阅读说明：本技术 基于多目标优化的蛋白质序列设计实现方法 (Protein sequence design implementation method based on multi-objective optimization ) 是由 沈红斌 李瑞祥 于 2020-04-29 设计创作，主要内容包括：一种基于多目标优化的蛋白质序列设计实现方法,通过融合目标蛋白的相似结构信息和基于局部结构的统计信息作为先验知识,并将现有离散蛋白质序列空间转换为连续的蛋白质序列空间,然后在其中通过多目标粒子群优化算法进行蛋白序列的迭代搜索并以排序最优的非支配解集实现蛋白质序列的优化。本发明通过集成结构、统计和物理力场三种信息引导蛋白质序列的搜索,得到的蛋白质序列无论是准确性和鲁棒性均由于其他方法,并且它的运行速度更快。本发明同样对设计的序列进行了生物实验验证,多个属于不同折叠类型的蛋白均能够正确表达,并且拥有与目标结构拟合度很高的二级结构比例和稳定的三级结构。(A protein sequence design realization method based on multi-objective optimization is characterized in that similar structure information of target protein and statistical information based on local structure are fused to serve as priori knowledge, an existing discrete protein sequence space is converted into a continuous protein sequence space, iterative search of protein sequences is conducted through a multi-objective particle swarm optimization algorithm, and optimization of the protein sequences is achieved through a non-dominated solution set with the optimal sequence. The invention guides the search of the protein sequence by integrating three information of structure, statistics and physical force field, the obtained protein sequence is subject to other methods no matter the accuracy and the robustness, and the operation speed is faster. The invention also carries out biological experiment verification on the designed sequence, a plurality of proteins belonging to different folding types can be correctly expressed, and the protein has a secondary structure proportion with high fitting degree with a target structure and a stable tertiary structure.)

1. A protein sequence design realization method based on multi-objective optimization is characterized in that similar structure information of target protein and statistical information based on local structure are fused to serve as priori knowledge, an existing discrete protein sequence space is converted into a continuous protein sequence space, iterative search of the protein sequence is conducted through a multi-objective particle swarm optimization algorithm, and optimization of the protein sequence is achieved through a non-dominated solution set with the optimal ordering.

2. The method for realizing Protein sequence design based on multi-objective optimization according to claim 1, wherein the similar structure information of the target Protein is obtained through a PDB database (Protein Data Bank); the statistical information based on the local Structure is obtained through a DSSP (Definition of Secondary Structure of Proteins).

3. The method for designing and implementing protein sequences based on multi-objective optimization of claim 2, wherein the similar structure information of the target protein satisfies all proteins with similar structures to the target protein, and the similarity satisfies TMscore >0.7, and then the amino acid sequences of all similar proteins are aligned to obtain a residue-position scoring matrix;

the statistical information based on the local structure meets the requirement of the probability distribution of the amino acids at different residue positions of the statistical target protein and obtains a residue-site scoring matrix;

the priori knowledge is obtained by fusing the obtained two residue-site scoring matrixes.

4. The method of claim 1, wherein the continuous protein sequence space is used for performing different amino acid quantitative ordering on each residue position of the target protein according to the fused information, thereby realizing the spatial transformation of the protein sequence.

5. The method for realizing protein sequence design based on multi-objective optimization of claim 1, wherein the multi-objective particle swarm optimization simultaneously optimizes two objective energy functions in a transformed space by using a multi-objective particle swarm optimization algorithm, specifically comprising: and extracting a final protein sequence from the non-dominated solution set as output after multiple iterations by adopting a Foldx energy function based on a physical force field and an energy function based on local structure information.

6. The method for realizing protein sequence design based on multi-objective optimization of claim 1 or 5, wherein in the multi-objective particle swarm optimization, each iteration maps the result to the original protein sequence space so as to calculate the values of different energy functions.

7. The method for realizing protein sequence design based on multi-objective optimization according to claim 1 or 5, wherein the multi-objective particle swarm optimization algorithm specifically comprises:

f₁＝w₁E_vdw+w₂E_solvH+w₃E_solvP+E_wb+E_hbond+E_el+E_Kon+w₄E_Smc+w₅E_Ssc，

f₂＝∑w₁ΔSS(p)+w₂ΔSA(p)+w₃(Δ φ (p) + Δ ψ (p)), wherein: Δ SS (p) is the secondary structural difference between the target protein and the designed sequence, specificallyWherein: Δ SA is the difference in solvent accessibility, Δ Φ (p) and Δ ψ (p) are the difference in twist angle;

the multi-objective optimization algorithm combining two energy functions is as follows:

the positions of the particles are as follows:wherein:

the initialization of the particles includes:

wherein: set (d) is an amino acid type set obtained by screening according to local structural information at the residue position d of the target protein;

the iterative formula for the particle includes:

8. the method for realizing protein sequence design based on multi-objective optimization according to claim 1, wherein the multi-objective particle swarm optimization algorithm is optimized by an accept-reject mechanism of the type Monte Carlo, and specifically comprises the following steps: probability of accepting the sequenceWherein: KT is a temperature parameter;

the updating mode of the non-dominant solution set is determined by convergence distance and dispersion distance, and comprises the following steps:

non-dominant solutions include:

① all particles are close to the ideal value and have good dispersibility_cv(x)<mean(d_cv)and d_dv(x)>mean(d_dv))，

② case where the particles are close to the ideal value but too dense (d)_cv(x)<mean(d_cv)and d_dv(x)<mean(d_dv))；

③ most of the examples are far from ideal, however, a part of the particles are distributed on the paret edge and may carry important information (d)_cv(x)>mean(d_cv)and d_dv(x)>mean(d_dv))；

④ are far from ideal and densely packed together so that their priority is lowest (d)_cv(x)>mean(d_cv)and d_dv(x)<mean(d_dv))；

The oscillation of the spatial transform of the non-dominated solution satisfies lambda₁+λ₂＝1,λ＝(λ₁,λ₂),||λ^*-λ||＞d_threshold。

9. The method as claimed in claim 8, wherein the optimal solution is selected from the non-dominated solution set and the solution is the solution that is farthest from the optimal solution and satisfies d_cv(x)<mean(d_cv) The two solutions are used as the protein optimized structure obtained by the method; said distance

10. A system for implementing the method of any one of claims 1 to 9, comprising: the system comprises a database information statistics module, a statistical information fusion module, a space transformation module, an iterative computation module and a sequence output module, wherein: the database information statistics module outputs information to the statistics information fusion module, the fused information is transmitted to the space transformation module for quantization transformation, the space transformation module transmits the information to the iterative computation module for iterative computation, and a solution set of the iterative computation is output to the sequence output module for final screening.

Technical Field

The invention relates to a technology in the field of bioengineering, in particular to a protein sequence design implementation method based on multi-objective optimization.

Background

In recent years, the artificial synthesis of proteins is a very attractive biotechnology, which has profound effects on pharmaceuticals, nanotechnology, catalytic chemistry, and many industries. In nature, proteins have a decisive role in physiological processes, such as the production of energy, the structural composition of cells and organ tissues, sensors, catalysts, etc. Although the amino acids are only 20 kinds, the protein molecules composed of the amino acids have inexhaustible complex structures in nature. In view of the current biotechnology, some proteins that do not exist in nature can be artificially synthesized, and thus a rapid and accurate protein design method is very important for rapidly developing protein engineering. Protein design is based on the tertiary structure of a target protein, and designs a corresponding protein sequence which can be folded into the target tertiary structure, and is the inverse process of protein sequence prediction. The designed novel protein has the application prospect of biological pharmacy and biological engineering, so that the protein design field draws wide attention. With the development of a large number of super computing devices, protein design algorithms have made outstanding progress, but many problems still need to be solved at present.

For the currently prevailing methods, protein design is solved as an optimization problem. The energy functions to be optimized are Rosetta, SEF _ V, CHARMM, etc. These energy functions are linearly composed of many different energy terms, which reflect different aspects of the mapping relationship between protein sequences and structures, respectively. These energy functions can be broadly classified into three categories: (1) an energy function based on a physical force field, (2) a function based on a statistical model, and (3) a function based on structural information. Due to the long length of protein (usually greater than 100), it is difficult for a single energy function to accurately capture information of complex structures, which is one of the reasons that the current protein design method is limited in effect. Most current protein design methods are based on single-objective optimization algorithms, and some multi-objective optimization algorithms are actually linear weighting of different energy functions, namely, describing functions with coarser granularity. And the multi-objective optimization method only prioritizes different energy functions, which is also a single-objective optimization mode in essence.

Although there are some proteinsThe quality design method adopts a partial multi-objective optimization idea, and a complete protein design method based on multi-objective optimization has not been proposed yet. Therefore, the feature of accurately combining different types of energy functions is the key to solving the problem of protein design. The energy function of the physical force field is usually time consuming and thus too many iterations in the protein design problem are very time consuming. Furthermore, a protein sequence is usually greater than 100, and thus 20¹⁰⁰This is a very high dimensional optimization problem. This problem is often not solved by enumeration, and a fast and accurate sequential search algorithm is therefore important to solve this problem.

Currently, many protein design approaches have made some progress: RosettaDesign uses the monte carlo stochastic algorithm to optimize Rosetta energy function, which was successful in designing a naturally non-existent protein called Top7, and which was verified to fit perfectly to the target structure in wet experiments; the ABACUS also adopts a Monte Carlo algorithm to optimize an energy function based on the statistical information of the local structure of the target protein; EvoDesign proposes an energy function based on homologous protein structural information.

The above methods are based on the Monte Carlo algorithm (REMC) of backup exchange to optimize a single target energy function. This method has some applicability to non-analytic energy functions (i.e., black box models), but it still has the following disadvantages:

1) REMC requires a long number of iterations to produce a more ideal solution. Because the combination of protein sequences is many, and the quantitative relationship between 20 residues is difficult to establish, a connection network between the protein sequences is difficult to establish in the searching process of the sequences, which means that a newly generated sequence cannot capture useful information from a previously ideal sequence, and all sequence searches are completely performed at random. The REMC algorithm is very computationally intensive, and if it is applied to multi-objective optimization, the search time will increase exponentially, since such an algorithm usually requires at least 10 paths to search simultaneously.

2) The ideal protein sequence exhibits a sparse distribution in the original protein sequence space. This is because there is no clear quantitative relationship between amino acids different at each residue position, and the sequence length of the protein is long, so that a small number of optimal solutions are sparsely distributed in this huge space, which causes great trouble in search.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides a protein sequence design implementation method based on multi-objective optimization, which guides the search of a protein sequence through three information of an integrated structure, statistics and a physical force field, so that the obtained protein sequence is both accurate and robust in other methods, and the running speed of the protein sequence is higher. The invention also carries out biological experiment verification on the designed sequence, a plurality of proteins belonging to different folding types can be correctly expressed, and the protein has a secondary structure proportion with high fitting degree with a target structure and a stable tertiary structure.

The invention is realized by the following technical scheme:

the invention relates to a protein sequence design realization method based on multi-objective optimization, which takes similar structure information of target protein and statistical information based on local structure as prior knowledge, converts the existing discrete protein sequence space into continuous protein sequence space, and then carries out iterative search of protein sequence by multi-objective particle swarm optimization algorithm and realizes optimization of protein sequence by ordering optimal non-dominated solution set.

The similar structural information of the target Protein is obtained by but not limited to PDB database (Protein Data Bank).

The statistical information based on the local Structure is obtained by, but not limited to, a DSSP database (Definition of secondary Structure of Proteins).

The similar structure information of the target protein satisfies all proteins with similar structures with the target protein, the similarity satisfies TMscore >0.7, and then the obtained amino acid sequences of all similar proteins are aligned to obtain a residue-position scoring matrix.

The statistical information based on the local structure meets the requirement of the probability distribution of the amino acids at different residue positions of the statistical target protein and obtains a residue-site scoring matrix.

The priori knowledge is obtained by fusing the obtained two residue-site scoring matrixes

The continuous protein sequence space carries out quantitative sequencing of different amino acids on each residue position of the target protein according to the fused information, thereby realizing the spatial transformation of the protein sequence.

The multi-objective particle swarm optimization simultaneously optimizes two objective energy functions by using a multi-objective particle swarm optimization algorithm in a transformed space, and specifically comprises the following steps: a physical force field based Foldx energy function and a local structure information based energy function are employed. And extracting a final protein sequence from the non-dominated solution set as an output after multiple iterations.

In the multi-objective particle swarm optimization, the result is preferably mapped to the original protein sequence space by each iteration so as to calculate the values of different energy functions.

The invention relates to a system for realizing the method, which comprises the following steps: the system comprises a database information statistics module, a statistical information fusion module, a space transformation module, an iterative computation module and a sequence output module, wherein: the database information statistics module outputs information to the statistics information fusion module, the fused information is transmitted to the space transformation module for quantization transformation, the space transformation module transmits the information to the iterative computation module for iterative computation, and a solution set of the iterative computation is output to the sequence output module for final screening.

Technical effects

The invention solves the technical problem of generating protein sequences which can be folded to target structures based on specific protein structure design.

Compared with the prior art, the method can optimize a plurality of objective functions in parallel, greatly reduce the iteration times by calculating in a transformed quantization space, and more accurately fit the designed protein sequence to the target protein structure.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a schematic diagram of spatial transformation;

FIG. 3 is a graph comparing the results of example similarity;

FIG. 4 is a diagram illustrating the comparison of the effects of the examples.

Detailed Description

As shown in FIG. 1, this embodiment relates to a protein sequence design implementation method based on multi-objective optimization, which includes the following steps:

step 1) establishing prior information, which specifically comprises the following steps:

1.1: and (3) carrying out TMalign comparison on the input protein PDB file and all files in the PDB database after screening to select all proteins with TMscore >0.7, and carrying out multi-sequence comparison on the amino acid sequences of the proteins to obtain a residue-site scoring matrix.

The scoring matrix is as follows:wherein: f (p, x) is the frequency of residue x at position p in the multiple sequence alignment, and B (x, r) is the elemental value of BLOSUM 62.

1.2: obtaining a DSSP file of the input PDB file, then counting the amino acid distribution condition which is the same as the local structure information of each residue position in a DSSP database, and obtaining a scoring matrix as follows:

S(r,p)＝-lnp(r|structure properties atpositionp)，

step 2) fusing prior information and performing spatial transformation, specifically comprising:

2.1) normalizing the two scoring matrices obtained previously:

obtaining the following information according to the normalized information matrix: f (r, p) ═ λ₁M(r,p)'+λ₂S (r, p)', wherein: m (r, p) 'and S (r, p)' are a structure information matrix and a statistical information matrix, respectivelyElement of (a)₁And λ₂Are two coefficients to balance the weights of the two kinds of information.

The coefficients will change as the iteration progresses, which may cause the transformed protein space to oscillate. This oscillation may lead to more exploratory possibilities because the correct quantitative relationship between different residue types is generally not known, and prevents the search from falling into a locally optimal solution.

2.2) spatial transformation: rⁿ＝R×R…×R＝{(x₁,x₂,…,x_n)|x_k∈ R, k is 1,2, …, n, where n is the length of the target protein, and the converted values are T (R, p) rank (F (R, p)), rank ∈ {1,2, …,20}, which are sorted from small to large, thus

The effect of the spatial transformation includes: 1) the search space for the optimal solution is greatly reduced; 2) the original discrete space optimization problem can be transformed into a continuous optimization problem.

Step 3) a Foldx energy function based on a physical force field and an energy function based on local structure information are adopted to realize a multi-objective particle swarm optimization algorithm, and the method specifically comprises the following steps:

f₁＝w₁E_vdw+w₂E_solvH+w₃E_solvP+E_wb+E_hbond+E_el+E_Kon+w₄E_Smc+w₅E_Ssc，

f₂＝∑w₁ΔSS(p)+w₂ΔSA(p)+w₃(Δ φ (p) + Δ ψ (p)), wherein: Δ SS (p) is the secondary structural difference between the target protein and the designed sequence, specificallyWherein: Δ SA is the difference in solvent accessibility, and Δ Φ (p) and Δ ψ (p) are the difference in twist angle.

The multi-objective optimization algorithm combining two energy functions is as follows:

the positions of the particles are as follows:wherein:

the initialization of the particles includes:

wherein: set (d) is a set of amino acid types screened for local structural information at residue position d of the target protein.

The iterative formula for the particle includes:

and 4) optimizing through an acceptance-rejection mechanism of the type Monte Carlo, which specifically comprises the following steps: probability of accepting the sequenceWherein: KT is a temperature parameter.

The updating mode of the non-dominant solution set is determined by convergence distance and dispersion distance, and comprises the following steps:

non-dominant solutions include:

① all particles are close to the ideal value and have good dispersibility_cv(x)<mean(d_cv)and d_dv(x)>mean(d_dv))，

② case where the particles are close to the ideal value but too dense (d)_cv(x)<mean(d_cv)and d_dv(x)<mean(d_dv))；

③ most of the examples are far from ideal, however, a part of the particles are distributed on the paret edge and may carry important information (d)_cv(x)>mean(d_cv)and d_dv(x)>mean(d_dv))；

④ are far from ideal and densely packed together so that their priority is lowest (d)_cv(x)>mean(d_cv)and d_dv(x)<mean(d_dv))。

The oscillation of the spatial transform of the non-dominated solution satisfies lambda₁+λ₂＝1,λ＝(λ₁,λ₂),||λ^*-λ||＞d_threshold。

Step 5) selecting an optimal solution from the non-dominated solution set and the solution which is farthest from the optimal solution and satisfies d_cv(x)<mean(d_cv) The two solutions of (a) are taken as the protein optimized structure obtained by the method.

Said distance

Through specific practical experiments, the method runs by matlab software under the linux system environment setting, and experimental data which can be obtained by using rosetta ab initio prediction tool verification is as follows: under 200 samples, the average proportion of the designed sequence folding structure and the target structure belonging to the same structure is 40.1 percent, wherein alpha protein can reach 70.5 percent; the most similar sample to the target protein of the 200 samples averaged 67.43 tmscore, which can reach 75 for alpha protein.

Compared with the prior art, the average iteration number of the method is reduced by 10³And in the order of magnitude, the accuracy of the result protein sequence fitting target protein is improved, the accuracy is improved by 30.1% through a rosetta ab initio tool verification, and the method is verified on a biological platform, wherein the rmsd between the designed protein structure after the target protein 1ubq sequence is folded and the structure of the original protein isCompared with the current mainstream method, the precision is improved

The foregoing embodiments may be modified in many different ways by those skilled in the art without departing from the spirit and scope of the invention, which is defined by the appended claims and all changes that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.