grid and uniformly sample 3D coordinates on the cubic grid;

[0013] 4) performing rotation and translation operations on the result obtained in step 3) based on the information obtained from the preprocessing;

[0014] 5) Encode the 3D coordinates obtained in 4) to generate a higher-dimensional embedding p i ;

[0015] 6) The initial experimental particle image I j Input encoder to extract latent variables for subsequent processing;

[0016] 7) Embed the high-dimensional embedding in 5) into p i The latent variables in 6) are integrated and the multi-layer perceptron is used to extract and decode the features to obtain the corresponding density value v of each cubic grid.i , generate a 3D density map in the spatial domain;

[0017] 8) The 3D density map generated in 7) is refined through a U-Net network to improve the resolution of the density map;

[0018] 9) using the differentiable model F to convert the 3D density map in the spatial domain generated in 7) into a 2D prediction image;

[0019] 10) The particle image I obtained after preprocessing in 2) j The loss function is calculated with the 2D predicted image obtained in 9), and the network parameters are optimized using back-propagation gradients to obtain a higher-resolution 3D density map.

[0020] 11) Calculate the resolution of 3D density maps by Fourier shell correlation.

[0021] Further, the process of 2) is as follows:

[0022] 2.1) Extracting particle images from cryo-EM images;

[0023] 2.2) Remove background noise and interference from the particle image, normalize pixel values, and estimate target pose.

[0024] Furthermore, the process of 3) is as follows:

[0025] 3.1) Divide the three-dimensional space into a regular and uniform three-dimensional cube grid. Each square (voxel) in the grid corresponds to a density value, and the 3D coordinates of each voxel are:

[0026] x={(x i ,y i ,z i )}

[0027] Among them, x i is the coordinate information on the x-axis, y i is the coordinate information on the y-axis, z i is the coordinate information on the z-axis;

[0028] 3.2) When generating the network, regularization constraints are introduced to it taking into account structural symmetry and physical rationality.

[0029] Furthermore, the process of 4) is as follows: according to the rotation matrix of the predicted posture in 2) and translation vectors j corresponds to the particle image I j , perform corresponding rotation and translation operations on the three-dimensional grid to obtain a new three-dimensional grid.

[0030] The process of 5) is as follows: each coordinate x is normalized and data enhanced using Fourier pose coding to generate a high-dimensional embedding. The encoding method is as follows:

[0031]

[0032] Where D is the image size and L is the high-dimensional embedding p i The embedding length.

[0033] The process of 7) is as follows:

[0034] 7.1) Set the number of MLP layers and parameters;

[0035] 7.2) The final density map is represented as:

[0036]

[0037] Where V j represents the density map corresponding to the j-th particle image, f θ represents a continuous mapping, Rotation and The translation is obtained from 2).

[0038] The process of 8) is as follows:

[0039] 8.1) The 3D density map is sliced and processed, and each block is processed through U-Net to enhance the local resolution. The process is as follows:

[0040] 8.1.1) The density map V obtained in 7) is j Input U-Net encoder part, extract features through a series of 3D convolution layers and pooling operations;

[0041] 8.1.2) In the U-Net decoder, upsampling is performed to initially restore the original resolution, and skip connections are used to combine low-dimensional features in the encoder with high-dimensional features in the decoder to optimize the density map.

[0042] 8.2) After U-Net refinement, a new high-resolution 3D density map is obtained

[0043] Furthermore, the process of 9) is as follows: the high-resolution 3D density map generated by U-Net in 8) is projected according to the physical model F, and corrected by the obtained CTF parameters to obtain a 2D predicted image:

[0044]

[0045] Among them, f -1 and f are inverse Fourier transform and Fourier transform respectively, H jrepresents the CTF of the j-th particle image, π z Indicates that the density map is projected along the z-axis.

[0046] Furthermore, the process of 10) is as follows;

[0047] 10.1) Process the particle image as follows:

[0048]

[0049] Among them, h j It represents the point spread function, which is obtained by the inverse Fourier transform of CTF. * represents the convolution operation. It represents the translation in 2D dimension, n j represents the added Gaussian distributed white noise;

[0050] 10.2) Calculate the loss based on the particle images in 10.1) and 9):

[0051]

[0052] Among them, σ 2 It represents the variance of Gaussian noise.

[0053] 10.3) Optimize the generated 3D density map by minimizing the loss using the back-propagation gradient optimization network.

[0054] Finally, the process of 11) is as follows: the cryo-electron microscopy data is divided into two independent groups, each of which is trained to generate two 3D density maps, and FSC analysis is performed on the two 3D density maps to generate GSFSC curves:

[0055]

[0056] Among them, m1 and m2 represent the density maps obtained from the two sets of training data respectively, and k is the frequency shell.

[0057] The technical concept of the present invention is as follows: first, the cryo-EM particle image, pose information and corresponding CTF information are input, and the particle image is preprocessed; then, the three-dimensional grid is initialized and subjected to rotation and translation operations; secondly, the image feature information is extracted into a high-dimensional representation through pose encoding, and then the high-dimensional representation is input into the MLP to improve the density value and generate a low-resolution 3D density map; then, the low-resolution 3D density map is cut into blocks and input into the U-Net to optimize the image quality and generate a high-resolution 3D density map; finally, the loss function is calculated for the projection of the processed particle image and the 3D density map, and the final 3D density map is generated by minimizing the loss function and continuously iteratively optimizing.

[0058] The beneficial effects of the present invention are as follows: in real space, a progressive construction strategy from a low-resolution density map to a high-resolution density map is adopted, thereby improving reconstruction efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is the basic flow chart of this method.

[0060] Figure 2 It is the particle image 5412 selected from the input cryo-electron microscope two-dimensional projection image.

[0061] Figure 3 Low-resolution density map generated by MLP for protein complex 8c8z.

[0062] Figure 4 It is the high-resolution density map of protein complex 8c8z after U-Net optimization.

[0063] Figure 5 is the GSFSC curve of the low-resolution density map.

[0064] Figure 6 is the GSFSC curve of the high-resolution density map. DETAILED DESCRIPTION

[0065] The present invention will be further described below with reference to the accompanying drawings.

[0066] Reference Figures 1 to 6 , a cryo-electron microscopy density map reconstruction method based on a multimodal deep learning framework, comprising the following steps:

[0067] 1) If Figure 2 As shown, the input particle image 5412 obtained from the cryo-EM experiment;

[0068] 2) Prepare preprocessing information and preprocess the image. The process is as follows:

[0069] 2.1) The preprocessing information includes the pose information processed by CryoSPARC (https: / / cryosparc.com) and the CTF file downloaded from the EMPIAR dataset (https: / / www.ebi.ac.uk / empiar / );

[0070] 2.2) Extract particle images from cryo-EM;

[0071] 2.3) Using CryoSPARC to remove background noise and interference from particle images, normalize pixel values, estimate target pose through consensus reconstruction, and obtain pose and CTF information;

[0072] 3) Discretize the initial density field into a regular three-dimensional grid and uniformly sample 3D coordinates on the cubic grid. The process is as follows:

[0073] 3.1) Divide the three-dimensional space into a regular and uniform three-dimensional cube grid, where each square (voxel) in the grid corresponds to a density value.

[0074] x={(x i ,y i ,z i )}

[0075] Among them, x i is the coordinate information on the x-axis, y i is the coordinate information on the y-axis, z i is the coordinate information on the z-axis;

[0076] 3.2) When generating the network, regularization constraints are introduced taking into account structural symmetry and physical rationality;

[0077] 4) Perform rotation and translation operations on the result obtained in step 3) based on the information obtained in the preprocessing, and calculate the rotation matrix of the predicted posture in step 2) and translation vectors j corresponds to the particle image I j , perform corresponding rotation and translation operations on the three-dimensional cube grid to obtain a new three-dimensional grid;

[0078] 5) The 3D coordinates obtained in 4) are encoded by Fourier pose encoding to generate a high-dimensional embedding p i The process is as follows: normalize each coordinate x to ensure that the range of x is between (-0.5, 0.5), and use Fourier pose coding to enhance the data and generate a high-dimensional embedding. The encoding method is as follows:

[0079]

[0080] Where D is the image size and L is the high-dimensional embedding p i Embedding length;

[0081] 6) Extracting the latent variables of the experimental image. The image is input to the encoder and compressed into a one-dimensional 8D latent vector, which is then reparameterized into a Gaussian distribution and regularized using the Kullback-Leibler (KL) divergence.

[0082] 7) Embed the high-dimensional embedding in 5) into p i After the latent variables in 6) are integrated, the features are extracted and decoded through a multi-layer perceptron (MLP), so that each voxel obtains the corresponding density value v i , the process is as follows:

[0083] 7.1) Divide the number of MLP layers and set MLP parameters;

[0084] 7.2) Generate a 3D density map in the spatial domain. The resulting density map is represented as:

[0085]

[0086] Where V j It represents the density map corresponding to the j-th particle image, f θ represents a continuous mapping, Rotation and

[0087] The translation is obtained from 2);

[0088] 8) The 3D density map generated in 7) is refined through the U-Net network to improve the resolution of the density map. The process is as follows:

[0089] 8.1) The 3D density map is sliced and processed, and each block is processed through U-Net to enhance the local resolution. The process is as follows:

[0090] 8.1.1) The density map V obtained in 7) is j Input encoder part, extract features through a series of 3D convolution layers and pooling operations;

[0091] 8.1.2) In the U-Net decoder, upsampling is performed to initially restore the original resolution, and skip connections are used to combine the low-dimensional features in the encoder with the high-dimensional features in the decoder to further optimize the density map.

[0092] 8.2) After U-Net refinement, a new high-resolution 3D density map is obtained

[0093] 9) The high-resolution 3D density map generated by U-Net in 8) is projected according to the physical model F and corrected using the CTF parameters obtained from the EMPIAR dataset to obtain a 2D predicted image:

[0094]

[0095] Among them, f -1 and f are inverse Fourier transform and Fourier transform respectively, H j represents the CTF of the jth particle, π z Indicates that the density map is projected along the z-axis;

[0096] 10) The particle image I obtained after preprocessing in 2) jThe loss function is calculated with the predicted 2D image obtained in 9), and the network parameters are optimized using back-propagation gradients to obtain a higher-resolution 3D density map. The process is as follows:

[0097] 10.1) Process the particle image as follows:

[0098]

[0099] Among them, n j It represents the point spread function, which is obtained by the inverse Fourier transform of CTF. * represents the convolution operation. It represents the translation in 2D dimension, n j represents the added Gaussian distributed white noise;

[0100] 10.2) Calculate the loss based on the particle images in 10.1) and 9):

[0101]

[0102] Among them, σ 2 It represents the variance of Gaussian noise.

[0103] 10.3) Optimize the generated 3D density map by minimizing the loss using the back-propagation gradient optimization network.

[0104] 11) Calculate the resolution of the low-resolution 3D density map and the high-resolution 3D density map separately. The process is as follows: divide the cryo-EM data into two independent groups, train them separately, generate two 3D density maps, perform FSC analysis on the two 3D density maps, and generate GSFSC curves:

[0105]

[0106] Among them, m1 and m2 represent the density maps obtained from the two sets of training data respectively, and k is the frequency shell.

[0107] This embodiment includes a protein complex 8c8z containing three single chains. A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework is provided. The method comprises the following steps:

[0108] 1) Input particle images 5412 obtained from cryo-EM experiments;

[0109] 2) Prepare preprocessing information and preprocess the image. The process is as follows:

[0110] 2.1) The preprocessing information includes the poses processed by CryoSPARC (https: / / cryosparc.com) and the CTF files downloaded from the EMPIAR dataset (https: / / www.ebi.ac.uk / empiar / );

[0111] 2.2) Extracting particle images from cryo-EM images;

[0112] 2.3) Use CryoSPARC to remove background noise and interference from the particle image, then normalize the pixel values, estimate the target pose through consensus reconstruction, and obtain pose and CTF information;

[0113] 3) Discretize the initial density field into a regular three-dimensional grid and uniformly sample 3D coordinates on the cubic grid. The process is as follows:

[0114] 3.1) Divide the three-dimensional space into a regular and uniform three-dimensional cube grid. Each square (voxel) in the grid corresponds to a density value, and the 3D coordinates of each voxel are:

[0115] x={(x i ,y i ,z i )}

[0116] Among them, x i is the coordinate information on the x-axis, y i is the coordinate information on the y-axis, z i is the coordinate information on the z-axis;

[0117] 3.2) When generating the network, regularization constraints are introduced to it, taking into account structural symmetry and physical rationality;

[0118] 4) Perform rotation and translation operations on the result obtained in step 3) based on the information obtained in the preprocessing, and calculate the rotation matrix of the predicted posture in step 2) and translation vectors j corresponds to the particle image I j , perform corresponding operations on the three-dimensional cube grid to obtain the rotated and translated three-dimensional cube grid;

[0119] 5) The 3D coordinates obtained in 4) are encoded into poses to represent higher-dimensional embedding p i , the process is as follows:

[0120] Each coordinate x is normalized to ensure that the range of coordinate x is between (-0.5, 0.5); and Fourier pose coding is used for data enhancement to achieve high-dimensional embedding. The coding method is as follows:

[0121]

[0122] Where D is the image size and L is the high-dimensional embedding p i Embedding length;

[0123] 6) The encoder uses 5 layers of convolution, 4 convolution kernels, a stride of 2, and a LeakyReLU activation function to extract the latent variables of the experimental image. The image is compressed into a one-dimensional vector of size 1×64×64 and input into the encoder to generate an 8D latent variable. The latent variable is then reparameterized into a Gaussian distribution. The variables in the latent space are regularized using the Kullback-Leibler (KL) divergence.

[0124] 7) Embed the high-dimensional embedding in 5) into p i After integrating the latent variables in 6), the features are extracted and decoded through a multi-layer perceptron (MLP), so that each coordinate obtains the corresponding density value v i , the process is as follows:

[0125] 7.1) The MLP is set up as a three-layer model. The first layer maps the 200-dimensional input features to 256 dimensions. The middle layers retain their dimensions. The last layer compresses the 256-dimensional features to 1 dimension. ReLU activation functions are added to each layer.

[0126] 7.2) Generate a 3D density map in the spatial domain. The resulting density map is represented as:

[0127]

[0128] Where V j It represents the density map corresponding to the j-th particle image, f θ represents a continuous mapping, Rotation and The translation is obtained from 2),

[0129] In order to speed up the training process, a static mask is introduced and the 3D density map is thresholded with a maximum density value of 20% and expanded to Soft cosine filling is

[0130] 8) The 3D density map generated in 7) is refined through the U-Net network to improve the resolution of the density map. The process is as follows:

[0131] 8.1) The 3D density map is sliced and processed, and each block is processed through U-Net to enhance the local resolution. The process is as follows:

[0132] 8.1.1) The density map V obtained in 7) is j Input U-Net encoder part, extract features through a series of 3D convolution layers and pooling operations;

[0133] 8.1.2) In the U-Net decoder, upsampling is performed to initially restore the original resolution. Skip connections combine lower-dimensional features from the encoder with higher-dimensional features from the decoder to further optimize the density map.

[0134] 8.2) After U-Net refinement, a new high-resolution 3D density map is obtained

[0135] 9) The high-resolution 3D density map generated by U-Net in 8) is projected according to the physical model F and corrected using the CTF parameters obtained from the EMPIAR dataset to obtain a 2D predicted image:

[0136]

[0137] Among them, f -1 and f are inverse Fourier transform and Fourier transform respectively, H j represents the CTF of the jth particle, π z Indicates that the density map is projected along the z-axis;

[0138] 10) The particle image I obtained after preprocessing in 2) j The loss function is calculated with the 2D predicted image obtained in 9), and the parameters of the MLP are optimized using back-propagation gradients to obtain a higher-resolution 3D density map. The process is as follows:

[0139] 10.1) Process the particle image as follows:

[0140]

[0141] Among them, h j It represents the point spread function, which is obtained by the inverse Fourier transform of CTF. * represents the convolution operation. It represents the translation in 2D dimension, n j represents the added Gaussian distributed white noise;

[0142] 10.2) Calculate the loss based on the particle images in 10.1) and 9):

[0143]

[0144] Among them, σ 2 It represents the variance of Gaussian noise;

[0145] 10.3) Optimize the generated 3D density map by minimizing the loss using the back-propagation gradient optimization network;

[0146] 11) Calculate the resolution of the low-resolution 3D density map and the high-resolution 3D density map respectively. The process is as follows:

[0147] The cryo-electron microscopy data set 5412 was divided into two independent groups, and trained separately to generate two 3D density maps. The two 3D density maps were subjected to FSC analysis to generate GSFSC curves:

[0148]

[0149] Among them, m1 and m2 represent the density maps obtained from the two sets of training data respectively, and k is the frequency shell.

[0150] After experimental calculation, it was found that when the FSC value is 0.143, the resolution of the low-resolution 3D density map is The resolution of the high-resolution 3D density map is

[0151] Taking the protein complex 8c8z containing three single chains as an example, the input particle image 5412 is as follows Figure 2 As shown, the predicted low-resolution density map and high-resolution density map are respectively as follows Figure 3 、 Figure 4 As shown, low-resolution GSFSC and high-resolution GSFSC are respectively as shown in Figure 5 、 Figure 6 As shown, when the FSC value is 0.143, the resolution of the density map constructed by the method of the present invention is

[0152] The above description is a 3D density map structure obtained by the present invention using the protein complex 8c8z as an example, and does not limit the scope of implementation of the present invention. Various modifications and improvements made thereto without departing from the scope of the basic content of the present invention should not be excluded from the scope of protection of the present invention.

Claims

1. A cryo-electron microscopy density map reconstruction method based on a multimodal deep learning framework, characterized in that: The method comprises the following steps: 1) Input cryo-EM experimental particle images; 2) Obtaining preprocessing preparation information and performing preprocessing; 3) Discretize the initial density field into a regular three-dimensional grid and uniformly sample 3D coordinates on the cubic grid; 4) performing rotation and translation operations on the result obtained in step 3) based on the information obtained from the preprocessing; 5) Encode the 3D coordinates obtained in 4) to generate a higher-dimensional embedding p i ; 6) The initial experimental particle image I j Input encoder to extract latent variables for subsequent processing; 7) Embed the high-dimensional embedding in 5) into p i The latent variables in 6) are integrated and the multi-layer perceptron is used to extract and decode the features to obtain the corresponding density value v of each cubic grid. i , generate a 3D density map in the spatial domain; 8) The 3D density map generated in 7) is refined through a U-Net network to improve the resolution of the density map; 9) using the differentiable model F to convert the 3D density map in the spatial domain generated in 7) into a 2D prediction image; 10) The particle image I obtained after preprocessing in 2) j The loss function is calculated with the 2D predicted image obtained in 9), and the network parameters are optimized using back-propagation gradients to obtain a higher-resolution 3D density map. 11) Calculate the resolution of 3D density maps by Fourier shell correlation.

2. The method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1, wherein: The process of 2) is as follows: 2.1) Extracting particle images from cryo-EM images; 2.2) Use cryoSPARC to remove background noise and interference from particle images, perform pixel value normalization, and estimate target pose.

3. A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, characterized in that: The process of 3) is as follows: 3.1) Divide the three-dimensional space into a regular and uniform three-dimensional cube grid. Each square in the grid corresponds to a density value, and the 3D coordinates of each voxel are: x={(x i ,y i ,z i )} Among them, x i is the coordinate information on the x-axis, y i is the coordinate information on the y-axis, z i is the coordinate information on the z-axis; 3.2) When generating the network, regularization constraints are introduced to it taking into account structural symmetry and physical rationality.

4. A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, characterized in that: The process of 4) is as follows: According to the rotation matrix of the predicted posture in 2) and translation vectors j corresponds to the particle image I j Perform corresponding rotation and translation operations on the three-dimensional grid to obtain a new three-dimensional grid.

5. The method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, wherein: The process of 5) is as follows: each coordinate x is normalized and data enhanced using Fourier pose coding to generate a high-dimensional embedding. The encoding method is as follows: Where D is the image size and L is the high-dimensional embedding p i The embedding length.

6. A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, characterized in that: The process of 7) is as follows: 7.1) Set the number of MLP layers and parameters; 7.2) The final density map is represented as: Where V j It represents the density map corresponding to the j-th particle image, f θ represents a continuous mapping, Rotation and The translation is obtained from 2).

7. The method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, wherein the process of 8) is as follows: 8.1) The 3D density map is sliced and processed, and each block is processed through U-Net to enhance the local resolution. The process is as follows: 8.1.1) The density map V obtained in 7) is j Input U-Net encoder part, extract features through a series of 3D convolution layers and pooling operations; 8.1.2) In the U-Net decoder, upsampling is performed to initially restore the original resolution, and skip connections are used to combine low-dimensional features in the encoder with high-dimensional features in the decoder to further optimize the density map. 8.2) After refinement by U-Net, a new high-resolution 3D density map V is obtained j * .

8. The method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, wherein: The process of 9) is as follows: the high-resolution 3D density map generated by U-Net in 8) is projected according to the physical model F and corrected by the CTF parameters obtained from the EMPIAR dataset: F(V j * )=f -1 (H j f(π z (V j * ))) Among them, f -1 and f are inverse Fourier transform and Fourier transform respectively, H j represents the CTF of the j-th particle image, π z Indicates that the density map is projected along the z-axis.

9. A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, characterized in that: The process of 10) is as follows: 10.1) Process the particle image as follows: in, h j It represents the point spread function, which is obtained by the inverse Fourier transform of CTF. * represents the convolution operation. It represents the translation in 2D dimension, n j represents the added Gaussian distributed white noise; 10.2) Calculate the loss based on the particle images in 10.1) and 90): Among them, σ 2 It represents the variance of Gaussian noise; 10.3) Optimize the generated 3D density map by minimizing the loss using the back-propagation gradient optimization network.

10. A method for reconstructing a cryo-electron microscopy density map based on a multimodal deep learning framework according to claim 1 or 2, characterized in that: The process of 11) is as follows: the cryo-electron microscopy data are divided into two independent groups, each of which is trained to generate two 3D density maps, and the two 3D density maps are subjected to FSC analysis to generate GSFSC curves: Among them, m1 and m2 represent the density maps obtained from the two sets of training data respectively, and k is the frequency shell.

Images