A deep learning-based method for predicting inter-residue distances of polymorphic proteins

Abstract

A kind of deep learning-based polymorphic protein inter-residue distance prediction method, first, the protein of set length and sequence redundancy is screened out in protein flexibility database PDBflex, find out the two structures of the biggest structural difference of polymorphic protein, for each structure, extract its one-dimensional, two-dimensional, three-dimensional features, different dimensions of features are spliced together by a series of convolution, unfolding, finally, through a kind of using residual network, triangular multiplication update and axial attention mechanism deep neural network, the absolute difference of polymorphic protein inter-residue distance is predicted, and then the polymorphic protein inter-residue distance is obtained.The present application can effectively predict the polymorphic protein inter-residue distance, and has important significance for further predicting the conformational ensemble of polymorphic protein.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics and computer applications, and in particular to a method for predicting distances between polymorphic protein residues based on deep learning. Background Art

[0002] The structure of polymorphic proteins is not static. Organisms are constantly in a dynamic, non-equilibrium state, and a variety of factors influence the structure and motion of polymorphic proteins. Furthermore, polymorphic proteins require unique conformations to perform specific functions. Therefore, the structure of polymorphic proteins represents a set of conformations rather than a static structure. Determining the distances between polymorphic protein residues is crucial for constructing a polymorphic protein conformational set. With the rise of deep learning and its successful application in protein structure prediction, its powerful learning capabilities have made it possible to predict distances between polymorphic protein residues. Summary of the Invention

[0003] In order to overcome the problem that traditional methods have difficulty in predicting the distances between polymorphic protein residues, the present invention proposes a method for predicting the distances between polymorphic protein residues based on deep learning. A new method for predicting the distances between polymorphic protein residues is designed. One-dimensional, two-dimensional, and three-dimensional information is extracted from the input protein structure as the input of the network, and the absolute difference of the distances between polymorphic protein residues is predicted. Finally, the predicted distances between polymorphic protein residues are obtained, which is of great significance for revealing the conformational set of proteins and the interaction forces between residues.

[0004] The technical solution adopted by the present invention to solve its technical problem is:

[0005] A method for predicting distances between polymorphic protein residues based on deep learning, the method comprising the following steps:

[0006] 1) Constructing the dataset: In the protein flexibility database PDBflex, the sequences were 100% identical and the length was

[0007] N polymorphic proteins were screened as a data set based on the condition of 50-500 residues, and N were randomly selected from the constructed data set. a proteins as training set, N b proteins as the test set, and the remaining proteins as the validation set;

[0008] 2) Generate label data: Among the multiple structures of each polymorphic protein, find the two structures with the smallest TM-score by comparing them with each other, and use them as the two structures of the polymorphic protein. β The three-dimensional coordinates of the atoms are used as the three-dimensional coordinates of the residues, and the Euclidean distance between each pair of residues is calculated. If the residue does not contain Cβ Atoms are C α Atoms replace, construct each

[0009] The absolute difference in the distance between residue pairs of two protein structures is calculated as follows:

[0010]

[0011] in, Represents the distance between residue pairs (i, j) in a protein structure, The distance between the residue pair (i, j) of another structure is divided into 11 intervals according to the size of the distance error value, respectively.

[0012] for

[0013] These 11 intervals are labeled 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10. The corresponding number of the interval to which the distance between the residue pairs belongs is taken as the label value;

[0014] 3) Feature extraction: For each protein structure in the dataset, extract its one-dimensional, two-dimensional, and three-dimensional features and splice them;

[0015] 4) Build the network model: The network consists of three parts. The first part consists of three 3D convolutional layers, the second part consists of a network based on triangular multiplication updates and axial attention mechanism, and the third part is a residual network with 10 residual blocks.

[0016] 5) Training model parameters: Use Xavier to initialize network parameters and Adam Optimizer to optimize network weights.

[0017] Re-optimize and use multivariate cross entropy CrossEntropyLoss as the loss function to optimize network parameters; 6) Obtain the distance between protein residues: First extract features from the input structure, input them into the network, and output the results

[0018] The distribution of the absolute difference of distances between all residues is obtained by using the argmax function to obtain the interval with the maximum probability. The average value of the interval is used as the absolute difference of distances between residues. The predicted distances between polymorphic protein residues are as follows:

[0019] Within the following range:

[0020]

[0021] Among them, d ij Represents the distance between residue pairs (i, j) of the input structure, pdif ij Represents the predicted residue pair (i,

[0022] j) Absolute difference in distance.

[0023] Further, the process of step 3) is as follows:

[0024] 3.1) Use pyRosetta to obtain the monomer energy terms (p_aa_pp, rama_prepro, omega, fa_dun) and secondary structure (L, E, F) of each protein structure, and obtain a one-dimensional feature of L×7;

[0025] 3.2) Use pyRosetta to obtain the residue pair energy terms (fa_atr, fa_rep, fa_sol, lk_ball_wtd, fa_elec, hbond_bb_sc, hbond_sc), orientations (ω,

[0026] θ), calculate the C between each pair of residues based on the three-dimensional coordinates of the atoms in each residue α Atom-C α Atom, C α Atom-Tip Atom, Tip Atom-C α The Euclidean distances of atoms are merged to obtain a two-dimensional feature of size L × L × 13;

[0027] 3.3) The direction from the C atom of the previous residue to the N atom of the current residue is Axis direction, perpendicular to Axis direction and make the C of the current residue α Atoms in a plane Established on Axis, by Determined direction definition Axis, establish the local three-dimensional coordinate system of the current residue, and obtain the current coordinate system The spatial coordinates of all atoms in The large cube space is divided into 20×20×20 equal sides. Each cube is assigned a value v, and the calculation formula is as follows:

[0028]

[0029] Where N is the number of atoms in the large cube, Represents the coordinates of the center of the small cube, represents the coordinates of the i-th atom, t i Represents a 167-dimensional unit vector, used to represent the type of atom, thus obtaining a three-dimensional feature of L×20×20×20×167;

[0030] 3.4) The three-dimensional features obtained in step 3.3) are flattened into one-dimensional features of L×640 after a series of three-dimensional convolutions, and then concatenated with the one-dimensional features obtained in step 3.1) to form features of L×647. After horizontal and vertical striping, they are concatenated with the two-dimensional features obtained in step 3.2) to obtain features with a final dimension of L×L×141.

[0031] Furthermore, the process of step 4) is as follows:

[0032] 4.1) The first part of step 3) consists of three 3D convolutional layers with 20, 30, and 20 kernels of 1×1×1, 3×3×3, and 4×4×4 kernel sizes, respectively. After passing through the elu activation function, the network finally passes through a 3D average pooling layer with a kernel size of 4×4×4 and a stride of 4.

[0033] 4.2) The second part consists of a network based on triangular multiplication update and axial attention mechanism. The triangular multiplication update rule is as follows:

[0034]

[0035]

[0036]

[0037]

[0038] Among them, x ij represents the residue pair (i, j), W1...W6 represent the linear transformation;

[0039] After the triangular multiplication update, there are two axial attention layers. The formula of axial attention is as follows:

[0040]

[0041] q ij =W q x ij

[0042] k ik =W k x ik

[0043] v ik =W v x ik

[0044] Among them, x ij represents the residue pair (i, j), W o 、W q 、Wk 、W v represents linear transformation, c represents the number of channels, h represents the number of attention heads, and f represents the softmax function;

[0045] 4.3) The third part is a residual network with 10 residual blocks, and the cyclic expansion rates are set to 1, 2, 4, 8, and 16; each residual block consists of three convolutional layers and three normalization layers; the output layer uses a softmax nonlinear function.

[0046] The beneficial effects of the present invention are mainly manifested in: utilizing the powerful information extraction ability of the deep learning network, capturing the information on the distance between polymorphic protein residues contained in the one-dimensional, two-dimensional, and three-dimensional characteristics of the protein structure, and being able to more accurately predict the distance between polymorphic protein residues. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is an overall flow chart of a deep learning-based method for predicting the distance between polymorphic protein residues.

[0048] Figure 2 It is a polymorphic protein residue distance prediction method based on deep learning to predict the distance between protein structures 1I74 and obtain the inter-residue distance deviation map. DETAILED DESCRIPTION

[0049] The present invention will be further described below with reference to the accompanying drawings.

[0050] Reference Figure 1 and Figure 2 , a method for predicting distances between polymorphic protein residues based on deep learning, comprising the following steps:

[0051] 1) Constructing the dataset: In the protein flexibility database PDBflex, the sequences were 100% identical and the length was

[0052] N polymorphic proteins were screened as a data set based on the condition of 50-500 residues, and N were randomly selected from the constructed data set. a proteins as training set, N b proteins as the test set, and the remaining proteins as the validation set;

[0053] 2) Generate label data: Among the multiple structures of each polymorphic protein, find the two structures with the smallest TM-score by comparing them with each other, and use them as the two structures of the polymorphic protein. β The three-dimensional coordinates of the atoms are used as the three-dimensional coordinates of the residues, and the Euclidean distance between each pair of residues is calculated. If the residue does not contain C β Atoms are C α Atoms replace, construct each

[0054] The absolute difference in the distance between residue pairs of two protein structures is calculated as follows:

[0055]

[0056] in, Represents the distance between residue pairs (i, j) in a protein structure, The distance between the residue pair (i, j) of another structure is divided into 11 intervals according to the size of the distance error value, respectively.

[0057] for

[0058] These 11 intervals are labeled 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10. The corresponding number of the interval to which the distance between the residue pairs belongs is taken as the label value;

[0059] 3) Feature extraction: For each protein structure in the dataset, extract its one-dimensional, two-dimensional, and three-dimensional features and splice them together. The process is as follows:

[0060] 3.1) Use pyRosetta to obtain the monomer energy terms (p_aa_pp, rama_prepro, omega, fa_dun) and secondary structure (L, E, F) of each protein structure, and obtain a one-dimensional feature of L×7;

[0061] 3.2) Use pyRosetta to obtain the residue pair energy terms (fa_atr, fa_rep, fa_sol, lk_ball_wtd, fa_elec, hbond_bb_sc, hbond_sc), orientations (ω,

[0062] θ), calculate the C between each pair of residues based on the three-dimensional coordinates of the atoms in each residue α Atom-C α Atom, C α Atom-Tip Atom, Tip Atom-C α The Euclidean distances of atoms are merged to obtain a two-dimensional feature of size L × L × 13;

[0063] 3.3) The direction from the C atom of the previous residue to the N atom of the current residue is Axis direction, perpendicular to Axis direction and make the C of the current residue α Atoms in a plane Established on Axis, by

[0064] Determined direction definition Axis, establish the local three-dimensional coordinate system of the current residue, and obtain the current coordinate system The spatial coordinates of all atoms in The large cube space is divided into 20×20×20 equal sides. Each cube is assigned a value v, and the calculation formula is as follows:

[0065]

[0066] Where N is the number of atoms in the large cube, represents the coordinates of the center of the small cube, b represents the coordinates of the i-th atom, t i Represents a 167-dimensional unit vector, used to represent the type of atom, thus obtaining a three-dimensional feature of L×20×20×20×167;

[0067] 3.4) The 3D features obtained in step 3.3) are flattened into a 1D feature of size L × 640 after a series of 3D convolutions. These features are then concatenated with the 1D features obtained in step 3.1) to form a 1D feature of size L × 647. After horizontal and vertical striping, they are concatenated with the 2D features obtained in step 3.2) to obtain a feature with a final dimension of size L × L × 141.

[0068] 4) Build the network model: The network consists of three parts, and the process is as follows:

[0069] 4.1) The first part of step 3) consists of three 3D convolutional layers with 20, 30, and 20 kernels of 1×1×1, 3×3×3, and 4×4×4 kernel sizes, respectively. After passing through the elu activation function, the network finally passes through a 3D average pooling layer with a kernel size of 4×4×4 and a stride of 4.

[0070] 4.2) The second part consists of a network based on triangular multiplication update and axial attention mechanism. The triangular multiplication update rule is as follows:

[0071]

[0072]

[0073]

[0074]

[0075] Among them, x ij represents the residue pair (i, j), W1...W6 represent the linear transformation;

[0076] After the triangular multiplication update, there are two axial attention layers. The formula of axial attention is as follows:

[0077]

[0078] q ij =W q x ij

[0079] k ik =W k x ik

[0080] v ik =W v x ik

[0081] Among them, x ij represents the residue pair (i, j), W o 、W q 、W k 、W v represents linear transformation, c represents the number of channels, h represents the number of attention heads, and f represents the softmax function;

[0082] 4.3) The third part is a residual network with 10 residual blocks, and the cyclic expansion rates are set to 1, 2, 4, 8, and 16. Each residual block consists of three convolutional layers and three normalization layers. The output layer uses a softmax nonlinear function.

[0083] 5) Training model parameters: Initialize network parameters using Xavier, optimize network weights using Adam Optimizer, and use multivariate cross entropy CrossEntropyLoss as the loss function to optimize network parameters; 6) Obtain the distance between protein residues: First, extract features from the input structure, input them into the network, and output the results

[0084] The distribution of the absolute difference of distances between all residues is obtained by using the argmax function to obtain the interval with the maximum probability. The average value of the interval is used as the absolute difference of distances between residues. The predicted distances between polymorphic protein residues are as follows:

[0085] Within the following range:

[0086] pdist ij ∈[d ij -pdif ij ,d ij +pdif ij ]

[0087] Among them, d ij Represents the distance between residue pairs (i, j) of the input structure, pdifij Represents the predicted residue pair (i,

[0088] j) Absolute difference in distance.

[0089] This example uses the protein structure 1174 with an amino acid sequence length of 309 as an example. A method for predicting distances between polymorphic protein residues based on deep learning includes the following steps:

[0090] 1) Constructing the dataset: In the protein flexibility database PDBflex, the sequences were 100% identical and the length was

[0091] N = 17430 polymorphic proteins were screened as a data set based on the condition of 50-500 residues, and N were randomly selected from the constructed data set. a = 15687 proteins as training set, N b =871 proteins as the test set, and the remaining 872 proteins as the validation set;

[0092] 2) Generate label data: Among the multiple structures of each polymorphic protein, find the two structures with the smallest TM-score by comparing them with each other, and use them as the two structures of the polymorphic protein. β The three-dimensional coordinates of the atoms are used as the three-dimensional coordinates of the residues, and the Euclidean distance between each pair of residues is calculated. If the residue does not contain C β Atoms are C α Atoms replace, construct each

[0093] The absolute difference in the distance between residue pairs of two protein structures is calculated as follows:

[0094]

[0095] in, Represents the distance between residue pairs (i, j) in a protein structure, The distance between the residue pair (i, j) of another structure is divided into 11 intervals according to the size of the distance error value, respectively.

[0096] for

[0097] These 11 intervals are labeled 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10. The corresponding number of the interval to which the distance between the residue pairs belongs is taken as the label value;

[0098] 3) Feature extraction: For each protein structure in the dataset, extract its one-dimensional, two-dimensional, and three-dimensional features and splice them together. The process is as follows:

[0099] 3.1) Use pyRosetta to obtain the monomer energy terms (p_aa_pp, rama_prepro, omega, fa_dun) and secondary structure (L, E, F) of each protein structure, and obtain a one-dimensional feature of L×7;

[0100] 3.2) Use pyRosetta to obtain the residue pair energy terms (fa_atr, fa_rep,

[0101] fa_sol, lk_ball_wtd, fa_elec, hbond_bb_sc, hbond_sc), orientations(ω,

[0102] θ), calculate the C between each pair of residues based on the three-dimensional coordinates of the atoms in each residue α Atom-C α Atom, C α Atom-Tip Atom, Tip Atom-C α The Euclidean distances of atoms are merged to obtain a two-dimensional feature of size L × L × 13;

[0103] 3.3) The direction from the C atom of the previous residue to the N atom of the current residue is Axis direction, perpendicular to Axis direction and make the C of the current residue α Atoms in a plane Established on Axis, by

[0104] Determined direction definition Axis, establish the local three-dimensional coordinate system of the current residue, and obtain the current coordinate system The spatial coordinates of all atoms in The large cube space is divided into 20×20×20 equal sides. Each cube is assigned a value v, and the calculation formula is as follows:

[0105]

[0106] Where N is the number of atoms in the large cube, represents the coordinates of the center of the small cube, b represents the coordinates of the i-th atom, t i Represents a 167-dimensional unit vector, used to represent the type of atom, thus obtaining a three-dimensional feature of L×20×20×20×167;

[0107] 3.4) The 3D features obtained in step 3.3) are flattened into a 1D feature of size L × 640 after a series of 3D convolutions. These features are then concatenated with the 1D features obtained in step 3.1) to form a 1D feature of size L × 647. After horizontal and vertical striping, they are concatenated with the 2D features obtained in step 3.2) to obtain a feature with a final dimension of size L × L × 141.

[0108] 4) Build the network model: The network consists of three parts, and the process is as follows:

[0109] 4.1) The first part of step 3) consists of three 3D convolutional layers with 20, 30, and 20 kernels of 1×1×1, 3×3×3, and 4×4×4 kernel sizes, respectively. After passing through the elu activation function, the network finally passes through a 3D average pooling layer with a kernel size of 4×4×4 and a stride of 4.

[0110] 4.2) The second part consists of a network based on triangular multiplication update and axial attention mechanism. The triangular multiplication update rule is as follows:

[0111]

[0112]

[0113]

[0114]

[0115] Among them, x ij represents the residue pair (i, j), W1...W6 represent the linear transformation;

[0116] After the triangular multiplication update, there are two axial attention layers. The formula of axial attention is as follows:

[0117]

[0118] q ij =W q x ij

[0119] k ik =W k x ik

[0120] v ik =W v x ik

[0121] Among them, x ij represents the residue pair (i, j), W o 、W q 、W k 、Wv represents linear transformation, c represents the number of channels, h represents the number of attention heads, and f represents the softmax function;

[0122] 4.3) The third part is a residual network with 10 residual blocks, and the cyclic expansion rates are set to 1, 2, 4, 8, and 16. Each residual block consists of three convolutional layers and three normalization layers. The output layer uses a softmax nonlinear function.

[0123] 5) Training model parameters: Use Xavier to initialize network parameters, use Adam Optimizer to optimize network weights, and use multivariate cross entropy CrossEntropyLoss as the loss function to optimize network parameters;

[0124] 6) Obtaining the distance between protein residues: First, extract features from the input structure and input them into the network. The output is the distribution of the absolute difference of distances between all residues. The argmax function is used to obtain the interval with the maximum probability. The average value of the interval is used as the absolute difference of distance between residues. The predicted distance between polymorphic protein residues is within the following interval:

[0125] pdist ij ∈[d ij -pdif ij ,d ij +pdif ij ]

[0126] Among them, d ij Represents the distance between residue pairs (i, j) of the input structure, pdif ij represents the absolute difference in distance between the predicted residue pair (i, j).

[0127] Taking the protein 1I74 with an amino acid sequence length of 309 as an example, the distance between the residues of the protein was obtained using the above method. The prediction results are as follows: Figure 2 shown.

[0128] The above description is about the excellent effect shown by an embodiment of the present invention. Obviously, the present invention is not only suitable for the above embodiment, but also can be implemented with various changes without departing from the basic spirit of the present invention and without exceeding the content involved in the essential content of the present invention.

Claims

1. A method for predicting distances between polymorphic protein residues based on deep learning, characterized in that: The method comprises the following steps: 1) Constructing a dataset: N polymorphic proteins with 100% sequence identity and length between 50 and 500 residues were selected from the protein flexibility database PDBflex as a dataset. N proteins were randomly selected from the constructed dataset. a proteins as training set, N b proteins as the test set, and the remaining proteins as the validation set; 2) Generate label data: Among the multiple structures of each polymorphic protein, find the two structures with the smallest TM-score by comparing them with each other, and use them as the two structures of the polymorphic protein. β The three-dimensional coordinates of the atoms are used as the three-dimensional coordinates of the residues, and the Euclidean distance between each pair of residues is calculated. If the residue does not contain C β Atoms are C α Atom replacement is performed to construct the absolute difference in the distance between the residue pairs of each protein structure, which is calculated as follows: in, Represents the distance between residue pairs (i, j) in a protein structure, The distance between the residue pair (i, j) of another structure is divided into 11 intervals according to the size of the distance error value, which are These 11 intervals are numbered 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, and the residue pairs The distance between the two intervals belongs to which interval and the corresponding label is taken as the label value; 3) Feature extraction: For each protein structure in the dataset, extract its one-dimensional, two-dimensional, and three-dimensional features and splice them; 4) Build the network model: The network consists of three parts. The first part consists of three 3D convolutional layers, the second part consists of a network based on triangular multiplication updates and axial attention mechanism, and the third part is a residual network with 10 residual blocks. 5) Training model parameters: Use Xavier to initialize network parameters, use Adam Optimizer to optimize network weights, and use multivariate cross entropy CrossEntropyLoss as the loss function to optimize network parameters; 6) Obtaining the distance between protein residues: First, extract features from the input structure and input them into the network. The output is the distribution of the absolute difference of distances between all residues. The argmax function is used to obtain the interval with the maximum probability. The average value of the interval is used as the absolute difference of distance between residues. The predicted distance between polymorphic protein residues is within the following interval: pdist ij ∈[d ij -pdif ij ,d ij +pdif ij ] Among them, d ij Represents the distance between residue pairs (i, j) of the input structure, pdif ij represents the absolute difference in distance between the predicted residue pair (i, j).

2. The method for predicting distances between polymorphic protein residues based on deep learning according to claim 1, wherein: The process of step 3) is as follows: 3.1) Use pyRosetta to obtain the monomer energy terms (p_aa_pp, rama_prepro, omega, fa_dun) and secondary structure (L, E, F) of each protein structure, and obtain a one-dimensional feature of L×7; 3.2) Use pyRosetta to obtain the residue pair energy terms (fa_atr, fa_rep, fa_sol, lk_ball_wtd, fa_elec, hbond_bb_sc, hbond_sc), orientations for each protein structure The C between each pair of residues is calculated based on the three-dimensional coordinates of the atoms in each residue. α Atom-C α Atom, C α Atom-Tip Atom, Tip Atom-C α The Euclidean distances of atoms are merged to obtain a two-dimensional feature of size L × L × 13; 3.3) The direction from the C atom of the previous residue to the N atom of the current residue is Axis direction, perpendicular to Axis direction and make the C of the current residue α Atoms in a plane Established on Axis, by Determined direction definition Axis, establish the local three-dimensional coordinate system of the current residue, and obtain the current coordinate system The spatial coordinates of all atoms in The large cube space is divided into 20×20×20 equal sides. Each cube is assigned a value v, and the calculation formula is as follows: Where N is the number of atoms in the large cube, Represents the coordinates of the center of the small cube, represents the coordinates of the i-th atom, t i Represents a 167-dimensional unit vector, used to represent the type of atom, thus obtaining a three-dimensional feature of L×20×20×20×167; 3.4) The three-dimensional features obtained in step 3.3) are flattened into one-dimensional features of L×640 after a series of three-dimensional convolutions, and then concatenated with the one-dimensional features obtained in step 3.1) to form features of L×647. After horizontal and vertical striping, they are concatenated with the two-dimensional features obtained in step 3.2) to obtain features with a final dimension of L×L×141.

3. The method for predicting distances between polymorphic protein residues based on deep learning according to claim 1 or 2, wherein: The process of step 4) is as follows: 4.1) The first part of step 3) consists of three 3D convolutional layers with 20, 30, and 20 kernels of 1×1×1, 3×3×3, and 4×4×4 kernel sizes, respectively. After passing through the elu activation function, the network finally passes through a 3D average pooling layer with a kernel size of 4×4×4 and a stride of 4. 4.2) The second part consists of a network based on triangular multiplication update and axial attention mechanism. The triangular multiplication update rule is as follows: Among them, x ij represents the residue pair (i, j), W1...W6 represent the linear transformation; After the triangular multiplication update, there are two axial attention layers. The formula of axial attention is as follows: q ij =W q x ij k ik =W k x ik v ik =W v x ik Among them, x ij represents the residue pair (i, j), W o 、W q 、W k 、W v represents linear transformation, c represents the number of channels, h represents the number of attention heads, and f represents the softmax function; 4.3) The third part is a residual network with 10 residual blocks, and the cyclic expansion rates are set to 1, 2, 4, 8, and 16; each residual block consists of three convolutional layers and three normalization layers; the output layer uses a softmax nonlinear function.

Images