Protein structure prediction method based on cryoelectron microscope and diffusion model

Abstract

A protein structure prediction method based on a cryoelectron microscope and a diffusion model comprises the following steps: extracting feature points from an electron density map of the cryoelectron microscope, converting the feature points into point cloud structures, and predicting corresponding amino acid types through a DAQ model according to the point cloud structures; one-dimensional and two-dimensional tensor features are obtained for an input amino acid sequence through an AlphaFold2 neural network and are used for training and constructing an obtained denoising neural network based on conditional diffusion, so that the denoising neural network supports input of point cloud structure information of a cryoelectron microscope and fusion of electron microscope information and sequence information; and the trained denoising neural network is matched with a sampling process based on conditional diffusion to realize protein skeleton structure prediction. According to the method, a more accurate prediction structure is given according to the information of the cryoelectron microscope, meanwhile, a change conformation space is reflected through the diffusion model, the prediction structure precision can be improved on the basis of a low-quality prediction structure prediction sample, meanwhile, the overall structure prediction precision based on the electron density map of the cryoelectron microscope is improved, and the change of the prediction structure conformation is reflected.

Description

Technical Field

[0001] The present invention relates to a technology in the field of protein structure prediction, specifically a method for predicting the skeleton structure of monomeric proteins based on amino acid sequence, cryo-electron microscopy electron density map and diffusion model. Background Art

[0002] The current mainstream protein structure prediction method relies solely on amino acid sequences and requires database searches to find multiple sequence alignments and template information. The accuracy will be severely reduced when there is insufficient information. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a protein structure prediction method based on cryo-electron microscopy and diffusion model, which relies on cryo-electron microscopy information to give a more accurate predicted structure, and reflects the changing conformational space through the diffusion model. It can improve the predicted structure accuracy on low-quality predicted structure prediction samples, and at the same time improve the overall structure prediction accuracy based on the cryo-electron microscopy electron density map, reflecting the changes in the predicted structure conformation.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a protein structure prediction method based on cryo-electron microscopy and diffusion model, wherein feature points are extracted from a cryo-electron microscopy electron density map and converted into a point cloud structure, and the corresponding amino acid types are predicted by a DAQ model according to the point cloud structure, and one-dimensional and two-dimensional tensor features are obtained for an input amino acid sequence through an AlphaFold2 neural network and used to train a constructed denoising neural network based on conditional diffusion, so that the denoising neural network supports the input of cryo-electron microscopy point cloud structure information and the fusion of electron microscopy information and sequence information, and then the trained denoising neural network is used in conjunction with a sampling process based on conditional diffusion to realize protein skeleton structure prediction.

[0006] The feature points are obtained by clustering the cryo-EM electron density map, specifically, filtering the cryo-EM electron density map by a given threshold and scaling the voxel size to Afterwards, the processed electron density map is clustered by the MeanShift++ clustering algorithm, and the class center points of each class are taken as feature points. Their coordinates are recorded in the form of point cloud, which is in the form of M×3 tensor, where M is the number of extracted feature points. After that, the point cloud is sampled at the farthest point and downsampled to N×3 point cloud tensor, where N is the length of the amino acid sequence.

[0007] The amino acid types are predicted according to the point cloud structure using a trained DAQ model, specifically: the cryo-EM electron density map is input into the trained DAQ neural network, and the probability of the amino acid type at each voxel is obtained. Then, the coordinates of each feature point of the electron density map are matched to the amino acid type distribution to obtain the amino acid type distribution of each feature point, and the amino acid type with the maximum probability is taken as the predicted amino acid type.

[0008] The DAQ model is implemented using, but not limited to, the technology described in Terashi, G. et al., "Residue-wise local quality estimation for protein models from cryo-EM maps" (Nature methods, 19(9), pp.1116-1125).

[0009] The one-dimensional and two-dimensional tensor features refer to: inputting the protein amino acid sequence into the AlphaFold2 neural network, and extracting the one-dimensional and two-dimensional features output by its network structure EvoFormer from the result, wherein the one-dimensional feature form is an N×384 tensor, and the two-dimensional feature form is an N×N×128 tensor, where N is the length of the amino acid sequence.

[0010] The AlphaFold2 neural network is implemented using but not limited to the technology described by Jumper, J. et al. in "Highly accurate protein structure prediction with AlphaFold" (Nature, 596 (7873), pp.583-589).

[0011] The denoising neural network based on conditional diffusion is based on the DDPM diffusion form, sets the maximum time step to T = 100, and adopts the cosine time table [α1, α2, ..., α N ], according to the input N×1 amino acid sequence, N×3 noise C α Carbon atom coordinate x t , constant time step t, N×3 cryo-EM feature point coordinates, N×1 cryo-EM feature point amino acid type, N×384 sequence corresponding to the AlphaFold2 one-dimensional features and N×N×128 sequence corresponding to the AlphaFold2 two-dimensional features, predicting the N×3 denoised C α Carbon atom coordinates Where: N is the length of the protein amino acid sequence, x i is the ith C α The three-dimensional coordinates of the carbon atom coordinates include the corresponding coordinates of the (x, y, z) axes. The denoising neural network is trained according to the input noise-containing protein Cα Carbon atom coordinates, predict the added coordinate noise; during the sampling process, the noise is gradually removed according to the initialized random noise to obtain the noise-free and accurate protein C α Carbon atom coordinates.

[0012] The training process is as follows: randomly sample the time step t from the time step range [1, T], and obtain the noise coordinates at the time step t according to the noise addition formula: ∈ t is a standard Gaussian noise of N×3, and the noise coordinate is x t , time step t and the coordinates of the cryo-EM feature points that are independent of t, the types of amino acids at the cryo-EM feature points, and the one-dimensional and two-dimensional features of AlphaFold2 are used as conditions to jointly input the denoising neural network to obtain its output Calculate the estimated noise Assume the loss function is Its specific form is each C α The square of the coordinate difference of the carbon atom is the same as the RMSD, a commonly used protein structure distance calculation indicator. The parameter update of the denoising neural network is achieved by minimizing the loss function and completing the gradient back propagation, and training is achieved through repeated updates.

[0013] The sampling process based on conditional diffusion is specifically as follows: Initialize protein C α The carbon atom coordinate is x T , where x T is a standard Gaussian noise of N×3. Starting from t=T, the noise coordinate is x t , time step t and the coordinates of the cryo-EM feature points that are independent of t, the types of amino acids at the cryo-EM feature points, and the one-dimensional and two-dimensional features of AlphaFold2 are used as conditions to jointly input the denoising neural network to obtain its output Calculate the estimated noise And calculate the noise coordinates of t-1 as in z is N×3 standard Gaussian noise, and then t=t-1, and the above process is repeated until t=0, and x0 is the protein C obtained by the final sampling. α Carbon atom coordinates.

[0014] The training process and sampling process of C α Each of the three axis coordinates in the carbon atom coordinate x is calculated independently.

[0015] The denoising neural network comprises: a feature embedding subnetwork, a feature fusion subnetwork and a structure construction subnetwork connected in series, wherein: the feature embedding subnetwork is based on the input N×1 amino acid sequence, N×3 noise C αCarbon atom coordinate x t , constant time step t, N×3 cryo-EM feature point coordinates, N×1 cryo-EM feature point amino acid type, N×384 AlphaFold2 one-dimensional features corresponding to the sequence, and N×N×128 AlphaFold2 two-dimensional features corresponding to the sequence, to obtain embedded N×384 one-dimensional features, N×N×128 two-dimensional features, N×N×128 cryo-EM features, and N×N×128 correlation features; the feature fusion subnetwork updates the input N×384 one-dimensional features, N×N×128 two-dimensional features, N×N×128 cryo-EM features, and N×N×128 correlation features through the modified AlphaFold2 EvoFormer structure to obtain four updated features with exactly the same input dimensions, and the final four updated features are obtained after repeated updating twice; the structure construction subnetwork fuses the N×384 one-dimensional features and N×N×128 two-dimensional features updated by the subnetwork according to the input features, as well as the noise C α Carbon atom coordinate x t , with x t Initialize the output coordinates for the coordinates, update the coordinates and features through the one-dimensional features and two-dimensional features, repeat the cycle four times, and output the final denoised C α Carbon atom coordinates

[0016] The N×1 amino acid sequence and time step t are encoded by the sine coding method for the amino acid sequence to obtain two N×384 codes. The time step t is sine-coded to obtain a 1×384 code and then expanded to N×384. The three codes are fused through the linear layer and the layer norm to obtain the amino acid sequence feature. This feature is fused with the AlphaFold2 one-dimensional feature through the linear layer and the layer norm to obtain a one-dimensional feature. The input noise C α Carbon atom coordinate x t Based on the spatial distance between two points and the index number relationship of each amino acid, two N×N×128 codes are obtained, and the initial two-dimensional features are obtained by fusion through linear layers and layer norms. This feature is fused with the AlphaFold2 two-dimensional features through linear layers and layer norms to obtain two-dimensional features; the coordinates of the cryo-EM feature points are input, encoded by the spatial distance between two points, and calculated through linear layers and layer norms to obtain cryo-EM features; the amino acid sequence and the amino acid type of the cryo-EM feature points are input, and the correlation characteristics are obtained by comparing the correlations between the two.

[0017] The present invention relates to a system for implementing the above method, comprising: a data processing unit, a model building unit and a result generating unit, wherein: the data processing unit extracts feature point cloud, downsamples and predicts corresponding amino acid types for cryo-electron microscopy electron density, and extracts one-dimensional and two-dimensional features of amino acid sequences using AlphaFold2; the model building unit completes formula reasoning, parameter setting, network architecture, and mathematical form determination for a denoising neural network based on conditional diffusion, and completes model parameter updates of the denoising neural network through a training process; the result generating unit processes the input data according to the data processing unit to obtain a standard form, and then inputs the data into the trained denoising neural network for sampling to obtain the final result. Technical Effects

[0018] The present invention uses a diffusion model to help complete the training of a protein structure prediction model based on cryo-electron microscopy information on small-scale data, thereby reflecting possible protein conformations, and uses deep learning to complete the cross-modal fusion of protein sequence information and cryo-electron microscopy structure information at a deep level, without the need for human intervention, and completes protein skeleton structure prediction efficiently and end-to-end; the introduction of cryo-electron microscopy information in the end-to-end protein structure prediction task improves the structure prediction accuracy of AlphaFold2 on its low-precision result prediction samples, and also performs better in overall prediction accuracy and has higher computational efficiency than discrete cryo-electron microscopy information-based structure construction methods in the field, and reflects possible conformational changes in protein structure through the diffusion model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the present invention;

[0020] Figure 2 It is a schematic diagram of embodiment data;

[0021] In the figure: the yellow curve is the single chain numbered F of the protein numbered 5h37 in the PDB, and the blue volume is the cryo-EM electron density map numbered 9575 in the EMDB database;

[0022] Figure 3 Schematic diagram of denoising neural network;

[0023] Figure 4 Schematic diagram of feature fusion sub-network;

[0024] Figure 5 Construct subnetwork diagrams for the structure;

[0025] Figure 6 The figure is a schematic diagram of the protein prediction skeleton structure results for the embodiment. DETAILED DESCRIPTION

[0026] like Figure 1 As shown, this embodiment is directed to Figure 2 The end-to-end deep protein monomer skeleton structure prediction method for the single chain No. F of protein No. 5h37 in the RCSB PDB database and its corresponding cryo-EM electron density map data is No. 9575 in the EMDB database, specifically including:

[0027] In the first step, the input cryo-EM electron density map is filtered by a given threshold and the voxel size is scaled to Finally, the filtered cryo-EM electron density map was clustered using the MeanShift++ clustering algorithm. The center points of each class were taken as feature points, and their coordinates were recorded in the form of a point cloud. A total of 177 feature points were extracted in the form of a 177×3 tensor. Finally, the point cloud was sampled at the farthest point and downsampled to a feature point tensor in the form of 75×3.

[0028] In the second step, the input cryo-EM electron density map is input into the trained DAQ neural network to obtain the probability of the amino acid type at each voxel. Then, the coordinates of each point of the feature point tensor obtained in the first step are matched to the amino acid type probability distribution to obtain the amino acid type probability of each feature point. The amino acid type with the highest probability is taken as the predicted amino acid type.

[0029] In the third step, the amino acid sequence of length 75 is input into the trained AlphaFold2 neural network, and the one-dimensional and two-dimensional features output by its network structure EvoFormer are extracted from the results. The one-dimensional feature form is a 75×384 tensor, and the two-dimensional feature form is a 75×75×128 tensor.

[0030] The fourth step is to set the maximum time step of the denoising neural network based on conditional diffusion to 100 and initialize protein C α The carbon atom coordinate is x 100 is a standard Gaussian noise of the form 75×3. Starting from t=100, the noise coordinates are x t , time step t, and the amino acid sequence generated in steps 1, 2, and 3, the feature point tensor, the amino acid type predicted by the feature point tensor, and the one-dimensional and two-dimensional features of AlphaFold2 are input into the trained denoising neural network, and the denoised C is output. α Carbon atom coordinates Calculate the estimated noise And calculate the noise coordinates at time step t-1 as in z is a standard Gaussian noise in the form of 75×3. Then, let t=t-1 and repeat the above process until t=0. x0 is the protein C obtained by the sampling process. α Carbon atom coordinates.

[0031] like Figure 3 As shown, the denoising neural network includes: a structure building sub-network, Figure 4 The feature fusion subnetwork shown and Figure 5 The structure shown builds the sub-network.

[0032] Step 5: After repeating step 4 to obtain several sampling structures, the best result is obtained based on the TM-score evaluation score (the larger the better) and the RMSD evaluation score (the smaller the better).

[0033] After specific practical experiments, the corresponding protein skeleton structure, namely C α Carbon atom coordinates and corresponding amino acid types, using AlphaFold2 model version 2.3.1, DAQ model is the official latest version.

[0034] like Figure 6 As shown, after multiple samplings, the present invention obtains 10 different sampling structures based on the numbered F single chain of the protein numbered 5h37 in the RCSB PDB database, and its TM-score evaluation score and RMSD evaluation score are shown in the figure. The result obtained by the benchmark method AlphaFold2 is TM-score 0.66 and RMSD 6.010. The present invention is placed on this implementation example to achieve higher structural prediction accuracy. The optimal structure is framed by a red rectangular frame. To facilitate the comparison of structural differences, the 10 sampling structures are simultaneously fixed with a yellow curve to show the true structure, and the best result TM-score is 0.7809 and RMSD is 2.320.

[0035] Compared with the prior art, this invention introduces cryo-EM information into the protein structure prediction task, which improves the structure prediction accuracy of AlphaFold2 on its low-precision result prediction samples as a whole, and also performs better in overall prediction accuracy than the structure construction method based on cryo-EM electron density map in the field. At the same time, the diffusion model is used to reflect the possible conformational changes of protein structure.

[0036] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A protein structure prediction method based on cryo-electron microscopy and diffusion model, characterized in that: Feature points are extracted from the cryo-EM electron density map and converted into a point cloud structure. The corresponding amino acid types are predicted by the DAQ model according to the point cloud structure. One-dimensional and two-dimensional tensor features are obtained from the input amino acid sequence through the AlphaFold2 neural network and used to train the constructed denoising neural network based on conditional diffusion, so that it supports the input of cryo-EM point cloud structure information and the fusion of EM information and sequence information. The trained denoising neural network is then used in conjunction with the sampling process based on conditional diffusion to realize protein skeleton structure prediction. The denoising neural network is trained according to the input noise-containing protein C α Carbon atom coordinates, predicting the added coordinate noise; During the sampling process, the noise-free and accurate protein C is obtained by gradually denoising the initialized random noise. α Carbon atomic coordinates, denoising neural network during training and sampling α Each of the three axis coordinates in the carbon atom coordinate x is calculated independently.

2. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The feature points are obtained by clustering the cryo-EM electron density map, specifically, filtering the cryo-EM electron density map by a given threshold and scaling the voxel size to Afterwards, the processed electron density map is clustered by the MeanShift++ clustering algorithm, and the class center points of each class are taken as feature points. Their coordinates are recorded in the form of point cloud, which is in the form of M×3 tensor, where M is the number of extracted feature points. After that, the point cloud is sampled at the farthest point and downsampled to N×3 point cloud tensor, where N is the length of the amino acid sequence.

3. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The amino acid types are predicted according to the point cloud structure using a trained DAQ model, specifically: the cryo-EM electron density map is input into the trained DAQ neural network, and the probability of the amino acid type at each voxel is obtained. Then, the coordinates of each feature point of the electron density map are matched to the amino acid type distribution to obtain the amino acid type distribution of each feature point, and the amino acid type with the maximum probability is taken as the predicted amino acid type.

4. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The one-dimensional and two-dimensional tensor features refer to: inputting the protein amino acid sequence into the AlphaFold2 neural network, and extracting the one-dimensional and two-dimensional features output by its network structure EvoFormer from the result, wherein the one-dimensional feature form is an N×384 tensor, and the two-dimensional feature form is an N×N×128 tensor, where N is the length of the amino acid sequence.

5. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The training process is as follows: randomly sample the time step t from the time step range [1, T], and obtain the noise coordinates at the time step t according to the noise addition formula: ∈ t is a standard Gaussian noise of N×3, and the noise coordinate is x t , time step t and the coordinates of the cryo-EM feature points that are independent of t, the types of amino acids at the cryo-EM feature points, and the one-dimensional and two-dimensional features of AlphaFold2 are used as conditions to jointly input the denoising neural network to obtain its output Calculate the estimated noise Assume the loss function is Its specific form is each C α The square of the coordinate difference of the carbon atom is the same as the RMSD, a commonly used protein structure distance calculation indicator. The parameter update of the denoising neural network is achieved by minimizing the loss function and completing the gradient back propagation, and training is achieved through repeated updates.

6. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The sampling process based on conditional diffusion is specifically as follows: Initialize protein C α The carbon atom coordinate is x T , where x T is a standard Gaussian noise of N×3. Starting from t=T, the noise coordinate is x t , time step t and the coordinates of the cryo-EM feature points that are independent of t, the types of amino acids at the cryo-EM feature points, and the one-dimensional and two-dimensional features of AlphaFold2 are used as conditions to jointly input the denoising neural network to obtain its output Calculate the estimated noise And calculate the noise coordinates of t-1 as in z is N×3 standard Gaussian noise, and then t=t-1, and the above process is repeated until t=0, and x0 is the protein C obtained by the final sampling. α Carbon atom coordinates.

7. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The denoising neural network based on conditional diffusion is based on the DDPM diffusion form, sets the maximum time step to T = 100, and adopts the cosine time table [α1, α2, ..., α N ], according to the input N×1 amino acid sequence, N×3 noise C α Carbon atom coordinate x t , constant time step t, N×3 cryo-EM feature point coordinates, N×1 cryo-EM feature point amino acid type, N×384 sequence corresponding to the AlphaFold2 one-dimensional features and N×N×128 sequence corresponding to the AlphaFold2 two-dimensional features, predicting the N×3 denoised C α Carbon atom coordinates Where: N is the length of the protein amino acid sequence, x i is the ith C α The three-dimensional coordinates of the carbon atom include the corresponding coordinates of the (x, y, z) axes.

8. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 1, characterized in that: The denoising neural network comprises: a feature embedding subnetwork, a feature fusion subnetwork and a structure construction subnetwork connected in series, wherein: the feature embedding subnetwork is based on the input N×1 amino acid sequence, N×3 noise C α Carbon atom coordinate x t , constant time step t, N×3 cryo-EM feature point coordinates, N×1 cryo-EM feature point amino acid type, N×384 AlphaFold2 one-dimensional features corresponding to the sequence, and N×N×128 AlphaFold2 two-dimensional features corresponding to the sequence, to obtain embedded N×384 one-dimensional features, N×N×128 two-dimensional features, N×N×128 cryo-EM features, and N×N×128 correlation features; the feature fusion subnetwork updates the input N×384 one-dimensional features, N×N×128 two-dimensional features, N×N×128 cryo-EM features, and N×N×128 correlation features through the modified AlphaFold2 EvoFormer structure to obtain four updated features with exactly the same input dimensions, and the final four updated features are obtained after repeated updating twice; the structure construction subnetwork fuses the N×384 one-dimensional features and N×N×128 two-dimensional features updated by the subnetwork according to the input features, as well as the noise C α Carbon atom coordinate x t , with x t Initialize the output coordinates for the coordinates, update the coordinates and features through the one-dimensional features and two-dimensional features, repeat the cycle four times, and output the final denoised C α Carbon atom coordinates 9. The protein structure prediction method based on cryo-electron microscopy and diffusion model according to claim 8, characterized in that: The N×1 amino acid sequence and time step t, for the amino acid sequence, the index number and the type of amino acid are encoded by the sine coding method to obtain two N×384 codes, and the time step t is sine-coded to obtain a 1×384 code and expanded to N×384, and the three codes are fused through a linear layer and a layer norm to obtain an amino acid sequence feature, and this feature is fused with the AlphaFold2 one-dimensional feature through a linear layer and a layer norm to obtain a one-dimensional feature; Input Noise C α Carbon atom coordinate x t Based on the spatial distance between two points and the index number relationship of each amino acid, two N×N×128 codes are obtained, and the initial two-dimensional features are obtained by fusion through linear layers and layer norms. This feature is fused with the AlphaFold2 two-dimensional features through linear layers and layer norms to obtain two-dimensional features; the coordinates of the cryo-EM feature points are input, encoded by the spatial distance between two points, and calculated through linear layers and layer norms to obtain cryo-EM features; the amino acid sequence and the amino acid type of the cryo-EM feature points are input, and the correlation characteristics are obtained by comparing the correlations between the two.

10. A protein structure prediction system based on cryo-electron microscopy and diffusion model for implementing any of the methods described in claims 1-9, characterized in that: include: A data processing unit, a model building unit and a result generating unit, wherein: the data processing unit extracts feature point cloud, downsamples and predicts the corresponding amino acid types of cryo-EM electron density, and extracts one-dimensional and two-dimensional features of the amino acid sequence using AlphaFold2; the model building unit completes formula reasoning, parameter setting, network architecture, and mathematical form determination for a denoising neural network based on conditional diffusion, and completes the model parameter update of the denoising neural network through a training process; the result generating unit processes the input data according to the data processing unit to obtain a standardized form, and then inputs the data into the trained denoising neural network for sampling to obtain the final result.

Images