CN105808972A - Method for predicting protein structure from local to global on basis of knowledge spectrum - Google Patents

Abstract

A protein structure prediction method based on spectral knowledge from local to global, including the following steps: first, for the query sequence, a high-quality fragment library is obtained through a multi-feature seamless threading method, and residue-residue is obtained through statistical consistency analysis based on the fragment library. The distance spectrum knowledge between the bases; then, the query sequence is divided into several segments according to the residue information recorded in the distance spectrum; after that, for each segment of the structure, the low energy and residue-residue between residues are obtained by fragment assembly The spatial distance approximates the predicted distance in the distance spectrum; finally, fragment assembly is performed on the unsegmented structure, and the global energy is calculated to obtain a metastable conformation with low energy and a more reasonable structure. The invention has good conformational space sampling capability and high prediction accuracy.

Description

A Local-to-Global Protein Structure Prediction Method Based on Spectral Knowledge

technical field

The invention relates to the field of bioinformatics and computer application, in particular to a protein structure prediction method based on spectrum knowledge from local to global.

Background technique

Protein molecules play a vital role in the process of biological and cellular chemical reactions. Their structural models and bioactive states have important implications for our understanding and cure of many diseases. Only when proteins are folded into a specific three-dimensional structure can they produce their unique biological functions. Therefore, to understand the function of a protein, it is necessary to obtain its three-dimensional structure.

Protein tertiary structure prediction is an important task in bioinformatics. The biggest challenge facing the protein conformation optimization problem is to search the extremely complex protein energy function surface. The protein energy model takes into account the bonding of molecular systems and non-bonding interactions such as van der Waals forces, electrostatics, hydrogen bonds, and hydrophobicity, resulting in extremely rough energy surfaces, and the number of local minimum solutions corresponding to conformations increases exponentially with sequence length . The mechanism by which the protein conformation prediction algorithm can find protein stable structures is that a large number of protein metastable structures constitute low-energy regions, so the key to finding the most stable protein structure globally is that the algorithm can find a large number of protein metastable structures, that is, increasing Algorithm population diversity. Therefore, for a more accurate protein force field model, selecting an effective conformational space optimization algorithm to make the new protein structure prediction algorithm more universal and efficient has become the focus of protein structure prediction in bioinformatics.

At present, protein structure prediction methods can be roughly divided into two categories, template-based methods and non-template-based methods. Among them, the non-template-based ab initio prediction (Ab-inito) method is the most widely used. It is suitable for most proteins whose homology is less than 25%, and only generates a new structure from the sequence, which is of great significance to the study of protein molecular design and protein folding. At present, there are several relatively successful ab initio prediction methods: the TASSER (Threading/Assembly/Refinement) method jointly developed by Zhang Yang and Jeffrey Skolnick, the Rosetta method designed by David Baker and his team, and the FeLTr method designed by Shehu et al. But so far there is no perfect method to predict the three-dimensional structure of proteins. Even if good prediction results are obtained, it is only for some proteins. At present, the main technical bottlenecks lie in two aspects. First, On the one hand, it lies in the sampling method, and the prior art is not strong in sampling the conformation space; on the other hand, it lies in the conformation updating method, and the accuracy of the prior art on updating the conformation is still insufficient.

Therefore, existing protein structure prediction methods are deficient and need to be improved.

Contents of the invention

In order to overcome the disadvantages of weak conformational space sampling ability and low prediction accuracy of existing protein structure prediction methods, the present invention proposes a protein structure based on spectral knowledge from local to global with better conformational space sampling ability and high prediction accuracy method of prediction.

The technical solution adopted by the present invention to solve its technical problems is:

A protein structure prediction method from local to global based on spectral knowledge, the optimization method includes the following steps:

1) given query sequence information;

2) Download the resolution less than from the protein database (PDB) website high-precision protein, where is the unit of distance, m, according to the sequence alignment algorithm NW-Align to remove amino acid chains with a sequence similarity greater than 30%, to obtain a non-redundant protein template library;

3) According to the multi-feature similarity function:

Compare the score f(i,j) of each residue position of the protein chain in the non-redundant template library relative to the query sequence by seamless threading method, where i is the residue position of the query sequence, and j is the fragment structure; In f(i,j), the subscript q represents the query sequence feature score item, the subscript t represents the template protein feature score item, P _q (i,k) is the sequence frequency spectrum of the query sequence obtained by PSI-BLAST, where k is the preset number of amino acid types; L _q (i,k) and L _t (j,k) are the query sequence and template sequence logarithmic spectrum obtained by PSI-BLAST; ss _t (j) is the template protein secondary structure classification , calculated by DSSP; ss _q (i) is the secondary structure classification of the query sequence, which can be obtained by two-layer neural network training; sa _t (j) and sa _q (i) are the template structure and solvent accessibility of the query sequence Indicators, trained by EDTSurf and neural network programs; ψ _q (j) is the query sequence dihedral angle pair can be obtained through two-layer neural network training; ψ _t (j) is obtained by querying the protein dictionary; SP _t (j,k) is the structural spectrum of the template protein; w ₁ , w ₂ , w ₃ , w ₄ and w ₅ are weight values;

4) According to the similarity score f(i, j), select the M fragments with the highest score at each position of the query sequence to obtain the fragment library file;

5) Statistical query sequence residue pairs come from the distance between the same template fragment, here only statistics less than The distance between the residue pairs, draw the histogram to get the distance spectrum, the distance interval of the abscissa of the histogram is When the distance between residue pairs in the template is within a certain interval, the total number of the interval will be increased by 1. If the line graph is in A peak appears in a certain distance interval within the peak, then the distance interval corresponding to the peak is the predicted distance from residue i to residue j in the target sequence, and the distribution is recorded as the distance profile between the two residues;

6) dividing the query sequence into n segments according to the positions of the residues in the obtained distance spectrum;

7) Let l = 1, l ∈ {1,2,3,...,n}, perform the following operations on the segmented structure:

7.1) Fragment assembly is carried out to the segment l fragment structure;

7.2) Calculate the distance between the residue positions containing the distance spectrum information, and calculate the deviation from the predicted distance, accumulate the deviation value and take the average and record it as ΔD;

7.3) If ΔD<R, store the conformation as a cell, where R is the structural accuracy constraint;

7.4) Repeat 7.1) to 7.3) until the number of stored cells reaches 100 cells, compare the energy of the segmented structure in the cell based on RosettaScore3, and select the structure with the lowest energy as the predicted structure of the segment;

7.5) l=l+1, judge whether l is greater than or equal to n, then enter 8), otherwise return to 7.1);

8) Set the number of iterations as G, let s=1, perform the following operations:

8.1) Calculate the target-induced conformational energy E(P _target );

8.2) Fragment assembly is performed on the unsegmented structure, and the energy value E(P _trail ) is calculated;

8.3) If E(P _target )>E(P _trail ), replace P _target with P _trail ;

8.4) s=s+1; judge whether s is greater than or equal to G, if so, enter 9), otherwise return to 8.1);

9) Output the induced conformation, and obtain the near-native state structure of the query sequence.

The technical concept of the present invention is as follows: firstly, for the query sequence, a high-quality fragment library is obtained through the multi-feature seamless threading method, and the residue-residue distance spectrum knowledge is obtained through statistical consistency analysis based on the fragment library; then, the query sequence is According to the residue information recorded in the distance spectrum, it is divided into several segments of structure; after that, for each segment of the structure, the energy is low and the spatial distance between residues is approximated by the predicted distance in the distance spectrum through fragment assembly; finally, for The unsegmented structure is assembled into fragments, the global energy is calculated, and a metastable conformation with low energy and a more reasonable structure is obtained.

The beneficial effect of the invention is that the conformational space sampling ability is good and the prediction precision is high. Description of drawings

Figure 1 is a schematic diagram of the relationship between the RMSD and the energy value of the test sequence during the population update process.

Figure 2 is a schematic diagram of the three-dimensional structure of the protein predicted by the 1ENH algorithm.

detailed description

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

Referring to Figures 1 to 2, a protein structure prediction method based on spectral knowledge from local to global, including the following steps:

1) given query sequence information;

2) Download the resolution less than from the protein database (PDB) website high-precision protein, where is the unit of distance, m, according to the sequence alignment algorithm NW-Align to remove amino acid chains with a sequence similarity greater than 30%, to obtain a non-redundant protein template library;

3) According to the multi-feature similarity function:

Compare the score f(i,j) of each residue position of the protein chain in the non-redundant template library relative to the query sequence by seamless threading method, where i is the residue position of the query sequence, and j is the fragment structure; In f(i,j), the subscript q represents the query sequence feature score item, the subscript t represents the template protein feature score item, P _q (i,k) is the sequence frequency spectrum of the query sequence obtained by PSI-BLAST, where k is the preset number of amino acid types; L _q (i,k) and L _t (j,k) are the query sequence and template sequence logarithmic spectrum obtained by PSI-BLAST; ss _t (j) is the template protein secondary structure classification , calculated by DSSP; ss _q (i) is the secondary structure classification of the query sequence, which can be obtained by two-layer neural network training; sa _t (j) and sa _q (i) are the template structure and solvent accessibility of the query sequence Indicators, trained by EDTSurf and neural network programs; ψ _q (j) is the query sequence dihedral angle pair can be obtained through two-layer neural network training; ψ _t (j) is obtained by querying the protein dictionary; SP _t (j,k) is the structural spectrum of the template protein; w ₁ , w ₂ , w ₃ , w ₄ and w ₅ are weight values;

4) According to the similarity score f(i, j), select the M fragments with the highest score at each position of the query sequence to obtain the fragment library file;

5) Statistical query sequence residue pairs come from the distance between the same template fragment, here only statistics less than The distance between the residue pairs, draw the histogram to get the distance spectrum, the distance interval of the abscissa of the histogram is When the distance between residue pairs in the template is within a certain interval, the total number of the interval will be increased by 1. If the line graph is in A peak appears in a certain distance interval within the peak, then the distance interval corresponding to the peak is the predicted distance from residue i to residue j in the target sequence, and the distribution is recorded as the distance profile between the two residues;

6) dividing the query sequence into n segments according to the positions of the residues in the obtained distance spectrum;

7) Let l = 1, l ∈ {1,2,3,...,n}, perform the following operations on the segmented structure:

7.1) Fragment assembly is carried out to the segment l fragment structure;

7.2) Calculate the distance between the residue positions containing the distance spectrum information, and calculate the deviation from the predicted distance, accumulate the deviation value and take the average and record it as ΔD;

7.3) If ΔD<R, store the conformation as a cell, where R is the structural accuracy constraint;

7.4) Repeat 7.1) to 7.3) until the number of stored cells reaches x cells, compare the energy of the segmented structure in the cell based on RosettaScore3, and select the structure with the lowest energy as the predicted structure of the segment;

7.5) l=l+1, judge whether l is greater than or equal to n, then enter 8), otherwise return to 7.1);

8) Set the number of iterations as G, let s=1, perform the following operations:

8.1) Calculate the target-induced conformational energy E(P _target );

8.2) Fragment assembly is performed on the unsegmented structure, and the energy value E(P _trail ) is calculated;

8.3) If E(P _target )>E(P _trail ), replace P _target with P _trail ;

8.4) s=s+1; judge whether s is greater than or equal to G, if so, enter 9), otherwise return to 8.1);

9) Output the induced conformation, and obtain the near-native state structure of the query sequence.

This example takes protein 1ENH with a sequence length of 54 as an example, a method for predicting protein structure from local to global based on spectral knowledge, which includes the following steps:

1) given query sequence information;

2) Download the resolution less than from the protein database (PDB) website high-precision protein, where is the unit of distance, m, according to the sequence comparison algorithm NW-Align to remove amino acid chains with a sequence similarity greater than 30%, and obtain 8619 protein chains to form a non-redundant template library;

3) According to the multi-feature similarity function:

Compare the score f(i,j) of each residue position of the protein chain in the non-redundant template library relative to the query sequence by seamless threading method, where i is the residue position of the query sequence, and j is the fragment structure; In f(i,j), the subscript q represents the query sequence feature score item, the subscript t represents the template protein feature score item, P _q (i,k) is the sequence frequency spectrum of the query sequence obtained by PSI-BLAST, where k is the preset number of amino acid types; L _q (i,k) and L _t (j,k) are the query sequence and template sequence logarithmic spectrum obtained by PSI-BLAST; ss _t (j) is the template protein secondary structure classification , calculated by DSSP; ss _q (i) is the secondary structure classification of the query sequence, which can be obtained by two-layer neural network training; sa _t (j) and sa _q (i) are the template structure and solvent accessibility of the query sequence Indicators, trained by EDTSurf and neural network programs; ψ _q (j) is the query sequence dihedral angle pair can be obtained through two-layer neural network training; ψ _t (j) is obtained by querying the protein dictionary; SP _t (j,k) is the structural spectrum of the template protein; w ₁ , w ₂ , w ₃ , w ₄ and w ₅ are weight values;

4) Select the 200 fragments with the highest score at each position of the query sequence according to the similarity score f(i,j) to obtain the fragment library file;

5) Statistical query sequence residue pairs come from the distance between the same template fragment, here only statistics less than The distance between the residue pairs, draw the histogram to get the distance spectrum, the distance interval of the abscissa of the histogram is When the distance between residue pairs in the template is within a certain interval, the total number of the interval will be increased by 1. If the line graph is in A peak appears in a certain distance interval within the peak, then the distance interval corresponding to the peak is the predicted distance from residue i to residue j in the target sequence, and the distribution is recorded as the distance profile between the two residues;

6) dividing the query sequence into n segments according to the positions of the residues in the obtained distance spectrum;

7) Let l = 1, l ∈ {1,2,3,...,n}, perform the following operations on the segmented structure:

7.1) Fragment assembly is carried out to the segment l fragment structure;

7.2) Calculate the distance between the residue positions containing the distance spectrum information, and calculate the deviation from the predicted distance, accumulate the deviation value and take the average and record it as ΔD;

7.3) If ΔD<R, store the conformation as a cell, where R is the structural accuracy constraint;

7.4) Repeat 7.1) to 7.3) until the number of stored cells reaches x cells, compare the energy of the segmented structure in the cell based on RosettaScore3, and select the structure with the lowest energy as the predicted structure of the segment;

7.5) l=l+1, judge whether l is greater than or equal to n, then enter 8), otherwise return to 7.1);

8) Set the number of iterations as G=50000, let s=1, perform the following operations:

8.1) Calculate the target-induced conformational energy E(P _target );

8.2) Fragment assembly is performed on the unsegmented structure, and the energy value E(P _trail ) is calculated;

8.3) If E(P _target )>E(P _trail ), replace P _target with P _trail ;

8.4) s=s+1; judge whether s is greater than or equal to G=50000, if so, enter 9), otherwise return to 8.1);

9) Output the induced conformation, and obtain the near-native state structure of the query sequence.

Taking the protein 1ENH with a sequence length of 54 as an example, the near-native conformation of the protein was obtained by using the above method. The conformation update map in the conformation ensemble is shown in Figure 1, and the three-dimensional structure of the protein predicted by the algorithm is shown in Figure 2. .

What set forth above is the excellent effect shown by an embodiment of the present invention. Obviously, the present invention is not only suitable for the above-mentioned embodiment, but it can be used under the premise of not departing from the basic spirit of the present invention and not exceeding the content involved in the essence of the present invention. Make changes and implement them.

Claims (1)

1. A protein structure prediction method from local to global based on spectral knowledge, characterized in that: the protein structure prediction method comprises the following steps: 1) given query sequence information; 2) Download the resolution less than high-precision protein, where is the unit of distance, According to the sequence comparison algorithm NW-Align, amino acid chains with a sequence similarity greater than 30% were removed to obtain a non-redundant protein template library; 3) According to the multi-feature similarity function: Compare the score f(i,j) of each residue position of the protein chain in the non-redundant template library relative to the query sequence by seamless threading method, where i is the residue position of the query sequence, and j is the fragment structure; In f(i,j), the subscript q represents the query sequence feature score item, the subscript t represents the template protein feature score item, P _q (i,k) is the sequence frequency spectrum of the query sequence obtained by PSI-BLAST, where k is the preset number of amino acid types; L _q (i,k) and L _t (j,k) are the query sequence and template sequence logarithmic spectrum obtained by PSI-BLAST; ss _t (j) is the template protein secondary structure classification , calculated by DSSP; ss _q (i) is the secondary structure classification of the query sequence, which can be obtained by two-layer neural network training; sa _t (j) and sa _q (i) are the template structure and solvent accessibility of the query sequence Indicators, trained by EDTSurf and neural network programs; ψ _q (j) is the query sequence dihedral angle pair can be obtained through two-layer neural network training; ψ _t (j) is obtained by querying the protein dictionary; SP _t (j,k) is the structural spectrum of the template protein; w ₁ , w ₂ , w ₃ , w ₄ and w ₅ are weight values; 4) According to the similarity score f(i, j), select the M fragments with the highest score at each position of the query sequence to obtain the fragment library file; 5) Statistical query sequence residue pairs come from the distance between the same template fragment, here only statistics less than The distance between the residue pairs, draw the histogram to get the distance spectrum, the distance interval of the abscissa of the histogram is When the distance between residue pairs in the template is within a certain interval, the total number of the interval will be increased by 1. If the line graph is in A peak appears in a certain distance interval within the peak, then the distance interval corresponding to the peak is the predicted distance from residue i to residue j in the target sequence, and the distribution is recorded as the distance profile between the two residues; 6) dividing the query sequence into n segments according to the positions of the residues in the obtained distance spectrum; 7) Let l = 1, l ∈ {1,2,3,...,n}, perform the following operations on the segmented structure: 7.1) Fragment assembly is carried out to the segment l fragment structure; 7.2) Calculate the distance between the residue positions containing the distance spectrum information, and calculate the deviation from the predicted distance, accumulate the deviation value and take the average and record it as ΔD; 7.3) If ΔD<R, store the conformation as a cell, where R is the structural accuracy constraint; 7.4) Repeat 7.1) to 7.3) until the number of stored cells reaches x cells, compare the energy of the segmented structure in the cell based on RosettaScore3, and select the structure with the lowest energy as the predicted structure of the segment; 7.5) l=l+1, judge whether l is greater than or equal to n, then enter 8), otherwise return to 7.1); 8) Set the number of iterations as G, let s=1, perform the following operations: 8.1) Calculate the target-induced conformational energy E(P _target ); 8.2) Fragment assembly is performed on the unsegmented structure, and the energy value E(P _trail ) is calculated; 8.3) If E(P _target )>E(P _trail ), replace P _target with P _trail ; 8.4) s=s+1; judge whether s is greater than or equal to G, if so, enter 9), otherwise return to 8.1); 9) Output the induced conformation, and obtain the near-native state structure of the query sequence.