A multi-domain protein assembly method based on multi-objective optimization

Abstract

A multi-domain protein assembly method based on multi-objective optimization, an initial full-chain conformation population is constructed by obtaining single-domain structure through AlphaFold2, and then a cross mutation operation is performed on the inter-domain dihedral angle of the conformation to generate a new conformation; secondly, the full-chain distance features are obtained by using AlphaFold2, and the inter-domain orientation information is combined with DeepAssembly to construct two energy functions; finally, a multi-objective optimization algorithm is designed to obtain a non-dominated solution set from the conformation space to update the population; after a certain number of cyclic iterations, including cross mutation, conformation sampling and population updating operation, the final conformation set is obtained; then, the conformation is selected by using the model quality evaluation technology. The present application provides rich information for structure modeling; and provides a higher quality conformation set for high-precision protein structure.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and computer applications, and in particular to a multi-domain protein assembly method based on multi-objective optimization. Background Technology

[0002] Proteins, as the main executors of life activities, undertake a variety of biological functions. Their sequence determines their three-dimensional structure, and the three-dimensional structure determines their function. Determining the three-dimensional structure of proteins is of great significance for a deeper understanding of protein functional mechanisms, protein-protein interactions, and the development of new drugs. However, the complex and diverse structures of proteins make their determination challenging. Traditional experimental methods, such as X-ray crystallography and nuclear magnetic resonance, require significant time and resources, limiting the speed and scale of protein structure determination. Therefore, determining protein structure has always been a major challenge in biology.

[0003] In recent years, with the rapid development of machine learning, AlphaFold2 has made milestone progress in the field of single-domain protein structure prediction, providing a solid foundation for many protein structure prediction works. Protein domains represent tightly packed and independent folded units within a protein structure; multi-domain proteins in nature are often formed by the repetition and recombination of single domains. Although AlphaFold2 has made breakthrough progress on single-domain proteins, it should be noted that modeling remains a challenge for analytical tasks involving multi-domain proteins with weak co-evolutionary information and large structures. In this regard, accurately capturing the orientation of interactions between domains and the distance features within domains is crucial.

[0004] The flexibility between multiple protein domains presents a significant challenge to the task of multi-domain protein structure resolution. While various multi-domain protein structure prediction methods exist, such as Rosetta, AIDA, DEMO, SADA, E2EDA, and DeepAssembly, these methods typically optimize only for a single objective and struggle to comprehensively address the complexity of multi-domain proteins. Therefore, coordinating the use of high-precision single-domain structure prediction and multi-domain orientation prediction methods offered by AlphaFold2 for multi-objective optimization is crucial for solving the problem of predicting the structure of large multi-domain proteins with insufficient evolutionary information. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies and to effectively utilize high-precision single-domain protein structure information and inter-domain orientation prediction methods, this invention proposes a multi-domain protein assembly method based on multi-objective optimization. By using AlphaFold2 to predict intra-domain distance features and DeepAssembly to predict inter-domain orientation information, a multi-objective optimization algorithm is designed to guide the conformation sampling process. Finally, model quality assessment technology is used to select the structure with the highest accuracy.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] A multi-domain protein assembly method based on multi-objective optimization, the method comprising the following steps:

[0008] 1) Given the amino acid sequence of the target protein;

[0009] 2) Based on the given sequence, the full-strand amino acid sequence is divided into single-domain sequences using the domain divider DomBpred. The obtained single-domain sequences and full-strand sequences are then input into AlphaFold2 to obtain single-domain structures and full-strand distance maps. In addition, the full-strand sequence is input into the DeepAssembly method to obtain inter-domain interactions.

[0010] 3) Set parameters: initial population size N, crossover factor p c Acceptance probability p a Maximum number of sampling iterations G max The initial iteration number g = 0;

[0011] 4) Constructing the initial conformation population: Based on the single structural domains obtained in step 2), randomly initialize the dihedral angles between structural domains to obtain N full-chain structures C. i i = 1, 2, 3, ..., N, and thus construct the initial conformation population;

[0012] 5) For each conformation C in the conformational population i To perform a crossover / mutation operation, follow these steps:

[0013] 5.1) For conformation C in the population i The dihedral angle Θ between its protein domains i ={θ i,j} i＝1,...,N,j＝1,...,M Mutation is performed using formula (1), where N represents the population size and M represents the dihedral dimension:

[0014]

[0015] Where i, f1, f2 ∈ {1, 2, ..., N}, i ≠ f1 ≠ f2, j ∈ {1, 2, ..., M}, Let θ represent the j-th dihedral angle in the i-th conformation. f1,j and θ f2,j It is the dihedral angle between two randomly selected conformations in the population;

[0016] 5.2) After the mutation operation, perform the conformational crossover operation using formula (2):

[0017]

[0018] It is the dihedral angle corresponding to the mutated conformation, F is a scaling factor; rand(0,1) is a random number between 0 and 1, p c Represents the cross factor;

[0019] After performing sub-steps 5.1) and 5.2), the conformation C after crossover mutation is obtained. mut ;

[0020] 6) Set the energy function, the process is as follows:

[0021] 6.1) Energy function f for interaction between structural domains inter (x), the formula is as follows:

[0022]

[0023] Where l is the number of connections between structural domains, d i,j This represents the predicted distance between the i-th and j-th residues in the target conformation. This represents the actual distance between the i-th and j-th residues;

[0024] 6.2) Energy function f of interaction within the structural domain full (x), the formula is as follows:

[0025]

[0026] Where L represents the protein sequence length, d w,q , The C-values ​​between the w-th and q-th residues of the target conformation are respectively α The predicted distance and the actual distance are related by ε, a constant set to prevent the denominator from being zero; ε is a very small constant.

[0027] 7) Multi-objective optimization algorithm, the process is as follows:

[0028] Using the energy function obtained in step 6), a multi-objective energy model f(x) is constructed, and the formula is as follows:

[0029] min x f(x) = [f1(x), f1(x), ..., fi (x)] (5)

[0030] stx∈Ω (6)

[0031] Among them, f i (x), i∈{1,2,3,...} is the objective energy function, and x is the decision vector in the decision space Ω. For two conformations x in the population a and x b ,if and Then x a Regarding the control of x b (x a <x b If there is no decision vector dominating x in the entire decision space, * Then x * This is called a Pareto optimal solution or a non-dominated solution;

[0032] 8) Population renewal, the process is as follows:

[0033] All conformations in the conformational population undergo crossover and mutation, followed by multi-objective optimization to generate non-dominated solutions. For conformation x... * If there is no decision space that satisfies x < x * If the conformation x is such that x is a given shape, then x is a given shape. * For non-dominated solutions, accept with probability 1. Furthermore, for dominated solutions x′, accept with probability p. a Receive with a probability of , and all obtained x * x′ and x′ are used to update the population to form a new conformation set, when the maximum number of iterations G is reached. max At that time, the final set of conformations is output;

[0034] 9) Loop: g = g + 1, if g < G max If the condition is met, return to step 5) and continue the loop; otherwise, proceed to step 10.

[0035] 10) Model quality assessment, the process is as follows:

[0036] The DeepUMQA-rank algorithm, used to evaluate model quality, was used to select the highest-ranking model in the final set.

[0037] The model;

[0038] 11) Output the results.

[0039] The technical concept of this invention is as follows: First, an initial full-chain conformation population is constructed using the single-domain structure obtained by AlphaFold2, followed by conformational crossover and mutation operations; second, full-chain distance features are obtained using AlphaFold2, and DeepAssembly is used to predict inter-domain orientation information, thereby constructing an energy function; finally, a multi-objective optimization algorithm is designed to sample conformations to update the population, iterating crossover and mutation, conformation sampling, and population update operations, and finally, model quality evaluation techniques are used to select conformations.

[0040] The beneficial effects of this invention are as follows: First, it cleverly integrates intra-domain distance features and inter-domain orientation information to provide rich information for structural modeling; second, it designs a multi-objective optimization strategy to sample protein conformations from all directions, providing a higher quality conformation set for high-precision protein structures. Attached Figure Description

[0041] Figure 1 This is the overall flowchart of the method.

[0042] Figure 2 This is a schematic diagram of a multi-objective optimization algorithm.

[0043] Figure 3 The three-dimensional structure of the 2xwbf protein was obtained by assembling a multi-domain protein assembly method based on multi-objective optimization. Detailed Implementation

[0044] The present invention will now be further described with reference to the accompanying drawings.

[0045] Reference Figures 1-3 A multi-domain protein assembly method based on multi-objective optimization, the method comprising the following steps:

[0046] 1) Given the amino acid sequence of the target protein;

[0047] 2) Based on the given sequence, the domain divider DomBpred is used to divide the full-strand amino acid sequence into single-domain sequences. The obtained single-domain sequences and full-strand sequences are then input into AlphaFold2 to obtain single-domain structures and full-strand distance maps. In addition, the full-strand sequence is input into the DeepAssembly method to obtain inter-domain interactions.

[0048] 3) Set parameters: initial population size N, crossover factor p c Acceptance probability p a Maximum number of sampling iterations G max The initial iteration number g = 0;

[0049] 4) Constructing the initial conformation population: Based on the single structural domains obtained in step 2), randomly initialize the dihedral angles between structural domains to obtain N full-chain structures C. ii = 1, 2, 3, ..., N, and thus construct the initial conformation population;

[0050] 5) For each conformation C in the conformational population i To perform a crossover / mutation operation, follow these steps:

[0051] 5.1) For conformation C in the population i The dihedral angle Θ between its protein domains i ={θ i,j} i＝1,...,N,j＝1,...,M Mutation is performed using formula (1), where N represents the population size and M represents the dihedral dimension:

[0052]

[0053] Where i, f1, f2 ∈ {1, 2, ..., N}, i ≠ f1 ≠ f2, j ∈ {1, 2, ..., M}, Let θ represent the j-th dihedral angle in the i-th conformation. f1,j and θ f2,j It is the dihedral angle between two randomly selected conformations in the population;

[0054] 5.2) After the mutation operation, perform the conformational crossover operation using formula (2):

[0055]

[0056] It is the dihedral angle corresponding to the mutated conformation, F is a scaling factor; rand(0,1) is a random number between 0 and 1, p c Represents the cross factor;

[0057] After performing steps 5.1) and 5.2), the conformation C after crossover mutation is obtained. mut ;

[0058] 6) Set the energy function, the process is as follows:

[0059] 6.1) Energy function f for interaction between structural domains inter (x), the formula is as follows:

[0060]

[0061] Where l is the number of connections between structural domains, d i,j This represents the predicted distance between the i-th and j-th residues in the target conformation. This represents the actual distance between the i-th and j-th residues;

[0062] 6.2) Energy function f of interaction within the structural domain full(x), the formula is as follows:

[0063]

[0064] Where L represents the protein sequence length, d w,q , The C-values ​​between the w-th and q-th residues of the target conformation are respectively... α The predicted distance and the actual distance are related by ε, which is a constant set to prevent the denominator from being zero; it is a very small constant.

[0065] 7) Multi-objective optimization algorithm, the process is as follows:

[0066] Construct a multi-objective energy model f(x) using the energy function obtained in step 6), such as Figure 2 As shown, the formula is constructed as follows:

[0067] min x f(x) = [f1(x), f1(x), ..., f i (x)] (11)

[0068] stx∈Ω (12)

[0069] Among them, f i (x), i∈{1,2,3,...} is the objective energy function, and x is the decision vector in the decision space Ω. For two conformations x in the population a and x b ,if and Then x a Regarding the control of x b (x a <x b If there is no decision vector dominating x in the entire decision space, * Then x * This is called a Pareto optimal solution or a non-dominated solution;

[0070] 8) Population renewal, the process is as follows:

[0071] All conformations in the conformational population undergo crossover and mutation, followed by multi-objective optimization to generate non-dominated solutions. For conformation x... * If there is no decision space that satisfies x < x * If the conformation x is such that x is a given shape, then x is a given shape. * For non-dominated solutions, accept with probability 1. Furthermore, for dominated solutions x′, accept with probability p. a Receive with a probability of , and all obtained x * x′ and x′ are used to update the population to form a new conformation set, when the maximum number of iterations G is reached. max At that time, the final set of conformations is output;

[0072] 9) Loop: g = g + 1, if g < G max If the condition is met, return to step 5) and continue the loop; otherwise, proceed to step 10.

[0073] 10) Model quality assessment:

[0074] The DeepUMQA-rank algorithm, used to evaluate model quality, was used to select the model with the highest ranking in the final conformation set.

[0075] 11) Output the results.

[0076] This implementation example uses a 764-byte protein 2xwbf as an example. A multi-domain protein assembly method based on multi-objective optimization includes the following steps:

[0077] 1) Given the amino acid sequence of the target protein;

[0078] 2) Based on the given sequence, the full-strand amino acid sequence is divided into single-domain sequences using the domain divider DomBpred. The obtained single-domain sequences and full-strand sequences are then input into AlphaFold2 to obtain single-domain structures and full-strand distance maps. In addition, the full-strand sequence is input into the DeepAssembly method to obtain inter-domain interactions.

[0079] 3) Setting parameters: Initial population size N = 10, crossover factor p c =0.5, acceptance probability p a =0.1, maximum number of sampling iterations G max =500, initial iteration algebra g=0;

[0080] 4) Constructing the initial conformation population: Based on the single structural domains obtained in step 2), randomly initialize the dihedral angles between structural domains to obtain N full-chain structures C. i i = 1, 2, 3, ..., N, and thus construct the initial conformation population;

[0081] 5) For each conformation C in the conformational population i To perform a crossover / mutation operation, follow these steps:

[0082] 5.1) For conformation C in the population i The dihedral angle Θ between its protein domains i ={θ i,j} i＝1,...,N,j＝1,...,M Mutation is performed using formula (1), where N represents the population size and M represents the dihedral dimension:

[0083]

[0084] Where i, f1, f2 ∈ {1, 2, ..., N}, i ≠ f1 ≠ f2, j ∈ {1, 2, ..., M}, Let θ represent the j-th dihedral angle in the i-th conformation. f1,j and θ f2,j It is the dihedral angle of two randomly selected conformations in the population.

[0085] 5.2) After the mutation operation, perform the conformational crossover operation using formula (2):

[0086]

[0087] It is the dihedral angle corresponding to the mutated conformation, F is a scaling factor; rand(0,1) is a random number between 0 and 1, p c Represents the cross factor;

[0088] After performing steps 5.1) and 5.2), the conformation C after crossover mutation is obtained. mut ;

[0089] 6) Set the energy function, the process is as follows:

[0090] 6.1) Energy function f for interaction between structural domains inter (x), the formula is as follows:

[0091]

[0092] Where l is the number of connections between structural domains, d i,j This represents the predicted distance between the i-th and j-th residues in the target conformation. This represents the actual distance between the i-th and j-th residues;

[0093] 6.2) Energy function f of interaction within the structural domain full (x), the formula is as follows:

[0094]

[0095] Where L represents the protein sequence length, d w,q , The C-values ​​between the w-th and q-th residues of the target conformation are respectively... α The predicted distance and the actual distance are represented by ε, which is a very small number set to prevent the denominator from being zero; here it is set to 1e-8.

[0096] 7) Multi-objective optimization algorithm, the process is as follows:

[0097] The multi-objective energy model f(x) is constructed using the energy function obtained in step 6), and the formula is as follows:

[0098] min x f(x) = [f1(x), f1(x), ..., f i (x)] (17)

[0099] stx∈Ω (18)

[0100] Among them, f i (x), i∈{1,2,3,...} is the objective energy function, and x is the decision vector in the decision space Ω. For two conformations x in the population a and x b ,if and Then x a Regarding the control of x b (x a <x b If there is no decision vector dominating x in the entire decision space, * Then x * This is called a Pareto optimal solution or a non-dominated solution;

[0101] 8) Population renewal, the process is as follows:

[0102] All conformations in the conformational population undergo crossover and mutation, followed by multi-objective optimization to generate non-dominated solutions. For conformation x... * If there is no decision space that satisfies x < x * If the conformation x is such that x is a given shape, then x is a given shape. * For non-dominated solutions, accept with probability 1. Furthermore, for dominated solutions x′, accept with probability p. a Receive with a probability of , and all obtained x * x′ and x′ are used to update the population to form a new conformation set, when the maximum number of iterations G is reached. max At that time, the final set of conformations is output;

[0103] 9) Loop: g = g + 1, if g < G max If the condition is met, return to step 5) and continue the loop; otherwise, proceed to step 10.

[0104] 10) Model quality assessment:

[0105] The DeepUMQA-rank algorithm, used to evaluate model quality, was used to select the model with the highest ranking in the final conformation set.

[0106] 11) Output the results.

[0107] Taking the 764-byte protein 2xwbf as an example, the conformation of this multi-domain protein was predicted using the method described above. After 500 iterations of a multi-objective optimization algorithm, the root mean square error between the predicted structure and the native structure was obtained. The TM-score is 0.974, and the resulting 3D structure diagram is as follows. Figure 3 As shown.

[0108] The above describes the results of an example provided by the present invention. Obviously, the present invention is not only suitable for the above embodiments, 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 multi-domain protein assembly method based on multi-objective optimization, characterized in that, The method includes the following steps: 1) Given the amino acid sequence of the target protein; 2) Based on the given sequence, the domain divider DomBpred is used to divide the full-strand amino acid sequence into single-domain sequences. The obtained single-domain sequences and full-strand sequences are then input into AlphaFold2 to obtain single-domain structures and full-strand distance maps. In addition, the full-strand sequence is input into the DeepAssembly method to obtain inter-domain interactions. 3) Set parameters: initial population size N, crossover factor p c Acceptance probability p a Maximum number of sampling iterations G max The initial iteration number g = 0; 4) Constructing the initial conformation population: Based on the single structural domains obtained in step 2), randomly initialize the dihedral angles between structural domains to obtain N full-chain structures C. i i = 1, 2, 3, ..., N, and thus construct the initial conformation population; 5) For each conformation C in the conformational population i To perform a crossover / mutation operation, follow these steps: 5.1) For conformation C in the population i The dihedral angle Θ between its protein domains i ={θ i,j } i＝1,...,N,j＝1,...,M Mutation is performed using formula (1), where N represents the population size and M represents the dihedral dimension: Where i, f1, f2 ∈ {1, 2, ..., N}, i ≠ f1 ≠ f2, j ∈ {1, 2, ..., M}, This represents the j-th dihedral angle in the i-th conformation. and It is the dihedral angle between two randomly selected conformations in the population; 5.2) After the mutation operation, perform the conformational crossover operation using formula (2): It is the dihedral angle corresponding to the mutated conformation, F is a scaling factor; rand(0,1) is a random number between 0 and 1, p c Represents the cross factor; After performing steps 5.1) and 5.2), the conformation C after crossover mutation is obtained. mut ; 6) Set the energy function, the process is as follows: 6.1) Energy function f for interaction between structural domains inter (x), the formula is as follows: in, l is the number of connections between structural domains, d i,j This represents the predicted distance between the i-th and j-th residues in the target conformation. This represents the actual distance between the i-th and j-th residues; 6.2) Energy function f of interaction within the structural domain full (x), the formula is as follows: Where L represents the protein sequence length, d w,q , The C-values ​​between the w-th and q-th residues of the target conformation are respectively... α The predicted distance and the actual distance are related by ε, which is a constant set to prevent the denominator from being zero; it is a very small constant. 7) Multi-objective optimization algorithm, the process is as follows: The multi-objective energy model f(x) is constructed using the energy function obtained in step 6), and the formula is as follows: min x f(x)=[f1(x),f1(x),...,f i (x)] (5) stx∈Ω(6) where, f i (x), i∈{1,2,3,...} is the objective energy function, and x is the decision vector in the decision space Ω. For two conformations x in the population a and x b If f i (x a )≤f i (x b ), And f i (x a )<f i (x b ), Then x a Regarding control If there is no decision vector dominating x in the entire decision space * Then x * This is called a Pareto optimal solution or a non-dominated solution; 8) Population renewal, the process is as follows: After crossover and mutation, all conformations in the conformational population undergo multi-objective optimization to generate non-dominated solutions. For conformation x... * If there is no condition in the decision space that satisfies If the conformation x is such that x is a given shape, then x is a given shape. * For non-dominated solutions, accept with probability 1; for dominated solutions x′, accept with probability p. a Receive with a probability of , and all obtained x * x′ and x′ are used to update the population to form a new conformation set, when the maximum number of iterations G is reached. max At that time, the final set of conformations is output; 9) Loop: g = g + 1, if g < G max If the condition is met, return to step 5) and continue the loop; otherwise, proceed to step 10. 10) Model quality assessment: The DeepUMQA-rank algorithm, used to evaluate model quality, was used to select the model with the highest ranking in the final conformation set. 11) Output the results.

Images