Method and apparatus for optimizing protein sequence on basis of genetic algorithm, and device and medium

Abstract

The present invention relates to the technical field of bioinformation, specifically to technical fields such as artificial intelligence and biocomputing. Provided are a method and apparatus for optimizing a protein sequence on the basis of a genetic algorithm, and a device and a medium. The method for optimizing a protein sequence on the basis of a genetic algorithm comprises: acquiring a first protein sequence to be optimized and at least one attribute, which is to be optimized, of said first protein sequence, wherein the first protein sequence comprises a plurality of first amino acids; mutating a preset number of first amino acids in the first protein sequence to obtain a second protein sequence; using a set composed of a plurality of second protein sequences as an initial population; determining at least one optimization objective of the genetic algorithm on the basis of the at least one attribute; for each second protein sequence in the initial population, determining a first fitness of the second protein sequence for each optimization objective of the at least one optimization objective, wherein the first fitness indicates a degree of improvement of the second protein sequence with respect to a corresponding optimization objective; and performing multiple updates on the initial population by using the genetic algorithm and on the basis of the first fitness respectively corresponding to the plurality of second protein sequences, so as to obtain a plurality of optimized target protein sequences.

Description

Methods, apparatus, equipment, and media for protein sequence optimization based on genetic algorithms

[0001] Cross-reference of related applications

[0002] This application claims priority to Chinese patent application No. 202510307410.9, filed on March 14, 2025, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to the field of bioinformatics, specifically to the fields of artificial intelligence and biological computing, and particularly to a method and apparatus for protein sequence optimization based on genetic algorithms, a computing device, a storage medium, a computer program product, a protein molecule, and a pharmaceutical composition. Background Technology

[0004] Protein sequence optimization refers to adjusting one or more amino acids in a protein sequence to achieve targeted enhancement of specific protein properties (such as activity, stability, water solubility, etc.).

[0005] In related technologies, the protein sequence to be optimized is typically subjected to saturation mutations to obtain multiple mutant sequences, from which the sequence with the most attribute improvement is selected as the optimal sequence. For a protein sequence of length N, the number of mutant sequences obtained by a single-point saturation mutation is N*19, and the number of mutant sequences obtained by a two-point saturation mutation is N*(N-1)*19*19 / 2. As the number of mutation sites increases, the number of mutant sequences obtained increases exponentially. Selecting the optimal sequence from the mutant sequences through exhaustive search would consume a significant amount of time and computational resources.

[0006] Furthermore, in some cases, it is desirable to achieve synergistic optimization of multiple protein properties. For example, improving (or maintaining) activity while enhancing stability, or improving activity, stability, and water solubility. The need for synergistic optimization of multiple properties further increases the difficulty of protein sequence optimization, as existing methods struggle to obtain protein sequences containing a specified number of mutations that meet optimization requirements within a short timeframe. Summary of the Invention

[0007] It would be beneficial to provide a mechanism to alleviate, reduce, or even eliminate one or more of the aforementioned problems.

[0008] According to one aspect of this disclosure, a protein sequence optimization method based on a genetic algorithm is provided, comprising: obtaining a first protein sequence to be optimized and at least one attribute to be optimized of the first protein sequence, the first protein sequence comprising a plurality of first amino acids; mutating a predetermined number of first amino acids in the first protein sequence to obtain a second protein sequence; using a set of a plurality of second protein sequences as an initial population; determining at least one optimization objective of the genetic algorithm based on the at least one attribute; for each second protein sequence in the initial population, determining a first fitness of the second protein sequence on each of the at least one optimization objective, the first fitness indicating the degree of improvement of the second protein sequence on the corresponding optimization objective; and updating the initial population multiple times using the genetic algorithm based on the first fitness corresponding to each of the plurality of second protein sequences to obtain a plurality of optimized target protein sequences.

[0009] In some implementations, mutating a predetermined number of first amino acids in the first protein sequence includes: selecting a predetermined number of first amino acids from the first protein sequence according to a predetermined selection algorithm; and for each of the predetermined number of first amino acids, mutating the first amino acid to a second amino acid based on a predetermined amino acid substitution scoring matrix.

[0010] In some implementations, the step of mutating the first amino acid into a second amino acid based on a preset amino acid substitution scoring matrix includes: obtaining a plurality of candidate mutant amino acids for the first amino acid; for any candidate mutant amino acid among the plurality of candidate mutant amino acids: obtaining a substitution score between the first amino acid and the candidate mutant amino acid from the amino acid substitution scoring matrix, the substitution score indicating the probability of the first amino acid and the candidate mutant amino acid substituting for each other; determining the probability of the first amino acid mutating into the candidate mutant amino acid based on the substitution score between the first amino acid and the candidate mutant amino acid; and selecting one candidate mutant amino acid from the plurality of candidate mutant amino acids as the second amino acid based on the probability corresponding to each of the plurality of candidate mutant amino acids.

[0011] In some embodiments, the amino acid substitution scoring matrix includes a BLOSUM matrix or a PAM matrix.

[0012] In some implementations, the at least one optimization objective includes at least one attribute optimization objective corresponding to each of the at least one attribute, and wherein determining the first fitness of the second protein sequence on each of the at least one optimization objective includes: for each of the at least one attribute: using a trained prediction model corresponding to the attribute to determine a predicted value of the second protein sequence on that attribute; and performing a numerical conversion on the predicted value to obtain the first fitness of the second protein sequence on the corresponding attribute optimization objective.

[0013] In some implementations, the genetic algorithm is a multi-objective genetic algorithm, and wherein determining at least one optimization objective of the genetic algorithm based on the at least one attribute includes: in response to the at least one attribute including only one attribute, taking the sequence similarity optimization objective and the attribute optimization objective corresponding to that attribute as multiple optimization objectives of the multi-objective genetic algorithm; and in response to the at least one attribute including two or more attributes, taking the attribute optimization objectives corresponding to each of the two or more attributes as multiple optimization objectives of the multi-objective genetic algorithm.

[0014] In some implementations, each of the multiple updates includes: obtaining a current population, the current population comprising a plurality of third protein sequences; for any third protein sequence in the current population, obtaining a first fitness of the third protein sequence on each of the plurality of optimization objectives; determining a non-dominated solution set corresponding to the current population based on the first fitness of each of the plurality of third protein sequences, the non-dominated solution set comprising a plurality of fourth protein sequences; and performing selection, crossover, and / or mutation operations on a subset of the union of the current population and the non-dominated solution set to generate an updated population.

[0015] In some implementations, the genetic algorithm includes the SPEA2 algorithm, and the selection operation performed on a subset of the union of the current population and the non-dominated solution set includes: truncating the non-dominated solution set or expanding the non-dominated solution set using the current population based on the relative size of the number of the plurality of fourth protein sequences to a preset archive number to obtain a parent population, the parent population including a plurality of fifth protein sequences, the number of the plurality of fifth proteins being equal to the archive number; for each fifth protein sequence in the parent population, obtaining a second fitness of the fifth protein sequence, the second fitness being determined based on a first fitness of the fifth protein sequence on each of the plurality of optimization objectives, the second fitness indicating the overall score of the fifth protein sequence; and selecting parent individuals from the plurality of fifth protein sequences for performing crossover and / or mutation operations based on the second fitness corresponding to each of the plurality of fifth protein sequences.

[0016] In some embodiments, the genetic algorithm includes the NSGA-II algorithm, and wherein performing a selection operation on a subset of the union of the current population and the non-dominated solution set includes: for any third protein sequence in the current population, determining the non-dominated level and crowding distance of the third protein sequence based on a first fitness of the third protein sequence on each of the plurality of optimization objectives, the non-dominated level indicating dominance and subdomination information of the third protein sequence, and the crowding distance indicating the density of the third protein sequence in the non-dominated level; and selecting parent individuals from the plurality of third protein sequences for performing crossover and / or mutation operations based on the non-dominated level and crowding distance corresponding to each of the plurality of third protein sequences.

[0017] In some implementations, the genetic algorithm includes the PESA2 algorithm, and wherein performing a selection operation on a subset of the union of the current population and the non-dominated solution set includes: dividing the solution space occupied by the non-dominated solution set into a plurality of grids of equal size; determining a compression factor for each of the plurality of grids, the compression factor indicating the number of fourth protein sequences contained in the grid; selecting a target grid from the plurality of grids based on the compression factors corresponding to each of the plurality of grids; and selecting parent individuals from the fourth protein sequences within the target grid for performing crossover and / or mutation operations.

[0018] In some implementations, the plurality of target protein sequences is the non-dominated solution set obtained from the last update in the plurality of updates.

[0019] In some implementations, the genetic algorithm is a single-objective genetic algorithm, the at least one attribute includes only one attribute, and the step of determining at least one optimization objective of the genetic algorithm based on the at least one attribute includes: taking the attribute optimization objective corresponding to the attribute as the optimization objective of the single-objective genetic algorithm.

[0020] In some implementations, each of the multiple updates includes: obtaining a current population, the current population comprising a plurality of sixth protein sequences; for any sixth protein sequence in the current population, obtaining the first fitness of the sixth protein sequence on the optimization target; and performing selection, crossover, and / or mutation operations on the plurality of sixth protein sequences based on the first fitness corresponding to each of the plurality of sixth protein sequences to generate an updated population, wherein the plurality of target protein sequences are the plurality of sixth protein sequences with the highest first fitness in the population obtained in the last update.

[0021] In some embodiments, the method further includes representing the second protein sequence as a binary code, the binary code including a first binary code for indicating a mutation site and a second binary code for indicating a mutated second amino acid.

[0022] In some embodiments, the at least one property includes at least one of the following properties: activity, stability, solubility, expression level, toxicity, degradability, and immunogenicity.

[0023] In some implementations, any one of the at least one optimization objective is a sequence similarity optimization objective or an attribute optimization objective corresponding to an attribute. The sequence similarity optimization objective is to minimize the similarity between the protein sequence and other protein sequences in the population, and the attribute optimization objective corresponding to the attribute is to maximize the improvement of the protein sequence in that attribute.

[0024] According to one aspect of this disclosure, a protein sequence optimization apparatus based on a genetic algorithm is provided, comprising: an acquisition unit configured to acquire a first protein sequence to be optimized and at least one attribute to be optimized of the first protein sequence, the first protein sequence comprising a plurality of first amino acids; a mutation unit configured to mutate a predetermined number of first amino acids in the first protein sequence to obtain a second protein sequence; an initialization unit configured to use a set of a plurality of second protein sequences as an initial population; a first determination unit configured to determine at least one optimization objective of the genetic algorithm based on the at least one attribute; a second determination unit configured to determine, for each second protein sequence in the initial population, a first fitness of the second protein sequence on each of the at least one optimization objective, the first fitness indicating the degree of improvement of the second protein sequence on the corresponding optimization objective; and an update unit configured to update the initial population multiple times using the genetic algorithm based on the first fitness corresponding to each of the plurality of second protein sequences to obtain a plurality of optimized target protein sequences.

[0025] According to one aspect of this disclosure, a computing device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the method described above.

[0026] According to one aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the above-described method.

[0027] According to one aspect of this disclosure, a computer program product is provided, including computer program instructions, wherein the computer program instructions, when executed by a processor, implement the above-described method.

[0028] According to one aspect of this disclosure, a protein molecule is provided, wherein the sequence of the protein molecule is prepared by the method described above.

[0029] According to one aspect of this disclosure, a pharmaceutical composition is provided comprising the aforementioned protein molecule and pharmaceutically acceptable excipients.

[0030] According to one or more embodiments of this disclosure, a genetic algorithm is used to optimize a protein sequence by a specified number of mutations for one or more attribute optimization objectives. This can generate an optimal protein sequence that satisfies the specified attribute optimization objectives in a short time, greatly improving the optimization efficiency of the protein sequence and realizing targeted and efficient optimization of the protein sequence.

[0031] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0032] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0033] Figure 1 shows a flowchart of a protein sequence optimization method based on a genetic algorithm according to an embodiment of the present disclosure;

[0034] Figure 2 illustrates a schematic diagram of a protein sequence optimization process based on a multi-objective genetic algorithm according to an embodiment of the present disclosure;

[0035] Figures 3A-3C illustrate the synergistic optimization effect of the three properties of three test proteins according to embodiments of the present disclosure;

[0036] Figures 4A-4O illustrate the dual-attribute synergistic optimization effect of three test proteins according to embodiments of the present disclosure;

[0037] Figures 5A-5O illustrate the single-attribute optimization effects of three test proteins according to embodiments of the present disclosure;

[0038] Figure 6 shows a structural block diagram of a protein sequence optimization device based on a genetic algorithm according to an embodiment of the present disclosure; and

[0039] Figure 7 shows a structural block diagram of an exemplary computing device that can be used to implement embodiments of the present disclosure. Detailed Implementation

[0040] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0041] In this disclosure, unless otherwise stated, the use of terms such as "first," "second," etc., to describe various elements is not intended to limit the positional, temporal, or importance relationships of these elements; such terms are merely used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of that element, while in other cases, based on the context, they may refer to different instances.

[0042] The terminology used in the description of the various examples described in this disclosure is for the purpose of describing particular examples only and is not intended to be limiting. Unless the context explicitly indicates otherwise, an element may be one or more unless the number of elements is specifically limited. Furthermore, the term "and / or" as used in this disclosure covers any one of the listed items and all possible combinations thereof. The term "multiple" refers to two or more.

[0043] In recent years, with the development of artificial intelligence, significant breakthroughs have been achieved in the field of protein design. Especially in protein structure and function prediction, AI technology has demonstrated enormous potential. Protein design methods typically involve sequence alterations, but early protein design methods, limited by algorithms and training sets, could only modify protein properties through single-point mutations. When designing, researchers generally choose the result with the highest prediction performance or score to pursue the highest success rate. Therefore, the optimal approach is to perform saturation mutations on the protein sequence to be designed, selecting the result with the highest prediction score to obtain the best sequence. For a protein with a sequence length of N, the number of variants from a single-point saturation mutation is N*19, which is sufficient for protein optimization algorithms. However, with the expansion of datasets and advancements in algorithms, various models capable of predicting the impact of multiple-point mutations on protein properties have emerged. However, even with just two-point mutations, the number of all possible combinations for a length of N has reached N*(N-1)*19*19 / 2, a number that is difficult to optimize using exhaustive methods. Developing a method to obtain relatively optimal multi-point mutation results within a limited time is an urgent problem to be solved.

[0044] Furthermore, proteins possess a variety of properties, such as stability, activity, solubility, and expression levels. When engineering proteins, researchers aim to optimize multiple properties simultaneously—that is, to improve protein stability while maintaining or enhancing activity, or to improve all properties. This requirement significantly increases the difficulty of optimization because protein optimization algorithms in related technologies use different prediction models based on different properties. These models are independent of each other, and when optimizing using one model, the optimization results of other models cannot be taken into account.

[0045] To address the aforementioned problems, this disclosure provides a protein sequence optimization method based on a genetic algorithm. This method utilizes a genetic algorithm to optimize a protein sequence with a specified number of mutations targeting one or more attribute optimization objectives. It can generate an optimal protein sequence that satisfies the specified attribute optimization objectives in a short time, significantly improving the optimization efficiency of protein sequences and achieving targeted and efficient optimization.

[0046] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0047] According to one aspect of this disclosure, a protein sequence optimization method based on a genetic algorithm is provided. Figure 1 shows a flowchart of a protein sequence optimization method 100 based on a genetic algorithm according to an embodiment of this disclosure. Method 100 can be executed by a computing device (e.g., computing device 700 described below). As shown in Figure 1, method 100 includes steps S110-S160.

[0048] In step S110, a first protein sequence to be optimized and at least one attribute of the first protein sequence to be optimized are obtained. The first protein sequence includes a plurality of first amino acids.

[0049] In step S120, a predetermined number of first amino acids in the first protein sequence are mutated to obtain a second protein sequence.

[0050] In step S130, the set consisting of multiple second protein sequences is used as the initial population.

[0051] In step S140, at least one optimization objective of the genetic algorithm is determined based on at least one of the above attributes.

[0052] In step S150, for each second protein sequence in the initial population, a first fitness of the second protein sequence on each of the at least one optimization objective is determined. The first fitness indicates the degree of improvement of the second protein sequence on the corresponding optimization objective.

[0053] In step S160, based on the first fitness corresponding to each of the multiple second protein sequences, the initial population is updated multiple times using a genetic algorithm to obtain multiple optimized target protein sequences.

[0054] According to embodiments of this disclosure, a genetic algorithm is used to optimize a protein sequence by a specified number of mutations for one or more attribute optimization objectives. This can generate an optimal protein sequence that satisfies the specified attribute optimization objectives in a short time, greatly improving the optimization efficiency of the protein sequence and achieving targeted and efficient optimization of the protein sequence.

[0055] In step S110, a first protein sequence to be optimized and at least one attribute of the first protein sequence to be optimized are obtained. The first protein sequence includes a plurality of first amino acids.

[0056] According to some embodiments, the first protein sequence can be the sequence information of any protein. The first protein sequence includes, but is not limited to: protein amino acid sequence information obtained by mass spectrometry using peptide mass spectrometry and tandem mass spectrometry; protein sequences deduced by transcribing complementary deoxyribonucleic acid (cDNA) from messenger ribonucleic acid (mRNA) and then sequencing deoxyribonucleic acid (DNA); artificially designed protein sequences; computer-designed protein sequences; and protein sequences from NCBI databases, EXProt databases, UniProtKB databases, and other common bioinformatics databases. It is understood that the first protein sequence consists of multiple amino acids. In the embodiments of this disclosure, the amino acid in the first protein sequence is referred to as the first amino acid.

[0057] The first protein sequence has multiple properties, such as activity, stability, solubility (water solubility), expression level, toxicity, degradability, and immunogenicity. One or more of these properties can be selected as the attributes to be optimized. The number and type of attributes to be optimized can be specified according to the optimization requirements.

[0058] It should be noted that attribute optimization refers to adjusting the attribute values ​​of a first protein sequence in a desired direction (i.e., the optimization direction, such as increasing or decreasing them). The direction of attribute optimization is determined based on the intended use of the first protein sequence. For example, for a first protein sequence intended for use as an enzyme, antibody, or drug to treat human diseases, it is generally desirable to increase its activity, stability, water solubility, expression level, etc., while decreasing its toxicity or immunogenicity; while for a first protein sequence intended for use as an insecticide, it is generally desirable to increase its toxicity.

[0059] In step S120, a predetermined number of first amino acids in the first protein sequence are mutated to obtain a second protein sequence. The predetermined number refers to the number of mutation sites, which can be set according to optimization requirements, such as 1, 2, 3, 5, 7, 9, 11, 13, 15, etc. That is, it can be a single amino acid mutation in the first protein sequence, or it can be multiple amino acid mutations at 2, 3, 5, 7, or 9 sites in the first protein sequence.

[0060] By repeatedly executing step S120, multiple second protein sequences can be obtained. Each second protein sequence includes the same number (i.e., the preset number) of mutation sites, but the locations of the mutation sites may differ. For example, if the length of the first protein sequence is 100 and the preset number of mutation sites is 3, the mutation sites of the second protein sequence prot1 can be 20, 30, or 50; the mutation sites of the second protein sequence prot2 can be 4, 22, or 60; and the mutation sites of the third protein sequence prot3 can be 1, 56, or 89, etc.

[0061] According to some embodiments, step S120 may include steps S121 and S122.

[0062] In step S121, a preset number of first amino acids are selected from the first protein sequence according to a preset selection algorithm. The preset selection algorithm can be any selection algorithm, such as completely random selection, grouped random selection, equidistant sampling, selection using a sliding window of the preset length, etc.

[0063] In step S122, for each of the first amino acids in the preset number of first amino acids, the first amino acid is mutated into a second amino acid based on the preset amino acid substitution scoring matrix.

[0064] An amino acid substitution score matrix indicates the probability of substitutions between specific amino acids during evolution. Using this score matrix to mutate each amino acid mutation site improves the likelihood of mutations occurring. By repeatedly executing steps S121-S122, multiple different and reasonable second protein sequences can be obtained. Using these obtained second protein sequences as the initial population in step S130 enhances the diversity and rationality of the initial population, enabling the genetic algorithm to effectively search the feasible solution space and thus improve its optimization performance.

[0065] According to some embodiments, for each of the first amino acids in the selected preset number of first amino acids, the first amino acid can be mutated into a second amino acid by the following steps S1221-S1223.

[0066] In step S1221, multiple candidate mutant amino acids of the first amino acid are obtained.

[0067] In step S1222, for any candidate mutant amino acid among a plurality of candidate mutant amino acids: the substitution score between the first amino acid and the candidate mutant amino acid is obtained from the amino acid substitution scoring matrix, the substitution score indicating the probability of the first amino acid and the candidate mutant amino acid substituting for each other; and based on the substitution score between the first amino acid and the candidate mutant amino acid, the probability of the first amino acid mutating into the candidate mutant amino acid is determined.

[0068] In step S1223, based on the probability of each of the multiple candidate mutant amino acids, one candidate mutant amino acid is selected from the multiple candidate mutant amino acids as the second amino acid.

[0069] According to the above embodiments, the probability of the first amino acid mutation being a candidate mutated amino acid is determined based on the amino acid substitution scoring matrix, and then the second amino acid after the mutation of the first amino acid is determined based on the mutation probability. This can improve the rationality of the mutation, thereby improving the rationality and feasibility of the initial population, and thus improving the optimization effect of the genetic algorithm.

[0070] According to some embodiments, the candidate mutant amino acids can be, for example, other amino acids from a predetermined set of amino acids besides the first amino acid. For example, the amino acid set may include 20 natural amino acids: A (alanine), V (valine), L (leucine), I (isoleucine), F (phenylalanine), W (tryptophan), M (methionine), P (proline), G (glycine), T (threonine), S (serine), C (cysteine), N (aspartic acid), Q (glutamine), Y (tyrosine), K (lysine), R (arginine), H (histidine), D (aspartic acid), and E (glutamic acid). The first amino acid can be any one of the amino acid sets, and correspondingly, the multiple candidate mutant amino acids of the first amino acid are another 19 amino acids from the same set.

[0071] According to some embodiments, the amino acid substitution scoring matrix can be a BLOSUM (BLOCKS SUbstitution Matrix) or a PAM (Point Accepted Mutation) matrix. BLOSUM includes multiple versions such as BLOSUM62 and BLOSUM80, while PAM matrices include multiple versions such as PAM1, PAM30, and PAM250. In practice, any version of the matrix (e.g., BLOSUM62) can be selected as the amino acid substitution scoring matrix for determining the mutated second amino acid, as needed.

[0072] According to some embodiments, the probability P(w,m) of the first amino acid w mutating into the candidate mutant amino acid m can be calculated according to the following formula:

[0073] Wherein, BLOSUM(w,m) is the substitution score between the first amino acid w and the candidate mutant amino acid m obtained from the BLOSUM matrix, and S is the set of candidate mutant amino acids for the first amino acid w. It should be understood that the BLOSUM matrix can be replaced with the PAM matrix, and correspondingly, BLOSUM(w,m) in equation (1) can be replaced with PAM(w,m), that is, the substitution score between the first amino acid w and the candidate mutant amino acid m is obtained from the PAM matrix.

[0074] According to some embodiments, there are various ways to determine the second amino acid based on probability in step 1223. For example, the candidate mutant amino acid with the highest probability can be directly selected as the second amino acid; alternatively, a candidate mutant amino acid can be selected as the second amino acid according to the probability distribution formed by the probabilities of each candidate mutant amino acid.

[0075] Step S120 yields the mutated second protein sequence. According to some embodiments, method 100 may further include representing the second protein sequence as a binary code, the binary code including a first binary code indicating the mutation site and a second binary code indicating the mutated second amino acid.

[0076] The second protein sequence may contain many amino acids. Directly representing the second protein sequence using amino acid characters would consume significant storage space and reduce computational efficiency. According to the above embodiments, by representing the second protein sequence as binary codes indicating mutation sites and the mutated second amino acid, the representation of the second protein sequence can be simplified, thereby saving storage space and improving computational efficiency, enabling rapid sequence optimization.

[0077] According to some embodiments, by analyzing N in a first protein sequence of length L... mut By mutating amino acids at several mutation sites, a second protein sequence of length L can be obtained. Each mutation site can be used to generate a second protein sequence of length L. The first binary code representation of a bit, where, This indicates rounding up. The second amino acid resulting from the mutation at each mutation site can be represented by a length of... The second binary code representation uses 5 bits to represent 20 common amino acids. Therefore, each mutation site in the second protein sequence can be represented as a sequence of length [length missing]. The binary encoding of bits, correspondingly, contains N mut The length of the binary code of the second protein sequence at each mutation site is [length missing].

[0078] In step S140, at least one optimization objective of the genetic algorithm is determined based on at least one attribute to be optimized. The at least one optimization objective is used to define the optimization problem to be solved by the genetic algorithm.

[0079] In some embodiments, each attribute to be optimized corresponds to an attribute optimization objective. The attribute optimization objective is to maximize the improvement of the protein sequence in that attribute. In addition to attribute optimization objectives, other optimization objectives can also be set, such as sequence similarity optimization objectives. The similarity optimization objective is to minimize the similarity between the protein sequence and other protein sequences in the population.

[0080] According to some embodiments, when the genetic algorithm is a single-objective genetic algorithm (i.e., a genetic algorithm with only one optimization objective), there is only one attribute to be optimized, and the attribute optimization objective corresponding to that attribute is the optimization objective of the single-objective genetic algorithm.

[0081] According to some embodiments, when the genetic algorithm is a multi-objective genetic algorithm (i.e., a genetic algorithm with multiple optimization objectives), step S140 may include steps S141 and S142.

[0082] In step S141, in response to the fact that at least one attribute to be optimized includes only one attribute, the sequence similarity optimization objective and the attribute optimization objective corresponding to that attribute are used as multiple optimization objectives of the multi-objective genetic algorithm. By combining the sequence similarity optimization objective, the optimization problem for a single attribute can be transformed into a multi-objective optimization problem, thereby enabling the optimization of a single attribute using the multi-objective genetic algorithm and improving the versatility of the multi-objective genetic algorithm for optimization problems with different numbers of attributes.

[0083] In step S142, in response to the fact that at least one attribute to be optimized includes two or more attributes, the attribute optimization objectives corresponding to each of these two or more attributes are taken as multiple optimization objectives of the multi-objective genetic algorithm. It can be understood that for optimization problems involving multiple attributes, the multi-objective genetic algorithm can be directly used to solve them, thus eliminating the need to combine sequence similarity optimization objectives.

[0084] After determining the initial population and optimization target through steps S130 and S140, fitness calculation and population update can be performed through steps S150 and S160 to obtain multiple optimized target protein sequences.

[0085] In step S150, for each second protein sequence in the initial population, a first fitness of the second protein sequence on each of the at least one optimization objective is determined. The first fitness indicates the degree of improvement of the second protein sequence on the corresponding optimization objective.

[0086] According to some embodiments, at least one optimization objective includes at least one attribute optimization objective corresponding to at least one attribute. Accordingly, in step S150, for each of the at least one attribute, a trained prediction model corresponding to that attribute can be used to determine the predicted value of the second protein sequence on that attribute. Subsequently, the predicted value is numerically converted to obtain the first fitness of the second protein sequence on the corresponding attribute optimization objective.

[0087] According to some embodiments, each attribute corresponds to a trained prediction model. By inputting the encoding of the second protein sequence (e.g., binary encoding, amino acid character encoding, etc. as described above) into the prediction model, the predicted value of that attribute can be obtained from the model's output. This prediction model can, for example, be trained using the sample protein sequence and the attribute labels corresponding to that sample protein sequence.

[0088] After determining the predicted values ​​of the second protein sequence for each attribute, the first fitness of the second protein sequence for each attribute optimization target can be obtained by numerically transforming the predicted values ​​for each attribute. This numerical transformation facilitates subsequent data processing with the obtained first fitness, thereby improving computational efficiency and optimization effectiveness.

[0089] According to some embodiments, numerical transformation may include one or more of the following transformation operations: normalization transformation, negative value transformation, logarithmic transformation, etc.

[0090] According to some embodiments, numerical transformation may include normalization transformation. It should be noted that the predicted values ​​output by prediction models for different attributes may have different ranges. In multi-attribute optimization problems, to improve computational efficiency and ensure the feasibility of the optimization solution, it is necessary to normalize the predicted values ​​output by each prediction model to the same target numerical range, so that the first fitness values ​​on different attribute optimization objectives have the same order of magnitude. The target numerical range may, for example, be the range of 0 to 1. Depending on different data processing requirements, this range may include boundary points or may not include boundary points; that is, the range of 0 to 1 may specifically be [0,1], [0,1], (0,1], or (0,1).

[0091] According to some embodiments, numerical transformation may include normalization transformation and negative value transformation. As described above, at least one optimization objective is used to define the optimization problem solved by the genetic algorithm. To improve solution efficiency and accuracy, the optimization problem is usually expressed as a minimum value problem. Accordingly, the first fitness corresponding to each attribute optimization objective is negatively correlated with the improvement of the second protein sequence on that attribute optimization objective. That is, the more the second protein sequence improves on that attribute optimization objective, the smaller its first fitness value on that attribute optimization objective. For example, for an attribute that is expected to improve (i.e., the larger the value of the attribute, the better), the predicted value output by the prediction model corresponding to the attribute can be normalized to the range of 0 to 1, and the negative value of the normalized value can be taken (i.e., the normalized value is converted to a negative value). This negative value is used as the first fitness of the attribute optimization objective, thereby making the first fitness of the attribute optimization objective negatively correlated with the improvement of the second protein sequence on that attribute optimization objective.

[0092] According to some embodiments, for the attribute A to be improved, numerical transformation may include normalization transformation and negative value transformation. The first fitness f of the second protein sequence on the optimization target of attribute A. A For example, it can be calculated using the following formula:

[0093] Where, p A This represents the predicted value of the second protein sequence for attribute A, output by the prediction model corresponding to attribute A. A min A These represent the maximum and minimum predicted values ​​of each second protein sequence in the initial population for attribute A, respectively.

[0094] According to some embodiments, for attribute B that is expected to be reduced (i.e., the smaller the value of this attribute, the better), the numerical transformation may include a normalization transformation. The first fitness f of the second protein sequence on the optimization objective of attribute B. B For example, it can be calculated using the following formula:

[0095] Where, p B This represents the predicted value of the second protein sequence for attribute B, output by the prediction model corresponding to attribute B. B min B These represent the maximum and minimum predicted values ​​of each second protein sequence in the initial population for attribute B, respectively.

[0096] According to some embodiments, numerical transformation may include logarithmic transformation. For example, the natural logarithm (ln) or the base-10 logarithm (lg) can be used to transform the predicted value of an attribute output by the predictive model into the first fitness on the corresponding attribute optimization objective. Logarithmic transformation can reduce data volatility and facilitate the interpretation of linear relationships between data.

[0097] According to some embodiments, when at least one optimization objective includes a sequence similarity objective, a first fitness of the second protein sequence on the sequence similarity objective can be calculated using a preset formula. Generally, it is desirable for the second protein sequence to have as little similarity as possible to other second protein sequences in the population. The first fitness f of the second protein sequence on the sequence similarity objective is... simi For example, it can be calculated using the following formula:

[0098] Where Similarity(i) represents the similarity between the second protein sequence i and other second protein sequences in the initial population P, and N pop Mut represents the number of second protein sequences included in the initial population P. i Mut j These are the sets of mutated amino acids for the second protein sequences i and j, respectively. Intersection(Mut i Mut j ) indicates Mut i and Mut j The number of elements included in the intersection of N mut This represents the number of mutation sites.

[0099] In step S160, based on the first fitness corresponding to each of the multiple second protein sequences, the initial population is updated multiple times using a genetic algorithm to obtain multiple optimized target protein sequences.

[0100] In each update of step S160, selection, crossover, and mutation operations are performed based on the first fitness of individuals in the current population to generate offspring populations. This step selects the optimal sequence that performs best overall across all objectives for evolution (crossover and mutation), thereby passing on superior genes to the next generation while maintaining population diversity by introducing random changes.

[0101] Genetic algorithms include multi-objective genetic algorithms and single-objective genetic algorithms. Multi-objective genetic algorithms include, but are not limited to, the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and its derivatives, the Strength Pareto Evolutionary Algorithm 2 (SPEA2), the Pareto Envelope-based Selection Algorithm 2 (PESA2), the NSGA-III algorithm and its derivatives, the SEMO2 algorithm (Simple Evolutionary Multi-objective Optimizer 2), the VEGA algorithm (Vector Evaluated Genetic Algorithm), the ESPEA algorithm (Evolutionary Search and Pareto Estimation Algorithm), and the PAES algorithm (Pareto Archived Evolution Strategy).

[0102] According to some embodiments, for a multi-objective genetic algorithm, each update in step S160 may include steps S161-S164.

[0103] In step S161, the current population is obtained, which includes multiple third protein sequences.

[0104] In step S162, for any third protein sequence in the current population, the first fitness of the third protein sequence on each of the multiple optimization objectives is obtained.

[0105] In step S163, based on the first fitness corresponding to each of the multiple third protein sequences, the non-dominated solution set corresponding to the current population is determined, and the non-dominated solution set includes multiple fourth protein sequences.

[0106] In step S164, selection, crossover, and / or mutation operations are performed on a subset of the union of the current population and the non-dominated solution set to generate an updated population.

[0107] Steps S161-S164 can be repeated multiple times until a preset termination condition is met. The preset termination condition may be, for example, that the number of population updates reaches a preset number (e.g., 100 times), the first fitness reaches a plateau and no better results are produced, or some solutions meet the minimum criteria for the optimal solution.

[0108] When the preset termination condition is met, the non-dominated solution set in the current population (i.e., the last generation population, the population obtained from the last update) is used as multiple optimized target protein sequences.

[0109] It should be understood that, for step S161, the current population in the first update is the initial population, and the multiple third protein sequences at this time are the multiple second protein sequences described above. The current population in the second and subsequent updates is the population after the previous update.

[0110] According to some embodiments, for step S162, the first fitness of the third protein sequence on each optimization target can be calculated with reference to the above equations (2)-(5). Specifically, the second protein sequence in the above equations (2)-(5) can be replaced with the third protein sequence to obtain the first fitness of the third protein sequence on each optimization target.

[0111] According to some embodiments, for step S163, the non-dominated solution set corresponding to the current population can be determined in different ways for different multi-objective genetic algorithms. For example, for the SPEA2 algorithm, the non-dominated solution set corresponding to the current population can be a set composed of the non-dominated solutions in the current population and the current external archive. For the NSGA-II and PESA2 algorithms, the non-dominated solution set corresponding to the current population can be a set composed of the non-dominated solutions in the current population.

[0112] It should be understood that a non-dominated solution set refers to a set in which, for any given individual, there is no other individual in the set that is superior to that individual in all optimization objectives. In simpler terms, each individual in a non-dominated solution set has its merits, and there is no other individual that is completely superior to it.

[0113] In step S164, a subset of the union of the current population and the non-dominated solution set is taken as the parent population. Selection, crossover, and / or mutation operations are performed on the parent population to generate an updated population, i.e., the next generation population. It should be noted that the method for determining the parent population and the method for selecting parent individuals from the parent population can differ for different multi-objective genetic algorithms.

[0114] According to some embodiments, for the SPEA2 algorithm, the selection operation in step S164 may include the following steps S1641-S1643.

[0115] In step S1641, based on the relative size of the number of multiple fourth protein sequences and the preset number of archives, the non-dominated solution set is truncated or expanded using the current population to obtain the parent population. This parent population includes multiple fifth protein sequences, and the number of fifth proteins is equal to the number of archives.

[0116] In step S1642, for each fifth protein sequence in the parent population, the second fitness of the fifth protein sequence is obtained. The second fitness is determined based on the first fitness of the fifth protein sequence on each of the multiple optimization objectives. The second fitness indicates the overall score of the fifth protein sequence.

[0117] In step S1643, based on the second fitness corresponding to each of the multiple fifth protein sequences, a parent individual for performing crossover and / or mutation operations is selected from the multiple fifth protein sequences.

[0118] According to some embodiments, in step S1641, the number of archives refers to the number of individuals included in the external archive.

[0119] If the number of fourth protein sequences included in the current non-dominated solution set is greater than the number of archives, the non-dominated solution set is truncated. Specifically, the distance between the first fitness vector of each fourth protein sequence in the non-dominated solution set and the first fitness vectors of other fourth protein sequences can be calculated. The fourth protein sequences with the smallest distances are removed sequentially until the number of fourth protein sequences in the non-dominated solution set equals the number of archives. The non-dominated solution set at this point is the parent population.

[0120] If the number of fourth protein sequences included in the current non-dominated solution set is less than the number of archived sequences, the non-dominated solution set can be expanded using the current population. Specifically, the second fitness of each dominant solution in the current population can be obtained. The second fitness is determined based on the first fitness and is used to indicate the overall score of each dominant solution. The specific calculation method of the second fitness will be detailed below. The dominant solutions with the lowest second fitness are sequentially added to the non-dominated solution set until the number of individuals in the filled set equals the number of archived sequences. The filled set at this point is the parent population.

[0121] According to some embodiments, in step S1643, an individual with a lower fitness can be selected from the parent population as a parent individual according to a preset selection operator. The selection operator for the parent individual includes, but is not limited to, binary tournament operator, roulette wheel operator, sorting selection operator, elite selection operator, etc.

[0122] By performing crossover and / or mutation operations on parent individuals, multiple offspring individuals can be obtained, thus creating the next generation population. The crossover rate for crossover operations and the mutation rate for mutation operations can both be set as needed.

[0123] The population update process of the SPEA2 algorithm will be described in detail in the examples below.

[0124] According to some embodiments, for the NSGA-II algorithm, the current population is used as the parent population. The selection operation in step S164 may include the following steps S1644-S1645.

[0125] In step S1644, for any third protein sequence in the current population, based on the first fitness of the third protein sequence on each of the multiple optimization objectives, the non-dominated level and crowding distance of the third protein sequence are determined. The non-dominated level indicates the dominance and subdominance information of the third protein sequence, and the crowding distance indicates the density of the third protein sequence within the non-dominated level. The calculation methods for the non-dominated level and crowding distance will be detailed below.

[0126] In step S1645, a parent individual for performing crossover and / or mutation operations is selected from the multiple third protein sequences based on the non-dominated hierarchy and crowding distance corresponding to each of the multiple third protein sequences.

[0127] According to some embodiments, in step S1645, individuals with higher non-dominant levels and larger crowding distances can be selected from the parent population as parent individuals according to a preset selection operator. The selection operator for parent individuals includes, but is not limited to, binary tournament operator, roulette wheel operator, sorting selection operator, elite selection operator, etc.

[0128] Multiple offspring individuals can be obtained by performing crossover and / or mutation operations on parent individuals. The crossover rate and mutation rate can be set as needed. Parent and offspring individuals are then merged, and a non-dominated sort is performed again. Individuals at the top of the non-dominated hierarchy are selected to form the next generation population.

[0129] The population update process of the NSGA-II algorithm will be described in detail in the examples below.

[0130] According to some embodiments, for the PESA2 algorithm, the non-dominated solution set corresponding to the current population is used as the parent population. The selection operation in step S164 may include the following steps S1646-S1649.

[0131] In step S1646, the solution space occupied by the non-dominated solution set is divided into multiple grids of the same size.

[0132] In step S1647, the compression factor of each of the multiple grids is determined. The compression factor indicates the number of fourth protein sequences contained in that grid. The calculation method of the compression factor will be described in detail below.

[0133] In step S1648, a target grid is selected from multiple grids based on the compression factors corresponding to each grid.

[0134] In step S1649, a parent individual is selected from the fourth protein sequence within the target grid to perform crossover and / or mutation operations.

[0135] According to some embodiments, in step S1648, the grid with the smaller compression factor can be selected as the target grid from multiple grids according to a preset selection operator. The selection operator includes, but is not limited to, the binary tournament operator, the roulette wheel operator, the sorting selection operator, and the elite selection operator.

[0136] According to some embodiments, in step S1649, a fourth protein sequence within the target grid can be randomly selected as the parent individual.

[0137] By performing crossover and / or mutation operations on parent individuals, multiple offspring individuals can be obtained, thus creating the next generation population. The crossover rate for crossover operations and the mutation rate for mutation operations can both be set as needed.

[0138] The population update process of the PESA2 algorithm will be described in detail in the examples below.

[0139] According to some embodiments, for a single-objective genetic algorithm, each update in step S160 may include steps S165-S167.

[0140] In step S165, the current population is obtained, which includes multiple sixth protein sequences. It should be understood that, for step S165, the current population in the first update is the initial population, and the multiple sixth protein sequences at this time are the multiple second protein sequences described above. The current population in the second and subsequent updates is the population after the previous update.

[0141] In step S166, for any sixth protein sequence in the current population, the first fitness of the sixth protein sequence on the optimization target is obtained. The first fitness of the sixth protein sequence on the optimization target can be calculated with reference to the above equations (2)-(5). Specifically, the second protein sequence in the above equations (2)-(5) can be replaced with the sixth protein sequence to obtain the first fitness of the sixth protein sequence on the optimization target.

[0142] In step S167, based on the first fitness corresponding to each of the multiple sixth protein sequences, selection, crossover, and / or mutation operations are performed on the multiple sixth protein sequences to generate an updated population. In step S167, a sixth protein sequence with a lower first fitness can be selected as a parent individual for crossover and / or mutation operations to generate the next generation population.

[0143] Steps S165-S167 can be repeated multiple times until a preset termination condition is met. The preset termination condition may be, for example, that the number of population updates reaches a preset number (e.g., 100 times), the first fitness reaches a plateau and no better results are produced, or some solutions meet the minimum criteria for the optimal solution.

[0144] When the preset termination condition is met, the multiple sixth protein sequences with the highest first fitness in the current population (i.e., the last generation population, the population obtained from the last update) are selected as multiple optimized target protein sequences.

[0145] Figure 2 illustrates a schematic diagram of a protein sequence optimization process based on a multi-objective genetic algorithm according to an embodiment of the present disclosure. As shown in Figure 2, the optimization process of the multi-objective genetic algorithm includes four steps: population initialization, fitness calculation, population update (selection, crossover, mutation), and iterative evolution.

[0146] In step 1, the optimized sequence population is initialized based on the initial sequence.

[0147] It should be understood that the initial sequence in Figure 2 corresponds to the first protein sequence described above; the initial population corresponds to the initial population described above, where optimized sequence 1, optimized sequence 2, ..., optimized sequence N correspond to the multiple second protein sequences described above.

[0148] Multi-objective genetic algorithms start with an initial population and mimic natural selection and genetic mechanisms to find the optimal solution. A high-quality initial population is a crucial foundation for the algorithm's iterative optimization and can significantly improve its performance. The initial population should ideally possess high diversity, strong feasibility, and broad coverage to help the algorithm effectively search the feasible solution space and avoid being confined to a finite or infeasible target space, thus preventing it from quickly getting stuck in local optima. Population initialization methods include random initialization, distributed sampling initialization, and heuristic initialization. Heuristic initialization, which incorporates domain knowledge, can be combined with other initialization methods such as random initialization to obtain an initial solution that better suits the characteristics of the target problem.

[0149] Currently, various bioinformatics methods are available for obtaining protein evolutionary information, including but not limited to using amino acid substitution scoring matrices (e.g., PAM matrices or BLOSUM62 matrices) to estimate the probability of specific amino acid substitutions occurring during evolution. In the embodiments of this disclosure, each individual in the population represents a possible protein mutant. During population initialization, initial mutation sites are selected using random initialization or distributed sampling initialization methods, and initial post-mutated amino acids are selected based on protein evolutionary information (e.g., PAM matrices or BLOSUM62 matrices), resulting in an initial optimized sequence population that combines diversity and feasibility.

[0150] In step 2, fitness is calculated.

[0151] With the continuous development of deep learning, there are now various models that can be used to predict various properties of proteins and mutation effects, including models such as EnzyAct and SCANEER that predict the effect of mutations on protein activity, models such as ProSTAGE, Mutate_everthing, DDGun, MAESTRO, and FoldX that predict the effect of mutations on protein stability, and models such as CamSol, Protein-Sol, PaRSnIP, DeepSol-S2, SKADE, SWI, SoluProt, GraphSol, EPSOL, NetSoIP, DeepSoluE, HybridGCN, and ProG-SOL that predict protein solubility.

[0152] Inputting the initial sequence obtained in step 1 into this type of model (corresponding to the prediction model described above) allows us to obtain the changes in various properties of these mutants compared to the initial sequence. By numerically transforming the predicted values ​​of the artificial intelligence model according to the specific task, such as protein activity, stability, or solubility, we can obtain the fitness score for each individual corresponding to the specific optimization problem. The numerical transformation methods involved include negative value transformation, logarithmic transformation, and normalization transformation.

[0153] In step 3, a population update is performed.

[0154] Selection, crossover, and mutation operations are performed based on individual fitness to generate a progeny population. This step selects the optimal sequence with the best overall performance across all objectives for evolution (crossover and mutation), thereby passing on superior genes to the next generation and maintaining population diversity by introducing random changes. Currently, various evolutionary-based algorithms have been used to solve multi-objective optimization problems. These algorithms employ diverse methods for population updates to improve optimization performance and convergence speed. Specific methods include, but are not limited to: NSGA-II, SPEA2, PESA2, NSGA-II, SEMO2, VEGA, ESPEA, and PAES. Among them, NSGA-II, SPEA2, and PESA2 are preferred algorithms.

[0155] In step 4, steps 2-3 are repeated until the termination condition is met. At this point, the population P of generation t is... t The non-dominated solutions are the optimization results of the initial sequence. Termination conditions include, for example, reaching a fixed number of generations, the optimal fitness reaching a plateau and no better results being produced, and some solutions meeting the minimum criteria for the optimal solution.

[0156] The following describes examples of using multi-objective genetic algorithms for three-attribute collaborative optimization, two-attribute collaborative optimization, and single-attribute optimization.

[0157] Example 1: Synergistic Optimization of Three Objectives: Protein Activity, Stability, and Solubility

[0158] 1. Methods and Materials

[0159] 1.1 Test Data

[0160] The test data included three representative proteins within the common amino acid length range: oleic acid cyclase (Uniprot ID: I6WU39) with a length of 101 amino acids, green fluorescent protein (Uniprot ID: P42212) with a length of 238 amino acids, and cytochrome P450 2C9 protein (Uniprot ID: P11712) with a length of 490 amino acids.

[0161] 1.2 Definition of Optimization Problem

[0162] The optimization objective of this embodiment is to find a set of sequences such that, for any given sequence, no other sequence in the set is superior to it in all three objectives: activity, stability, and solubility. Such a solution set is called a non-dominated solution set, representing the optimal solution after weighing different objectives in a multi-objective optimization problem. In simpler terms, each sequence in the non-dominated solution set has its merits, and no other sequence is completely superior. Based on this, the goal is to maximize the improvement (and similarly, reduce or maintain) of the sequences in the final solution set in all three objectives: activity, stability, and solubility.

[0163] Given an individual x, the optimization problem is defined as: min V(x)=(f act (x),f stab (x),f sol (x)). Where, f act (x),f stab (x),f sol (x) represents the fitness values ​​(corresponding to the first fitness mentioned above) for the activity, stability, and solubility of sequence x, respectively. The smaller the fitness value, the greater the improvement of sequence x in that attribute. See 1.5 for details on the calculation of fitness scores.

[0164] 1.3 Individual Coding

[0165] Instead of directly representing specific mutation types, the algorithm encodes all individuals involved in the algorithm using binary strings, thereby using less computer memory and achieving faster computation speed.

[0166] Each individual in the population is composed of a concatenation of the binary codes of all the mutations involved; the binary code of each mutation consists of a concatenation of the binary codes of the mutation site and the mutated amino acids. For wild-type proteins of length L (which can also be artificially designed or machine-designed proteins), N is performed. mut For sequence optimization at mutation sites, a length of is used. Bits are represented by binary codes to indicate mutation sites, using a length of 160 ... That is, 5 bits of binary code represent 20 common amino acids, and the final binary code length for each individual is... In this embodiment, L takes values ​​of 101, 238, and 490; N mut The value can be 2, 3, 5, 7, or 9. N mut The value can be freely defined according to the specific problem.

[0167] 1.4 Population Initialization

[0168] Given a wild-type protein of length L, perform N mutSequence optimization at each mutation site. Initially, N are randomly selected for each individual. mut There are 10 unique mutation sites. For each mutation site, the mutated amino acid is randomly selected based on the substitution frequency of the wild-type amino acid with other amino acids at that site in the BLOSUM62 matrix (Henikoff and Henikoff 1992). Specifically, given a wild-type amino acid w and a set S of its selectable mutated amino acids, the probability that amino acid w mutates into amino acid m is:

[0169] Repeat this process to generate a sequence containing N pop An initial population P0 of mutant sequences. Additionally, an external archive P0 is set. ′ P0 is used to store non-dominated solutions. ′ The size is limited to N archive In one embodiment, N can be set. pop and N archive All are 100.

[0170] 1.5 Fitness Calculation

[0171] Three classifiers M based on protein property prediction methods were selected respectively. act M stab M sol This study predicts the changes in protein activity, stability, and solubility of each mutant sequence in the current population relative to the initial sequence, and standardizes the predicted values. Specifically, the maximum predicted value of the stability and solubility of the sequences in the initial population P0 is set to max. stab_P0 max sol_P0 The minimum value is min stab_P0 min sol_P0 Given a sequence, assume its activity model predicts a value of p. act The stability model predicts a value of p. stab The soluble model predicts a value of p. sol Then its fitness f takes the following values:

[0172] It should be noted that, in this embodiment, due to classifier M act The output p act The value range of p is [0,1], therefore there is no need to use the prediction of the maximum and minimum values ​​to adjust p. act Normalization is performed, requiring only negative value transformation, thus transforming the optimization problem into finding the minimum fitness f.

[0173] The predicted values ​​of the three classifiers for each mutant sequence are transformed according to Equation (7) above to obtain the fitness, and then the population is updated based on the fitness. In Equation (7), f...act f stab f sol This can be understood as three objective functions.

[0174] 1.6 Population Update

[0175] Population updates were performed using three multi-objective genetic algorithms: SPEA2 (Corne et al., 2001), NSGA-II (Deb et al., 2002), and PESA2 (Zitzler, 2001). The update process selected dominant individuals in the population as parents, and then performed crossover and mutation operations on these parents according to preset crossover and mutation rates to obtain offspring individuals, forming a new generation of the population.

[0176] 1.6.1 Population Update Based on SPEA2 Algorithm

[0177] 1.6.1.1 Calculate Raw Fitness score and density value

[0178] Based on the fitness f obtained in 1.5, the dominance and subordination information of each individual is calculated, and then their Raw Fitness score and density value are calculated. The fitness is then updated to F (corresponding to the second fitness mentioned above). The specific formula for calculating fitness F is as follows: S(i)=|{j|j∈P t ∪P t ′ ∧j<i}| (8) F(i)=R(i)+D(i) (11)

[0179] Where j < i indicates that individual i dominates individual j, meaning that individual i is superior to individual j in all objectives, and vice versa. S(i) represents the strength value of individual i, representing the population P of individual i at time t (or generation t). t and external archive P t ′ The number of solutions dominated by the middle; R(i) represents the Raw Fitness score of individual i, representing P t and P ′ t The sum of the strength values ​​of all individuals dominating individual i, R(i) = 0 represents a non-dominated solution, not dominated by any other solution. Furthermore, the relationship between i and P is calculated. t ∪P t ′ The distance between the remaining individuals. The distance between individual i and individual j can be, for example, the objective function vector V(i) = (f... act (i),f stab (i),f sol (i)) and V(j)=(fact (j),f stab (j),f sol The Euclidean distance of (j) is calculated. These distances are sorted in ascending order, and the k-th distance is selected. This is used to estimate the density value D(i) of individual i. In one embodiment, k can be set to 20. Generally, the value of k is usually initially set to... Based on this, adjustments are made and multiple tests are conducted. The updated fitness F(i) = R(i) + D(i), where F(i) < 1 indicates that individual i is a non-dominated solution.

[0180] 1.6.1.2 Update P′ t Obtain P′ t+1

[0181] Based on the fitness F calculated in step 1.6.1.1, select P. t and P′ t All non-dominated solutions form P′ t+1 If P′ at this time t+1 Size (i.e., P′) t+1 The number of individuals included is greater than N. archive Then calculate the relationship between each individual and P′. t+1 The distances between other individuals are considered, and the individual with the smallest distance is eliminated sequentially until the number of non-dominated solutions equals N. archive If P′ at this time t+1 The size is less than N archive Then, sort the remaining dominant solutions in ascending order according to F, and select individuals with smaller F to fill P′. t+1 until P′ t+1 The size is equal to N archive .

[0182] 1.6.1.3 Selection Operator

[0183] Using the binary tournament operator from P′ t+1 The operator selects individuals as parents for subsequent evolutionary operations. Specifically, it randomly selects two individuals from the current population each time, compares their fitness scores, and chooses the individual with the better fitness value for crossover / mutation until the number of individuals in the population reaches a preset value. Depending on the specific problem, the selection operator can also be set as a roulette wheel operator, an elite selection operator, etc.

[0184] 1.6.1.4 Crossover Operator

[0185] In one embodiment, a binary crossover method can be used to crossover individuals, with the crossover rate set to, for example, 0.8. Depending on the specific problem, the crossover rate can be freely chosen within the range of 0.7-0.9, and the optimal crossover rate can be selected after multiple tests.

[0186] 1.6.1.5 Mutation Operator

[0187] Due to factors such as the redundancy of the genetic code, the frequency of changes at mutation sites in proteins does not entirely correspond to the frequency of changes in amino acids following mutations at specific mutation sites. In the embodiments of this disclosure, binary encoding is used to represent the mutated protein sequence. Knowing the length of the protein sequence and the number of mutation sites, the positions of the bits representing the mutation sites and the bits representing the mutated amino acids in the individual binary encoding can be obtained. Therefore, different mutation rates can be applied to both, set to 0.01 and 0.05 respectively. The two mutation rates are defined as MR... pos MR aa MR pos and MR aa The values ​​can be freely selected and combined between 0.001 and 0.1.

[0188] Update P to version 1.6.1.6 t Get P t+1

[0189] At P′ t+1 Individuals are selected based on their fitness F for crossover and mutation operations, and then added to the next generation population P. t+1 In, until P t+1 The size is equal to N pop .

[0190] 1.6.2 Population Update Based on NSGAII Algorithm

[0191] 1.6.2.1 Quick Non-Dominated Sort

[0192] Based on the fitness f obtained in section 1.5, a fast non-dominated sort is performed on the current population to obtain the non-dominated level corresponding to each individual. Individuals not dominated by any solution are classified as L0, solutions dominated only by L0 individuals are classified as L1, and so on to obtain the non-dominated level of all individuals.

[0193] 1.6.2.2 Calculating congestion

[0194] Given an individual i, define its crowding degree i. distance It is the sum of the distance differences between the two adjacent individuals on each objective function.

[0195] Update 1.6.2.3 obtained P t+1

[0196] Parent individuals are selected based on the non-dominated hierarchy obtained in 1.6.2.1 using the selection operator, and then offspring individuals are generated through crossover and mutation operators. The selection, crossover, and mutation operators are the same as those in Sections 1.6.1.3-1.6.1.5 above. For individuals in the same non-dominated hierarchy, their crowding distance is compared, and the individual with the larger crowding distance wins. This operation is repeated to obtain an N-generation hierarchy. pop The population of offspring of each individual. Then, merge the parent and offspring individuals, perform non-dominated sorting again, and select the Nth child from the highest non-dominated hierarchy. pop Individuals form the next generation population P. t+1 .

[0197] 1.6.3 Population Update Based on PESA2 Algorithm

[0198] 1.6.3.1 Mesh Generation

[0199] Based on the fitness f obtained in 1.5, the current population is non-dominated and sorted, and the non-dominated solution set at this time is taken as P′. t P′ t The occupied target space (i.e., the space consisting of the value range of each target in the non-dominated solution set) is uniformly divided into several hypercubes, and the compression factor of each hypercube is calculated. The compression factor is defined as the number of individuals contained in the hypercube.

[0200] Update 1.6.3.2 to obtain P t+1

[0201] Selection operators based on compression factors are used to select hypercubes for species evolution. Within the selected hypercube, individuals are randomly chosen as parents, and then offspring individuals are generated through crossover and mutation operators to join the next generation population P. t+1 until P t+1 The size is N pop The selection, crossover, and mutation operators are the same as those in Sections 1.6.1.3-1.6.1.5.

[0202] 1.7 Performance Indicators

[0203] In one embodiment, the hypervolume (HV) metric can be used to evaluate the performance of solving multi-objective optimization problems. This metric first selects the points corresponding to the worst values ​​of each objective in the objective space as reference points. Then, it calculates the volume of the hypercube formed by the non-dominated solution set obtained by the algorithm and the reference points in the objective space. A larger HV value indicates that the solution set obtained by the method is closer to the optimization direction for each objective, and the better the overall performance of the method. Here, the values ​​of each objective at the reference points are selected as the known worst values ​​of each objective during the calculation process.

[0204] 2 Results

[0205] To evaluate the optimization performance of three multi-objective genetic algorithms (SPEA2, NSGA-II, and PESA2), five tasks were tested for three test proteins, including 2-point, 3-point, 5-point, 7-point, and 9-point mutations. Simultaneously, three attributes—activity, stability, and solubility—were optimized. The resulting sequences obtained using the three different multi-objective genetic algorithms were compared with those obtained through random sampling. The evolutionary generation number and population size for all three genetic algorithms were set to 100 generations, meaning the genetic algorithms needed to predict the mutation effects of up to 10,000 mutations using a deep learning model. Since the optimization tasks in the examples could not exhaustively enumerate all mutations, a random sampling method was used to randomly sample 10,000 mutations from the solution space for comparison, corresponding to the genetic algorithms. This resulted in an overall computation time of 2-3 hours for all four methods. Specifically, the random sampling method randomly selects N mutation sites from all sites in a given protein sequence each time. For these sites, mutated residues are randomly selected indiscriminately from the remaining 19 common amino acids. This process is repeated 10,000 times, and three properties (activity, stability, and solubility) are predicted for these mutations. The sequences used for comparison are the non-dominated solution sets generated by the three genetic algorithms and the random sampling method, containing less than or equal to 100 mutated sequences. For the above four non-dominated solution sets, the overvolume index is calculated, and the comparison results are shown in Figures 3A-3C.

[0206] Figure 3A shows a comparison of the performance of three genetic algorithms and a random sampling algorithm for the sequence obtained from the I6WU39 protein (101 aa). Figure 3B shows a comparison of the performance of three genetic algorithms and a random sampling algorithm for the sequence obtained from the P42212 protein (238 aa). Figure 3C shows a comparison of the performance of three genetic algorithms and a random sampling algorithm for the sequence obtained from the P11712 protein (490 aa).

[0207] As shown in Figures 3A-3C, the non-dominated solution set obtained by the optimization method based on the multi-objective genetic algorithm in this disclosure has a much larger supervolume than that of the random sampling method, and is closer to the desired optimization direction in each objective attribute. Among them, the SPEA2 algorithm has the best overall performance in HV value across 15 tasks involving 2-point, 3-point, 5-point, 7-point, and 9-point mutation optimization for three proteins. It ranks first in 8 tasks and second in the remaining tasks, with HV values ​​close to those of the top-ranked method. Therefore, the SPEA2 algorithm can be preferred for sequence optimization problems involving three-attribute collaborative optimization.

[0208] Example 2: Dual-objective optimization of any combination of protein activity, stability, and solubility

[0209] 1. Methods and Materials

[0210] 1.1 Test Data

[0211] Similar to Example 1, three proteins, I6WU39, P42212, and P11712, were used for testing.

[0212] 1.2 Definition of Optimization Problem

[0213] Similar to Example 1, the optimization objectives are as follows:

[0214] 1) Find a set of sequences such that, for any given sequence, no other sequence in the set is superior to it in both activity and stability. Furthermore, this set of sequences should achieve the greatest possible improvement in both activity and stability.

[0215] 2) Find a set of sequences such that for any given sequence, no other sequence in the set is superior to it in both activity and solubility. Furthermore, this set of sequences should achieve the greatest possible improvement in both activity and solubility.

[0216] 3) Find a set of sequences such that, for any given sequence, no other sequence in the set is superior to it in both stability and solubility. Furthermore, this set of sequences should achieve the greatest possible improvement in both stability and solubility.

[0217] 1.3 Individual Coding

[0218] Consistent with Example 1.

[0219] 1.4 Population Initialization

[0220] Consistent with Example 1.

[0221] 1.5 Fitness Calculation

[0222] Similar to Example 1, three classifiers M based on artificial intelligence methods are selected. act M stab M sol These are used to predict changes in activity, stability, and solubility of the optimized sequence compared to the initial sequence, and the predicted values ​​are standardized. For optimization problems (1)-(3), the fitness f values ​​are as follows:

[0223] Optimization problem (1):

[0224] Optimization problem (2):

[0225] Optimization problem (3):

[0226] 1.6 Population Update

[0227] Consistent with the embodiments.

[0228] 1.7 Performance Indicators

[0229] 1.7.1 Hypervolume (HV)

[0230] Consistent with Example 1.

[0231] 1.7.2 Generational distance (GD)

[0232] The generation distance metric is primarily used to evaluate how close an algorithm's generated non-dominated solution set is to the true Pareto front. A smaller distance indicates better convergence and that the solution set found is closer to the ideal Pareto front. Suppose a method computes a final non-dominated solution set PF, and the true Pareto front is PF... * Then we have:

[0233] in,

[0234] For most optimization problems, due to computational limitations, it is often impossible to exhaustively predict all possible mutation sequences, thus the true Pareto front is often unknown, making accurate generational distance calculation impossible. However, the I6WU39 protein is only 101 units long, and exhaustively predicting two-point mutations in it can generate a total of 1,823,050 mutation sequences. Using M... act M stab M sol Three models are used to predict the mutational effects of these mutations, which can be calculated in about 5 days, thus obtaining the true Pareto front for two-point mutations in the I6WU39 protein. Therefore, for the optimization problem of two-point mutations in the I6WU39 protein, generational distance can be added as an evaluation metric to more accurately evaluate the actual optimization performance of the three multi-objective genetic algorithms involved in this embodiment.

[0235] 2 Results

[0236] Similar to Example 1, three test proteins were subjected to 2, 3, 5, 7, and 9 point mutations, respectively, with optimization objectives of activity + stability, activity + solubility, and stability + solubility, respectively. The resulting sequences obtained using three different multi-objective genetic algorithms were compared with those obtained through random sampling. Similarly, the evolutionary generations and population size for all three genetic algorithms were set to 100, thus requiring the genetic algorithms to predict the mutation effects of up to 10,000 mutations using a deep learning model. The random sampling method randomly sampled 10,000 mutations from the solution space for comparison, ensuring that the overall computation time for all four methods was between 2 and 3 hours. The sequences used for comparison were the non-dominated solution sets generated by the three genetic algorithms and the random sampling method, containing no more than 100 mutation sequences. The comparison results of the four non-dominated solution sets are shown in Figures 4A-4O.

[0237] Figures 4A-4E show the optimization effects of the I6WU39 protein (101 aa) under 2-, 3-, 5-, 7-, and 9-point mutations, respectively. Figures 4F-4J show the optimization effects of the P42212 protein (238 aa) under 2-, 3-, 5-, 7-, and 9-point mutations, respectively. Figures 4K-4O show the optimization effects of the P11712 protein (490 aa) under 2-, 3-, 5-, 7-, and 9-point mutations, respectively.

[0238] As shown in Figures 4A-4O, the non-dominated solution set occupied by the optimization method based on the multi-objective genetic algorithm in this embodiment of the present disclosure has a much larger supervolume than that obtained by the random sampling method, and is closer to the optimization direction in each objective attribute compared to the random sampling prediction method. Among them, the SPEA2 algorithm has the smallest generation distance from the true optimal sequence in the optimization task of the 2-point mutation of the I6WU39 protein, demonstrating the best optimization effect. In terms of the supervolume index, in the 45 optimization tasks involved in this embodiment, the three optimization algorithms have achieved the top supervolume in some optimization tasks. Among them, the SPEA2 algorithm ranks the top 2 in supervolume in 32 tasks, demonstrating a relatively stable optimization ability. Therefore, the SPEA2 algorithm can be preferred for sequence optimization problems of this type of dual-attribute collaborative optimization. The NSGA-II algorithm follows closely behind, showing optimization ability comparable to the SPEA2 algorithm in most tasks, while the PESA2 algorithm has a larger fluctuation in optimization effect, showing a more prominent HV value in some problems. In practical applications, the NSGA-II algorithm and PESA2 algorithm can be used as supplements to help find the optimal sequence.

[0239] Example 3: Single-objective optimization of any one of the following properties of protein: activity, stability, and solubility.

[0240] 1. Methods and Materials

[0241] 1.1 Test Data

[0242] Similar to Example 1, three proteins, I6WU39, P42212, and P11712, were used for testing.

[0243] 1.2 Definition of Optimization Problem

[0244] Similar to Example 1, the modification and optimization objectives are as follows:

[0245] 1) Find a set of sequences that, compared to the initial sequence, should provide the greatest possible improvement in activity while maintaining high sequence diversity.

[0246] 2) Find a set of sequences that, compared to the initial sequence, should have the greatest possible improvement in stability while maintaining high sequence diversity.

[0247] 3) Find a set of sequences that, compared to the initial sequence, should have the greatest possible improvement in solubility while maintaining high sequence diversity.

[0248] 1.3 Individual Coding

[0249] Consistent with Example 1.

[0250] 1.4 Population Initialization

[0251] Consistent with Example 1.

[0252] 1.5 Fitness Calculation

[0253] Select three classifiers M based on prediction methods act M stab M sol These are used to predict changes in activity, stability, and solubility of the optimized sequence compared to the initial sequence, and the predicted values ​​are standardized. Since the optimization problem has only one objective, sequence similarity is added as an optimization objective, transforming it into a bi-objective optimization problem. This allows a multi-objective genetic algorithm framework to be applied to single-attribute optimization problems. Given an individual i and a population P, the set of amino acid mutations contained in i is defined as Mut. i The mutation similarity between i and other individuals in the population is:

[0254] in,

[0255] For optimization problems (1)-(3), the fitness f values ​​of individual i are as follows:

[0256] Optimization problem (1):

[0257] Optimization problem (2):

[0258] Optimization problem (3):

[0259] 1.6 Population Update

[0260] Consistent with the embodiments.

[0261] 1.7 Performance Indicators

[0262] Since this embodiment only involves one protein property, the focus is on the degree of fit between the final result and the protein property to be optimized, and a supervolume comparison similar to that in Examples 1 and 2 is not possible. Therefore, box plots can be used to compare the maximum, minimum, average, upper quartile, and lower quartile values ​​of the final sequences obtained by each method in terms of activity, stability, and solubility. The more concentrated the numerical distribution is in the high-value region, the more the sequence obtained by this method matches the optimization target of the corresponding protein property.

[0263] 2 Results

[0264] Similar to Example 1, three test proteins were subjected to 2, 3, 5, 7, and 9 point mutations, respectively, with optimization objectives of activity, stability, and solubility. The resulting sequences obtained using three different multi-objective genetic algorithms were compared with those obtained through random sampling. Similarly, the evolutionary generations for all three genetic algorithms were set to 100, and the population size was set to 100. For the random sampling method, 10,000 mutations were randomly sampled from the solution space for comparison, ensuring that the overall computation time for all four methods was between 2 and 3 hours. Since only one protein attribute was optimized here, a large-volume comparison similar to that in Examples 1 and 2 could not be performed for this single attribute. The sequences obtained by the four methods were sorted according to the predicted values ​​of their corresponding optimization objectives, and the top 100 sequences were used for comparison, resulting in box plots. The more concentrated the box plot is in the high-value region, the better the sequence obtained by the corresponding method matches the optimization objective for that attribute. The optimization effects of the four methods are compared in Figures 5A-5O.

[0265] Figures 5A-5E show the optimization effects of 2-point, 3-point, 5-point, 7-point, and 9-point mutations on the I6WU39 protein (101 aa). Figures 5F-5J show the optimization effects of 2-point, 3-point, 5-point, 7-point, and 9-point mutations on the P52212 protein (238 aa). Figures 5K-5O show the optimization effects of 2-point, 3-point, 5-point, 7-point, and 9-point mutations on the P11712 protein (590 aa).

[0266] As shown in Figures 5A-5E, the sequences obtained by the optimization method based on the multi-objective genetic algorithm in this embodiment of the present disclosure are more concentrated in the high-value region of the corresponding protein attributes, and are closer to the optimization direction in each target attribute compared with the random sampling prediction method. The three optimization algorithms achieved the top-ranked maximum value in some optimization tasks, indicating that the optimal sequence found by the algorithm is closest to the true optimal sequence. Specifically, the NSGA-II algorithm ranked first in maximum value (including ties) in 23 tasks, followed closely by the PESA2 algorithm, which ranked first in maximum value (including ties) in 22 tasks. Furthermore, the PESA2 algorithm ranked first in average value in 26 tasks. Therefore, the PESA2 algorithm can be preferred for this type of single-attribute sequence optimization problem. In practical applications, the NSGA-II algorithm and the SPEA2 algorithm can be used as supplements to assist in finding the optimal sequence.

[0267] When engineering proteins, it is essential to optimize multiple properties simultaneously. This disclosure proposes a method for multi-attribute collaborative optimization of protein sequences with a specified number of mutation sites based on a multi-objective genetic algorithm and a prediction model. This method can find the optimal sequence that simultaneously satisfies multiple attribute optimization objectives within a limited time, greatly improving the efficiency of protein sequence optimization. It exhibits good optimization performance, strong versatility, and can be applied to various biological fields, including but not limited to enzyme engineering, antibody engineering, protein design, biopharmaceuticals, and disease diagnosis.

[0268] This disclosure provides an alternative, comprehensive sequence design framework that more tightly guides the protein sequence generation process and fully considers obtaining a relatively optimal solution within a limited timeframe. Taking a 100-amino acid protein sequence as an example, performing two point mutations on this sequence could generate 1,786,950 mutants. Predicting and ranking these mutants based on various properties would take a week. Further extending the protein sequence length and the number of mutation sites, when the protein sequence length is extended to 500 amino acids, only three point mutations would generate 1.4 * 103 mutants. 11There are countless possible mutants, and the number of combinations is too large to be optimized and predicted using an exhaustive approach. Using the method of this disclosure, with consistent computational resources, a batch of feasible solutions simultaneously satisfying multiple attribute optimization objectives can be obtained in less than 3 hours for a protein sequence of 500 amino acids, such as a batch of protein sequences optimizing stability, activity, and solubility. Some embodiments of this disclosure employ multi-objective genetic algorithms for the synergistic optimization of multiple protein attributes, achieving relatively good optimization results without exhaustive searching, saving significant time. Because multiple attributes are considered simultaneously during the optimization process, its application scenarios are broadened, avoiding the limitation of optimizing only a single attribute, which may prevent other attributes from meeting practical application requirements.

[0269] According to one aspect of this disclosure, a protein sequence optimization device based on a genetic algorithm is provided. Figure 6 shows a structural block diagram of a protein sequence optimization device 600 based on a genetic algorithm according to an embodiment of this disclosure. As shown in Figure 6, the device 600 includes an acquisition unit 610, a mutation unit 620, an initialization unit 630, a first determination unit 640, a second determination unit 650, and an update unit 660.

[0270] The acquisition unit 610 is configured to acquire a first protein sequence to be optimized and at least one attribute of the first protein sequence to be optimized, the first protein sequence comprising a plurality of first amino acids.

[0271] The mutation unit 620 is configured to mutate a predetermined number of first amino acids in the first protein sequence to obtain a second protein sequence.

[0272] The initialization unit 630 is configured to use a set of multiple second protein sequences as the initial population.

[0273] The first determining unit 640 is configured to determine at least one optimization objective of the genetic algorithm based on the at least one attribute.

[0274] The second determining unit 650 is configured to, for each second protein sequence in the initial population, determine a first fitness of the second protein sequence on each of the at least one optimization objective, the first fitness indicating the degree of improvement of the second protein sequence on the corresponding optimization objective.

[0275] The update unit 660 is configured to update the initial population multiple times using the genetic algorithm based on the first fitness corresponding to each of the plurality of second protein sequences, so as to obtain multiple optimized target protein sequences.

[0276] It should be understood that units 610-660 in device 600 correspond to steps S110-S160 in method 100 above, respectively. The specific operation of units 610-660 can be referred to the description of steps S110-S140 above, and will not be repeated here.

[0277] According to one aspect of this disclosure, a computing device is provided, including a memory, a processor, and a computer program stored in the memory. The processor is configured to execute the computer program to implement the steps of any of the method embodiments described above.

[0278] According to one aspect of this disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of any of the method embodiments described above.

[0279] According to one aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of any of the method embodiments described above.

[0280] The following description, in conjunction with Figure 7, provides illustrative examples of such computing devices, non-transitory computer-readable storage media, and computer program products.

[0281] Figure 7 illustrates an example configuration of a computing device 700 that can be used to implement the methods described herein. The aforementioned device 600 can also be implemented wholly or at least partially by the computing device 700 or similar devices or systems.

[0282] The computing device 700 can be a variety of different types of devices. Examples of the computing device 700 include, but are not limited to: desktop computers, server computers, laptop or netbook computers, mobile devices (e.g., tablet computers, cellular or other wireless phones (e.g., smartphones), notebook computers, mobile stations), wearable devices (e.g., glasses, watches), entertainment devices (e.g., entertainment appliances, set-top boxes communicatively coupled to a display device, game consoles), televisions or other display devices, automotive computers, and so on.

[0283] The computing device 700 may include at least one processor 702, memory 704, multiple communication interfaces 706, display device 708, other input / output (I / O) devices 710, and one or more mass storage devices 712 capable of communicating with each other, such as via a system bus 714 or other suitable connection.

[0284] Processor 702 may be a single processing unit or multiple processing units, and all processing units may include single or multiple computing units or multiple cores. Processor 702 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operating instructions. Among other capabilities, processor 702 may be configured to acquire and execute computer-readable instructions stored in memory 704, mass storage device 712, or other computer-readable media, such as program code of operating system 716, program code of application program 718, program code of other program 720, etc.

[0285] Memory 704 and mass storage device 712 are examples of computer-readable storage media for storing instructions executed by processor 702 to perform the various functions described above. For example, memory 704 can generally include both volatile and non-volatile memory (e.g., RAM, ROM, etc.). Furthermore, mass storage device 712 can generally include hard disk drives, solid-state drives, removable media, including external and removable drives, memory cards, flash memory, floppy disks, optical disks (e.g., CDs, DVDs), storage arrays, network-attached storage, storage area networks, etc. Both memory 704 and mass storage device 712 can be collectively referred to herein as memory or computer-readable storage media, and can be non-transitory media capable of storing computer-readable, processor-executable program instructions as computer program code, which can be executed by processor 702 as a specific machine configured to perform the operations and functions described in the examples herein.

[0286] Multiple programs can be stored on mass storage device 712. These programs include operating system 716, one or more application programs 718, other programs 720, and program data 722, and they can be loaded into memory 704 for execution.

[0287] Although illustrated in Figure 7 as stored in memory 704 of computing device 700, modules 716, 718, 720, and 722, or portions thereof, may be implemented using any form of computer-readable medium accessible by computing device 700. As used herein, “computer-readable medium” includes at least two types of computer-readable media: computer-readable storage media and communication media.

[0288] Computer-readable storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, DVD, or other optical storage devices, magnetic cassettes, magnetic tapes, disk storage devices or other magnetic storage devices, or any other non-transmission medium that can be used to store information for access by a computing device. In contrast, communication media can embody computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms. Computer-readable storage media as defined herein do not include communication media.

[0289] One or more communication interfaces 706 are used for exchanging data with other devices, such as via a network, direct connection, etc. Such communication interfaces can be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), wired or wireless (such as an IEEE 802.7 wireless LAN (WLAN)) interface, Wi-MAX interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth. TM Interfaces include near-field communication (NFC) interfaces. Communication interface 706 facilitates communication across various network and protocol types, including wired networks (e.g., LAN, cable, etc.) and wireless networks (e.g., WLAN, cellular, satellite, etc.), the Internet, etc. Communication interface 706 can also provide communication with external storage devices (not shown) such as storage arrays, network-attached storage, storage area networks, etc.

[0290] In some examples, a display device 708, such as a monitor, may be included for displaying information and images to the user. Other I / O devices 710 may be devices that receive various inputs from the user and provide various outputs to the user, and may include touch input devices, gesture input devices, cameras, keyboards, remote controls, mice, printers, audio input / output devices, and so on.

[0291] The technologies described herein can be supported by these various configurations of computing device 700, and are not limited to specific examples of the technologies described herein. For example, the functionality can also be implemented wholly or partially on a “cloud” using a distributed system. A cloud includes and / or represents a platform for resources. The platform abstracts the underlying functionality of the cloud’s hardware (e.g., servers) and software resources. Resources may include applications and / or data that can be used when performing computational processing on servers remote from computing device 700. Resources may also include services provided via the Internet and / or via subscriber networks such as cellular or Wi-Fi networks. The platform can abstract resources and functionality to connect computing device 700 to other computing devices. Therefore, the implementation of the functionality described herein can be distributed throughout the cloud. For example, the functionality can be implemented partly on computing device 700 and partly through a platform that abstracts the functionality of the cloud.

[0292] According to one aspect of this disclosure, a protein molecule is provided, the sequence of which is prepared by the method 100 described above.

[0293] According to one aspect of this disclosure, a pharmaceutical composition is provided comprising protein molecules of embodiments of this disclosure and pharmaceutically acceptable excipients.

[0294] The pharmaceutical compositions disclosed herein can be formulated in any manner known in the art, including but not limited to dosage forms such as tablets, capsules, small capsules, suspensions, powders, lyophilized preparations, suppositories, eye drops, skin patches, oral soluble preparations, sprays, aerosols, and other solid, semi-solid, or liquid systems.

[0295] The pharmaceutical composition may be an immediate-release and / or modified-release formulation, including delayed-release, sustained-release, pulsatile-release, controlled-release, targeted-release, and programmed-release formulations.

[0296] In this document, "pharmaceuticalally acceptable excipients" refers to components in a pharmaceutical composition other than the active ingredient that are non-toxic to the subject. Pharmaceutically acceptable excipients include, but are not limited to, excipients (e.g., diluents, carriers, etc.) and additives (e.g., stabilizers, preservatives, solubilizers, buffers, etc.). Excipients may include polyvinylpyrrolidone, gelatin, hydroxypropyl cellulose (HPC), gum arabic, polyethylene glycol, mannitol, sodium chloride, and sodium citrate. For injectable formulations or other liquid dosage forms, water containing at least one or more buffering components is preferred, and stabilizers, preservatives, and solubilizers may also be used. For solid dosage forms, any of a variety of thickeners, fillers, extenders, and carrier additives may be used, such as starch, sugars, cellulose derivatives, fatty acids, etc. For topical dosage forms, any of a variety of creams, ointments, gels, lotions, etc., may be used. For most pharmaceutical formulations, the inactive ingredient may constitute a significant portion of the formulation by weight or volume. For pharmaceutical formulations, this also covers any of a variety of metered-release, sustained-release or continuous-release formulations and additives, such that a dose can be formulated for delivery of the compositions disclosed over a period of time.

[0297] The compositions disclosed herein may be administered via mucosal administration, buccal administration, oral administration, transdermal administration, inhalation, intranasal administration, urethral administration, vaginal administration, and intravenous, subcutaneous, intramuscular, or intraperitoneal injection. The excipients in the pharmaceutical compositions are adapted to their route of administration.

[0298] In some embodiments, the compositions of this disclosure can be delivered orally, for example, in tablets or capsules. The compositions can be packaged in an intestinal protectant, preferably such that the composition is not released before the tablets or capsules are delivered to the stomach and optionally further to a portion of the small intestine.

[0299] In some embodiments, the compositions of this disclosure can be administered by injection. Suitable pharmaceutical forms for injection include sterile aqueous solutions or dispersions and sterile powders for the immediate preparation of sterile injectable solutions or dispersions. In all cases, the form must be sterile and its flowability must allow it to be administered via a syringe. The form must be stable under preparation and storage conditions and must be preserved to prevent contamination by microorganisms such as bacteria and fungi. The carrier may be a solvent or dispersion medium containing, for example, water, ethanol, polyols (e.g., glycerol, propylene glycol, or liquid polyethylene glycol), suitable mixtures thereof, and vegetable oils.

[0300] Therapeutic administration can also be achieved through injection of sustained-release formulations, such as those that allow for subcutaneous injection, including: nanospheres / microspheres, liposomes, emulsions, gels, insoluble salts, or suspensions.

[0301] In some embodiments, the compositions of this disclosure can be administered intranasally. The pharmaceutical compositions may be in aqueous solution form, such as a solution comprising saline, citrate, or other commonly used excipients or preservatives. They may also be in dry formulation or powder form.

[0302] Although this disclosure has been described and illustrated in detail in the accompanying drawings and the foregoing description, such description and illustration should be considered illustrative and suggestive, not restrictive; this disclosure is not limited to the disclosed embodiments. By studying the drawings, the disclosure, and the appended claims, those skilled in the art will be able to understand and implement variations of the disclosed embodiments in practice with respect to the claimed subject matter. In the claims, the word "comprising" does not exclude other elements or steps not listed, the indefinite article "a" or "an" does not exclude a plurality, the term "a plurality" means two or more, and the term "based on" should be interpreted as "at least partially based on". The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be beneficial.

Claims

1. A protein sequence optimization method based on genetic algorithms, comprising: Obtain a first protein sequence to be optimized and at least one attribute of the first protein sequence to be optimized, wherein the first protein sequence comprises a plurality of first amino acids; A predetermined number of first amino acids in the first protein sequence are mutated to obtain a second protein sequence; The initial population is a set consisting of multiple second protein sequences; Based on the at least one attribute, at least one optimization objective of the genetic algorithm is determined; For each second protein sequence in the initial population, a first fitness of the second protein sequence on each of the at least one optimization objective is determined, the first fitness indicating the degree of improvement of the second protein sequence on the corresponding optimization objective; as well as Based on the first fitness corresponding to each of the multiple second protein sequences, the initial population is updated multiple times using the genetic algorithm to obtain multiple optimized target protein sequences.

2. The method of claim 1, wherein, The mutation of a predetermined number of first amino acids in the first protein sequence includes: A predetermined number of first amino acids are selected from the first protein sequence according to a preset selection algorithm; and For each of the first amino acids in the preset number of first amino acids, based on the preset amino acid substitution scoring matrix, the first amino acid is mutated into a second amino acid.

3. The method of claim 2, wherein, The mutation of the first amino acid into the second amino acid based on the preset amino acid substitution scoring matrix includes: Obtain multiple candidate mutant amino acids for the first amino acid; For any one of the plurality of candidate mutant amino acids: The substitution score between the first amino acid and the candidate mutant amino acid is obtained from the amino acid substitution scoring matrix, the substitution score indicating the probability of the first amino acid and the candidate mutant amino acid substituting for each other; and Based on the substitution score between the first amino acid and the candidate mutant amino acid, the probability of the first amino acid mutating into the candidate mutant amino acid is determined; as well as Based on the probability of each of the plurality of candidate mutant amino acids, one candidate mutant amino acid is selected from the plurality of candidate mutant amino acids as the second amino acid.

4. The method of claim 2 or 3, wherein, The amino acid substitution scoring matrix includes either a BLOSUM matrix or a PAM matrix.

5. The method of any one of claims 1-4, wherein, The at least one optimization objective includes at least one attribute optimization objective corresponding to each of the at least one attribute, and wherein determining the first fitness of the second protein sequence on each of the at least one optimization objective includes: For each of the at least one attribute: Using the trained prediction model corresponding to this attribute, determine the predicted value of the second protein sequence for this attribute; and The predicted value is numerically converted to obtain the first fitness of the second protein sequence on the corresponding attribute optimization target.

6. The method of any one of claims 1-5, wherein, The genetic algorithm is a multi-objective genetic algorithm, and wherein determining at least one optimization objective of the genetic algorithm based on the at least one attribute includes: In response to the fact that the at least one attribute includes only one attribute, the sequence similarity optimization objective and the attribute optimization objective corresponding to that attribute are used as multiple optimization objectives of the multi-objective genetic algorithm; and In response to the fact that the at least one attribute includes two or more attributes, the attribute optimization objective corresponding to each of the two or more attributes is used as multiple optimization objectives of the multi-objective genetic algorithm.

7. The method of claim 6, wherein, Each of the multiple updates includes: Obtain the current population, which includes multiple third protein sequences; For any third protein sequence in the current population, obtain the first fitness of the third protein sequence on each of the plurality of optimization objectives; Based on the first fitness corresponding to each of the plurality of third protein sequences, a non-dominated solution set corresponding to the current population is determined, wherein the non-dominated solution set includes a plurality of fourth protein sequences; and Select, crossover, and / or mutation operations are performed on a subset of the union of the current population and the non-dominated solution set to generate an updated population.

8. The method of claim 7, wherein, The genetic algorithm includes the SPEA2 algorithm, and the selection operation performed on a subset of the union of the current population and the non-dominated solution set includes: Based on the relative size of the number of the plurality of fourth protein sequences and the preset number of archives, the non-dominated solution set is truncated or the non-dominated solution set is expanded using the current population to obtain a parent population, the parent population including a plurality of fifth protein sequences, the number of the plurality of fifth proteins being equal to the number of archives. For each fifth protein sequence in the parent population, a second fitness of that fifth protein sequence is obtained. This second fitness is determined based on the first fitness of the fifth protein sequence on each of the plurality of optimization objectives. The second fitness indicates the overall score of the fifth protein sequence. Based on the second fitness corresponding to each of the plurality of fifth protein sequences, a parent individual is selected from the plurality of fifth protein sequences to perform crossover and / or mutation operations.

9. The method of claim 7, wherein, The genetic algorithm includes the NSGA-II algorithm, and wherein performing a selection operation on a subset of the union of the current population and the non-dominated solution set includes: For any third protein sequence in the current population, based on the first fitness of the third protein sequence on each of the plurality of optimization objectives, the non-dominated level and crowding distance of the third protein sequence are determined, wherein the non-dominated level indicates the dominance and subdominance information of the third protein sequence, and the crowding distance indicates the density of the third protein sequence in the non-dominated level; and Based on the non-dominated hierarchy and crowding distance corresponding to each of the plurality of third protein sequences, a parent individual is selected from the plurality of third protein sequences to perform crossover and / or mutation operations.

10. The method of claim 7, wherein, The genetic algorithm includes the PESA2 algorithm, and the selection operation performed on a subset of the union of the current population and the non-dominated solution set includes: The solution space occupied by the non-dominated solution set is divided into multiple grids of the same size; Determine the compression factor for each of the plurality of grids, the compression factor indicating the number of fourth protein sequences contained in that grid; Based on the compression factors corresponding to each of the plurality of grids, a target grid is selected from the plurality of grids; and Select a parent individual from the fourth protein sequence within the target grid to perform crossover and / or mutation operations.

11. The method of any one of claims 7-10, wherein, The multiple target protein sequences are the non-dominated solution set obtained from the last update in the multiple updates.

12. The method of any one of claims 1-5, wherein, The genetic algorithm is a single-objective genetic algorithm, the at least one attribute includes only one attribute, and wherein determining at least one optimization objective of the genetic algorithm based on the at least one attribute includes: The attribute optimization objective corresponding to the attribute is used as the optimization objective of the single-objective genetic algorithm.

13. The method of claim 12, wherein, Each of the multiple updates includes: Obtain the current population, which includes multiple sixth protein sequences; For any sixth protein sequence in the current population, obtain the first fitness of that sixth protein sequence on the optimization objective; and Based on the first fitness corresponding to each of the plurality of sixth protein sequences, selection, crossover, and / or mutation operations are performed on the plurality of sixth protein sequences to generate an updated population. The multiple target protein sequences are the multiple sixth protein sequences with the highest first fitness in the population obtained from the last update.

14. The method according to any one of claims 1-13, further comprising: The second protein sequence is represented as a binary code, which includes a first binary code for indicating the mutation site and a second binary code for indicating the mutated second amino acid.

15. The method of any one of claims 1-14, wherein, The at least one attribute includes at least one of the following attributes: activity, stability, solubility, expression level, toxicity, degradability, and immunogenicity.

16. The method of any one of claims 1-15, wherein, Any one of the at least one optimization objective is a sequence similarity optimization objective or an attribute optimization objective corresponding to an attribute. The sequence similarity optimization objective is to minimize the similarity between the protein sequence and other protein sequences in the population, and the attribute optimization objective corresponding to the attribute is to maximize the improvement of the protein sequence on that attribute.

17. A protein sequence optimization device based on a genetic algorithm, comprising: The acquisition unit is configured to acquire a first protein sequence to be optimized and at least one attribute of the first protein sequence to be optimized, wherein the first protein sequence comprises a plurality of first amino acids; The mutation unit is configured to mutate a predetermined number of first amino acids in the first protein sequence to obtain a second protein sequence; The initialization unit is configured to use a set of multiple second protein sequences as the initial population; The first determining unit is configured to determine at least one optimization objective of the genetic algorithm based on the at least one attribute; The second determining unit is configured to, for each second protein sequence in the initial population, determine a first fitness of the second protein sequence on each of the at least one optimization objective, wherein the first fitness indicates the degree of improvement of the second protein sequence on the corresponding optimization objective; as well as The update unit is configured to update the initial population multiple times using the genetic algorithm based on the first fitness corresponding to each of the plurality of second protein sequences, so as to obtain multiple optimized target protein sequences.

18. A computing device, comprising: At least one processor; as well as A memory that is communicatively connected to the at least one processor; in The memory stores instructions that can be executed by the at least one processor to cause the at least one processor to perform the method of any one of claims 1-16.

19. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-16.

20. A computer program product comprising computer program instructions, wherein, The computer program instructions, when executed by a processor, implement the method of any one of claims 1-16.

21. A protein molecule, wherein, The sequence of the protein molecule is prepared by the method described in any one of claims 1-16.

22. A pharmaceutical composition comprising the protein molecule of claim 21 and pharmaceutically acceptable excipients.

Images