Molecular evolution optimization algorithm based on fragment similarity tree

Abstract

The invention relates to a molecular evolution optimization algorithm based on a fragment similarity tree, which is used for accelerating discovery and optimization of new drug molecules, and comprises the following steps of: screening molecules with excellent attributes from a public data set to construct an excellent molecule library, decomposing the molecules into a plurality of fragments, and constructing the fragment similarity tree based on valley similarity and Lelwinstein distance; the method comprises the following steps of: performing embedding coding on fragments by using a fragment similarity tree as a search space for molecular optimization, performing variation optimization on the basis of fragment embedding coding by using an evolutionary algorithm in an optimization stage, selecting a parent population by combining a Pareto optimal algorithm and a population crowding sorting method, and guiding molecular variation through a diversity operator and a convergence operator, according to the method, the chemical space is explored, the local structure is optimized, the weights of the two mutation operators are dynamically balanced in the evolution process, the optimization efficiency and effect are improved, the random walk problem can be effectively relieved, the molecule optimization efficiency is improved, and new molecules with better properties are generated.

Description

Technical Field

[0001] The present invention relates to the field of drug discovery, and particularly to a molecular evolution optimization algorithm based on a fragment similarity tree. Background Art

[0002] Challenges faced in drug discovery, such as high cost, long cycle, and the difficulty of exploring chemical space, aim to develop efficient and innovative methods to accelerate the discovery and optimization process of new drug molecules. In recent years, with the development of big data and artificial intelligence technologies, the latest research on molecular optimization is conducive to discovering new compounds with ideal properties, and using the existing experience to save a large amount of manpower and material resources required for drug molecule discovery, promoting the development of drug discovery.

[0003] The molecular optimization method based on combinatorial optimization has unique advantages in dealing with discrete and uncertain chemical space, but there are also significant disadvantages, especially the random walk problem. In the exploration based on random processes, the optimization path of molecules has a certain randomness, which may lead to low efficiency in the search process, spending a large amount of time but being difficult to quickly converge to the ideal molecular structure. Therefore, designing new methods to eliminate the random walk problem and redesigning the fragment library composed of molecular fragments is of great significance for molecular optimization.

[0004] In the prior art, for example, Chinese Patent Application No. 202410771966.9 discloses a method for generating and screening lead active molecules based on artificial intelligence fragmentation technology. The deficiency of this patent is that it does not set an optimization path for molecules to solve the random walk problem, resulting in low efficiency in the optimization process. Summary of the Invention

[0005] The object of the present invention is to use high-quality molecular fragments to construct a fragment library, design a fragment similarity tree to establish the connection between fragments, combine an evolutionary algorithm with population and iterative characteristics to achieve global search of the chemical space composed of fragment trees, and use two different mutation operators during the evolution process to guide the mutation of molecules during molecular evolution, effectively alleviating the random walk problem during molecular optimization and achieving the optimization of molecules.

[0006] To solve the above problems, an embodiment of this specification proposes a molecular evolution optimization algorithm based on a fragment similarity tree. A molecular evolution optimization algorithm based on a fragment similarity tree includes a common molecular data set and a lead molecule. This algorithm realizes the optimization of molecules through the following steps:

[0007] S1: Construct a search space for molecular optimization;

[0008] S1.1: Select task-related molecules from the common molecular database and decompose them into fragments together with the lead molecule;

[0009] S1.2: Use the most similar fragment similarity matrix as the leaf node to generate the basic tree structure. Calculate the fragment similarity matrix based on the maximum common substructure Tanimoto similarity and the Levenshtein distance. The formula is as follows:

[0010] A sim (i,j) = max(Tan(f i , f j ), Lev(f i , f j )) (1)

[0011] Where A sim is the similarity matrix, f i and f j are the i-th and j-th fragments respectively. Tan and Lev represent the Tanimoto similarity and the Levenshtein distance respectively;

[0012] S1.3: Finally form a single similarity tree T0 by iteratively merging the most similar subtrees. The selection criterion for the most similar subtrees is the maximum similarity between subtrees. The trees of the most similar subtrees are denoted as T i and T j , T i and T j The similarity matrix between them is defined as follows:

[0013]

[0014] The maximum similarity between two trees T i and T j is defined as follows:

[0015]

[0016] S1.4: Perform a depth-first traversal encoding on the final similarity tree T0. The left path of non-leaf nodes is marked with "0", and the right path is marked with "1";

[0017] S2: Evolutionary optimization process;

[0018] S2.1: Construct the initial parental population based on the lead molecule: Encode the lead molecule fragments as the initial embedding groups, randomly generate 2M embedding groups, and select M high-quality embedding groups as the parental population using Pareto optimality and crowding degree sorting;

[0019] S2.2: Generate the offspring population of the parental population selected in S2.1 using two mutation operators. The two mutation operators are the diversity operator and the convergence operator respectively:

[0020] Diversity operator: Uniformly and randomly flip each binary bit;

[0021] Convergence operator: Adjust the binary value at each position with a probability of p + qi, where p and q are parameters set in the experiment, and i represents the i-th position in the fragment embedding;

[0022] S2.3: Combine the parent and offspring populations, update the population through Pareto front screening and crowding sorting, and repeatedly execute the mutation-selection process until the target property is satisfied, and output the optimized molecule.

[0023] Preferably, the molecule corresponding to the optimization task described in S1.1 has the following molecular properties:

[0024] QED: Quantitative Estimation of Drug-likeness, a quantitative index of drug likeness;

[0025] DRD2: Dopamine receptor, a molecular bioactivity index targeting the dopamine type 2 receptor of a biological target;

[0026] PLogPim: The logarithm of the solute partition ratio between octanol and water minus the improvement value of the synthetic accessibility score and the number of long cycles.

[0027] Preferably, for the decomposition of the lead molecule described in S1.1, given a molecule, the molecule is split into r fragments, and there are t attachment points in total. Then, the r fragments can be replaced by any r fragments, and it is required that there are also t attachment points.

[0028] Preferably, for calculating the similarity between fragments in the excellent fragment library described in S1.2, the formula (1) is used, and the two-dimensional similarity matrix is used to record the similarity between all fragments. The abscissa and ordinate of the two-dimensional similarity matrix are each fragment, and the value stored inside the matrix is the similarity between the fragment corresponding to the abscissa and the fragment corresponding to the ordinate.

[0029] Preferably, if the fragment similarity matrix described in S1.2 is an odd number, there exists a tree that only contains a fragment similarity matrix in its left leaf node.

[0030] Preferably, in S1.4, the connection between fragments is established through the fragment similarity tree and fragment encoding. The more similar the fragments are, the longer the maximum common prefix of the encoding is.

[0031] Preferably, in the Pareto optimal algorithm described in S2.1, there exists an embedding group A, and there is no other embedding group that can comprehensively outperform embedding group A in all evaluation dimensions. Then, embedding group A is considered to be Pareto optimal. By comparing the performance of each embedding group in these dimensions, the set of all embedding groups that satisfy the Pareto optimal conditions is found. The all evaluation dimensions include QED, DRD2, PLogP_imp, SIM, and SR:

[0032] Similarity: SIM, the Tanimoto similarity between the lead molecule and the optimized molecule based on molecular fingerprints;

[0033] Success Rate: SR, the percentage of lead molecules that can be optimized into molecules with satisfactory properties;

[0034] Among them, SIM is the similarity between the optimized molecule and the original molecule; SR is the success rate of the optimized molecule.

[0035] Preferably, for the molecular evolution optimization algorithm based on the fragment similarity tree according to claim 1 and claim 6, it is characterized in that for the population crowding sorting, there are M embedding groups in the population. For the value of the i-th embedding in a certain evaluation dimension denoted as v(i), first sort the population separately in each dimension. In a certain evaluation dimension, the difference between the objective values of two adjacent embeddings is used to measure the crowding degree. For the i-th embedding, its crowding distance in a certain dimension is calculated as follows:

[0036]

[0037] where v max and v min are the maximum value and the minimum value in this dimension respectively. Add the crowding distances of each embedding in all dimensions to obtain the comprehensive crowding distance value. Then, combine the Pareto optimality and the crowding distance to select the parental population: first select from the Pareto optimal set. When the number of embeddings in the Pareto optimal set exceeds M, screen in the Pareto optimal set from large to small according to the crowding distance until M embeddings are selected.

[0038] Preferably, for the two mutation operators in S2.2, assign two weights

[0039] and to change the weights of the two mutation operators, where corresponds to the diversity operator, corresponds to the convergence operator. Iter and Itermax represent the current iteration number and the maximum iteration number respectively.

[0040] Preferably, for the two mutation operators in S2.2, since the embedding of the fragment is based on its position in the fragment similarity tree, the influence of the diversity operator mutation on the position at the rear of the fragment similarity tree is much smaller than that on the position at the front of the fragment similarity tree. The convergence operator tends to mutate at the position at the rear of the fragment similarity tree.

[0041] The main advantages of the present invention are:

[0042] The present invention has certain advantages in molecular optimization. When constructing the search space for the molecular evolutionary optimization algorithm of the fragment similarity tree, molecules with higher attributes are screened from the public dataset to construct an excellent molecular library. Then, these molecules and the lead molecule are decomposed into fragments, and the similarity between fragments is calculated using the Tanimoto similarity and the Levenshtein distance to construct a fragment similarity tree as the search space. This construction method based on fragment similarity allows similar fragments to aggregate, providing an orderly and efficient search environment for molecular evolution. Moreover, the 0 / 1 coding based on the tree structure can clearly establish fragment connections. The longer the maximum common prefix of the 0 / 1 coding of similar fragments, the more conducive it is to controlling the direction of fragment variation during evolution and improving the optimization efficiency. In the molecular evolutionary optimization stage, using the evolutionary algorithm, the parent population is selected through the Pareto optimal algorithm and the population crowding sorting. While maintaining similarity with the lead molecule, molecules with improved properties are found. Two mutation operators, namely the diversity operator and the convergence operator, are designed and combined with the weights of different iteration times, enabling more exploration of the chemical space in the early stage of evolution and focusing on local structure optimization in the later stage, thus balancing the similarity and diversity in molecular optimization. The molecular evolutionary optimization algorithm of the fragment similarity tree of the present invention can effectively optimize molecules and alleviate the problem that random walks in traditional combinatorial optimization algorithms affect the molecular optimization efficiency. Brief Description of the Drawings

[0043] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments recorded in this specification. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0044] Figure 1 It is a flowchart of the specific implementation process of the method of the present invention;

[0045] Figure 2 It is a diagram of the construction process of the fragment similarity tree of the present invention;

[0046] Figure 3 It is an example diagram of the fragment similarity tree of the present invention. Detailed Embodiments

[0047] The following will clearly and completely describe the technical solutions in the embodiments of this specification in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0048] First, the experimental settings, comparison baselines, and multi-objective optimization tasks are elaborated. The objectives of experimental optimization, evaluation metrics, the models to be compared, and the set multi-objective comparison tasks are described respectively. Secondly, comparative experiments and convergence operator probability difference experiments are conducted respectively to specifically analyze the effectiveness of each aspect of the molecular evolutionary optimization algorithm for fragment similarity trees.

[0049] 1. Experimental Settings

[0050] In the present invention, the following molecular properties are followed, and the higher all the properties are, the better. Molecular property targets such as QED, DRD2, and PLogP are concerned. The QED, DRD2, and PLogP values of chemically effective molecules are evaluated by the RDkit toolkit. The QED and DRD2 values of molecules are always in the range of [0, 1], while the PLogP value has no upper limit. Therefore, the improved value of PLogP (i.e., PLogP_imp) is used as an index to measure the molecular optimization performance.

[0051] Quantitative Estim ate of Drug likeness (QED): A quantitative index of drug likeness.

[0052] Dopamine Receptor (DRD2): A molecular bioactivity index targeting the biological target dopamine receptor type 2.

[0053] Penalized LogP (PLogP) and the improvement of PLogP (PLogPimp): The logarithm of the solute partition ratio between octanol and water minus the synthetic accessibility score and the number of long cycles.

[0054] At the same time, the following metrics for evaluation are also concerned:

[0055] Similarity (SIM): The Tanimoto similarity between the lead molecule and the optimized molecule based on molecular fingerprints.

[0056] Success Rate (SR): The percentage of lead molecules that can be optimized into molecules with satisfactory properties.

[0057] Among them, SIM is the similarity between the optimized molecule and the original molecule, which is an important index to judge whether the optimized molecule deviates from the original molecule; SR is the success rate of the optimized molecule, which judges the usability of the algorithm from an overall perspective.

[0058] 2. Comparison Baselines

[0059] The present invention is compared with the following four models, including two combinatorial optimization models and two deep learning models.

[0060] MSO: A molecular optimization model combining deep learning and particle swarm optimization algorithm.

[0061] QMO: A query-based molecular generation model.

[0062] GA-DNN: A discriminator-based genetic algorithm model that constructs the search space using SELFIES characters.

[0063] GB-GA: A genetic algorithm model based on molecular graphs that constructs the search space based on the substructures and atoms of the first 1000 molecules in the ZINC dataset.

[0064] 3. Multi-objective optimization tasks

[0065] To explore the performance of the present invention in molecular optimization, two multi-objective optimization tasks are designed to evaluate the performance of the method.

[0066] Task 1: QED + PLogPimp + SIM. If the output molecule of a molecule meets the following conditions, then the molecule is successfully optimized: the drug-likeness index (QED) ≥ 0.85, the improved logarithm of the octanol-water partition coefficient (PLogPimp) ≥ 3, and the similarity index (SIM) ≥ 0.3. In this work, the selected lead molecules are those with QED ≤ 0.8 and PLogP ≤ -2 from the ZINC database.

[0067] Task 2: QED + DRD2 + SIM. If the output molecule of a molecule meets the following conditions, then the molecule is successfully optimized: quantitative estimation of drug-likeness (QED) ≥ 0.8, dopamine D2 receptor affinity (DRD2) ≥ 0.4, and similarity index (SIM) ≥ 0.3. In this work, the selected lead molecules are those with QED ≤ 0.8 and DRD2 ≤ 0.02 from the ZINC database.

[0068] In the experiment, all datasets were downloaded from the 15th version of the ZINC database. It consists of 250,000 molecules. Based on the objectives of each task, a molecular library and a corresponding fragment library were constructed from the ZINC database. Each library contains molecules with large characteristics. Specifically, according to Task 1, molecules with QED ≥ 0.85 and PLogP ≥ 0 were selected, and then they were decomposed into fragments and Fragment Library 1 was constructed. The number of molecules is approximately 20,000, while the number of fragments is 5136. In Task 2, molecules with DRD2 ≥ 0.03 were selected to construct Fragment Library 2. The number of molecules is approximately 13,000, while the number of fragments is 4884.

[0069] 4. Comparative experiments

[0070] The experimental results of Task 1 and Task 2 of the present invention are shown in Table 1a and Table 1b, which respectively show the average value and standard deviation of the properties of all output molecules after optimization, as well as the average value and standard deviation of the similarity with the corresponding lead molecules, and the percentage SR of lead molecules that can be optimized into molecules with satisfactory properties.

[0071] First, the success rate SR is compared. It represents the utilization potential of an algorithm. The higher the SR, the stronger its ability to optimize lead molecules successfully. As shown in the two tables, the proposed molecular evolutionary optimization algorithm method based on fragment similarity tree achieved the highest SR in both tasks. Especially in Task 1, the SR value was 25.4% higher than that of the second-best method. In Task 2, the increase in the SR value was not particularly significant compared to Task 1. The reason is that the property conditions in Task 2 were relaxed when constructing the molecular library, and the selected molecules were not particularly excellent compared to Task 1. Then, the molecular properties and similarities are compared. In Task 1, compared with the molecular optimization models GA-DNN, GB-GA, MSO, and QMO, the molecular evolutionary optimization algorithm based on fragment similarity tree improved in almost all indicators. Although its SIM is not the largest, it is still higher than the preset baseline. In Task 2, compared with GA-DNN, GB-GA, MSO, and QMO, all properties of the molecular evolutionary optimization algorithm based on fragment similarity tree are better.

[0072] From the experimental results, using an excellent fragment library as the search space for molecular optimization and using fragment similarity tree and mutation operator to alleviate the random walk problem is effective, which can enhance the properties and success rate of molecular optimization. Next, further experiments will be carried out on the probability difference of the convergence operator to judge the influence degree and convergence situation of different mutation probabilities on molecular optimization during the mutation process.

[0073]

[0074] Table 1a

[0075]

[0076]

[0077] Table 1b

[0078] 5. Experiment on the Probability Difference of the Convergence Operator

[0079] The present invention proposes two different operators, namely, "diversity operator" and "convergence operator". In the early stage of mutation iteration, it is more inclined to mutate with the "diversity operator", changing the binary values at each position in the embedding with the same probability to help explore more spaces during the evolution process; in the later stage of mutation iteration, it is more inclined to mutate with the "convergence operator", preferentially changing the binary values at the later positions in the fragment embedding with different probabilities, so as to be more likely to mutate to more similar fragments during the evolution process, helping the evolution process to mutate more excellent molecules based on the excellent molecules that have been iterated out, so as to obtain more molecules that meet the optimization conditions in the offspring and complete rapid convergence.

[0080] In this experiment, the influence of different probability differences during the mutation process of the "convergence operator" on the molecular optimization process was studied. Specifically, the purpose of this experiment was to determine whether different probabilities adopted by this operator could obtain more successful molecules that meet the optimization conditions. To achieve this purpose, two evaluation metrics were added to Task 1 of the comparative experiment to judge the number of molecules in the last generation population of the evolutionary algorithm, namely:

[0081] (1) The average number of molecules success_avg (Savg) that meet the optimization conditions in the population of optimized successful molecules;

[0082] (2) The ratio success_all_rate (SArate) between the number of optimized successful molecules and the total number of all optimized molecules among all molecules.

[0083] The evaluation metric success_avg can evaluate the exploration ability of optimized successful molecules during the molecular optimization process. The more the average number, the more excellent molecules the algorithm can explore; the evaluation metric success_all_rate can evaluate the exploration ability of excellent molecules from the overall perspective of the algorithm, so as to judge the ability of the algorithm to guide molecules to mutate towards excellent molecules. Through these two metrics, the influence of the "convergence operator" with different probability differences on the convergence ability during the mutation process can be analyzed, and it can be judged whether the algorithm has the ability to alleviate the random walk problem. In this experiment, the probability difference ratio of the "convergence operator" was set to 1:1, 1:2, and 1:3 from left to right, while keeping other experimental parameters the same as those in Task 1.

[0084]

[0085] Table 2

[0086] According to the experimental results in Table 2, the results of molecular property optimization under the "convergence operator" with different probability differences and the exploration ability for optimized successful molecules are respectively shown. At different probability differences, no significant differences are found in the success rate and parameter values of molecular properties under different probability differences. However, in terms of the exploration ability of optimized successful molecules, the higher the probability difference, the larger the values of Savg and SArate, that is, the stronger the exploration ability, there is an effective convergence ability for the mutation process, and the random walk problem in the optimization process can be alleviated.

Claims

1. A molecular evolution optimization algorithm based on a fragment similarity tree, including a common molecular dataset and a lead molecule, characterized in that, The algorithm realizes the optimization of molecules through the following steps: S1: Construct the search space for molecular optimization; S1.1: Select task-related molecules from the public molecular database and decompose them into fragments together with the lead molecule; S1.2: Use the most similar fragment similarity matrix as the leaf node to generate the basic tree structure. Calculate the fragment similarity matrix based on the maximum common substructure Tanimoto similarity and the Levenshtein distance. The formula is as follows: A sim (i,j) = max(Tan(f i , f j ), Lev(f i , f j )) (1) Among them, A sim is a similar matrix, f i and f j are the i-th and j-th segments respectively, and Tan and Lev represent the Tanimoto similarity and the Levenshtein distance respectively; S1.3: Finally, form a single similarity tree T0 by iteratively merging the most similar subtrees, where the selection criterion for the most similar subtrees is the maximum similarity between subtrees. The trees of the most similar subtrees are denoted as T i and T j , T i and T j The similarity matrix between them is defined as follows: Two trees T i and T j The maximum similarity between them is defined as follows: S1.4: Perform a depth-first traversal encoding on the final similarity tree T0. Mark the left path of non-leaf nodes with "0" and the right path with "1"; S2: The evolutionary optimization process; S2.1: Construct the initial parental population based on the lead molecule: Encode the lead molecule fragments as the initial embedding group, randomly generate 2M embedding groups, and select M high-quality embedding groups as the parental population by using Pareto optimality and crowding degree sorting; S2.2: Use two mutation operators to generate the offspring population of the parental population selected in S2.

1. The two mutation operators are the diversity operator and the convergence operator respectively: Diversity operator: Randomly flip each binary bit uniformly; Convergence operator: Adjust the binary value at each position with a probability of p + qi, where p and q are experimentally set parameters, and i represents the i-th position in the fragment embedding; S2.3: Merge the parental and offspring populations, update the population through Pareto front screening and crowding sorting, and loop through the mutation-selection process until the target property is satisfied, and output the optimized molecule.

2. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, wherein The molecules corresponding to the optimization tasks in S1.1 include the following molecular properties: QED: Refers to the quantitative estimation of drug likeness, a quantitative index of drug likeness; DRD2: Dopamine receptor, a molecular bioactivity index targeting the dopamine type 2 receptor of biological targets; PLogPim: The logarithm of the solute partition ratio between octanol and water minus the improvement value of the synthetic accessibility score and the long cycle number.

3. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, characterized in that For the decomposition of the lead molecule in S1.1, given a molecule, the molecule is split into r fragments, and there are t attachment points in total. Then the r fragments can be replaced by any r fragments, and it is required that they also have t attachment points.

4. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, characterized in that In S1.2, the formula (1) is used to calculate the similarity between fragments in the excellent fragment library, and a two-dimensional similarity matrix is used to record the similarity between all fragments. The abscissa and ordinate of the two-dimensional similarity matrix are each fragment, and the value stored inside the matrix is the similarity between the fragment corresponding to the abscissa and the fragment corresponding to the ordinate.

5. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, characterized in that, If the fragment similarity matrix in S1.2 is an odd number, there is a tree that contains only one fragment similarity matrix in its left leaf node.

6. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, wherein In S1.4, the connection between fragments is established through the fragment similarity tree and fragment encoding. The more similar the fragments are, the longer the maximum common prefix of the encoding is.

7. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, characterized in that In the Pareto optimal algorithm described in S2.1, there is an embedding group A, and there is no other embedding group that can comprehensively outperform embedding group A in all evaluation dimensions. Then, embedding group A is considered Pareto optimal. By comparing the performance of each embedding group in these dimensions, a set of embedding groups that meet the Pareto optimal conditions is found. All the evaluation dimensions include QED, DRD2, PLogP_imp, SIM, and SR: Similarity: SIM, based on molecular fingerprints, the Tanimoto similarity between the lead molecule and the optimized molecule; Success Rate: SR, the percentage of lead molecules that can be optimized into molecules with satisfactory properties; Among them, SIM is the similarity between the optimized molecule and the original molecule; SR is the success rate of the optimized molecule.

8. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1 or claim 6, characterized in that For the population crowding sorting, there are M embedding groups in the population. The value of the i-th embedding in a certain evaluation dimension is denoted as v(i). First, the population is sorted separately in each dimension. In a certain evaluation dimension, the difference between the target values of two adjacent embeddings is used to measure the degree of crowding. For the i-th embedding, its crowding distance in a certain dimension is calculated as follows: where v max and v min are the maximum and minimum values in this dimension respectively. The crowding distances of each embedding in all dimensions are added up to obtain the comprehensive crowding distance value. Then, the parent population is selected by combining Pareto optimality and crowding distance: first, select from the Pareto optimal set. When the number of embeddings in the Pareto optimal set exceeds M, screen in the Pareto optimal set according to the crowding distance from large to small until M embeddings are selected.

9. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, characterized in that For the two mutation operators described in S2.2, two weights are assigned and to change the weights of two mutation operators, where corresponds to the diversity operator, corresponds to the convergence operator, and Iter and Itermax represent the current iteration number and the maximum iteration number respectively.

10. The molecular evolution optimization algorithm based on fragment similarity tree according to claim 1, characterized in that For the two mutation operators described in S2.2, since the embedding of fragments is based on their positions in the fragment similarity tree, the impact of the diversity operator mutation on the positions at the rear of the fragment similarity tree is much smaller than that on the positions at the front of the fragment similarity tree. The convergence operator tends to mutate at the positions at the rear of the fragment similarity tree.

Images