A protein model quality assessment method based on self-attention graph neural network

Abstract

A method for evaluating the quality of protein structure models based on deep neural networks. A self-attention graph neural network is designed to predict the quality of protein models. Features are extracted based on the input model: heavy atom coding, local space coordinates, Gaussian distance spectrum and hydrogen bond coding in the protein model. The similarity score between each pair of amino acids in the protein model and the natural protein and the overall similarity score between the two are predicted, thereby determining the quality of the protein model. The present invention sets the evaluation score in [0,1], and the higher the score, the closer it is to the true structure. Using huge data to learn structural similarity and estimating the quality of each residue based only on the predicted model without relying on the true structure is of great significance for further model refinement and selection.

Description

Technical Field

[0001] The technology of the present invention relates to the fields of artificial intelligence, computer application and bioinformatics, and in particular to a protein model quality assessment method based on neural network prediction. Background Art

[0002] Protein is the material basis of life, the basic organic matter that constitutes cells, and the main bearer of life activities. Without protein, there is no life. Therefore, protein is the top priority in the field of life science research. Proteins must form a certain structure to perform corresponding functions. With the development of sequencing technology, the determination of protein three-dimensional structure through time-consuming, laborious, and expensive experimental methods (X-ray crystal diffraction, nuclear magnetic resonance, and cryo-electron microscopy) cannot meet the needs of high-throughput acquisition of protein structure. Therefore, predicting the three-dimensional structure of a protein from its amino acid sequence has become one of the most challenging and significant research directions at present.

[0003] The Critical Assessment of Protein Structure Competition (CASP) is held every two years and is a worldwide competition for protein structure prediction techniques and methods. Protein model quality assessment is one of the important events and a key component of protein structure prediction, which helps in further model refinement and selection. Usually in protein structure prediction, the distance between the experimentally measured structure and the predicted structure corresponding to each residue is calculated, and the root mean square deviation (RMSD) is calculated to evaluate the accuracy of the predicted structure. However, the RMSD indicator is very sensitive to the deviation of local residues. If the matching degree of some residues is poor, it may lead to a significant decrease in the overall score. Therefore, RMSD cannot accurately reflect the accuracy of the predicted protein model.

[0004] In recent years, a local model quality assessment index lDDT (Local Distance Difference Test) has been proposed. This method calculates the local distance difference in protein structure. The residue pair distance within 100 is compared with the real structure and the predicted model at different distance thresholds. The prediction quality of each residue is evaluated by the distance difference under the condition of lddt. The advantage of lDDT is that it can accurately evaluate the accuracy of local structure. Evaluating the quality of protein structure model is a key component of protein structure prediction. Model quality assessment helps to verify and evaluate the predicted protein model at multiple stages of structure prediction. Therefore, how to design a model evaluation method to estimate the lDDT of each residue based on the predicted model without relying on the real structure is of great significance for further model refinement and selection. Summary of the invention

[0005] In order to overcome the deficiency of existing protein structure prediction methods that the quality of protein models cannot be accurately evaluated, the present invention provides a protein model quality evaluation method based on self-attention graph neural network, and designs a new protein model quality evaluation method, which extracts spatial information, sequence information, and interaction forces between residues from the input protein model as the input of the network, and predicts the quality of the protein model with high precision.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A protein model quality assessment method based on a self-attention graph neural network, the method comprising the following steps:

[0008] 1) Dataset construction: First, protein sequences with a length of 50-300 residues were screened from the PDB database using the PISCES server, and redundancy was removed with a maximum sequence similarity of 35%. Then, a model dataset is constructed by folding and generating a global distance test total score GDT-TS between [50, 100] using comparative modeling, native structure perturbation, and deep learning-based modeling methods, and the dataset is divided into a training set, a validation set, and a test set;

[0009] 2) Constructing network features: encoding representation of the heavy atoms of each residue in the protein sequence, constructing the local space coordinate system of three adjacent heavy atoms, the distance between adjacent heavy atoms in Euclidean space and the hydrogen bond force constraints. The process is as follows:

[0010] 2.1) Proteins contain 20 kinds of amino acids, each of which has heavy atoms with different properties, totaling 167 heavy atoms. Each heavy atom is uniquely encoded;

[0011] 2.2) Construct a local spatial coordinate system on each heavy atom of the protein. The construction method is: for each heavy atom in the sequence, find two adjacent heavy atoms in space with itself as the center to form a plane and generate a vector in the plane Calculate the vector perpendicular to the plane Then calculate the perpendicular vector again With vector Vector

[0012]

[0013]

[0014]

[0015] Among them, B is the central atom, A and C are two adjacent atoms in space, and the spatial coordinates of A and B form Vectors are constructed in the same way vector, is the unit vector of the plane where A, B, and C lie. is the unit vector perpendicular to the plane, is perpendicular to the vector With vector Unit vector;

[0016] 2.3) In the protein structure, each heavy atom uses a distance constraint Dist to represent its proximity to other heavy atoms. To this end, the Gaussian function of the distance constraint at different thresholds is calculated to form a multi-interval Gaussian distribution;

[0017]

[0018] Among them, Dist ij is the distance between heavy atom i and heavy atom j, N atoms are all the heavy atoms in the protein, N Datoms is the nearest D heavy atoms, μ k With σ k are the Gaussian parameters at different thresholds [0, 0.4, 0.8...15.2], which are set the same by default, D ij is the Gaussian distribution value of heavy atom i and heavy atom j at different thresholds;

[0019] 2.4) If there is a hydrogen bond between two heavy atoms in the protein structure, it is expressed as 1, otherwise it is 0;

[0020] 3) Build a self-attention graph neural network. The process is as follows:

[0021] 3.1) Overall framework of self-attention graph neural network: Feature input is extracted from proteins and passed through the embedding layer, followed by 4 layers of multi-head attention mechanism layers, and then through the residue-level pooling layer and protein-level pooling layer respectively. These two pooling layers are then connected to the attention mechanism layer and then fully connected to predict the local and global lDDT scores of the protein model;

[0022] 3.2) In the graph neural network, the embedding layer takes as input the one-hot encoding matrix of the heavy atoms in the protein sequence. After passing through the emb layer, it then passes through the fully connected linear layer to form the embedding layer.

[0023] 3.3) In the multi-head attention mechanism layer of the graph neural network, the number of heads in each layer is set to 8, 4, 2, and 1. The new matrix obtained by combining the embedded atomic features generated by the embedding layer with other features including the local spatial coordinate system of heavy atoms, the Gaussian distribution of distances, and the force constraints of hydrogen bonds constitutes the nodes of the graph neural network and is input into the multi-head attention mechanism layer. The node update method of the multi-head attention layer is as follows;

[0024]

[0025]

[0026]

[0027] In the formula, α i,j represents the attention coefficient, Represents the characteristics of the atom, u (i,j) represents the characteristics between atoms, σ represents the Sigmoid function, g represents the Relu function, represents the union of feature dimensions, ⊙ represents the multiplication of corresponding elements, · represents the multiplication of a number and a matrix, and represents the rotation matrix of the linear layer, and is its corresponding bias, N i represents all neighboring atoms of atom i, and j is a neighboring atom;

[0028] 3.4) Residue-level pooling layer and protein-level pooling layer. The residue-level pooling layer adds up the feature channels of the heavy atoms on each residue to output the residue-level features, while the whole sequence pooling layer adds up the feature channels of the entire protein to form the features of the entire sequence. Finally, the local lDDT score and the global score are predicted through the fully connected layer respectively.

[0029] 3.5) Glorot was used for initialization, the Adam optimizer was used as the gradient descent optimizer, the learning rate was set to 0.001, and the root mean square error MSELoss function was used as the Loss method;

[0030] 4) Label the data and calculate the true global lDDT score and the true local lDDT score: Calculate the lDDT score for each residue of the model protein in the dataset, and sum the lDDT scores of all residues and divide them by the number of residues to form the global lDDT score. β The standard calculation process of -lDDT score is as follows:

[0031] per_residue_lDDT=0.25×(p1×p2×p3×p4) / p0,

[0032]

[0033] p1 is the distance between the ith residue and all other residues that is less than And all residue pairs C β Distance error is less than The probability score of p2, p3, and p4 are p0 means that the distance between the ith residue and all other residues is less than The probability of N is the total number of residues. res is all the residues, per_residue_lDDT is the score for each residue in the protein, and global_lDDT is the global score of the true protein model. The quality of the protein model is determined according to the global score and the local residue score.

[0034] Furthermore, in step 1), 90% of the protein models in the data set are used as training sets, 5% of the models are used as validation sets, and 5% of the models are used as test sets.

[0035] The beneficial effects of the present invention are mainly manifested in that a novel network architecture is used to evaluate the quality of protein models, spatial information, sequence information, and the interaction force between atoms are extracted from the input model, which are used as the input of the network, and the lDDT score is directly predicted through the self-attention graph neural network. Without relying on the real structure, the lDDT of each residue is estimated with high precision and high throughput based only on the predicted model, which is of great significance for further model refinement and selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is an overall flow chart of a protein model quality assessment method based on self-attention graph neural network.

[0037] Figure 2 It is a protein model quality assessment method based on self-attention graph neural network to predict the local residue LDDT distribution map of protein 1evya model.

[0038] Figure 3 It is a protein model quality assessment method based on self-attention graph neural network to obtain the global LDDT score for protein 1evya model structure prediction. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings.

[0040] Reference Figures 1 to 3 , a protein model quality assessment method based on self-attention graph neural network, comprising the following steps:

[0041] 1) Dataset construction: First, protein sequences with a length of 50-300 residues were screened in the PDB (https: / / www.rcsb.org / ) database using the PISCES (http: / / dunbrack.fccc.edu / pisces) server, and redundancy was removed with a maximum sequence similarity of 35%. Then, comparative modeling, native structure perturbation and deep learning-based modeling methods were used to fold and generate models with a global distance test total score (GDT-TS) between [50,100] to form a dataset. 90% of the protein models in the dataset were used as training sets, 5% of the models were used as validation sets, and 5% of the models were used as test sets;

[0042] 2) Constructing network features: encoding representation of the heavy atoms of each residue in the protein sequence, constructing the local space coordinate system of three adjacent heavy atoms, the distance between adjacent heavy atoms in Euclidean space and the hydrogen bond force constraints. The process is as follows:

[0043] 2.1) Proteins contain 20 kinds of amino acids, each of which has different properties of heavy atoms (atoms other than hydrogen atoms). There are a total of 167 heavy atoms. Each heavy atom is uniquely encoded. For example, ALA (α-linolenic acid) amino acids contain C, N, O, C α ,C β , among these five heavy atoms, the C heavy atom is encoded as [1,0,0...,0], and the encoding length is 167; similarly, other atoms can be represented similarly. Therefore, a 167×167 matrix can be obtained, which corresponds to each heavy atom one by one;

[0044] 2.2) Construct a local spatial coordinate system on each heavy atom of the protein. The construction method is: for each heavy atom in the sequence, find two adjacent heavy atoms in space with itself as the center to form a plane and generate a vector in the plane Calculate the vector perpendicular to the plane Then calculate the perpendicular vector again With vector Vector

[0045]

[0046]

[0047]

[0048] Among them, B is the central atom, A and C are two adjacent atoms in space, and the spatial coordinates of A and B form Vectors are constructed in the same way vector, is the unit vector of the plane where A, B, and C lie. is the unit vector perpendicular to the plane, is perpendicular to the vector With vector Unit vector;

[0049] 2.3) In the protein structure, each heavy atom uses a distance constraint Dist to represent its proximity to other heavy atoms. To this end, the Gaussian function of the distance constraint at different thresholds is calculated to form a multi-interval Gaussian distribution;

[0050]

[0051] Among them, Dist ij is the distance between heavy atom i and heavy atom j, N atoms are all the heavy atoms in the protein, N Datoms is the nearest D heavy atoms, μ k With σ k are the Gaussian parameters at different thresholds [0, 0.4, 0.8...15.2], which are set the same by default, D ij is the Gaussian distribution value of heavy atom i and heavy atom j at different thresholds;

[0052] 2.4) If there is a hydrogen bond between two heavy atoms in the protein structure, it is expressed as 1, otherwise it is 0;

[0053] 3) Build a self-attention graph neural network. The process is as follows:

[0054] 3.1) Overall framework of self-attention graph neural network: Feature input is extracted from proteins and passed through the embedding layer, followed by 4 layers of multi-head attention mechanism layers, and then through the residue-level pooling layer and protein-level pooling layer respectively. These two pooling layers are then connected to the attention mechanism layer and then fully connected to predict the local and global lDDT scores of the protein model;

[0055] 3.2) In the graph neural network, the embedding layer takes as input the one-hot encoding matrix of the heavy atoms in the protein sequence (feature dimension 167), passes through the emb layer (built-in in the Porch environment), and then passes through the fully connected linear layer (feature dimension 64) to form the embedding layer;

[0056] 3.3) In the multi-head attention mechanism layer of the graph neural network, the number of heads in each layer is set to 8, 4, 2, and 1. The new matrix (feature dimension 171) obtained by combining the embedded atomic features (feature dimension 64) generated by the embedding layer with other features including the local spatial coordinate system of heavy atoms, the Gaussian distribution of distances, and the force constraints of hydrogen bonds (feature dimension 107) constitutes the nodes of the graph neural network and is input into the multi-head attention mechanism layer. The node update method of the multi-head attention layer is as follows;

[0057]

[0058]

[0059]

[0060] In the formula, α i,j represents the attention coefficient, Indicates the characteristics of the atom (node), u (i,j) represents the feature (edge) between atoms, σ represents the Sigmoid function, g represents the Relu function, represents the union of feature dimensions, ⊙ represents the multiplication of corresponding elements, · represents the multiplication of a number and a matrix, and represents the rotation matrix of the linear layer, and is its corresponding bias, N i represents all neighboring atoms of atom i, and j is a neighboring atom.

[0061] 3.4) Residue-level pooling layer and protein-level pooling layer. The residue-level pooling layer adds up the feature channels of the heavy atoms on each residue to output the residue-level features (number of residues, 64), while the whole sequence pooling layer adds up the feature channels of the entire protein to form the features of the entire sequence (number of proteins, 64). Finally, the local lDDT score (number of residues, 1) and the global score (1, 1) are predicted by the fully connected layer respectively.

[0062] 3.5) Glorot was used for initialization, the Adam optimizer was used as the gradient descent optimizer, the learning rate was set to 0.001, and the root mean square error MSELoss function was used as the Loss method;

[0063] 4) Label the data and calculate the true global lDDT score and the true local lDDT score: Calculate the lDDT score for each residue of the model protein in the dataset, and sum the lDDT scores of all residues and divide them by the number of residues to form the global lDDT score. β The standard calculation process of -lDDT score is as follows:

[0064] per_residue_lDDT=0.25×(p1×p2×p3×p4) / p0,

[0065]

[0066] p1 is the distance between the ith residue and all other residues that is less than And all residue pairs C β Distance error is less than The probability score of . Similarly, the thresholds of p2, p3, and p4 are p0 means that the distance between the ith residue and all other residues is less than The probability of N is the total number of residues. res is all the residues. per_residue_lddt is the score for each residue of the protein, global_lddt is the global score of the real protein model, and the quality of the protein model is determined according to the global score and the local residue score; the present invention sets the evaluation score in [0,1], and the higher the score, the closer it is to the real structure.

[0067] Taking the protein model 1evya with a sequence length of 130 residues and 1018 heavy atoms as an example, a protein model quality assessment method based on self-attention graph neural network includes the following steps:

[0068] 1) Dataset construction: First, protein sequences with a length of 50-300 residues were screened in the PDB (https: / / www.rcsb.org / ) database using the PISCES (http: / / dunbrack.fccc.edu / pisces) server, and redundancy was removed with a maximum sequence similarity of 35%. Then, comparative modeling, native structure perturbation and deep learning-based modeling methods were used to fold and generate models with a global distance test total score (GDT-TS) between [50,100] to form a dataset, and 90% of the protein models in the dataset were used as training sets, 5% of the models as validation sets, and 5% of the models as test sets;

[0069] 2) Construct network features: the encoding representation of the heavy atoms of each residue in the protein sequence, the construction of the local space coordinate system by three adjacent heavy atoms, the distance between adjacent heavy atoms in Euclidean space and the constraints of hydrogen bond forces, the process is as follows;

[0070] 2.1) Proteins contain 20 kinds of amino acids, each of which has different properties of heavy atoms (atoms other than hydrogen atoms). There are a total of 167 heavy atoms. Each heavy atom is uniquely encoded. For example, ALA (α-linolenic acid) amino acids contain C, N, O, C α ,C β , among these five heavy atoms, the C heavy atom is encoded as [1,0,0...,0], and the encoding length is 167. Similarly, other atoms can be represented similarly. Therefore, a 167×167 matrix can be obtained, which corresponds to each heavy atom one by one;

[0071] 2.2) Construct a local spatial coordinate system on each heavy atom of the protein. The construction method is: for each heavy atom in the sequence, find two adjacent heavy atoms in space with itself as the center to form a plane and generate a vector in the plane Calculate the vector perpendicular to the plane Then calculate the perpendicular vector again With vector Vector

[0072]

[0073]

[0074]

[0075] Among them, B is the central atom, A and C are two adjacent atoms in space, and the spatial coordinates of A and B form Vectors are constructed in the same way vector, is the unit vector of the plane where A, B, and C lie. is the unit vector perpendicular to the plane, is perpendicular to the vector With vector Unit vector;

[0076] 2.3) In the protein structure, each heavy atom uses a distance constraint (Dist) to represent its proximity to other heavy atoms. To this end, the Gaussian function of the distance constraint at different thresholds is calculated to form a multi-interval Gaussian distribution;

[0077]

[0078] Among them, Dist ij is the distance between heavy atom i and heavy atom j, N atoms are all the heavy atoms in the protein, N Datoms is the nearest D heavy atoms, μ k With σ k are the Gaussian parameters at different thresholds [0, 0.4, 0.8...15.2], which are set the same by default, D ij is the Gaussian distribution value of heavy atom i and heavy atom j at different thresholds;

[0079] 2.4) If there is a hydrogen bond between two heavy atoms in the protein structure, it is expressed as 1, otherwise it is 0;

[0080] 3) Build a self-attention graph neural network. The process is as follows:

[0081] 3.1) Overall framework of self-attention graph neural network: Feature input is extracted from proteins and passed through the embedding layer, followed by 4 layers of multi-head attention mechanism layers, and then through the residue-level pooling layer and protein-level pooling layer respectively. These two pooling layers are then connected to the attention mechanism layer and then fully connected to predict the local and global lDDT scores of the protein model;

[0082] 3.2) Embedding layer in graph neural network. The input is the one-hot encoding matrix of heavy atoms in the protein sequence (feature dimension 167), which passes through the emb layer (built-in in the Porch environment) and then passes through the fully connected linear layer (feature dimension 64) to form the embedding layer;

[0083] 3.3) In the multi-head attention mechanism layer of the graph neural network, the number of heads in each layer is set to 8, 4, 2, and 1. The embedded atomic features (feature dimension 64) generated by the embedding layer are combined with other features including the local spatial coordinate system of heavy atoms, the Gaussian distribution of distances, and the force constraints of hydrogen bonds (feature dimension 107) to obtain a new matrix (feature dimension 171) to form the nodes of the graph neural network and input them into the multi-head attention mechanism layer. Among them, the node update method of the multi-head attention layer;

[0084]

[0085]

[0086]

[0087] In the formula, α i,j represents the attention coefficient, Indicates the characteristics of the atom (node), u(i,j) represents the feature (edge) between atoms, σ represents the Sigmoid function, g represents the Relu function, represents the union of feature dimensions, ⊙ represents the multiplication of corresponding elements, · represents the multiplication of a number and a matrix, and represents the rotation matrix of the linear layer, and is its corresponding bias, N i represents all neighboring atoms of atom i, and j is a neighboring atom.

[0088] 3.4) Residue-level pooling layer and protein-level pooling layer. The residue-level pooling layer adds up the feature channels of the heavy atoms on each residue to output the residue-level features (number of residues, 64), while the whole sequence pooling layer adds up the feature channels of the entire protein to form the features of the entire sequence (number of proteins, 64). Finally, the local lDDT score (number of residues, 1) and the global score (1, 1) are predicted by the fully connected layer respectively;

[0089] 3.5) Glorot was used for initialization, the Adam optimizer was used as the gradient descent optimizer, the learning rate was set to 0.001, and the root mean square error MSELoss function was used as the Loss method;

[0090] 4) Label the data and calculate the true global lDDT score and the true local lDDT score: Calculate the lDDT score for each residue of the model protein in the dataset, and sum the lDDT scores of all residues and divide them by the number of residues to form the global lDDT score. β The standard calculation process of -lDDT score is as follows:

[0091] per_residue_lDDT=0.25×(p1×p2×p3×p4) / p0,

[0092]

[0093] p1 is the distance between the ith residue and all other residues that is less than And all residue pairs C β Distance error is less than The probability score of p2, p3, and p4 are p0 means that the distance between the ith residue and all other residues is less than The probability of. Where N is the total number of residues, N resare all residues, per_residue_lDDT is the score for each residue of the protein, global_lDDT is the global score of the true protein model, and the quality of the protein model is determined according to the global score and the local residue score; the present invention sets the evaluation score in [0,1], and the higher the score, the closer it is to the true structure.

[0094] The present invention is not only suitable for the above-mentioned embodiments, but also can be implemented with various modifications without departing from the basic spirit of the present invention and without exceeding the essential content of the present invention.

Claims

1. A protein model quality assessment method based on self-attention graph neural network, characterized in that: The method comprises the following steps: 1) Dataset construction: First, protein sequences with a length of 50-300 residues were screened from the PDB database using the PISCES server, and redundancy was removed with a maximum sequence similarity of 35%. Then, a model dataset is constructed by folding and generating a global distance test total score GDT-TS between [50, 100] using comparative modeling, native structure perturbation, and deep learning-based modeling methods, and the dataset is divided into a training set, a validation set, and a test set; 2) Constructing network features: encoding representation of the heavy atoms of each residue in the protein sequence, constructing the local space coordinate system of three adjacent heavy atoms, the distance between adjacent heavy atoms in Euclidean space and the hydrogen bond force constraints. The process is as follows: 2.1) Proteins contain 20 kinds of amino acids, each of which has heavy atoms with different properties, totaling 167 heavy atoms. Each heavy atom is uniquely encoded; 2.2) Construct a local spatial coordinate system on each heavy atom of the protein. The construction method is: for each heavy atom in the sequence, find two adjacent heavy atoms in space with itself as the center to form a plane and generate a vector in the plane Calculate the vector perpendicular to the plane Then calculate the perpendicular vector again With vector Vector Among them, B is the central atom, A and C are two adjacent atoms in space, and the spatial coordinates of A and B form Vectors are constructed in the same way vector, is the unit vector of the plane where A, B, and C lie. is the unit vector perpendicular to the plane, is perpendicular to the vector With vector Unit vector; 2.3) In the protein structure, each heavy atom uses a distance constraint Dist to represent its proximity to other heavy atoms. To this end, the Gaussian function of the distance constraint at different thresholds is calculated to form a multi-interval Gaussian distribution; Among them, Dist ij is the distance between heavy atom i and heavy atom j, N atoms are all the heavy atoms in the protein, N Datoms is the nearest D heavy atoms, μ k With σ k are the Gaussian parameters for different thresholds [0, 0.4, 0.8...15.2], which are set the same by default, D ij is the Gaussian distribution value of heavy atom i and heavy atom j at different thresholds; 2.4) If there is a hydrogen bond between two heavy atoms in the protein structure, it is expressed as 1, otherwise it is 0; 3) Build a self-attention graph neural network. The process is as follows: 3.1) Overall framework of self-attention graph neural network: Feature input is extracted from proteins and passed through the embedding layer, followed by 4 layers of multi-head attention mechanism layers, and then through the residue-level pooling layer and protein-level pooling layer respectively. These two pooling layers are then connected to the attention mechanism layer and then fully connected to predict the local and global lDDT scores of the protein model; 3.2) In the graph neural network, the embedding layer takes as input the one-hot encoding matrix of the heavy atoms in the protein sequence. After passing through the emb layer, it then passes through the fully connected linear layer to form the embedding layer. 3.3) In the multi-head attention mechanism layer of the graph neural network, the number of heads in each layer is set to 8, 4, 2, and 1. The embedded atomic features generated by the embedding layer are combined with other features including the local spatial coordinate system of heavy atoms, the Gaussian distribution of distances, and the force constraints of hydrogen bonds to obtain a new matrix to form the nodes of the graph neural network and input it into the multi-head attention mechanism layer, where: How to update nodes in multi-head attention layer; In the formula, α i,j represents the attention coefficient, Represents the characteristics of the atom, u (i,j) represents the characteristics between atoms, σ represents the Sigmoid function, g represents the Relu function, represents the union of feature dimensions, ⊙ represents the multiplication of corresponding elements, · represents the multiplication of a number and a matrix, and represents the rotation matrix of the linear layer, and is its corresponding bias, N i represents all neighboring atoms of atom i, and j is a neighboring atom; 3.4) Residue-level pooling layer and protein-level pooling layer. The residue-level pooling layer adds up the feature channels of the heavy atoms on each residue to output the residue-level features, while the whole sequence pooling layer adds up the feature channels of the entire protein to form the features of the entire sequence. Finally, the local lDDT score and the global score are predicted through the fully connected layer respectively. 3.5) Glorot was used for initialization, the Adam optimizer was used as the gradient descent optimizer, the learning rate was set to 0.001, and the root mean square error MSELoss function was used as the Loss method; 4) Label the data and calculate the true global lDDT score and the true local lDDT score: Calculate the lDDT score for each residue of the model protein in the dataset, and sum the lDDT scores of all residues and divide them by the number of residues to form the global lDDT score. β The standard calculation process of -lDDT score is as follows: per_residue_lDDT=0.25×(p1×p2×p3×p4) / p0, p1 is the distance between the ith residue and all other residues that is less than And all residue pairs C β Distance error is less than The probability score of p2, p3, and p4 are p0 means that the distance between the ith residue and all other residues is less than The probability of, where N is the total number of residues, N res are all the residues, per_residue_lDDT is the score for each residue in the protein, and global_lDDT is the global score of the true protein model. The quality of the protein model is determined according to the global score and the local residue score.

2. A protein model quality assessment method based on self-attention graph neural network as claimed in claim 1, characterized in that: In step 1), 90% of the protein models in the dataset are used as training sets, 5% of the models are used as validation sets, and 5% of the models are used as test sets.

Images