A multi-modal protein structure prediction method based on fitness sharing

Abstract

The application discloses a multi-modal protein structure prediction method based on fitness sharing. In the evolutionary algorithm framework, firstly, the population is divided into species according to the similarity between conformations; then, a fitness sharing function is constructed to change the fitness value of individuals in the population, so that the individuals have the ability to cross the energy barrier, avoid falling into the same energy trap repeatedly, and facilitate the exploration of other potential low-energy regions; finally, a plurality of conformations with the lowest energy in the population are output as the final prediction result. The multi-modal protein structure prediction method based on fitness sharing can not only alleviate the problem of inaccurate protein energy model, but also enhance the conformation sampling ability, improve the sampling efficiency, and thus improve the prediction accuracy. The application provides a multi-modal protein structure prediction method based on fitness sharing with high prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics and computer application, and particularly relates to a multi-modal protein structure prediction method based on fitness sharing. BACKGROUND

[0002] Proteins are the cornerstone of life. They maintain the metabolism of living beings, and the three-dimensional structure of each protein determines its unique function. Therefore, studying the structure of proteins is crucial to understanding how they perform biological functions, which not only helps to accelerate the development of innovative drugs, but also provides a reliable basis for the design of new proteins or the modification of existing proteins.

[0003] Currently, the three-dimensional structure of proteins is mainly determined by biological wet experiments, including X-ray crystal diffraction, nuclear magnetic resonance, and cryo-electron microscopy. X-ray crystal diffraction is one of the most effective methods for determining protein structure, and its accuracy is higher than other methods. However, this method requires crystallization, and the period of crystal structure determination is long, making the prediction process time-consuming and labor-intensive. The nuclear magnetic resonance method can directly determine the structure of proteins in solution, but it requires a large amount of sample and high purity, and is only suitable for small molecular proteins with a small number of residues. Cryo-electron microscopy technology can determine larger proteins, but its determination accuracy is limited, and the equipment is difficult to maintain and expensive. Experimental structure determination methods are time-consuming, labor-intensive, and costly, and some protein structures cannot be determined by experimental methods. In contrast, sequencing technology has made significant breakthroughs, and the gap between the number of known protein sequences and the number of protein structures obtained by experiments is widening. Therefore, there is an urgent need for an efficient, fast, and simple method to predict the structure of unknown proteins. According to the Anfinsen criterion, the three-dimensional structure of a protein is determined by its amino acid sequence. Therefore, directly predicting the three-dimensional structure of a protein from its amino acid sequence has become one of the hot issues in the field of structural bioinformatics for more than 50 years, and is an urgent problem to be solved in current biological research.

[0004] Protein structure prediction is extremely challenging, as it requires an accurate energy representation of protein structure and an efficient conformational sampling algorithm, the latter of which is one of the main bottlenecks that restricts the accuracy of ab initio protein structure prediction. Traditional conformational sampling algorithms are prone to fall into local basin regions during sampling, making it difficult to cross energy barriers and leading to premature convergence, which makes it difficult to improve prediction accuracy. On the other hand, the energy model of protein is often inaccurate, which means that even if the region with the lowest energy of the protein potential energy surface is found, it may not be able to explore the near-native structure.

[0005] Therefore, the existing protein structure prediction method has problems of inaccurate energy function, insufficient sampling capacity, low sampling efficiency, insufficient prediction accuracy and the like, and needs to be improved. SUMMARY

[0006] In order to overcome the problems of inaccurate energy function, insufficient sampling capacity, low sampling efficiency, insufficient prediction accuracy and the like of the existing protein structure prediction method, the present application provides a multi-modal protein structure prediction method based on fitness sharing.

[0007] The present application changes the individual fitness value by constructing a fitness sharing function and applying it to each species, so that it has the ability to cross the energy barrier, avoids repeated sampling of a single energy basin, and at the same time increases the probability of sampling potential energy basins that may appear in the natural state structure, thereby enhancing the sampling capacity, improving the sampling efficiency and improving the overall prediction accuracy.

[0008] The technical scheme adopted by the present application to solve its technical problems is:

[0009] A multi-modal protein structure prediction method based on fitness sharing, the method comprising the following steps:

[0010] 1) inputting sequence information of a target protein;

[0011] 2) obtaining 3-fragment and 9-fragment fragment library files from the ROBETTA server according to the sequence of the target protein;

[0012] 3) setting parameters: maximum iteration number G, population size N, initial value of niche radius d init , cutoff value of niche radius d cut , simulated annealing rate e, temperature coefficient β, and output prediction model number N out ;

[0013] 4) population initialization: generating a population C={C1,C2,...,C N} with a size of N using the first and second stages of the Rosetta protocol;

[0014] 5) setting g=1, g∈{1,2,...,G};

[0015] 6) population individual sorting: calculating the energy value E score3 (C i ) of each individual C i , i∈{1,2,...,N} in the population C using the score3 energy function in Rosetta, sorting the population in ascending order according to the energy value, and forming a population C′={C′1,C′2,...,C′ N};

[0016] 7) Species division operation, the process is as follows: traverse all individuals in population C', if individual C' i , i∈{1,2,...,N} has been marked as a certain species, do not handle; otherwise mark the individual as a new species and set the number of individuals of the population to 1, take the individual as the target individual, traverse all unmarked individuals after the target individual, if unmarked individual C' j , j∈{i+1,...,N} and target individual C' i Conformation similarity d(C' i , C' j ) meets formula (1), (2), (3), then mark C' j and C' i as the same species, and increase the number of individuals of the species, traverse to complete to obtain T populations P1, P2,..., Pt T , formula (1), (2), (3) are as follows:

[0017] d(C' i , C' j )<r g (1)

[0018]

[0019] r g =d init ×(d cut / d init ) g / e (3)

[0020] Wherein, r g is the radius of the niche in the gth iteration, are the dihedral angles of the lth residue of individual C' i , respectively, are the dihedral angles of the lth residue of individual C' j , L is the sequence length of the target protein;

[0021] 8) Individual fitness adjustment operation: traverse all individuals in population C', calculate the fitness of individual C' i , i∈{1,2,...,N} as follows:

[0022]

[0023]

[0024]

[0025] Wherein, E score3 (C' i ) is the fitness of individual C'the Rosetta score3 energy value of C i , E max is the largest energy value in the population, m g (C′ i ) is the number of individuals sharing fitness with individual C′ i in the gth iteration, C′ k is the population P i in which individual C′ t is located, t∈{1,2,...,T}, Sh(C′ i ,C′ k ) is the fitness decay ratio of individual C′ i and C′ k , d(C′ i ,C′ k ) is the conformation similarity of individual C′ i and C′ k ;

[0026] 9) Population updating operation, the process is as follows:

[0027] 9.1) Select individual C′ i in the population as the target conformation, and then randomly select an individual C′ i different from C′ a , a≠i, recombine a random fragment in C′ a to the corresponding position of individual C′ i to form a recombined individual C i ″;

[0028] 9.2) Use Rosetta to perform fragment assembly on the recombined individual C i ″ to generate a mutated individual C i ″′;

[0029] 9.3) Calculate the fitness f i (C g ″′) of the mutated individual C i ″′ according to formula (4), if f g (C i ″′) < f g (C′ i ), then the mutated individual C i ″′ replaces C′ i ; otherwise, calculate the acceptance probability accept and generate a uniform random number rand, rand∈[0,1], if rand < P i , then the individual C i ″′ replaces C′ out ; otherwise, do not process;

[0030] 10) set g = g + 1, if g > G, execute step 11), otherwise, go to step 6);

[0031] 11) output N individuals with the lowest fitness value in the Gth generation population as the final prediction result. out

[0032] The technical concept of the present application is that, under the framework of the evolutionary algorithm, first, the population is divided into species according to the similarity between conformations; then, a fitness sharing function is constructed to change the fitness value of individuals in the population, so that it has the ability to cross the energy barrier, avoids repeatedly falling into the same energy trap, and is beneficial to explore other potential low-energy regions; finally, output several conformations with the lowest energy in the population as the final prediction result. The multi-modal protein structure prediction method based on fitness sharing can not only alleviate the problem of inaccurate protein energy model, but also enhance the conformation sampling ability, improve the sampling efficiency, and thus improve the prediction accuracy.

[0033] The present application has the advantages that: the sampling efficiency is improved according to the fitness sharing strategy; the output of multiple conformations alleviates the defect of inaccurate protein energy model, increases the conformation diversity, and thus improves the overall prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a conformation sampling schematic diagram when the method of the present application is used to predict the structure of protein 1FJRA.

[0035] Figure 2 is a three-dimensional structure diagram obtained by using the method of the present application to predict the structure of protein 1FJRA. DETAILED DESCRIPTION

[0036] The present application will be further described below with reference to the accompanying drawings.

[0037] Referring to Figure 1 and Figure 2 , a multi-modal protein structure prediction method based on fitness sharing, the method comprising the following steps:

[0038] 1) input the sequence information of the target protein;

[0039] 2) obtain the fragment library files of 3 fragments and 9 fragments according to the sequence of the target protein from the ROBETTA server;

[0040] 3) set parameters: maximum iteration number G, population size N, initial value of niche radius d init , cutoff value of niche radius d cut , simulated annealing rate e, temperature coefficient β, and output prediction model number N out ; ​

[0041] 4) Population initialization: Use the first and second phases of the Rosetta protocol to generate a population of size N, C = {C1, C2, ..., C N};

[0042] 5) Set g = 1, g∈{1,2,...,G};

[0043] 6) Population individual ranking: Use the score3 energy function in Rosetta to calculate the score of each individual C in population C. i , the energy value E of i∈{1,2,...,N} score3 (C i ), sort the population in ascending order according to the energy value to form a population C′={C′1,C′2,...,C′ N};

[0044] 7) Species division operation, the process is as follows: traverse all individuals in population C', if individual C' i , i∈{1,2,...,N} has been marked as a species, it will not be processed; otherwise, the individual is marked as a new species and the number of individuals in the population is set to 1. The individual is taken as the target individual, and all unmarked individuals after the target individual are traversed. If the unmarked individual C′ j , j∈{i+1,...,N} and the target individual C′ i The conformational similarity d(C′ i ,C′ j ) satisfies formulas (1), (2), and (3), then mark C′ j and C′ i For the same species, and increase the number of individuals of the species, traverse

[0045] After completion, we get T populations P1, P2, ..., P T , formulas (1), (2), and (3) are as follows:

[0046] d(C′ i ,C′ j ) <r g (1)

[0047]

[0048] r g =d init ×(d cut / d init ) g / e (3)

[0049] Among them, r g is the radius of the niche in the g-th iteration, Individual C′i The dihedral angle of the lth residue, Individual C′ j The dihedral angle of the lth residue of , L is the sequence length of the target protein;

[0050] 8) Individual fitness adjustment operation: traverse all individuals in the population C′ and calculate individual C′ according to the following formula i , the fitness of i∈{1,2,...,N}:

[0051]

[0052]

[0053]

[0054] Among them, E score3 (C′ i ) is individual C′ i Rosetta score3 energy value, E max is the maximum energy value in the population, m g (C′ i ) is the number of individual C′ in the g-th iteration i The number of individuals sharing fitness, C′ k For individual C′ i Population P t , other individuals in t∈{1,2,...,T}, Sh(C′ i ,C′ k ) is individual C′ i and C′ k The fitness decay ratio, d(C′ i ,C′ k ) is individual C′ i and C′ k The conformational similarity of

[0055] 9) Population update operation, the process is as follows:

[0056] 9.1) Select individuals C′ in the population one by one i As the target conformation, randomly select a conformation that is similar to C′ i Different individuals C′ a , a≠i, C′ a A random fragment in the recombinant individual C' i The corresponding position of the recombinant individual C i ″;

[0057] 9.2) Using Rosetta to analyze the recombinant individual C i Perform a fragment assembly to generate variant individual Ci ″′;

[0058] 9.3) Calculate the variant individual C according to formula (4) i The fitness f of ″′ g (C i ″′), if f g (C i ″′) <f g (C′ i ), then the mutant individual C i ″′ replaces C′ i ; Otherwise calculate the reception probability And generate a uniform random decimal rand, rand∈[0,1], if rand <P accept , then individual C i ″′ replaces C′ i ; Otherwise, no processing;

[0059] 10) Set g = g + 1. If g > G, proceed to step 11); otherwise, go to step 6);

[0060] 11) Output the N with the lowest fitness value in the G generation population out The individuals are taken as the final prediction results.

[0061] This example uses the protein 1FJRA with a sequence length of 188 as an example, and provides a multimodal protein structure prediction method based on fitness sharing. The method includes the following steps:

[0062] 1) Input the sequence information of the target protein;

[0063] 2) Obtain 3-fragment and 9-fragment fragment library files from the ROBETTA server according to the target protein sequence;

[0064] 3) Set parameters: maximum number of iterations G, population size N, initial value of microhabitat radius d init , microhabitat radius cutoff value d cut , simulated annealing rate e, temperature coefficient β, number of output prediction models N out ;

[0065] 4) Population initialization: Use the first and second phases of the Rosetta protocol to generate a population of size N, C = {C1, C2, ..., C N};

[0066] 5) Set g = 1, g∈{1,2,...,G};

[0067] 6) Population individual ranking: Use the score3 energy function in Rosetta to calculate the score of each individual C in population C. i, the energy value E of i∈{1,2,...,N} score3 (C i ), sort the population in ascending order according to the energy value to form a population C′={C′1,C′2,...,C′ N};

[0068] 7) Species division operation, the process is as follows: traverse all individuals in population C', if individual C' i , i∈{1,2,...,N} has been marked as a species, it will not be processed; otherwise, the individual is marked as a new species and the number of individuals in the population is set to 1. The individual is taken as the target individual, and all unmarked individuals after the target individual are traversed. If the unmarked individual C′ j , j∈{i+1,...,N} and the target individual C′ i The conformational similarity d(C′ i ,C′ j ) satisfies formulas (1), (2), and (3), then mark C′ j and C′ i For the same species, and increase the number of individuals of the species, traverse

[0069] After completion, we get T populations P1, P2, ..., P T , formulas (1), (2), and (3) are as follows:

[0070] d(C′ i ,C′ j ) <r g (1)

[0071]

[0072] r g =d init ×(d cut / d init ) g / e (3)

[0073] Among them, r g is the radius of the niche in the g-th iteration, Individual C′ i The dihedral angle of the lth residue, Individual C′ j The dihedral angle of the lth residue of , where L is the sequence length of the target protein;

[0074] 8) Individual fitness adjustment operation: traverse all individuals in the population C′ and calculate individual C′ according to the following formula i , the fitness of i∈{1,2,...,N}:

[0075]

[0076]

[0077]

[0078] Among them, E score3 (C′ i ) is individual C i ′’s Rosetta score3 energy value, E max is the maximum energy value in the population, m g (C′ i ) is the number of individual C′ in the g-th iteration i The number of individuals sharing fitness, C′ k For individual C′ i Population P t , other individuals in t∈{1,2,...,T}, Sh(C′ i ,C′ k ) is individual C′ i and C′ k The fitness decay ratio, d(C′ i ,C′ k ) is individual C′ i and C′ k The conformational similarity of

[0079] 9) Population update operation, the process is as follows:

[0080] 9.1) Select individuals C′ in the population one by one i As the target conformation, randomly select a conformation that is similar to C′ i Different individuals C′ a , a≠i, C′ a A random fragment in the recombinant individual C' i The corresponding position of the recombinant individual C i ″;

[0081] 9.2) Using Rosetta to analyze the recombinant individual C i Perform a fragment assembly to generate variant individual C i ″′;

[0082] 9.3) Calculate the variant individual C according to formula (4) i The fitness f of ″′ g (C i ″′), if f g (C i ″′) <f g (C′ i ), then the mutant individual C iReplace C' i ; otherwise, calculate the receiving probability and generate a uniform random number rand, rand ∈ [0, 1], if rand < P accept , then the individual C i Replace C' i ; otherwise, do not process;

[0083] 10) Set g = g + 1, if g > G, execute step 11); otherwise, go to step 6);

[0084] 11) Output N out individuals with the lowest fitness value in the Gth generation population as the final prediction result.

[0085] Taking the protein 1FJRA with a sequence length of 188 as an example, the near-natural conformation of the protein is obtained by using the above method, the conformation sampling schematic diagram is as shown in Figure 1 , and the root mean square deviation between the output 5 models and the natural state structure is respectively The predicted three-dimensional structure is as shown in Figure 2 .

[0086] The above is a good result exhibited by one embodiment given by the present application, obviously, the present application is not only suitable for the above embodiment, and various improvements can be made to the present application without deviating from the basic idea of the present application and without exceeding the substantial content of the present application.

Claims

1. A multi-modal protein structure prediction method based on fitness sharing, characterized in that: The method comprises the following steps: 1) inputting sequence information of a target protein; 2) obtaining a 3-segment and 9-segment fragment library file from a ROBETTA server according to the sequence of the target protein; 3) Set parameters: maximum iteration number G, population size N, initial niche radius d init , cutoff value of niche radius d cut , simulated annealing rate e, temperature coefficient β, output prediction model number N out ; 4) Population initialization: The first and second stages of the Rosetta protocol are used to generate a population C = {C1, C2,..., CN} of size N N}; 5) setting g = 1, g ∈ {1, 2,..., G}; 6) Sort the population: calculate the energy value E i of each individual C score3 in the population C using the score3 energy function in Rosetta i , i ∈ {1, 2,..., N}, and sort the population in ascending order of energy value, resulting in the population C' = {C'1, C'2,..., C'N} N . 7) Species division operation, the process is as follows: traverse all individuals in population C', if individual C' i , i∈{1,2,...,N} has been marked as a species, it will not be processed; otherwise, the individual is marked as a new species and the number of individuals in the population is set to 1. The individual is taken as the target individual, and all unmarked individuals after the target individual are traversed. If the unmarked individual C′ j , j∈{i+1,...,N} and the target individual C′ i The conformational similarity d(C′ i ,C′ j ) satisfies formulas (1), (2), and (3), then mark C′ j and C′ i For the same species, and increase the number of individuals of the species, after the traversal is completed, T populations P1, P2, ..., P T , formulas (1), (2), and (3) are as follows: d(C i ,C j )<r g (1) r g = d init × (d cut / d init ) g / e (3) wherein r g is the niche radius in the gth iteration, is the dihedral angle of the lth residue of the individual C' i is the dihedral angle of the lth residue of the individual C' is the dihedral angle of the lth residue of the individual C' j is the dihedral angle of the lth residue of the individual C', and L is the sequence length of the target protein. 8) Individual fitness adjustment operation: traverse all individuals in population C', calculate the fitness of individual C' i , i e {1, 2,..., N} according to the following formula: wherein E score3 (C′ i ) is the Rosetta score3 energy value of individual C′ i , E max is the largest energy value in the population, m g (C′ i ) is the number of individuals sharing fitness with individual C′ i in the gth iteration, C′ k is the population P i where individual C′ t is located, t∈{1,2,...,T}, Sh(C′ i ,C′ k ) is the fitness decay ratio of individual C′ i and C′ k , d(C′ i ,C′ k ) is the conformation similarity of individual C′ i and C′ k ; 9) population updating operation, the process is as follows: 9.1) sequentially selecting individuals C' from the population i As the target conformation, randomly select another individual C' different from C' i , a a≠i, recombine a random fragment in C' to the corresponding position of individual C' to form a recombined individual C" a , i i ;​ 9.2) Recombinant individual C" is generated using Rosetta i Once the assembly of fragments is performed, the variant individual C'" is generated i ; 9.3) Calculate the variant individual C″′ according to formula (4) i The fitness f g (C″′ i ), if f g (C″′ i ) <f g (C′ i ), then the variant individual C″′ i Replace C′ i ; Otherwise calculate the reception probability And generate a uniform random decimal rand, rand∈[0,1], if rand <P accept , then individual C″′ i Replace C′ i ; Otherwise, no processing; 10) setting g = g + 1, if g > G, executing step 11); otherwise, going to step 6); 11) output the N individuals with the lowest fitness values in the Gth generation population as the final prediction result. out 11) output the N individuals with the lowest fitness values in the Gth generation population as the final prediction result.

Images