A deep learning inter-domain constraint assisted multi-domain protein structure assembly method

Abstract

The application discloses a deep learning inter-domain constraint auxiliary multi-domain protein structure assembly method, which comprises the following steps: firstly, inputting the three-dimensional structures of all single domains of a multi-domain protein to be assembled; then, scoring and searching local templates by aligning each template in a multi-domain protein structure library with each two continuous domains, and scoring and searching global templates according to the alignment of each template with each single domain; selecting T local templates and global templates with the highest scores as assembly templates; predicting the constraint information between the domains by using DeepPotential, so as to calculate a mixed energy function of the multi-domain protein; for each template, performing N-step independent L-BFGS simulation according to the mixed energy function; generating a final full-length model by using the translation vector and the rotation angle of each domain when the energy is the lowest; and finally selecting the structure with the lowest energy in the global templates of the T local templates as the final assembly structure. The application provides a multi-domain protein structure assembly method with high precision.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics, intelligent optimization, and computer applications, and relates to a deep learning-assisted method for assembling multi-domain protein structures. Background Technology

[0002] Proteins are the building blocks of biological processes and perform a variety of key functions in biological systems, such as DNA transcription, mRNA translation, and signal transduction. Moreover, protein formation sometimes activates or inhibits one or more other proteins. Damage to or dysregulation of these proteins often leads to disease.

[0003] Elucidating the three-dimensional structure of proteins is a crucial step in understanding their full functions and interaction mechanisms. This not only advances clinical and basic research but also expands our understanding of the functional roles of proteins in normal biological processes and under pathological conditions. In current protein libraries, the majority of proteins are considered single-domain proteins, with only 34.7% being multi-domain proteins. Multi-domain proteins have more complex structures than single-domain proteins, and their functions play a more important role in biological systems. However, almost all current protein structure prediction methods are designed for single-domain proteins, and our understanding and research on multi-domain proteins are far less extensive. Therefore, predicting the structure of multi-domain proteins has become an important and unavoidable problem in computational biology.

[0004] Starting from the sequence, the entire protein sequence is divided into multiple domains. Then, the structure of each domain is predicted using methods for predicting single-domain protein structures. Finally, the single-domain structures are assembled together to obtain the corresponding full-length structure. This is one of the effective methods for predicting multi-domain protein structures. However, because the above method is not very precise in terms of the constraint information between the domains and lacks effective energy guidance, it is impossible to accurately determine the assembly direction and rotation angle of the domain proteins during the assembly process, resulting in low prediction accuracy.

[0005] Therefore, existing multi-domain protein structure assembly methods have shortcomings in terms of assembly accuracy and need to be improved. Summary of the Invention

[0006] To overcome the shortcomings of existing multi-domain protein structure assembly methods in terms of low accuracy, this invention provides a deep learning-assisted inter-domain constraint multi-domain protein structure assembly method with high assembly accuracy.

[0007] The technical solution adopted by this invention to solve its technical problem is:

[0008] A deep learning-assisted method for assembling multi-domain protein structures with inter-domain constraints is proposed. First, the three-dimensional structures of all single domains of the multi-domain protein to be assembled are input. Then, local templates are obtained by scoring the alignment of each template with every two consecutive domains in the multi-domain protein structure library, and global templates are obtained by scoring the alignment of each template with each single domain. The T highest-scoring local templates and the global template are selected as assembly templates. DeepPotential is used to predict the constraint information between the domains, thereby calculating the mixture energy function of the multi-domain protein. For each template, N independent L-BFGS simulations are performed based on the mixture energy function. The translation vector and rotation angle of each domain at the lowest energy are used to generate the final full-length model. Finally, the structure with the lowest energy among the T local templates and the global template is selected as the final assembled structure.

[0009] Furthermore, the method includes the following steps:

[0010] 1) Input the three-dimensional structures of all single domains of the multi-domain protein to be assembled;

[0011] 2) Setting parameters: Conflict distance threshold d cut The L-BFGS quasi-Newton algorithm performs N independent simulations and assembles T templates.

[0012] 3) Using the protein structure alignment tool TM-align (https: / / zhanggroup.org / TM-align / ), the alignment of each template with every two consecutive domains in the multi-domain protein structure library is scored. The score of each local template is the average of the TM-scores between each domain and the template.

[0013] 4) Scoring is performed on the alignment of each template with each single domain in the multi-domain protein structure library. The score of each global template is the harmonic mean of the TM-score of the alignment between each domain and the template. The T local templates and the global template with the highest scores are selected as assembly templates respectively.

[0014] 5) Use the residue constraint prediction tool DeepPotential (https: / / zhanggroup.org / DeepPotential / ) to predict the rotation angle, hydrogen bond network, residue distance, and the confidence level U corresponding to the residue constraints between each domain. ij i = 1, 2, ..., L m j = 1, 2, ..., L n L m L n U represents the sequence length of the m-th and n-th structural domains. ijThis represents the confidence level of the constraint between the i-th residue in the m-th domain and the j-th residue in the n-th domain;

[0015] 6) Assemble according to each selected template, as follows:

[0016] 6.1) Calculate the structural domain C β Energy of distance between atoms:

[0017]

[0018] Where d ij C is in the i-th structural domain β Atom and C in the j-th structural domain β The distance between atoms, (Pi, j, k(d) ij )) is predicted by DeepPotential d ij The probability of falling into the k-th interval is ε = 10. -4 A very small positive number prevents the probability from being 0;

[0019] 6.2) Calculate the directional energy of the structural domain:

[0020]

[0021] Among them O ij P(i,j,k(d) represents the dihedral angle between residues in the domain. ij The O predicted by DeepPotential is... ij The probability of falling into the k-th angle interval;

[0022] 6.3) Calculate the contact energy between structural domains:

[0023]

[0024] 6.4) Calculate the hydrogen bond confinement energy between structural domains:

[0025]

[0026] in The calculation is as follows:

[0027]

[0028] in This represents the angle of the hydrogen bond between the i-th residue of the m-th domain and the j-th residue of the n-th domain. Indicates the corresponding angle is The probability of;

[0029] 6.5) Calculate the conflict energy between structural domains:

[0030] 6.6) Calculate the general contact energy between structural domains:

[0031]

[0032] Where, if the i-th residue and the j-th residue are hydrophobic groups, then u ij =0.1, if the i-th residue and the j-th residue are hydrophilic groups, then u ij =0.01, otherwise u ij =0.05;

[0033] 6.7) Calculate the inter-domain distance spectrum energy:

[0034]

[0035] Where T ij It is the number of templates; where a i and a j It is the index of the i-th residue and the j-th residue aligned to the template, and satisfies: 0.6|ij|<|a i -a j |<1.5|ij|,D t ij is in the t-th template a i and a j C in residues α The distance between atoms;

[0036] 6.8) Calculate the domain connectivity energy: E db (m, n) = (b mn -b0) 2 , where b mn C is the connection between the endpoints of two continuous domains. α The distance between atoms It is C α The standard distance between atoms;

[0037] 6.9) Calculate the template constraint energy term: Where d(S) i S′ i ) represents the i-th C in the initial structure generated during the template alignment process. α Atom (S) i ) and the corresponding C in the conformation generated during the assembly process α Atom (S′) i The distance between the two proteins, L, is the protein length;

[0038] 6.10) Calculate the energy at the radius of rotation:

[0039]

[0040] Where R decoy R is the radius of rotation of the conformation. min =2.849L 0.319 R is the estimated minimum radius of rotation. max =max{R min +7.5, 0.55N mh} represents the estimated maximum radius of rotation, N. mh It is the number of the longest helical residues in the conformation;

[0041] 6.11) Calculate the total energy E of the current structure:

[0042]

[0043] 6.12) Assign weights to the mixed energy function;

[0044] 6.13) Perform N independent L-BFGS simulations on the hybrid energy function to find the optimal rotation angles φ, θ, ψ and translation vectors (t1, t2, t3) for each structural domain, and generate the final full-length model using the translation vector and rotation matrix at the lowest total energy.

[0045] 7) Select the full-length structure with the lowest energy generated by the global template based on T local templates as the final assembly structure.

[0046] Furthermore, the process of step 6.12) is as follows:

[0047] 6.12.1) If the score of this template is less than 0.85, then the weights are: w1 = 5, w2 = 1, w3 = 3, w4 = 1.2, w5 = 0.2, w6 = 1.0, w7 = 0.02, w8 = 0.01, w9 = 0.15, w 10 =0.13;

[0048] 6.12.2) Otherwise, the weights are: w1 = 1, w2 = 0.2, w3 = 0.2, w4 = 0.15, w5 = 0.15, w6 = 0.1, w7 = 0.02, w8 = 0.01, w9 = 1.2, w 10 =0.12.

[0049] Furthermore, the process of step 6.13) is as follows:

[0050] 6.13.1) Generate the rotation matrix T based on the inter-domain rotation angles φ, θ, and ψ. rot :

[0051]

[0052] Where a iji = 1, 2, 3, j = 1, 2, 3 are the elements of the rotation matrix;

[0053] 6.13.2) Generate the translation matrix T = (x m y m , z m ):

[0054]

[0055] Where (t1, t2, t3) is the translation vector of the domain, and (x0, y0, z0) is the initial position of the m-th atom. m y m , z m ) is the new position of the m-th atom after the transformation, (x c y c , z c ) is the center point of the structural domain;

[0056] 6.12.3) Calculate the position of each atom in each domain according to the corresponding rotation angle and translation vector to generate a new full-length structure.

[0057] The beneficial effects of this invention are as follows: on the one hand, by using local templates and global templates simultaneously as assembly templates, the initial structure is generated in a coordinated manner, thereby improving the accuracy of the initial structure; on the other hand, by using the residue constraint prediction tool DeepPotential to predict the constraint information between various domains to guide the entire assembly process, the accuracy of the final assembled structure is improved. Attached Figure Description

[0058] Figure 1 This is a basic flowchart of a deep learning-assisted inter-domain protein structure assembly method.

[0059] Figure 2 This is the result of assembling a protein 1h88C containing three domains using a deep learning-assisted inter-domain constraint method for assembling multi-domain protein structures. Detailed Implementation

[0060] The present invention will now be further described with reference to the accompanying drawings.

[0061] Reference Figure 1 and Figure 2 A deep learning-assisted method for assembling multi-domain protein structures based on inter-domain constraints includes the following steps:

[0062] 1) Input the three-dimensional structures of all single domains of the multi-domain protein to be assembled;

[0063] 2) Setting parameters: Conflict distance threshold d cutThe L-BFGS quasi-Newton algorithm performs N independent simulations and assembles T templates.

[0064] 3) Using the protein structure alignment tool TM-align (https: / / zhanggroup.org / TM-align / ), the alignment of each template with every two consecutive domains in the multi-domain protein structure library is scored. The score of each local template is the average of the TM-scores between each domain and the template.

[0065] 4) Scoring is performed on the alignment of each template with each single domain in the multi-domain protein structure library. The score of each global template is the harmonic mean of the TM-score of the alignment between each domain and the template. The T local templates and the global template with the highest scores are selected as assembly templates respectively.

[0066] 5) Use the residue constraint prediction tool DeepPotential (https: / / zhanggroup.org / DeepPotential / ) to predict the rotation angle, hydrogen bond network, residue distance, and the confidence level U corresponding to the residue constraints between each domain. ij i = 1, 2, ..., L m j = 1, 2, ..., L n L m L n U represents the sequence length of the m-th and n-th structural domains. ij This represents the confidence level of the constraint between the i-th residue in the m-th domain and the j-th residue in the n-th domain;

[0067] 6) Assemble according to each selected template, as follows:

[0068] 6.1) Calculate the structural domain C β Energy of distance between atoms:

[0069]

[0070] Where d ij C is in the i-th structural domain β Atom and C in the j-th structural domain β The distance between atoms, P(i,j,k(d) ij )) is predicted by DeepPotential d ij The probability of falling into the k-th interval is ε = 10. -4 A very small positive number prevents the probability from being 0;

[0071] 6.2) Calculate the directional energy of the structural domain:

[0072]

[0073] Among them O ij P(i,j,k(d) represents the dihedral angle between residues in the domain. ij The O predicted by DeepPotential is... ij The probability of falling into the k-th angle interval;

[0074] 6.3) Calculate the contact energy between structural domains:

[0075]

[0076] 6.4) Calculate the hydrogen bond confinement energy between structural domains:

[0077]

[0078] in The calculation is as follows:

[0079]

[0080] in This represents the angle of the hydrogen bond between the i-th residue of the m-th domain and the j-th residue of the n-th domain. Indicates the corresponding angle is The probability of;

[0081] 6.5) Calculate the conflict energy between structural domains:

[0082] 6.6) Calculate the general contact energy between structural domains:

[0083]

[0084] Where, if the i-th residue and the j-th residue are hydrophobic groups, then u ij =0.1, if the i-th residue and the j-th residue are hydrophilic groups, then u ij =0.01, otherwise u ij =0.05;

[0085] 6.7) Calculate the inter-domain distance spectrum energy:

[0086]

[0087] Where T ij It is the number of templates; where a i and a j It is the index of the i-th residue and the j-th residue aligned to the template, and satisfies: 0.6|ij|<|a i -a j |<1.5|ij|,Dt ij is in the t-th template a i and a j C in residues α The distance between atoms;

[0088] 6.8) Calculate the domain connectivity energy: E db (m, n) = (b mn -b0) 2 , where b mn C is the connection between the endpoints of two continuous domains. α The distance between atoms It is C α The standard distance between atoms;

[0089] 6.9) Calculate the template constraint energy term: Where d(S) i S′ i ) represents the i-th C in the initial structure generated during the template alignment process. α Atom (S) i ) and the corresponding C in the conformation generated during the assembly process α Atom (S′) i The distance between the two proteins, L, is the protein length;

[0090] 6.10) Calculate the energy at the radius of rotation:

[0091]

[0092] Where R decoy R is the radius of rotation of the conformation. min =2.849L 0.319 R is the estimated minimum radius of rotation. max =max{R min +7.5, 0.55N mh For the estimated maximum radius of rotation, N mh It is the number of the longest helical residues in the conformation;

[0093] 6.11) Calculate the total energy E of the current structure:

[0094]

[0095] 6.12) Assign weights to the mixed energy function;

[0096] 6.13) Perform N independent L-BFGS simulations on the hybrid energy function to find the optimal rotation angles φ, θ, ψ and translation vectors (t1, t2, t3) for each structural domain, and generate the final full-length model using the translation vector and rotation matrix at the lowest total energy.

[0097] 7) Select the full-length structure with the lowest energy generated by the global template based on T local templates as the final assembly structure.

[0098] Furthermore, the process of step 6.12) is as follows:

[0099] 6.12.1) If the score of this template is less than 0.85, then the weights are: w1 = 5, w2 = 1, w3 = 3, w4 = 1.2, w5 = 0.2, w6 = 1.0, w7 = 0.02, w8 = 0.01, w9 = 0.15, w 10 =0.13;

[0100] 6.12.2) Otherwise, the weights are: w1 = 1, w2 = 0.2, w3 = 0.2, w4 = 0.15, w5 = 0.15, w6 = 0.1, w7 = 0.02, w8 = 0.01, w9 = 1.2, w 10 =0.12.

[0101] Furthermore, the process of step 6.13) is as follows:

[0102] 6.13.1) Generate the rotation matrix T based on the inter-domain rotation angles φ, θ, and ψ. rot :

[0103]

[0104] Where a ij i = 1, 2, 3, j = 1, 2, 3 are the elements of the rotation matrix;

[0105] 6.13.2) Generate the translation matrix T = (x m y m , z m ):

[0106]

[0107] Where (t1, t2, t3) is the translation vector of the domain, and (x0, y0, z0) is the initial position of the m-th atom. m y m , z m ) is the new position of the m-th atom after the transformation, (x c y c , z c ) is the center point of the structural domain;

[0108] 6.12.3) Calculate the position of each atom in each domain according to the corresponding rotation angle and translation vector to generate a new full-length structure.

[0109] This embodiment uses the multi-domain protein 1h88C, which contains three structural domains, as an example. A deep learning-assisted inter-domain constraint method for assembling multi-domain protein structures includes the following steps:

[0110] 1) Input the three-dimensional structures of all single domains of the multi-domain protein to be assembled;

[0111] 2) Setting parameters: Conflict distance threshold The L-BFGS quasi-Newton algorithm has 200 independent simulations and 10 assembled templates.

[0112] 3) Using the protein structure alignment tool TM-align (https: / / zhanggroup.org / TM-align / ), the alignment of each template with every two consecutive domains in the multi-domain protein structure library is scored. The score of each local template is the average of the TM-scores between each domain and the template.

[0113] 4) Scoring is performed on the alignment of each template with each single domain in the multi-domain protein structure library. The score of each global template is the harmonic mean of the TM-score of the alignment between each domain and the template. The T local templates and the global template with the highest scores are selected as assembly templates respectively.

[0114] 5) Use the residue constraint prediction tool DeepPotential (https: / / zhanggroup.org / DeepPotential / ) to predict the rotation angle, hydrogen bond network, residue distance, and the confidence level U corresponding to the residue constraints between each domain. ij i = 1, 2, ..., L m j = 1, 2, ..., L n L m L n U represents the sequence length of the m-th and n-th structural domains. ij This represents the confidence level of the constraint between the i-th residue in the m-th domain and the j-th residue in the n-th domain;

[0115] 6) Assemble according to each selected template, as follows:

[0116] 6.1) Calculate the structural domain C β Energy of distance between atoms:

[0117]

[0118] Where d ij C is in the i-th structural domain β Atom and C in the j-th structural domain βThe distance between atoms, P(i,j,k(d) ij )) is predicted by DeepPotential d ij The probability of falling into the k-th interval is ε = 10. -4 A very small positive number prevents the probability from being 0;

[0119] 6.2) Calculate the directional energy of the structural domain:

[0120]

[0121] Among them O ij P(i,j,k(d) represents the dihedral angle between residues in the domain. ij The O predicted by DeepPotential is... ij The probability of falling into the k-th angle interval;

[0122] 6.3) Calculate the contact energy between structural domains:

[0123]

[0124] 6.4) Calculate the hydrogen bond confinement energy between structural domains:

[0125]

[0126] in The calculation is as follows:

[0127]

[0128] in This represents the angle of the hydrogen bond between the i-th residue of the m-th domain and the j-th residue of the n-th domain. Indicates the corresponding angle is The probability of;

[0129] 6.5) Calculate the conflict energy between structural domains:

[0130] 6.6) Calculate the general contact energy between structural domains:

[0131]

[0132] Where, if the i-th residue and the j-th residue are hydrophobic groups, then u ij =0.1, if the i-th residue and the j-th residue are hydrophilic groups, then u ij =0.01, otherwise u ij =0.05;

[0133] 6.7) Calculate the inter-domain distance spectrum energy:

[0134]

[0135] Where T ij It is the number of templates; where a i and a j It is the index of the i-th residue and the j-th residue aligned to the template, and satisfies: 0.6|ij|<|a i -a j |<1.5|ij|,D t ij is in the t-th template a i and a j C in residues α The distance between atoms;

[0136] 6.8) Calculate the domain connectivity energy: E db (m, n) = (b mn -b0) 2 , where b mn C is the connection between the endpoints of two continuous domains. α The distance between atoms It is C α The standard distance between atoms;

[0137] 6.9) Calculate the template constraint energy term: Where d(S) i S′ i ) represents the i-th C in the initial structure generated during the template alignment process. α Atom (S) i ) and the corresponding C in the conformation generated during the assembly process α Atom (S′) i The distance between the two proteins, L, is the protein length;

[0138] 6.10) Calculate the energy at the radius of rotation:

[0139]

[0140] Where R decoy R is the radius of rotation of the conformation. min =2.849L 0.319 R is the estimated minimum radius of rotation. max =max{R min +7.5, 0.55N mh For the estimated maximum radius of rotation, N mh It is the number of the longest helical residues in the conformation;

[0141] 6.11) Calculate the total energy E of the current structure:

[0142]

[0143] 6.12) Assign weights to the mixed energy function;

[0144] 6.13) Perform N independent L-BFGS simulations on the hybrid energy function to find the optimal rotation angles φ, θ, ψ and translation vectors (t1, t2, t3) for each structural domain, and generate the final full-length model using the translation vector and rotation matrix at the lowest total energy.

[0145] 7) Select the full-length structure with the lowest energy generated by the global template based on T local templates as the final assembly structure.

[0146] Furthermore, the process of step 6.12) is as follows:

[0147] 6.12.1) If the score of this template is less than 0.85, then the weights are: w1 = 5, w2 = 1, w3 = 3, w4 = 1.2, w5 = 0.2, w6 = 1.0, w7 = 0.02, w8 = 0.01, w9 = 0.15, w 10 =0.13;

[0148] 6.12.2) Otherwise, the weights are: w1 = 1, w2 = 0.2, w3 = 0.2, w4 = 0.15, w5 = 0.15, w6 = 0.1, w7 = 0.02, w8 = 0.01, w9 = 1.2, w 10 =0.12.

[0149] Furthermore, the process of step 6.13) is as follows:

[0150] 6.13.1) Generate the rotation matrix T based on the inter-domain rotation angles φ, θ, and ψ. rot :

[0151]

[0152] Where a ij i = 1, 2, 3, j = 1, 2, 3 are the elements of the rotation matrix;

[0153] 6.13.2) Generate the translation matrix T = (x m y m , z m ):

[0154]

[0155] Where (t1, t2, t3) is the translation vector of the domain, and (x0, y0, z0) is the initial position of the m-th atom. m y m , z m ) is the new position of the m-th atom after the transformation, (xc y c , z c ) is the center point of the structural domain;

[0156] 6.12.3) Calculate the position of each atom in each domain according to the corresponding rotation angle and translation vector to generate a new full-length structure.

[0157] Taking the multi-domain protein 1h88C, which contains three structural domains, as an example, the near-native conformation of this protein was obtained using the above method, and the predicted structure is as follows. Figure 2 As shown.

[0158] The above describes the excellent effects of one embodiment of the present invention. Obviously, the present invention is not only suitable for the above embodiment, but can also be implemented in various ways without departing from the basic spirit of the present invention and without exceeding the content involved in the substantive content of the present invention.

Claims

1. A method for assembling multi-domain protein structures with deep learning-assisted inter-domain constraints, characterized in that, First, the three-dimensional structures of all single domains of the multi-domain protein to be assembled are input. Then, local templates are obtained by scoring the alignment of each template with every two consecutive domains in the multi-domain protein structure library, and global templates are obtained by scoring the alignment of each template with each single domain. The T highest-scoring local templates and the global template are selected as assembly templates. The constraint information between the structural domains is predicted using DeepPotential to calculate the mixture energy function of the multi-domain protein. For each template, N independent L-BFGS simulations are performed based on the mixture energy function. The translation vector and rotation angle of each structural domain at the lowest energy are used to generate the final full-length model. Finally, the structure with the lowest energy among the T local templates and the global template is selected as the final assembled structure. The method includes the following steps: 1) Input the three-dimensional structures of all single domains of the multi-domain protein to be assembled; 2) Setting parameters: Conflict distance threshold d cut The L-BFGS quasi-Newton algorithm performs N independent simulations and assembles T templates. 3) Using the protein structure alignment tool TM-align, each template in the multi-domain protein structure library is scored against every two consecutive domains. The score of each local template is the average of the TM-scores between each domain and the template. 4) Scoring is performed on the alignment of each template with each single domain in the multi-domain protein structure library. The score of each global template is the harmonic mean of the TM-score of the alignment between each domain and the template. The T local templates and the global template with the highest scores are selected as assembly templates respectively. 5) Utilize the deepPotential tool to predict the confidence levels U corresponding to rotation angles, hydrogen bond networks, inter-residue distances, and residue constraints between domains. ij i = 1, 2, ..., L m j = 1, 2, ..., L n L m L n U represents the sequence length of the m-th and n-th structural domains. ij This represents the confidence level of the constraint between the i-th residue in the m-th domain and the j-th residue in the n-th domain; 6) Assemble according to each selected template; 7) Select the full-length structure with the lowest energy generated by the global template based on T local templates as the final assembly structure.

2. The deep learning inter-domain constraint-assisted multi-domain protein structure assembly method as described in claim 1, characterized in that, The process described in 6) is as follows: 6.1) Calculate the structural domain C β Energy of distance between atoms: Where d ij C is in the i-th structural domain β Atom and C in the j-th structural domain β The distance between atoms, P(i,j,k(d) ij )) is predicted by DeepPotential d ij The probability of falling into the k-th interval. ε = 10 -4 A very small positive number prevents the probability from being 0; 6.2) Calculate the directional energy of the structural domain: Among them O ij P(i,j,k(d) represents the dihedral angle between residues in the domain. ij The O predicted by DeepPotential is... ij The probability of falling into the k-th angle interval; 6.3) Calculate the contact energy between structural domains: 6.4) Calculate the hydrogen bond confinement energy between structural domains: in The calculation is as follows: in This represents the angle of the hydrogen bond between the i-th residue of the m-th domain and the j-th residue of the n-th domain. Indicates the corresponding angle is The probability of; 6.5) Calculate the conflict energy between structural domains: 6.6) Calculate the general contact energy between structural domains: in, If the i-th residue and the j-th residue are hydrophobic groups, then u ij =0.1, if the i-th residue and the j-th residue are hydrophilic groups, then u ij =0.01, otherwise u ij =0.05; 6.7) Calculate the inter-domain distance spectrum energy: Where T ij It is the number of templates; where a i and a j It is the index of the i-th residue and the j-th residue aligned to the template, and satisfies: 0.6|ij|<|a i -a j |<1.5|ij|,D t ij is in the t-th template a i and a j C in residues α The distance between atoms; 6.8) Calculate the domain connectivity energy: E db (m, n) = (b mn -b0) 2 , where b mn C is the connection between the endpoints of two continuous domains. α The distance between atoms It is C α The standard distance between atoms; 6.9) Calculate the template constraint energy term: Where d(S) i S′ i ) represents the i-th C in the initial structure generated during the template alignment process. α Atom (S) i ) and the corresponding C in the conformation generated during the assembly process α Atom (S′) i The distance between the two proteins, L, is the protein length. 6.10) Calculate the energy at the radius of rotation: Where R decoy R is the radius of rotation of the conformation. min =2.849L 0.319 R is the estimated minimum radius of rotation. max =max{R min +7.5, 0.55N mh } represents the estimated maximum radius of rotation, N. mh It is the number of the longest helical residues in the conformation; 6.11) Calculate the total energy E of the current structure: 6.12) Assign weights to the mixed energy function; 6.13) Perform N independent L-BFGS simulations on the hybrid energy function to find the optimal rotation angles φ, θ, ψ and translation vectors (t1, t2, t3) for each structural domain. Use the translation vector and rotation matrix at the lowest total energy to generate the final full-length model.

3. The deep learning inter-domain constraint-assisted multi-domain protein structure assembly method as described in claim 2, characterized in that, The process of step 6.12) is as follows: 6.12.1) If the score of this template is less than 0.85, then the weights are: w1 = 5, w2 = 1, w3 = 3, w4 = 1.2, w5 = 0.2, w6 = 1.0, w7 = 0.02, w8 = 0.01, w9 = 0.15, w 10 =0.13; 6.12.2) Otherwise, the weights are: w1 = 1, w2 = 0.2, w3 = 0.2, w4 = 0.15, w5 = 0.15, w6 = 0.1, w7 = 0.02, w8 = 0.01, w9 = 1.2, w 10 =0.

12.

4. The deep learning inter-domain constraint-assisted multi-domain protein structure assembly method as described in claim 2 or 3, characterized in that, The process of step 6.13) is as follows: 6.13.1) Generate the rotation matrix T based on the inter-domain rotation angles φ, θ, and ψ. rot : Where a ij i = 1, 2, 3, j = 1, 2, 3 are the elements of the rotation matrix; 6.13.2) Generate the translation matrix T = (x m ,y m ,z m ): Where (t1, t2, t3) is the translation vector of the domain, and (x0, y0, z0) is the initial position of the m-th atom. m ,y m ,z m ) is the new position of the m-th atom after the transformation, (x c y c , z c ) is the center point of the structural domain; 6.12.3) Calculate the position of each atom in each domain according to the corresponding rotation angle and translation vector to generate a new full-length structure.

Images