Protein complex model quality evaluation method based on Voronoi diagram

Abstract

The invention discloses a protein complex model quality evaluation method based on a Voronoi diagram. The method comprises the following steps: firstly, screening 7099 structures in a PDB structure database through a PISCES server; and then generating about 150 structural models for each protein in the data set by adopting three methods of structural disturbance, template modeling and deep learning modeling. Extracting features based on a Voronoi diagram, such as a contact surface area, a solvent accessible area, an energy item based on a contact surface, a residue-level contact surface direction and the like, from an amino acid sequence and a protein model structure by each structure model, and extracting a Roselle feature and a language model feature; the invariant point attention mechanism neural network module captures feature sets of different scales, and the residual attention neural network decodes and outputs residue-by-residue lDDT scores. According to the method, the quality evaluation of the protein compound model can be quickly and specifically carried out, and the problems of low precision and poor reliability of the quality evaluation of the protein compound model are effectively solved.

Description

Technical Field

[0001] The invention relates to the fields of bioinformatics and computer application, and in particular to a protein complex model quality assessment method based on a Voronoi diagram. Background Art

[0002] The human body is a complex system. Enzymes, hormones, antibodies, muscles, bones and other important tissues and structures in the body are all composed of proteins. Various physiological activities and various classification traits of the human body are directly or indirectly related to proteins. Biologists evaluate the biological functions of proteins and then obtain more reliable protein targets and protein-protein interaction models to greatly accelerate the speed of structure-based innovative drug design and vaccine development. Protein model quality assessment helps to select the most accurate prediction model for subsequent work. Initially, the function of evaluating the degree of fit between the model and the natural structure was included as a component in the structure prediction method. However, in recent years, more and more independent methods have emerged, which are collectively referred to as model quality assessment methods. Traditional model quality assessment methods mainly consider energy-based, physicochemical, or statistical properties, while machine learning-based model quality assessment methods are mainly based on the analysis of the structure and characteristics of a single protein, and only evaluate the quality of one model at a time. This model is usually a structure obtained by experimental or computational methods during the three-dimensional structure analysis of the protein. The advantages of this method are fast evaluation speed and simple calculation. In terms of evaluation indicators, researchers have developed some evaluation indicators, the most widely used ones are the global distance test score (GDT-TS, The Global Distance Test-Total Score), template modeling (TM-score, Template Modeling Score) score, local distance difference test (lDDT, local Distance Difference Test score) score, contact area difference (CAD, Contact Area Difference), etc.

[0003] The future challenge for model quality assessment methods is how to distinguish high-quality models (such as those generated by AlphaFold2) from low-quality models, and to assess which areas in high-quality models are relatively incorrect and need to be improved. Since AlphaFold2 has basically solved the problem of tertiary structure prediction for protein monomers, the problem of protein structure prediction will shift to quaternary structure (complex) prediction, and the corresponding complex model quality assessment methods will also be a hot topic in the future. Summary of the invention

[0004] In order to overcome the deficiency of low accuracy in quality assessment of protein complex models in the prior art, the present invention proposes a protein complex model quality assessment method based on the Voronoi diagram, which aims to more accurately assess the quality of the protein complex model.

[0005] The technical solution adopted by the present invention to solve the technical problem is:

[0006] A method for evaluating the quality of a protein complex model based on a Voronoi diagram, the method comprising the following steps:

[0007] 1) Dataset construction: The protein dataset was constructed from the PDB structure database. n structures were screened from the PDB library (as of March 19, 2022) through the PISCES server. The selection criteria were sequence similarity less than the preset similarity threshold, structure length between 50 and 400, and minimum resolution of In the process of constructing the dataset, three methods, namely structural perturbation, template modeling and deep learning modeling, were used to generate M structural models for each protein in the dataset: the structural perturbation method was to fine-tune the dihedral angles of the native structure, and select the m1 structural model according to the Rosetta energy of the perturbed structure; the template modeling method was to use different template structures and fragment libraries in RosettaCM and Modeler to generate the m2 structural model; the deep learning modeling method was to use the RocketX method to generate the m2 structural model through constraints of different weights of short, medium and long lengths, m1+m2+m3=M; finally, the structural models generated by all methods were clustered, similar structures were filtered out, and m structural models were obtained, and the ratio of the training set and the validation set was set;

[0008] 2) Encoding of amino acids: Alanine A: 1, Cysteine ​​C: 2, Aspartic acid D: 3, Glutamic acid E: 4, Phenylalanine F: 5, Glycine G: 6, Histidine H: 7, Isoleucine I: 8, Lysine K: 9, Leucine L: 10, Methionine M: 11, Asparagine N: 12, Proline P: 13, Glutamine Q: 14, Arginine R: 15, Serine S: 16, Threonine T: 17, Valine V: 18, Tryptophan W: 19, Tyrosine Y: 20;

[0009] 3) Protein semantic feature embedding: First, search UniRef30 and BFD through HHblits to obtain the MSA (Multi Sequence Alignment) of the input structure model, then send the searched MSA to the ESM-MSA-1b language model to obtain the residual level sequence embedding and the row attention embedding between the residuals (1*L*1280), and use ESM-IF1 to generate the structural embedding of the backbone atomic coordinates (1*L*512);

[0010] 4) Ultra-fast shape recognition features: For a given protein structure model, extract the USR features of each residue. For the i-th residue, take the current residue as a reference and determine the distance from the residue r in sequence. i The furthest residue b i and residue b i The furthest residue g i , the topological relationship between the current residue and the overall structure is expressed as r i ,b i ,g i The three residue distance sets defined are represented as follows. i , calculate its distance to all other residues in the structure to form a distance set. i and residue g i , use the same method to get the corresponding distance set, the distance set is defined as follows:

[0011]

[0012] Where L represents the length of the protein sequence, is the C of the ith residue β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates, and Represents residue r i ,b i ,g i C β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates;

[0013] 5) Physicochemical properties: Eight physical energy characteristics are statistically analyzed for each amino acid, including: protein dihedral angle deviation energy term, protein conformation potential energy term, protein peptide bond torsion energy term, hydrogen bond, potential energy term describing dihedral angle, octanol-water partition coefficient and topological polar surface area, as well as a statistical scoring function based on the physical contact surface between atoms;

[0014] 6) Protein atomic-level, residue-level inter-chain contact area, intra-chain contact area, and solvent accessible surface area characteristics:

[0015] Contact is defined physically as the space between the two atomic surfaces that is not large enough for a water molecule (with a van der Waals radius of ) enters, the two atoms are in contact, and then the contact surface area is calculated. The contact area between atoms is calculated by the Voronoi diagram (Voronoi-diagram) and triangulation algorithm (Delaunay) in computer graphics. The contact surface function and contact surface area calculation method of the two atoms A (x1, y1, z1) and B (x2, y2, z2) of the Voronoi algorithm are as follows:

[0016] ((x-x1) 2 +(y-y1) 2 +(z-z1) 2 ) 0.5 -r1=((x-x2) 2 +(y-y2) 2 +(z-z2) 2 ) 0.5 -r2

[0017]

[0018] The Voronoi contact surface is defined as a surface composed of points with equal distances from the van der Waals surfaces of two atoms. AB is the contact surface or solvent accessible surface area of ​​atoms A and B, n is the number of surface triangles obtained by triangulation algorithm, a, b, c are the sides of a surface triangle, after the contact area and solvent accessible area are calculated for each pair of heavy atoms (C, N, O, P, S…) in the protein, the residue-level contact area and solvent accessible area are obtained by numerical summing according to the amino acid to which the atoms belong;

[0019] 7) Inter-residue distance feature, defined as the Cβ distance between residue pairs. The Cβ distance expression is as follows:

[0020]

[0021] Where d represents the Cβ distance, (x1, x2, x3), (y1, y2, y3) represent the three-dimensional coordinates of any pair of heavy atoms;

[0022] 8) Construct a neural network module based on the invariant point attention mechanism and the residual attention neural network module to evaluate the quality of the protein complex model; the overall network architecture includes two parts: the invariant point attention mechanism neural network module and the residual attention neural network module: the invariant point attention mechanism neural network module, a protein graph is represented as in is the set of residues, ε={ε ij} i≠j is the set of edges between residues, represents the C, O, N, C α ,Cβ The atomic coordinates are used to embed residues obtained from the protein language model and residual attention The input is sent to the encoding module, and then the features are embedded through the graph attention network and the triangular update mechanism to deeply explore the spatial geometric information between the sequence and the structure; the triangular multiplication operation in AlphaFold2 is used to update the edge information, and the invariant point attention module is combined to establish the intrinsic relationship between the protein sequence and the structure in the high-dimensional space; finally, the geometric space constraints strongly related to the sequence are obtained; the residual attention neural network module extracts the node representation from the output of the encoding module and edge representation Use the new network function W * To regenerate the node and edge Combining all of these features generates the following representation of structural quality:

[0023] in Indicates the concatenation of feature vectors; tile indicates horizontal striping of node features into edge features; in the structural quality decoding module, a residual network based on transformation strategy is adopted, which consists of a main residual block and a branch residual block. Each residual block consists of three two-dimensional convolutional layers with different expansion rate coefficients and a normalization operation, and uses the GELU activation function. In addition, a convolutional network layer is added to the branch residual block to improve the prediction of distance error and threshold, as follows:

[0024] Among them, Conv r With different expansion coefficients r = p 2 (p=1,2,3,4), IB represents the inverted bottleneck operation, and finally the distance error can be obtained from the network and distance threshold is the prediction distance error between the true structure and the model structure, It is the distance threshold defined in lDDT. The local quality score is calculated by these two values, and the quality of the prediction model is evaluated according to the score;

[0025] 9) lDDT score calculation: lDDT (local distance difference test) is a non-overlapping score that evaluates the local distance difference of all atoms in the protein structure. It is very suitable for evaluating the quality of local protein models. The deviation predicted by the deep learning network and the contact map threshold are used to calculate the lDDT score of each residue in the protein model. The lDDT score of the i-th residue and the global lDDT score are defined as follows:

[0026]

[0027] Where p0 represents the distance to the ith residue The ratio of the number of residues within to the total number of residues, p1 represents the distance from the ith residue The ratio of the number of residues within to the total number of residues, p2, p3, p4 represent the distance from the ith residue and The ratio of the number of residues within to the total number of residues;

[0028] 10) Network parameters, the learning rate of the network is set to 0.001, and the exponential decay factor of the learning rate is set to 0.99.

[0029] The technical concept of the present invention is: in the field of protein model quality assessment, a protein complex model quality assessment method based on the Voronoi diagram is proposed. After extracting the contact surface area, solvent accessible area, energy terms based on the contact surface, residue-level contact surface direction and other Voronoi diagram-based features of the amino acid sequence and protein model structure, the invariant point attention mechanism neural network module captures feature sets of different scales, and the residual attention neural network decodes and outputs the residue-by-residue lDDT score.

[0030] The beneficial effects of the present invention are mainly manifested in: powerful feature extraction of amino acid sequences and protein model structures, capturing feature sets of different scales by the invariant point attention mechanism neural network module, and finally evaluating the quality of the complex model by the residual attention neural network module. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is the overall flow chart of the invariant point attention mechanism, residual attention neural network and loss function.

[0032] Figure 2 Flowchart of a method for assessing the quality of protein complex models based on the Voronoi diagram. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings.

[0034] Reference Figure 1 and Figure 2 , a protein complex model quality assessment method based on Voronoi diagram, comprising the following steps:

[0035] 1) Dataset construction: The protein dataset was constructed from the PDB structure database. 7099 structures were screened from the PDB library (as of March 19, 2022) through the PISCES server. The selection criteria were sequence similarity less than 35%, structure length between 50 and 400, and minimum resolution of In the process of constructing the dataset, three methods, structure perturbation, template modeling and deep learning modeling, were used to generate about 150 structural models for each protein in the dataset: the structure perturbation method was to fine-tune the dihedral angles of the native structure, and select about 60 structural models based on the Rosetta energy of the perturbed structure. The template modeling method was to use different template structures and fragment libraries in RosettaCM and Modeler to generate about 40 structural models. The deep learning modeling method was to use the RocketX method to generate about 50 structural models through constraints of different weights of short, medium and long. Finally, the structural models generated by all methods were clustered and similar structures were filtered out to obtain about 1,064,850 structural models, 70% of which were used as training sets and 30% were used as validation sets.

[0036] 2) Encode the amino acids: Alanine A: 1, Cysteine ​​C: 2, Aspartic acid D: 3, Glutamic acid E: 4, Phenylalanine F: 5, Glycine G: 6, Histidine H: 7, Isoleucine I: 8, Lysine K: 9, Leucine L: 10, Methionine M: 11, Asparagine N: 12, Proline P: 13, Glutamine Q: 14, Arginine R: 15, Serine S: 16, Threonine T: 17, Valine V: 18, Tryptophan W: 19, Tyrosine Y: 20.

[0037] 3) Protein semantic feature embedding: First, search UniRef30 and BFD through HHblits to obtain the MSA (Multi Sequence Alignment) of the input structure model, then send the searched MSA to the ESM-MSA-1b language model to obtain the residual level sequence embedding and the row attention embedding between the residuals (1*L*1280), and use ESM-IF1 to generate the structural embedding of the backbone atomic coordinates (1*L*512); 4) Ultra-fast shape recognition features: For a given protein structure model, extract the USR features of each residue. For the i-th residue, take the current residue as a reference and determine the distance from the residue r in sequence. i The furthest residue b i and residue b i The furthest residue g i , the topological relationship between the current residue and the overall structure can be expressed by r i ,b i ,g i The three residue distance sets defined are represented as follows. i , calculate its distance to all other residues in the structure to form a distance set. For residue b i and residue g i, use the same method to get the corresponding distance set, the distance set is defined as follows:

[0038]

[0039]

[0040] Where L represents the length of the protein sequence, is the C of the ith residue β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates, and Represents residue r i ,b i ,g i C β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates;

[0041] 5) Physicochemical properties: Eight physical energy characteristics are statistically analyzed for each amino acid, including: protein dihedral angle deviation energy term, protein conformation potential energy term, protein peptide bond torsion energy term, hydrogen bond, potential energy term describing dihedral angle, octanol-water partition coefficient and topological polar surface area, as well as a statistical scoring function based on the physical contact surface between atoms;

[0042] 6) Protein atomic and residue-level interchain and intrachain contact areas and solvent accessible surface area characteristics: Contact is defined physically as: the space between the surfaces of two atoms is not large enough for a water molecule (van der Waals radius is ) enters, the two atoms are in contact, and then the contact surface area is calculated. The contact area between atoms is calculated by the Voronoi diagram (Voronoi-diagram) and triangulation algorithm (Delaunay) in computer graphics. Taking two atoms A (x1, y1, z1) and B (x2, y2, z2) as an example, the contact surface function and contact surface area calculation method of the Voronoi algorithm are as follows:

[0043] ((x-x1) 2 +(y-y1) 2 +(z-z1) 2 ) 0.5 -r1=((x-x2) 2 +(y-y2) 2 +(z-z2) 2 ) 0.5 -r2

[0044]

[0045] The definition of the Voronoi contact surface is a surface composed of points with equal distances from the van der Waals surfaces of two atoms, areaAB is the contact surface of atoms A and B or the solvent accessible surface area, n is the number of surface triangles obtained by the triangulation algorithm, a, b, c are the sides of a surface triangle, and after the contact area and solvent accessible area are calculated for each pair of heavy atoms (C, N, O, P, S…) in the protein, the residue-level contact area and solvent accessible area are obtained by numerical summation according to the amino acid to which the atoms belong;

[0046] 7) Inter-residue distance feature, defined as the Cβ distance between residue pairs. The Cβ distance expression is as follows:

[0047]

[0048] Where d represents the Cβ distance, (x1, x2, x3), (y1, y2, y3) represent the three-dimensional coordinates of any pair of heavy atoms;

[0049] 8) Construct a neural network module based on the invariant point attention mechanism and the residual attention neural network module to evaluate the quality of the protein complex model. The overall network architecture includes two parts: the invariant point attention mechanism neural network module and the residual attention neural network module: In the invariant point attention mechanism neural network module, a protein graph can be represented as in is the set of residues, ε={ε ij} i≠j is the set of edges between residues, represents the C, O, N, C α ,C β The atomic coordinates are used to embed residues obtained from the protein language model and residual attention Input to the encoding module, and then embed features through the graph attention network and triangular update mechanism, which can deeply mine the spatial geometric information between sequence and structure, use the triangular multiplication operation in AlphaFold2 to update the edge information, and combine the invariant point attention module to establish the intrinsic relationship between protein sequence and structure in high-dimensional space, and finally obtain geometric space constraints that are strongly related to the sequence; residual attention neural network module. Extract node representation from the output of the encoding module and edge representation In addition, using the new network function W * To regenerate the node and edge Combining all of these features generates the following representation of structural quality: in Indicates the concatenation of feature vectors; tile indicates horizontal striping of node features into edge features. In the structural quality decoding module, a residual network based on transformation strategy is adopted. The network consists of a main residual block and a branch residual block. Each residual block consists of three two-dimensional convolutional layers with different expansion rate coefficients and a normalization operation. The GELU activation function is used, and a convolutional network layer is added to the branch residual block to improve the prediction of distance error and threshold, as follows: Among them, Conv r With different expansion coefficients r = p 2 (p=1,2,3,4), IB represents the inverted bottleneck operation. Finally, the distance error can be obtained from the network and distance threshold is the prediction distance error between the true structure and the model structure, It is the distance threshold defined in lDDT. The local quality score is calculated by these two values, and the quality of the prediction model is evaluated according to the score;

[0050] 9) lDDT score calculation: lDDT (local distance difference test) is a non-overlapping score that evaluates the local distance difference of all atoms in the protein structure. It is very suitable for evaluating the quality of local protein models. The deviation predicted by the deep learning network and the contact map threshold can be used to calculate the lDDT score of each residue in the protein model. The lDDT score of the i-th residue and the global lDDT score are defined as follows:

[0051]

[0052] Where p0 represents the distance to the ith residue The ratio of the number of residues within to the total number of residues, p1 represents the distance from the ith residue The ratio of the number of residues within to the total number of residues, p2, p3, p4 represent the distance from the ith residue and The ratio of the number of residues within to the total number of residues;

[0053] 10) Network parameters, the learning rate of the network is set to 0.001, and the exponential decay factor of the learning rate is set to 0.99.

[0054] Taking the model H1106TS390_1 predicted by AF2-multimer in the protein complex H1106 with a sequence length of 236 as an example, a protein complex model quality assessment method based on Voronoi diagram includes the following steps:

[0055] 1) Dataset construction: The protein dataset was constructed from the PDB structure database. 7099 structures were screened from the PDB library (as of March 19, 2022) through the PISCES server. The selection criteria were sequence similarity less than 35%, structure length between 50 and 400, and minimum resolution of In the process of constructing the dataset, three methods, structure perturbation, template modeling and deep learning modeling, were used to generate about 150 structural models for each protein in the dataset: the structure perturbation method was to fine-tune the dihedral angles of the native structure, and select about 60 structural models based on the Rosetta energy of the perturbed structure; the template modeling method was to generate about 40 structural models using different template structures and fragment libraries in RosettaCM and Modeler; the deep learning modeling method was to use the RocketX method to generate about 50 structural models through constraints of different weights of short, medium and long lengths; finally, the structural models generated by all methods were clustered, and similar structures were filtered out to obtain about 1,064,850 structural models, 70% of which were used as training sets and 30% were used as validation sets;

[0056] 2) Encoding of amino acids: Alanine A: 1, Cysteine ​​C: 2, Aspartic acid D: 3, Glutamic acid E: 4, Phenylalanine F: 5, Glycine G: 6, Histidine H: 7, Isoleucine I: 8, Lysine K: 9, Leucine L: 10, Methionine M: 11, Asparagine N: 12, Proline P: 13, Glutamine Q: 14, Arginine R: 15, Serine S: 16, Threonine T: 17, Valine V: 18, Tryptophan W: 19, Tyrosine Y: 20;

[0057] 3) Protein semantic feature embedding: First, search UniRef30 and BFD through HHblits to obtain the MSA (Multi Sequence Alignment) of the input structure model, then send the searched MSA to the ESM-MSA-1b language model to obtain the residual level sequence embedding and the row attention embedding between the residuals (1*L*1280), and use ESM-IF1 to generate the structural embedding of the backbone atomic coordinates (1*L*512);

[0058] 4) Ultra-fast shape recognition features: For a given protein structure model, extract the USR features of each residue. For the i-th residue, take the current residue as a reference and determine the distance from the residue r in sequence. i The furthest residue b i and residue b i The furthest residue g i , the topological relationship between the current residue and the overall structure can be expressed by r i,b i ,g i The three residue distance sets defined are used to represent the residue r i , calculate its distance to all other residues in the structure to form a distance set. i and residue g i , use the same method to get the corresponding distance set, the distance set is defined as follows:

[0059]

[0060] Where L represents the length of the protein sequence, is the C of the ith residue β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates, and Represents residue r i ,b i ,g i C β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates;

[0061] 5) Physicochemical properties: Eight physical energy characteristics are statistically analyzed for each amino acid, including: protein dihedral angle deviation energy term, protein conformation potential energy term, protein peptide bond torsion energy term, hydrogen bond, potential energy term describing dihedral angle, octanol-water partition coefficient and topological polar surface area, as well as a statistical scoring function based on the physical contact surface between atoms;

[0062] 6) Protein atomic and residue-level interchain and intrachain contact areas and solvent accessible surface area characteristics: Contact is defined physically as: the space between the surfaces of two atoms is not large enough for a water molecule (van der Waals radius is ) enters, the two atoms are in contact. Then the contact surface area is calculated, and the interatomic contact area is calculated by the Voronoi diagram (Voronoi-diagram) and triangulation algorithm (Delaunay) in computer graphics. Taking two atoms A (x1, y1, z1) and B (x2, y2, z2) as an example, the contact surface function and contact surface area calculation method of the Voronoi algorithm are as follows:

[0063] ((x-x1) 2 +(y-y1) 2 +(z-z1) 2 ) 0.5 -r1=((x-x2) 2 +(y-y2) 2 +(z-z2) 2 ) 0.5 -r2

[0064]

[0065] Among them, the Voronoi contact surface is defined as a surface composed of points with equal distances from the van der Waals surfaces of two atoms, areaAB is the contact surface of atoms A and B or the solvent accessible surface area, n is the number of surface triangles obtained by the triangulation algorithm, a, b, c are the sides of a surface triangle, and after calculating the contact area and solvent accessible area for each pair of heavy atoms (C, N, O, P, S...) in the protein, the residue-level contact area and solvent accessible area are obtained by numerical summation according to the amino acids to which the atoms belong.

[0066] 7) Inter-residue distance feature, defined as the Cβ distance between residue pairs. The Cβ distance expression is as follows:

[0067]

[0068] Where d represents the Cβ distance, (x1, x2, x3), (y1, y2, y3) represent the three-dimensional coordinates of any pair of heavy atoms;

[0069] 8) Construct a neural network module based on the invariant point attention mechanism and the residual attention neural network module to evaluate the quality of the protein complex model. The overall network architecture includes two parts: the invariant point attention mechanism neural network module and the residual attention neural network module: In the invariant point attention mechanism neural network module, a protein graph can be represented as in is the set of residues, ε={ε ij} i≠j is the set of edges between residues, represents the C, O, N, C α ,C β The atomic coordinates are used to embed residues obtained from the protein language model and residual attention The input is sent to the encoding module, and then the features are embedded through the graph attention network and the triangular update mechanism, which can deeply mine the spatial geometric information between the sequence and the structure, use the triangular multiplication operation in AlphaFold2 to update the edge information, and combine the invariant point attention module to establish the intrinsic relationship between the protein sequence and the structure in the high-dimensional space, and finally obtain the geometric space constraints that are strongly related to the sequence; the residual attention neural network module extracts the node representation from the output of the encoding module and edge representation In addition, using the new network function W * To regenerate the node and edge Combining all of these features generates the following representation of structural quality: in Indicates the concatenation of feature vectors; tile indicates horizontal striping of node features into edge features. In the structural quality decoding module, a residual network based on transformation strategy is adopted. The network consists of a main residual block and a branch residual block. Each residual block consists of three two-dimensional convolutional layers with different expansion rate coefficients and a normalization operation. The GELU activation function is used, and a convolutional network layer is added to the branch residual block to improve the prediction of distance error and threshold, as follows: Among them, Conv r With different expansion coefficients r = p 2 (p=1,2,3,4), IB represents the inverted bottleneck operation, and finally the distance error can be obtained from the network and distance threshold is the prediction distance error between the true structure and the model structure, It is the distance threshold defined in lDDT. The local quality score is calculated by these two values, and the quality of the prediction model is evaluated according to the score;

[0070] 9) lDDT score calculation: lDDT (local distance difference test) is a non-overlapping score that evaluates the local distance difference of all atoms in the protein structure. It is very suitable for evaluating the quality of local protein models. The deviation predicted by the deep learning network and the contact map threshold can be used to calculate the lDDT score of each residue in the protein model. The lDDT score of the i-th residue and the global lDDT score are defined as follows:

[0071]

[0072] Where p0 represents the distance to the ith residue The ratio of the number of residues within to the total number of residues, p1 represents the distance from the ith residue The ratio of the number of residues within to the total number of residues, p2, p3, p4 represent the distance from the ith residue and The ratio of the number of residues within to the total number of residues;

[0073] 10) Network parameters, the learning rate of the network is set to 0.001, and the exponential decay factor of the learning rate is set to 0.99.

[0074] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. A method for evaluating the quality of a protein complex model based on a Voronoi diagram, characterized in that: The method comprises the following steps: 1) Dataset construction: The protein dataset was constructed from the PDB structure database. n structures were screened from the PDB library through the PISCES server. The selection criteria were that the sequence similarity was less than the preset similarity threshold, the structure length was between 50 and 400, and the minimum resolution was Cluster the generated structural models, filter out similar structures, obtain m structural models, and set the ratio of training set and validation set; 2) Encoding of amino acids: Alanine A: 1, Cysteine ​​C: 2, Aspartic acid D: 3, Glutamic acid E: 4, Phenylalanine F: 5, Glycine G: 6, Histidine H: 7, Isoleucine I: 8, Lysine K: 9, Leucine L: 10, Methionine M: 11, Asparagine N: 12, Proline P: 13, Glutamine Q: 14, Arginine R: 15, Serine S: 16, Threonine T: 17, Valine V: 18, Tryptophan W: 19, Tyrosine Y: 20; 3) Protein semantic feature embedding: First, the MSA of the input structure model is searched by HHblits for UniRef30 and BFD, and then the searched MSA is fed into the ESM-MSA-1b language model to obtain the residual level sequence embedding and the row attention embedding between the residuals (1*L*1280), and ESM-IF1 is used to generate the structural embedding of the backbone atomic coordinates (1*L*512); 4) Ultra-fast shape recognition features: For a given protein structure model, extract the USR features of each residue: for the i-th residue, take the current residue as a reference and determine the distance from the residue r in sequence. i The furthest residue b i and residue b i The furthest residue g i , the topological relationship between the current residue and the overall structure is expressed as r i ,b i ,g i The three residue distance sets defined are used to represent the residue r i , calculate its distance to all other residues in the structure to form a distance set. For residue b i and residue g i , use the same method to get the corresponding distance set, the distance set is defined as follows: Where L represents the length of the protein sequence, is the C of the ith residue β Atomic coordinates, if the residue is glycine, use C α Atomic coordinates, and Represents residue r i ,b i ,g i C β Atomic coordinates, if If the residue is glycine, use C α Atomic coordinates; 5) Physicochemical properties: Eight physical energy characteristics are statistically analyzed for each amino acid, including: protein dihedral angle deviation energy term, protein conformation potential energy term, protein peptide bond torsion energy term, hydrogen bond, potential energy term describing dihedral angle, octanol-water partition coefficient and topological polar surface area, as well as a statistical scoring function based on the physical contact surface between atoms; 6) Protein atomic-level, residue-level inter-chain contact area, intra-chain contact area, and solvent accessible surface area characteristics: Contact is defined in a physical way: if the space between the surfaces of two atoms is not enough for water molecules to enter, the two atoms are in contact, and then the contact surface area is calculated. The interatomic contact area is calculated by the Voronoi diagram and triangulation algorithm in computer graphics. The contact surface function and contact surface area calculation method of the two atoms A(x1,y1,z1) and B(x2,y2,z2) of the Voronoi algorithm are as follows: ((x-x1) 2 +(y-y1) 2 +(z-z1) 2 ) 0.5 -r1=((x-x2) 2 +(y-y2) 2 +(z-z2) 2 ) 0.5 -r2 The Voronoi contact surface is defined as a surface composed of points with equal distances from the van der Waals surfaces of two atoms. AB is the contact surface or solvent accessible surface area of ​​atoms A and B, n is the number of surface triangles obtained by triangulation algorithm, a, b, c are the sides of a surface triangle, after the contact area and solvent accessible area are calculated for each pair of heavy atoms (C, N, O, P, S…) in the protein, the residue-level contact area and solvent accessible area are obtained by numerical summing according to the amino acid to which the atoms belong; 7) Inter-residue distance feature, defined as the Cβ distance between residue pairs. The Cβ distance expression is as follows: Where d represents the Cβ distance, (x1, x2, x3), (y1, y2, y3) represent the three-dimensional coordinates of any pair of heavy atoms; 8) Construct a neural network module based on the invariant point attention mechanism and a residual attention neural network module to evaluate the quality of protein complex models; 9) lDDT score calculation: lDDT (local distance difference test) is a non-overlapping score that evaluates the local distance difference of all atoms in the protein structure. It is very suitable for evaluating the quality of local protein models. The deviation predicted by the deep learning network and the contact map threshold are used to calculate the lDDT score of each residue in the protein model. The lDDT score of the i-th residue and the global lDDT score are defined as follows: Where p0 represents the distance to the ith residue The ratio of the number of residues within to the total number of residues, p1 represents the distance from the ith residue The ratio of the number of residues within to the total number of residues, p2, p3, p4 represent the distance from the ith residue and The ratio of the number of residues within to the total number of residues; 10) Network parameters, set the network's learning rate and learning rate exponential decay factor.

2. The method for evaluating the quality of a protein complex model based on a Voronoi diagram according to claim 1, characterized in that: In the step 1), during the construction of the data set, three methods, namely, structural perturbation, template modeling and deep learning modeling, are used to generate M structural models for each protein in the data set: the structural perturbation method is to fine-tune the dihedral angles of the native structure, and select the m1 structural model according to the Rosetta energy of the perturbed structure; the template modeling method is to use different template structures and fragment libraries in RosettaCM and Modeler to generate the m2 structural model; the deep learning modeling method is to use the RocketX method to generate the m2 structural model through constraints of different weights of short, medium and long, m1+m2+m3=M.

3. A protein complex model quality assessment method based on Voronoi diagram according to claim 1 or 2, characterized in that: In the step 8), the overall network architecture includes two parts: the invariant point attention mechanism neural network module and the residual attention neural network module: the invariant point attention mechanism neural network module, a protein graph is represented as in is the set of residues, ε={ε ij } i≠j is the set of edges between residues, represents the C, O, N, C α ,C β The atomic coordinates are used to embed residues obtained from the protein language model and residual attention The input is sent to the encoding module, and then the features are embedded through the graph attention network and the triangular update mechanism to deeply explore the spatial geometric information between the sequence and the structure; the triangular multiplication operation in AlphaFold2 is used to update the edge information, and the invariant point attention module is combined to establish the intrinsic relationship between the protein sequence and the structure in the high-dimensional space; finally, the geometric space constraints strongly related to the sequence are obtained; the residual attention neural network module extracts the node representation from the output of the encoding module and edge representation Use the new network function W * To regenerate the node and edge Combining all of these features generates the following representation of structural quality: in Indicates the concatenation of feature vectors; tile means horizontally striping node features into edge features; In the structural quality decoding module, a residual network based on a transformation strategy is used. The network consists of a main residual block and a branch residual block. Each residual block consists of three two-dimensional convolutional layers with different expansion rate coefficients and a normalization operation, and uses a GELU activation function. In addition, a convolutional network layer is added to the branch residual block to improve the prediction of the distance error and threshold, as follows: Among them, Conv r With different expansion coefficients r = p 2 (p=1,2,3,4), IB represents the inverted bottleneck operation, and finally the distance error can be obtained from the network and distance threshold is the prediction distance error between the true structure and the model structure, It is the distance threshold defined in lDDT. The local quality score is calculated by these two values, and the quality of the prediction model is evaluated based on this score.

4. A protein complex model quality assessment method based on Voronoi diagram according to claim 1 or 2, characterized in that: In the step 10), the learning rate of the network is set to 0.001, and the exponential decay factor of the learning rate is set to 0.99.

Images