Method for constructing protein structure from cryoelectron microscope density map by combining de novo modeling and structure prediction, computer device, readable storage medium and program product

Abstract

The invention belongs to the field of construction of a protein structure on a cryoelectron microscope density map, and relates to a method for constructing a protein structure from a cryoelectron microscope density map by combining de novo modeling and structure prediction, a computer device, a readable storage medium and a program product. According to the method, the atomic probability and the amino acid type are predicted through the three deep neural networks respectively, full-atom optimization is carried out, the output result is combined with the graph theory, optimization and geometric algorithms to jointly assist protein structure modeling, and protein structure information can be mined to the maximum extent. According to the method, de novo modeling can be carried out in the absence of full-length structure data of the protein monomer, integrated modeling can also be carried out under the condition of inputting the full-length structure data of the protein monomer, the application scene is wide, the structure of the protein compound can be automatically constructed in a cryoelectron microscope density map with middle and high resolution, and the method is suitable for popularization and application. And for a region with poor local resolution in a traditional method, the modeling precision can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of constructing protein structure from cryo-EM density map, and more particularly, relates to a method, computer device, readable storage medium and program product for constructing protein structure from cryo-EM density map by combining de novo modeling and structure prediction. BACKGROUND

[0002] Proteins play important roles in numerous biological processes, and therefore, determining the three-dimensional atomic structure of proteins is crucial for understanding molecular life processes. In recent years, cryo-electron microscopy (Cryo-EM) has made significant progress in resolving the structure of biological macromolecules, and current technology has been able to distinguish individual atoms at the level of partial structures, although it is still a time-consuming, labor-intensive and expertise-dependent work to model structures from cryo-EM density maps (referred to as density maps).

[0003] Some methods have also been proposed in the field in recent years to reduce the difficulty of modeling from density maps. Current cryo-EM structure modeling strategies can be roughly divided into two categories: de novo modeling and template fitting. For de novo modeling methods (such as phenix.map_to_model, MAINMAST, DeepM M, C-CNN, A2-Net, DeepTracer, Cryo2Struct and ModelAngelo), it is generally necessary to identify the positions of central carbon atoms from the density map and connect them into chains, and ModelAngelo, as the most advanced de novo modeling tool currently, has received widespread attention worldwide. Although these methods can directly establish atomic models from density maps without the aid of prior knowledge and other additional information, they are still prone to failure in areas of poor local quality of density maps. For template fitting methods, it is usually necessary to first fit the template structure into the density map as the initial position, and then perform multiple rounds of manual fine-tuning. With the progress of protein structure prediction algorithms such as AlphaFold2 (AF2), it has also become a common practice in the field to fit predicted structures into cryo-EM maps. Currently, there are some automated tools to assist in this process, such as gmfit, Situs, EMBuild, DEMO-EM, phenix.dock_and_rebuild, etc. If the predicted protein structure itself is accurate enough, these automated methods can construct reliable structures; but if the predicted structure is significantly different from the true structure, they often fail to model.

[0004] The two types of modeling methods are not independent because some methods attempt to fuse the dual information of density map and template structure. For example, the CR-I-TASSER method uses the central carbon atom positions predicted from the density map to select appropriate templates from the template library, but it can only be used for single-chain structure modeling and does not utilize the predicted structure information of AF2. Another method DeepMainmast combines the density map and the predicted structure of AF2, but the fitting of the structure to the density map is not sufficient, and it consumes a large amount of time in actual application.

[0005] With more and more medium-high resolution density maps being resolved, there is an urgent need in the field for an integrated automatic high-precision modeling method that combines the high-resolution regional information of the density map and the predicted structure information of AF2 to improve the modeling precision, especially in the regions with poor local resolution. SUMMARY

[0006] In view of the above defects or improvement needs of the prior art, the present application provides a method, a computer device, a readable storage medium and a program product for constructing a protein structure from a cryo-EM density map by combining de novo modeling and structure prediction, which aims to solve the technical problem of insufficient precision of constructing a protein structure in the prior art.

[0007] To achieve the above-mentioned purpose, according to one aspect of the present application, a method for constructing a protein structure from a cryo-EM density map by combining de novo modeling and structure prediction is provided, comprising the following steps:

[0008] After the cryo-EM density map is preprocessed into a grid and divided into overlapping density sub-blocks according to a preset step size and size, it is sent to an atomic probability prediction neural network M1 to predict the N, Cα and C atom probabilities of each sub-block, obtain the corresponding probability sub-graph, and reassemble all the probability sub-graphs into a total atomic probability graph; all the overlapping density sub-blocks are sent to an amino acid type prediction neural network M2 to predict an amino acid type graph;

[0009] For the Cα atom probability graph, local maximum points are found by a mean shift algorithm to obtain Cα atom positions; for each Cα atom, N and C atom positions are randomly initialized, and the N, Cα and C atoms are orthogonalized to obtain a local framework of the amino acid; an amino acid and a plurality of nearest neighbor amino acids thereof are randomly selected and sent to a all-atom optimization neural network M3 to predict the real N, Cα and C atom positions and the side chain torsion angles of 20 kinds of amino acids, and the process is repeated until all the amino acids are optimized;

[0010] For all optimized amino acids, randomly select a starting amino acid i, find the amino acid j from the remaining N atoms that is closest to the C atom of amino acid i and within a predetermined range, repeat this operation to find the next amino acid, until no next amino acid can be found, extract amino acids i, j,... as a fragment; repeat this operation until all amino acids are accessed, and obtain all fragments;

[0011] In each fragment, for each amino acid, find the grid point closest to its Cα atom in the amino acid type prediction neural network M2 prediction amino acid type map, obtain the 20 amino acid type probabilities on the grid point and convert them into amino acid scores, sequence align each amino acid score of the fragment with the target sequence to find the best matching subsequence in the target sequence, and take this subsequence as the sequence of the fragment.

[0012] According to the sequence of the fragment, extract the torsion angle of the corresponding amino acid from the side chain torsion angle predicted by the all-atom optimization neural network M3 and convert it into a rotation matrix, and obtain the positions of all atoms of the amino acid by rotating and translating the predefined standard template amino acid rigid group, thereby obtaining the de novo modeled protein structure.

[0013] Further, it also includes a predicted protein monomer full-length structure from the target sequence, and:

[0014] Divide the predicted protein monomer full-length structure into monomer fragments, align the monomer fragments with the de novo modeled protein structure to obtain amino acid correspondence, transformation matrix and RMSD, if the RMSD is less than a predetermined threshold, find the missing part in the de novo modeled protein structure in the amino acid correspondence, and apply the transformation matrix to the monomer fragment, fill the corresponding part of the monomer fragment to the missing part of the de novo modeled structure, and obtain the repaired de novo modeled structure.

[0015] Add the N, Cα, C atom probability maps predicted by the neural network M1 to obtain the main chain probability map, and fit the monomer fragment to the main chain probability map to obtain the fitted structure.

[0016] Construct a structure library with the repaired de novo modeled structure and the fitted structure, find all monomer fragments that have no conflicts and meet the matching requirements from the structure library, and input them into the all-atom optimization neural network M3 for structure optimization, and finally output the optimized protein structure.

[0017] Further, the atom probability prediction neural network M1 includes three encoders, one transition layer and three decoders, and there is a residual connection between the encoder and the decoder; the channel number output by the atom probability prediction neural network M1 is 3, representing N, Cα, and C atoms.

[0018] Furthermore, the amino acid type prediction neural network M2 contains three encoders, one transition layer, and three decoders, with residual connections between the encoders and decoders; the output of the amino acid type prediction neural network M2 has 20 channels, representing 20 amino acids.

[0019] Furthermore, the all-atom optimized neural network M3 includes a feature initialization module, a three-track attention module, and a twist angle prediction module;

[0020] In the feature initialization module, a density sub-block of a predetermined size centered on each amino acid is extracted from the local framework and density map of amino acids as a one-dimensional feature, and a density sub-block of a predetermined size between any two amino acids is extracted as a two-dimensional feature. After being processed by the convolutional layer, they are used as the encoded one-dimensional and two-dimensional features; among them, the local framework of amino acids is a three-dimensional feature.

[0021] In the three-track attention module, one-dimensional, two-dimensional, and three-dimensional features interact through an attention mechanism. Specifically, two-dimensional features are added as attention biases to the self-attention update mechanism of one-dimensional features; three-dimensional features are transformed into two-dimensional distance features and then added as attention biases to the axial attention update mechanism of two-dimensional features. One-dimensional and three-dimensional features interact through an invariant self-attention update mechanism.

[0022] In the torsion angle prediction module, the final one-dimensional feature is converted into the predicted amino acid torsion angle through a linear layer.

[0023] Furthermore, the training process of the all-atom optimized neural network M3 is as follows:

[0024] ① Select a PDB structure model from the training set; ② Randomly select an amino acid from the PDB structure model; ③ Extract the amino acid and its k spatial nearest neighbor residues; ④ For each extracted amino acid, add random noise to its N, Cα, and C atoms; ⑤ Feed the cryo-electron microscopy density map and the amino acid with added random noise into the all-atom optimization neural network M3, so that the neural network can predict the true local framework and side chain torsion angle of these amino acids.

[0025] According to another aspect of the invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the method as described in any of the preceding claims.

[0026] According to another aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method as described in any of the preceding claims.

[0027] According to another aspect of the invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method as described in any of the preceding claims.

[0028] In summary, the technical solutions conceived in this invention, compared with the prior art, can achieve the following beneficial effects:

[0029] 1. Modeling protein structures from cryo-electron microscopy density maps is a complex process. This invention uses three deep neural networks to predict atom probabilities, amino acid types, and perform full-atom optimization, respectively. The outputs are then combined with graph theory, optimization, and geometric algorithms to jointly assist in protein structure modeling, maximizing the extraction of protein structural information. It is particularly suitable for accurate analytical modeling of medium- to high-resolution density maps, significantly improving modeling accuracy in areas where traditional methods have poor local resolution.

[0030] 2. This invention can perform de novo modeling when full-length protein monomer structure data is lacking, and can also perform integrated modeling when full-length protein monomer structure data is input, making it applicable to a wide range of scenarios.

[0031] 3. This invention utilizes two deep neural networks to extract the type maps of 20 amino acids in proteins and the probability maps of N, Cα, and C atoms in proteins from the original density map, and then obtains the Cα atom position and main chain probability map from them. This fully takes into account the different atomic positions and structures in different amino acids, making the prediction results more accurate.

[0032] 4. This invention proposes for the first time a neural network based on three-track attention. Through an attention update mechanism that links three different tracks (one-dimensional, two-dimensional, and three-dimensional), it fully mines protein structural information. It can directly predict the main chain atomic positions (N, Cα, C, O) and side chain atomic positions (Cβ, Cγ, etc.) of amino acids by starting only from the Cα atom and combining density map information, which greatly improves the accuracy of the modeling results.

[0033] 5. This invention utilizes the spatial geometric relationships that amino acids must satisfy to divide segments, and combines amino acid type prediction and sequence alignment to allocate sequences of segments, making it easier to discover missing segments. Based on this, this invention uses inter-segment structural alignment for structural repair; utilizes semi-flexible fitting to fit the structure to the main chain probability graph; and uses a graph theory-based constraint optimization algorithm to find non-conflicting segment structures from all segments and output the structure, thereby further ensuring the completeness and accuracy of the modeling results.

[0034] 6. The EMPert method of this invention has a resolution better than 177 resolutions. The results were evaluated on experimental density plots and compared with previous state-of-the-art methods: DeepMainmast, ModelAngelo, Phenix, and AlphaFold3, as shown in Tables 3 and 4. The EMPert method of this invention is comprehensively superior to other methods. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the workflow of the EMPert method according to a preferred embodiment of the present invention.

[0036] Figure 2 These are schematic diagrams illustrating two ways of utilizing AF2 structural information in the EMPert method of a preferred embodiment of the present invention.

[0037] Figure 3 This is a fully atomic optimization network of the EMPert method in a preferred embodiment of the present invention, which includes three main modules: 1-feature initialization module, 2-three-track attention module, and 3-torsion angle prediction module.

[0038] Figure 4 This is a schematic diagram of the core "three-track attention" module in the all-atom optimization network of a preferred embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] Given the significant local resolution heterogeneity of density maps and the inconsistent quality of AF2-predicted structures, this invention aims to leverage the strengths of both methods by fully integrating density map information and AF2-predicted fragment structure information for automated protein structure modeling from density maps, thereby improving modeling accuracy and reducing manual intervention. To this end, this invention proposes a method for constructing protein structures that can be used for both de novo modeling and integrative modeling: EMProt, a method combining de novo modeling and structure prediction to construct protein structures from cryo-electron microscopy density maps. EMProt takes as input a density map, a target protein sequence, and optional protein monomer structures predicted from the sequence (obtainable using the traditional prediction methods mentioned in the background section), and outputs the protein complex structure. Based on this method, this invention can automatically construct the structure of a protein complex from medium- to high-resolution cryo-electron microscopy density maps.

[0041] A preferred embodiment of the present invention (EMProt) is as follows:

[0042] The inputs to EMProt are: ① cryo-electron microscopy density map (provided in .mrc or .map format), ② target sequence (provided in .fasta format), and ③ protein monomer structure predicted from the target sequence (such as prediction results from AlphaFold2, AlphaFold3, RosettaFold, ESMFold, etc.) (where ③ is an optional input).

[0043] If no input ③ is provided, meaning there is no reference protein monomer structure, then EMPot will perform de novo modeling:

[0044] The cryo-electron microscopy density map was preprocessed to a grid size of [size missing]. Then, with 1 The step size is divided into segments of size 48×48×48. The overlapping density sub-blocks are fed into neural network M1 to predict the N, Cα, and C atom probabilities of each sub-block. The predicted probability sub-graphs are then reassembled into a total atom probability graph (averaging the values ​​for overlapping areas). Similarly, these sub-blocks are fed into neural network M2 to predict the amino acid type graph.

[0045] For the probability map of Cα atoms, the local maxima are found using the mean-shift algorithm to obtain the positions of Cα atoms.

[0046] For each Cα atom, the positions of N and C atoms are randomly initialized. The N, Cα, and C atoms are then orthogonalized using the Gram-Schmidt method to obtain the local frame of the amino acid. A single amino acid and its k=128 nearest-neighbor amino acids are randomly selected and fed into a neural network M3 to predict the actual positions of N, Cα, and C atoms and the side chain torsion angles of 20 amino acids. This process is repeated until all amino acids have been optimized. The local frame is a rotation and translation matrix in a local coordinate system constructed for each amino acid N, Cα, and C. Using orthogonalization methods (such as the Gram-Schmidt method), given the N, Cα, and C atoms of an amino acid, the corresponding local frame can be obtained. Applying the local frame to a rigid residue group (some constants) yields the positions of N, Cα, and C. Therefore, the local frame is equivalent to the positions of the N, Cα, and C atoms.

[0047] For all optimized amino acids, arbitrarily select a starting amino acid i, and find the one from the remaining amino acids whose N atom is closest to the C atom of amino acid i and whose distance is within a certain range. For each amino acid i, j, ..., repeat this process until no more amino acids are found. Extract each amino acid i, j, ... as a fragment. Repeat this process until all amino acids have been accessed, thus obtaining all the fragments.

[0048] In each fragment, for each amino acid, the grid point closest to its Cα atom is found in the amino acid type map predicted by the neural network M2. The probabilities of 20 amino acid types on the grid point are obtained, and the negative logarithm of the amino acid probabilities is converted into an amino acid score. The score of each amino acid in the fragment is compared with the input target sequence using the HMMER algorithm to find the best matching subsequence in the input sequence, and this subsequence is used as the sequence of the fragment.

[0049] According to the sequence of the fragment, the torsion angle of the corresponding amino acid is extracted from the side chain torsion angle predicted by the neural network M3. The torsion angle is converted into a rotation matrix. By rotating and translating the predefined standard template amino acid rigid group, the positions of all atoms of the amino acid are obtained, which is the de novo modeling structure of EMPert.

[0050] If input ③ is provided, the present invention can be further integrated and modeled:

[0051] First, the predicted full-length protein monomer structure is divided into fragments using the UniDoc program. These fragments are then compared with the de novo modeled structure using TM-align to obtain amino acid correspondences, transformation matrices, and RMSD. The missing parts of the de novo modeling structure are found in the correspondence, and the transformation matrix output by TM-align is applied to the fragment structure. Parts of the fragment structure are then used to fill in the missing parts of the de novo modeling structure, resulting in the repaired de novo modeling structure (e.g., ...). Figure 2 (As shown in the "Structural Repair" section).

[0052] Second, the probability maps of N, Cα, and C atoms predicted by neural network M1 are summed to obtain the main chain probability map. The segments segmented above are then matched to the main chain probability map using a semi-flexible fitting accelerated by FFT (e.g., ...). Figure 2 As shown in the "Structure Fitting" section).

[0053] The repaired and fitted structures are used to build a structure library. From the structure library, a graph optimization algorithm is used to find all segments that are non-conflicting and well-matched. All segments are then fed into a neural network M3 for structure optimization, and finally the optimized structure is output.

[0054] Preferably, the EMPert method of the present invention uses three deep neural networks. The following is the neural network training process:

[0055] ① Training the atomic probability prediction neural network M1 for EMPert:

[0056] The M1 network is actually a neural network prediction model that takes overlapping density sub-blocks of cryo-electron microscopy density maps as input and the N, Cα, and C atom probabilities of each sub-block as output. It can be trained using the aforementioned input and output data according to conventional training methods. To further improve the model's prediction accuracy, this invention provides the following more preferred training method:

[0057] The neural network architecture is Swin-Conv-UNet, which consists of three encoders, one transition layer, and three decoders, with residual connections between the encoders and decoders. Each encoder and decoder contains a parallel SwinT (Swin-Transformer) module and an RConv (Residual-Conversion) module, sandwiched between two 1×1 convolutional layers. The SwinT module uses a window size of 3. The network input is a 48×48×48... The density block outputs the same size as the input, and the number of output channels n = 3 (representing N, Cα, and C atoms).

[0058] For the experimental density map: first, trilinear interpolation is used to normalize the grid size of the density map to 1. The values ​​in the experimental density map were cropped to the range of ≥0.0 to the 99.999th quantile of the density value and normalized to [0,1]. All Cα atoms were extracted from the PDB (Protein Data Bank) structure, and the probability value of the atom at grid point x in the experimental density map was calculated according to a Gaussian distribution to serve as the tag map. The tag map preparation process for N and C atoms was the same as for Cα. Subsequently, the experimental density map and tag map were divided into 60×60×60... Step size is The density blocks. During training, from 60×60×60 48×48×48 blocks were randomly cut from the density blocks. The sub-blocks are randomly rotated by 90°. Training uses smoothing L1 loss and structural similarity loss as loss functions, and the Adam optimizer is used to update parameters. The initial learning rate is set to 5e-5, and the weight decay coefficient is 1e-4. During training, if the average training loss does not decrease for 4 consecutive epochs, the learning rate is halved. The minimum learning rate is limited to 5e-6. Each network is trained using two 40GB NVIDIA A100 GPUs with a batch size of 96. Finally, the model with the lowest loss on the validation set is selected as the final model.

[0059] ② Training the amino acid type prediction neural network M2 for EMProt:

[0060] The M2 network is essentially a neural network prediction model that takes overlapping density sub-blocks of cryo-electron microscopy density maps as input and the amino acid type map corresponding to each sub-block as output. It can be trained using the aforementioned input and output data according to conventional training methods. To further improve the model's prediction accuracy, this invention provides the following more preferred training method:

[0061] The neural network architecture of the amino acid type prediction neural network M2 is consistent with that of M1, the only difference being the number of output channels n = 20 (representing 20 amino acids). The label map of the amino acid type network is generated as follows: all heavy atoms are extracted from the PDB structure, and each voxel in the experimental density map is labeled according to its... The amino acid to which the nearest heavy atom within the radius belongs (represented by an integer between 0 and 19), if in The voxel containing no amino acids is labeled as background (represented by the integer 20). Training uses cross-entropy as the loss function, ignoring predictions for regions labeled 20, and other hyperparameters remain consistent with those used in M1 training.

[0062] ③ Training the all-atom optimized neural network M3 for EMPert:

[0063] The M3 network is actually a neural network prediction model that takes multiple neighboring amino acid structures as input and outputs the positions of N, Cα, and C atoms and the side chain torsion angles of 20 amino acids. It can be trained using the aforementioned input and output data according to conventional training methods. To further improve the model's prediction accuracy, this invention provides the following preferred training method:

[0064] The all-atom optimized neural network M3 includes a feature initialization module, a three-track attention module, and a torsion angle prediction module.

[0065] In the feature initialization module, the input amino acid local framework (3D feature) and density map are used to extract features centered on each amino acid with a size of 23×23×23. The density sub-blocks are used as one-dimensional features, and the 17×3×3 intervals between any two amino acids are considered. The density sub-blocks are used as two-dimensional features, which are then fed into a convolutional layer for processing and used as encoded one-dimensional and two-dimensional features.

[0066] In the three-track attention module, one-dimensional, two-dimensional, and three-dimensional features interact through an attention mechanism. Specifically, two-dimensional features are added as attention biases to the self-attention update mechanism of one-dimensional features; three-dimensional features are transformed into two-dimensional distance features and then added as attention biases to the axial attention update mechanism of two-dimensional features. One-dimensional and three-dimensional features interact through an invariant self-attention update mechanism.

[0067] In the torsion angle prediction module, the final one-dimensional feature is converted into the predicted amino acid torsion angle through a linear layer.

[0068] The training process includes the following steps: ① Select a PDB structure model from the training set; ② Randomly select one amino acid from it; ③ Extract the amino acid and its k spatial nearest neighbors (including itself); ④ Add random noise to the N, Cα, and C atoms of each amino acid; ⑤ Feed the density map and the noisy amino acid into the neural network so that the neural network can predict the true position of these amino acids.

[0069] The network training loss function includes three types: ①RMSD loss; ②torsion angle loss; ③structural conflict loss, and the total loss is the sum of the three. During training, the AdamW optimizer is used to update the network parameters. The initial learning rate is 1e-4, and the learning rate is reduced to 0.95 every 1000 steps. The effective batch size is 8, and k is set to 128. Training uses two Nvidia A100 GPUs with 40GB of video memory each.

[0070] The training sets for neural networks M1, M2, and M3 were obtained from the public databases PDB and EMDB (Electron Microscopy Data Bank). After searching, downloading, and filtering in the databases, a total of 803 density map-structure pairs were obtained. 90% of these were used as the training set, and the remaining 10% were used as the validation set to monitor the network training process.

[0071] Table 1 - Training Set

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] Table 2 - Validation Set

[0079]

[0080] Table 3 - Comparison of this invention with DeepMainmast, ModelAngelo, and Phenix in terms of coverage, matching rate, and recall (higher is better).

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] Table 4 - Comparison of this invention with DeepMainmast, Phenix, and AlphaFold3 in terms of TM-score and RMSD (higher TM-score is better, lower RMSD is better).

[0087]

[0088]

[0089]

[0090]

[0091]

[0092] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing protein structures from cryo-electron microscopy density maps by combining de novo modeling and structure prediction, characterized in that, Includes the following steps: The cryo-electron microscopy density map is preprocessed into grid points and then divided into overlapping density sub-blocks according to a preset step size and size. The sub-blocks are then fed into an atomic probability prediction neural network M1 to predict the N, Cα, and C atomic probabilities of each sub-block, resulting in a corresponding probability sub-map. All probability sub-maps are then reassembled into a total atomic probability map. All overlapping density sub-blocks are fed into an amino acid type prediction neural network M2 to predict the amino acid type map. For the Cα atom probability map, the local maxima are found using the mean-shift algorithm to obtain the Cα atom position. For each Cα atom, the N and C atom positions are randomly initialized, and the N, Cα, and C atoms are orthogonalized to obtain the local framework of the amino acid. An amino acid and its multiple nearest neighbor amino acids are randomly selected and fed into the all-atom optimization neural network M3 to predict the true N, Cα, and C atom positions and the side chain torsion angles of 20 amino acids. This process is repeated until all amino acids have been optimized. For all optimized amino acids, arbitrarily select a starting amino acid i. From the N atoms of the remaining amino acids, find the amino acid j that is closest to the C atom of amino acid i and within a preset range. Repeat this operation to find the next amino acid until no more amino acids can be found. Extract amino acids i, j, ... as a fragment. Repeat this operation until all amino acids have been visited to obtain all fragments. In each fragment, for each amino acid, the grid point closest to its Cα atom is found in the amino acid type prediction map predicted by the amino acid type prediction neural network M2. The probabilities of 20 amino acid types on the grid point are obtained and converted into amino acid scores. The scores of each amino acid in the fragment are compared with the target sequence to find the best matching subsequence in the target sequence. This subsequence is used as the sequence of the fragment. Based on the sequence of the fragment, the torsion angles of the corresponding amino acids are extracted from the side chain torsion angles predicted by the all-atom optimization neural network M3 and converted into rotation matrices. By rotating and translating the predefined standard template amino acid rigid group, the positions of all atoms of the amino acid are obtained, thus obtaining the de novo modeled protein structure.

2. The method for constructing protein structures from cryo-electron microscopy density maps by combining de novo modeling and structure prediction as described in claim 1, characterized in that, It also includes the full-length structure of the protein monomer predicted from the target sequence, and: The predicted full-length structure of the protein monomer is divided into monomer fragments. The monomer fragments are compared with the de novo modeled protein structure to obtain the amino acid correspondence, transformation matrix and RMSD. If the RMSD is less than a preset threshold, the missing part of the de novo modeled protein structure is found in the amino acid correspondence. The transformation matrix is ​​applied to the monomer fragments and the part of the monomer fragment corresponding to the missing structure is filled into the missing part of the de novo modeled structure to obtain the repaired de novo modeled structure. The probability maps of N, Cα, and C atoms predicted by the neural network M1 are added together to obtain the main chain probability map. The single-unit fragment is then fitted to the main chain probability map to obtain the fitted structure. The repaired de novo modeling structure and fitted structure are used to build a structure library. All monomer fragments that do not conflict and meet the matching requirements are found in the structure library and fed into the all-atom optimization neural network M3 for structure optimization. Finally, the optimized protein structure is output.

3. A method for constructing protein structures from cryo-electron microscopy density maps by combining de novo modeling and structure prediction as described in claim 1 or 2, characterized in that, The atomic probability prediction neural network M1 contains three encoders, one transition layer and three decoders, with residual connections between the encoders and decoders; the atomic probability prediction neural network M1 outputs 3 channels, representing three types of atoms: N, Cα and C.

4. A method for constructing protein structures from cryo-electron microscopy density maps by combining de novo modeling and structure prediction as described in claim 1 or 2, characterized in that, The amino acid type prediction neural network M2 contains three encoders, one transition layer, and three decoders, with residual connections between the encoders and decoders. The output of the amino acid type prediction neural network M2 has 20 channels, representing 20 amino acids.

5. A method for constructing protein structures from cryo-electron microscopy density maps by combining de novo modeling and structure prediction as described in claim 1 or 2, characterized in that, The all-atom optimized neural network M3 includes a feature initialization module, a three-track attention module, and a torsion angle prediction module; In the feature initialization module, a density sub-block of a predetermined size centered on each amino acid is extracted from the local framework and density map of the amino acid as a one-dimensional feature, and a density sub-block of a predetermined size between any two amino acids is extracted as a two-dimensional feature. These are then processed by a convolutional layer to obtain the encoded one-dimensional and two-dimensional features. The local framework of amino acids is a three-dimensional feature; In the three-track attention module, one-dimensional, two-dimensional, and three-dimensional features interact through an attention mechanism, where two-dimensional features are added as attention biases to the self-attention update mechanism of one-dimensional features. After the three-dimensional features are transformed into two-dimensional distance features, they are added as attention biases to the axial attention update mechanism of the two-dimensional features. The one-dimensional and three-dimensional features interact through an invariant self-attention update mechanism. In the torsion angle prediction module, the final one-dimensional feature is converted into the predicted amino acid torsion angle through a linear layer.

6. The method for constructing protein structures from cryo-electron microscopy density maps by combining de novo modeling and structure prediction as described in claim 5, characterized in that, The training process of the all-atom optimized neural network M3 is as follows: ① Select a PDB structure model from the training set; ② Randomly select an amino acid from the PDB structure model; ③ Extract the amino acid and its k spatial nearest neighbor residues; ④ For each extracted amino acid, add random noise to its N, Cα, and C atoms; ⑤ Feed the cryo-electron microscopy density map and the amino acid with added random noise into the all-atom optimization neural network M3, so that the neural network can predict the true local framework and side chain torsion angle of these amino acids.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 6.

Images