Gene feature selection method

Abstract

The invention provides a gene feature selection method. Relates to the technical field of data mining. The method comprises the following steps: taking high-dimensional medical gene expression data to be analyzed as a search space; generating a population in the search space, and constructing an elite archive for storing a Pareto optimal solution; storing the Pareto optimal solution of the current population into an elite file; based on the difference between the occurrence frequency of the gene feature combination in the elite archive and the occurrence frequency of the gene feature combination in the current population as an information gain, selecting a high-quality gene feature combination in the elite archive to replace a non-high-quality gene feature combination in the current population so as to iteratively update the current population; when the number of times of continuous non-updating of the elite archive exceeds a stagnation threshold value, performing local redundant gene feature elimination operation and local gene feature replacement operation on the elite archive; judging whether a preset termination condition is met or not, and if not, returning to update; and if so, outputting the optimal gene feature combination in the elite archive. And gene feature selection with low gene number and high precision can be realized.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, and in particular to a gene feature selection method. Background Technology

[0002] With the rapid development of high-throughput sequencing technology, the medical field has accumulated massive amounts of gene expression profiling data. This data provides important evidence for disease diagnosis and precision medicine. However, due to the extremely high dimensionality of gene features and the limited clinical sample size, directly applying all genes to construct diagnostic models is neither economical nor practical. Therefore, gene feature selection has become an important research direction in the field of precision medicine.

[0003] Although multi-objective optimization algorithms provide a theoretical framework for gene feature selection, in actual clinical applications, the need to develop portable diagnostic devices and control testing costs necessitates the selection of a very small subset of feature genes. However, multi-objective optimization algorithms have significant limitations when dealing with high-dimensional medical data: they not only have slow convergence speeds and are prone to getting trapped in local optima, but also have low computational efficiency and fail to balance diagnostic accuracy with the number of features, making it difficult to meet the needs of clinical application. Summary of the Invention

[0004] Therefore, it is necessary to provide a gene feature selection method to address the above-mentioned technical problems, so as to solve the defects of slow convergence speed, easy getting trapped in local optima and low computational efficiency in the existing technology, and realize high-precision screening of clinical diagnostic biomarkers with few genes.

[0005] The following technical solution is adopted in this specification: This specification provides a method for selecting gene traits, including: S100. Use the high-dimensional medical gene expression data to be analyzed as the search space for gene feature selection.

[0006] S200. Generate a population containing multiple gene feature combinations within the search space, and construct an elite archive for storing Pareto optimal solutions, wherein a Pareto optimal solution is a high-quality gene feature combination that has no other gene feature combination that can simultaneously outperform it on all preset optimization objectives.

[0007] S300. Obtain the Pareto optimal solution for the current population, store the Pareto optimal solution for the current population into the elite archive to update the elite archive; after updating the elite archive, perform population update and stagnation monitoring.

[0008] S400. Iteratively update the current population, including: The information gain is calculated by using the difference between the frequency of each gene trait in the updated elite archive and its frequency in the current population. This information gain is then converted into gene trait selection probabilities using the Sigmoid function. ; ; in, It is the first in the elite archive. The frequency of occurrence of each genetic trait in elite archives It is the first in the elite archive. The frequency of occurrence of a gene trait in the current population; It is the first in the elite archive. Information gain of individual gene features; It is the first in the elite archive. The probability of gene feature selection corresponding to each gene feature; It is a natural constant; Based on the gene feature selection probability, select high-quality gene features from the elite archive and replace the low-quality gene features in the current population with the selected high-quality gene features. The low-quality gene features are the gene features in the gene feature combination with the smallest crowding distance and the lowest non-dominant level in the current population.

[0009] This also includes dynamically adjusting the selection probability of Pareto-guided binary operators, extreme vector fusion operators, and random mutation operators in the current iteration period based on the success rate of offspring generation by Pareto-guided binary operators, extreme vector fusion operators, and random mutation operators in the previous iteration period, and dynamically executing each operator based on the dynamically adjusted selection probability of Pareto-guided binary operators, extreme vector fusion operators, and random mutation operators in the current iteration period. ; ; in, Indicates the first k The operator at the th ... t The percentage of successfully generated offspring within the iteration cycle. A successful offspring is an offspring population that has an optimal combination of gene traits that satisfies at least one preset optimization objective. Indicates the first k The operator in the th... t The number of times a successful offspring is generated within the iteration cycle; Indicates the first k The operator in the th... t Total number of uses within the iteration cycle; For memory factors; Indicates the first t Operator probability vector within the iteration period, This represents the Pareto-guided binary operator selection probability. This represents the selection probability of the extreme vector fusion operator. This represents the selection probability of a random mutation operator; This is represented as a vector showing the ratio of successfully generated offspring. This represents the percentage of successful offspring generated by Pareto-guided binary operators. This represents the ratio of successful offspring generated by the extreme vector fusion operator. This represents the percentage of offspring successfully generated by the random mutation operator. This is represented as L1 norm normalization.

[0010] The steps for Pareto-guided binary operators to iteratively update the current population include: The first initial offspring population is generated by combining high-quality gene features from a pre-defined probability differential inheritance guide template. ; in, It is the first in the first initial offspring population A combination of gene characteristics; It is the first in the guide template A combination of gene characteristics; It is the first in the current population One parent population; It is obedience The first uniform probability distribution 1 independent random variable; Indicates the first initial offspring population. The combination of gene traits was selected from the paternal population. A combination of gene characteristics; For gene feature combinations in the first initial offspring population with the same dimension as those in the parent population, a random bit flipping perturbation is introduced to generate a second initial offspring population. ; in, It is the first in the second initial offspring population A combination of gene characteristics; It is obedience The second uniform probability distribution 1 independent random variable; Perform a global bit-flipping mutation on the second initial offspring population to generate the first offspring population: ; in, It is the first generation of the first offspring population. A combination of gene characteristics; It is obedience The third uniform probability distribution in the 1 independent random variable; It is the dimension of the combination of gene characteristics in the second initial offspring population.

[0011] The steps of the extreme vector fusion operator to iteratively update the current population include: The optimal combination of gene features in the elite archive that performs best on a single preset optimization target is taken as the extreme solution, and the difference vector formed by the extreme solution guides the mutation of the baseline vector. ; in, It is a combination of mutated gene characteristics; and These are two combinations of genetic traits randomly selected from the union of the parent population and the elite archives. As a reference vector; This is an adaptive scaling factor; and The optimal gene feature combinations, which perform best under the first and second preset optimization objectives, are used to guide the second generation population to converge toward the boundary and middle of the Pareto front. Combining mutated gene characteristics Crossing with the parent population generates a second offspring population: ; in, It is the first in the second generation population A combination of gene characteristics; express Random numbers between; For crossover probability, For the first One parent population.

[0012] The steps for iteratively updating the current population using the random mutation operator include: performing a random perturbation operation on the selected gene feature combination to generate the third generation population.

[0013] The first, second, and third offspring populations and the parent population are merged. The merged population is then subjected to non-dominated sorting and crowding distance assessment. Based on the sorting and assessment results, the current population is iteratively updated.

[0014] S500. Stagnation monitoring includes: recording the number of stagnant iterations in which the elite archive does not undergo substantial changes; when the number of stagnant iterations exceeds a preset stagnation threshold, performing local redundant gene feature elimination and local gene feature replacement operations on the elite archive.

[0015] S600. Determine whether the following conditions are met: the number of iterations reaches the preset maximum value or the population update difference is less than the preset convergence threshold. If not, return to step S300; if yes, stop updating and output the optimal gene feature combination in the current elite archive.

[0016] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This invention constructs information gain based on the difference in gene frequencies between the elite archive and the current population. It identifies frequently occurring key gene feature combinations as high-value genes and replaces non-high-quality genes in the population, quickly identifying high-potential search regions. This not only improves the overall quality of the population but also avoids blind searching, significantly increasing iteration efficiency. To prevent the algorithm from getting stuck in local optima due to a single rapid guide, this invention monitors the iteration stagnation state and performs local redundancy elimination and replacement operations. Redundant gene elimination reduces the number of features without affecting the optimization objective, while gene replacement introduces potentially high-quality new combinations, effectively overcoming the local optimum trap caused by a single guide. This deeply explores the redundant relationships between genes while maximizing feature compression while maintaining the optimization objective, which is beneficial for developing low-cost clinical testing kits. The combination of globally efficient directional search based on information gain and the refinement of local redundancy elimination synergistically improves the speed and accuracy of gene feature selection, adapting to the needs of clinical testing cost control and portable device development. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is one of the flowcharts illustrating a gene feature selection method provided in this specification; Figure 2 This is the second flowchart illustrating a gene feature selection method provided in this specification. Figure 3 This is the third flowchart illustrating a gene feature selection method provided in this specification; Figure 4 This is the fourth flowchart illustrating a gene feature selection method provided in this specification. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0019] The technical solution provided by this invention can be applied to multi-target feature selection scenarios for high-dimensional medical data, such as tumor classification and gene biomarker screening. With the development of high-throughput sequencing technology, the medical field has generated massive amounts of gene expression data (such as microarrays and RNA-Seq). These data typically exhibit the curse of dimensionality, meaning the sample size is relatively small, while the number of features (genes) is extremely large. When constructing disease diagnostic models, directly using all genes not only incurs enormous computational costs, but the large number of redundant or noisy genes contained within them can severely reduce diagnostic accuracy. Furthermore, in clinical applications, to develop portable diagnostic kits, the number of biomarkers needs to be controlled to a very small range to reduce testing costs.

[0020] Therefore, gene feature selection is essentially a multi-objective optimization problem: on the one hand, it aims to minimize the classification error rate of the diagnostic model, and on the other hand, it aims to minimize the number of selected genes. Existing methods have significant shortcomings in handling this type of problem: traditional filtering methods, while fast, ignore the combinatorial effects between genes; wrapping methods have high accuracy but extremely high computational costs because a classifier needs to be trained for each evaluation; conventional multi-objective evolutionary algorithms are inefficient in searching high-dimensional gene spaces and are prone to getting trapped in local optima, resulting in the selected gene subset still containing redundant information.

[0021] Existing multi-objective evolutionary algorithms perform well in handling low-dimensional problems, but when faced with high-dimensional feature selection problems, they often suffer from slow convergence, susceptibility to local optima, and insufficient population diversity. Traditional wrapper-based feature selection methods, while highly accurate, require frequent calls to the classifier for evaluation, resulting in extremely high computational costs. Although some research has introduced surrogate models to assist the search, constructing high-accuracy surrogate models themselves also requires significant computational resources, which to some extent offsets their advantages.

[0022] Therefore, there is an urgent need for a multi-objective feature selection method for high-dimensional medical genomic data that can reduce computational costs, effectively escape local optima, and quickly converge to the Pareto front. To address the problems of high computational costs, difficulty in removing redundant genes, and bottlenecks in improving diagnostic accuracy in feature selection for high-dimensional medical genomic data, this invention proposes a surrogate-assisted multi-objective feature selection method for medical genomic data.

[0023] The following is combined Figures 1-3 The gene feature selection method of the present invention is described.

[0024] Figure 1 This is one of the flowcharts illustrating a gene feature selection method provided in this specification, such as... Figure 1 As shown, the method includes: S100. Use the high-dimensional medical gene expression data to be analyzed as the search space for gene feature selection.

[0025] For example, the high-dimensional medical gene expression data to be analyzed can be obtained through various technical means. For instance, it can be acquired using microarray technology based on nucleic acid hybridization, or detected using RNA-Seq technology based on high-throughput sequencing. The appropriate data source can be selected according to the actual research needs. This application does not limit the specific method of acquiring high-dimensional medical gene expression data.

[0026] For example, the high-dimensional medical gene expression data to be analyzed is preprocessed data. Data preprocessing includes missing value imputation and normalization; let the initial training sample population be... X =( , ..., ), E = ( , ..., ), i =1, 2, ..., N ; k =1, 2, ...; N , j =1, 2, ..., D .in, N The number of individuals in the training sample. D The number of dimensions for each training sample. Indicates the first i The first individual j 1 eigenvalue, Indicates the first k The first elite individual j Each feature value.

[0027] S200. Generate a population containing multiple gene feature combinations within the search space, and construct an elite archive for storing Pareto optimal solutions, where a Pareto optimal solution is a solution for which no other gene feature combination can simultaneously outperform it on all preset optimization objectives.

[0028] For example, generating a population containing multiple gene feature combinations within the search space means: based on all gene features contained in the high-dimensional medical gene expression data to be analyzed, multiple gene subsets are generated through a standard initialization strategy. Each gene subset constitutes a gene feature combination (i.e., an individual), and multiple individuals together form the initial population. Each gene feature combination may contain the same or different numbers of gene features. The population size and the number of gene features in each individual can be set according to the data dimension, computational resources, and problem complexity. All gene features originate from the preset search space, i.e., the set of gene features contained in the high-dimensional medical gene expression data to be analyzed.

[0029] For example, a medical gene expression profile dataset is divided into a training set and a test set for training a pre-defined classification model; and an initial subset population of gene features and an initial elite archive are generated, the first step being the archive. A And Elite Saves E The latter is initialized to an empty set to store all subsets of genetic features found by the algorithm at the Pareto front during evolution, for the population. P Generate using standard initialization methods. N The initial population of individuals.

[0030] S300. Obtain the Pareto optimal solution in the current population, store the Pareto optimal solution in the current population into the elite archive to update the elite archive; after updating the elite archive, perform population update and stagnation monitoring.

[0031] For example, for each gene feature combination in the current population, its fitness value under multiple preset optimization objectives is calculated; the current population is stratified by non-dominated sorting method and crowding distance calculation, and all gene feature combinations in the first non-dominated level are identified as Pareto optimal solutions for the current population. Specifically, the process is as follows: Initialize the dominance count of all individuals to zero; traverse each pair of individuals in the population, determine their dominance relationship, and increment the dominance count for solutions dominated by other individuals; assign individuals with a dominance count of zero to the first non-dominated level, temporarily remove these individuals, and repeat the above process for the remaining individuals until all individuals are assigned to the corresponding level; subsequently, within each non-dominated level, calculate the crowding distance for each gene feature combination, which is defined as the sum of the average distances between the solution and its neighboring solutions in the target space. The calculation method is as follows: sort each optimization objective separately, assign infinite crowding distances to boundary solutions, and accumulate the differences between the solutions on both sides of the normalized objective value for the crowding distance of intermediate solutions; finally, combine the non-dominated level index and the crowding distance within the level to construct a complete hierarchical structure, where gene feature combinations in the first non-dominated level with large crowding distances are identified as Pareto optimal solutions in the current population, and these solutions achieve the best balance among multiple optimization objectives.

[0032] For example, several preset optimization objectives include minimizing the classification error rate and minimizing the number of gene features. The classification error rate is obtained by inputting gene feature combinations into a preset classification model, which uses a K-nearest neighbor (KNN) classifier for cross-validation on the training set, and calculating the deviation between the model's prediction and the sample's true label. The number of gene features is obtained by counting the number of genes corresponding to the gene feature combinations.

[0033] S400. Population update includes: using the difference between the occurrence frequency of each gene trait combination in the updated elite archive and its occurrence frequency in the current population as information gain, selecting high-quality gene trait combinations in the updated elite archive to replace low-quality gene trait combinations in the current population based on the information gain, so as to iteratively update the current population.

[0034] For example, the average feature expression frequency of each gene locus in the elite gene subset is calculated, and its average feature expression frequency in the non-elite gene subset is subtracted. A lightweight information gain agent is periodically activated to archive the elite gene locus. E The solutions in the population are considered positive samples, and the population is... P The dominated solutions are treated as negative samples, and each feature is calculated. j The information gain value of a gene is obtained by calculating the frequency difference between positive and negative samples. This information gain value is then converted into a feature selection probability using the sigmoid function. Based on this selection probability, a batch of new gene features are selected from the elite archive, and these new gene features replace the gene features in the lowest-ranking, least crowded individuals in the current population. The formula for calculating the gene feature combination selection probability is as follows: ; ; in, It is the first in the elite archive. The frequency of occurrence of each genetic trait in elite archives It is the first in the elite archive. The frequency of occurrence of a gene trait in the current population; It is the first in the elite archive. Information gain of individual gene features; It is the first in the elite archive. The probability of gene feature selection corresponding to each gene feature; It is a natural constant. A value greater than 0 indicates that the feature is more likely to be selected in the elite gene subset, while a value not greater than 0 indicates that the feature is more common in the non-elite gene subset.

[0035] S500. Stagnation monitoring includes: recording the number of stagnant iterations in which the elite archive does not undergo substantial changes; when the number of stagnant iterations exceeds a preset stagnation threshold, performing local redundant gene feature elimination and local gene feature replacement operations on the elite archive.

[0036] For example, the stagnation generation refers to the size of the elite archive or the number of iterations in which its elite subset remains unchanged. When the stagnation generation exceeds a preset stagnation threshold, a local search operation is triggered to reduce or swap redundant genes in the elite archive. The calculation is a stagnation detection function: ; in It is a non-negative integer, representing a continuously stagnant algebraic number. , Indicates the first t The Archive set for the iteration cycle is a counting process that resets when the Archive size changes. Denotes the difference between the two sets, that is, in In but not in Individuals in the population. , is the threshold for the stagnant solution update rate. When When = 0, Archive has no new solutions and is completely stalled; when 0 < < In the case of stagnation, the Archive is rarely updated and may remain stagnant; in other cases, the Archive is updated frequently, indicating convergence or strong convergence. Once stagnation is detected, adaptive local search refinement is triggered. (From the archive...) E The solution is randomly selected for refinement; it is simplified by eliminating local redundant gene features and exchanged by replacing local gene features.

[0037] For example, the local redundant gene feature elimination operation randomly removes a gene feature from the elite archive. If the corresponding preset optimization target does not decrease after removal, the simplified combination is retained; if the preset optimization target decreases after removal and causes the overall preset optimization target to deteriorate, the removal operation is abandoned, and the original gene feature combination remains unchanged. Specifically: a gene is randomly removed from the current elite gene subset. If the subset after removal reduces the number of genes while maintaining diagnostic accuracy, the simplified subset is accepted. ;in, It is a simplified new subset; It is an individual set of atoms X ; It is the first k The standard basis vectors, i.e. D A subset of gene features in a 3D feature space; For example, a local gene feature replacement operation involves randomly removing a gene from the elite archive and adding a new gene to form a new combination. If the preset optimization objective corresponding to the new combination is better, the combination is accepted; if the preset optimization objective corresponding to the new combination is lower, the replacement operation is rejected, and the original gene feature combination is retained. Specifically: a gene is randomly removed from the current elite gene subset and a new gene is added to the unused gene library. If the diagnostic accuracy of the new subset is higher than that of the original set, the new subset is accepted. ;in, It is the new subset after the exchange; It is a subset of unselected genetic features; It is a selected subset of gene features; This indicates that no feature set was selected. This indicates that feature set X has been selected. If the solution after the operation is better than the original solution, then the elite archive is updated.

[0038] S600. Determine whether the following conditions are met: the number of iterations reaches the preset maximum value or the population update difference is less than the preset convergence threshold. If not, return to step S300; if yes, stop updating and output the optimal gene feature combination in the current elite archive.

[0039] For example, preset termination conditions include, but are not limited to, the maximum number of iterations, or the difference between individuals in the population falling below a predetermined threshold. Obtaining the optimal combination of gene features can serve as candidate biomarkers for the high-dimensional medical gene expression data to be analyzed.

[0040] As can be seen, this invention acquires and preprocesses high-dimensional medical gene expression data. Through a lightweight information gain proxy mechanism, it statistically analyzes the distribution differences of pathogenic genes in elite and ordinary gene subsets, constructs a gene importance probability model to quickly screen potential biomarker combinations, and then uses Pareto-guided binary operators to generate new gene subsets by combining known high-precision gene combinations with random exploration. Furthermore, it optimizes both the diagnostic error rate and the number of genes through multi-objective environment selection. Finally, it uses an adaptive local search refinement mechanism to monitor optimization stagnation in real time, removing or replacing redundant genes in the gene subsets. This invention can quickly identify key pathogenic genes in massive gene data, effectively reducing the complexity of diagnostic models and detection costs, and improving the accuracy of disease diagnosis.

[0041] The beneficial effects of this invention include: (1) Efficiently identifying pathogenic genes: By using a lightweight information gain proxy mechanism, key pathogenic genes that frequently occur can be quickly identified without frequently calling expensive classifiers, significantly shortening the screening time. (2) Reducing clinical testing costs: By using a simplification strategy based on an adaptive local search refinement mechanism, redundant relationships between genes can be deeply explored, minimizing the number of biomarkers without reducing diagnostic accuracy, which is beneficial for developing low-cost clinical test kits. (3) Enhancing the generalization ability of diagnostic models: The gene combination generation operator based on Pareto guidance combines elite guidance and random mutation, effectively avoiding overfitting, making the selected gene subset more robust on independent test sets. (4) Adapting to high-dimensional small sample characteristics: This invention is designed specifically for high-dimensional small sample medical data. By combining proxy assistance and local refinement, it effectively overcomes the problem of insufficient search ability of traditional algorithms in the 10,000-dimensional gene space.

[0042] Based on the above Figure 1 In the illustrated embodiment, for example, during the iterative update of the current population in S400, the process further includes generating a first-generation population using a Pareto-guided binary operation operator, generating a second-generation population using an extreme vector fusion operator, generating a third-generation population using a random mutation operator, merging the first-generation, second-generation, and third-generation populations with the parent population, performing non-dominated sorting and crowding distance evaluation on the merged population, and iteratively updating the current population based on the sorting and evaluation results. The order in which the first-generation population is generated using the Pareto-guided binary operation operator, the second-generation population is generated using the extreme vector fusion operator, and the third-generation population is generated using the random mutation operator is not limited; the specific execution order can be flexibly configured to achieve efficient searching for biomarkers in high-dimensional medical gene expression data.

[0043] Figure 2 This is the second flowchart illustrating a gene feature selection method provided in this specification, as shown below. Figure 2As shown, in one implementation, after selecting high-quality gene feature combinations from the elite archive based on information gain to replace non-high-quality gene feature combinations in the current population, the following steps are performed simultaneously: updating the population using three operators and determining whether the number of stagnant iterations in the elite archive that have not been updated for several consecutive times exceeds a preset stagnation threshold. This sequence emphasizes that after feature replacement is completed based on information gain, the population should be prioritized to complete the full evolution of this round using the three operators. At the same time, stagnation monitoring is based on the latest elite archive state after the operator update, rather than on outdated or incomplete intermediate states, thereby ensuring the real-time performance and accuracy of monitoring. This synchronous execution method not only effectively avoids the waste of iteration time caused by monitoring lag or pre-judgment, but also achieves the best balance between the stability of gene feature combination solutions and the diversity of exploring new high-quality solutions.

[0044] Figure 3 This is the second flowchart illustrating a gene feature selection method provided in this specification, as shown below. Figure 3 As shown, in another implementation, after selecting high-quality gene feature combinations from the elite archive to replace the non-high-quality gene feature combinations in the current population based on information gain, it is determined whether the number of stagnant iterations in the elite archive that have not been updated for a continuous period exceeds a preset stagnant threshold. When the preset stagnant threshold is not exceeded, the three operators can be executed in parallel to generate the first, second, and third generation populations independently. The three are independent of each other and can be processed in parallel on different computing threads or processors, which significantly improves the overall computational efficiency of the algorithm.

[0045] For example, based on Figure 2 or Figure 3 The illustrated embodiment, based on information gain, selects superior gene feature combinations from the elite archive to replace inferior gene feature combinations in the current population, and then executes an operator probability adaptive update mechanism. This mechanism is based on the ratio of successful offspring generated by the Pareto-guided binary operation operator, extreme vector fusion operator, and random mutation operator in the previous iteration cycle. Successful offspring are offspring populations with gene feature combinations that satisfy at least one preset optimization objective. The selection probabilities of the Pareto-guided binary operation operator, extreme vector fusion operator, and random mutation operator in the current iteration cycle are dynamically adjusted. ; ; in, Indicates the first k The operator in the th... t The percentage of successfully generated offspring within the iteration cycle. A successful offspring is an offspring population that has an optimal combination of gene traits that satisfies at least one preset optimization objective. Indicates the first k The operator in the th...t The number of times a successful offspring is generated within the iteration cycle; Indicates the first k The operator in the th... t Total number of uses within the iteration cycle; For memory factors; Indicates the first t Operator probability vector within the iteration period, This represents the Pareto-guided binary operator selection probability. This represents the selection probability of the extreme vector fusion operator. This represents the selection probability of a random mutation operator; This is represented as a vector showing the ratio of successfully generated offspring. This represents the percentage of successful offspring generated by Pareto-guided binary operators. This represents the ratio of successful offspring generated by the extreme vector fusion operator. This represents the percentage of offspring successfully generated by the random mutation operator. This is represented as L1 norm normalization.

[0046] Based on the selection probabilities of Pareto-guided binary operators, extreme vector fusion operators, and random mutation operators, each operator is dynamically executed.

[0047] For example, selecting the optimal combination of high-quality gene traits that satisfies a preset optimization objective from the elite archive using a Pareto-guided binary operator as a guiding template, and then performing differential inheritance, consistency perturbation, and global mutation on the guiding template to generate the first generation population, means: from the elite archive... E Select a subset of genes with high diagnostic accuracy. g As a guiding template and from the current population P Randomly select a subset of parent genes p Offspring inherit the selected state of the guide template with a preset probability to preserve key disease-causing genes. This includes: S410. Generate the first initial offspring population by using high-quality gene feature combinations in the pre-set probability differential inheritance guide template: ; in, It is the first in the first initial offspring population A combination of gene characteristics; It is the first in the guide template A combination of gene characteristics; It is the first in the current population One parent population; It is obedience The first uniform probability distribution 1 independent random variable; It is the first in the parent population A combination of gene characteristics. Among them, = Indicates the first initial offspring population. The combination of gene traits was selected from the paternal population. No. A combination of gene characteristics.

[0048] S411. For gene feature combinations in the first initial offspring population with the same dimension as the parent population, introduce random bit flipping perturbation to generate a second initial offspring population. ; in, It is the first in the second initial offspring population A combination of gene characteristics; It is obedience The second uniform probability distribution 1 independent random variable; S412. Perform a global bit-flipping mutation on the second initial offspring population to generate the first offspring population: ; in, It is the first generation of the first offspring population. A combination of gene characteristics; It is obedience The third uniform probability distribution in the 1 independent random variable; It is the dimension of the combination of gene characteristics in the second initial offspring population.

[0049] For example, a combination of gene features is randomly selected from the current population using an extreme vector fusion operator to serve as the parent population, and the parent population is mutated to generate a second offspring population. This includes: S420: The best combination of gene features in the elite archive that performs best on a single preset optimization target is taken as the extreme solution, and the difference vector formed by the extreme solution guides the mutation of the benchmark vector. ; in, It is a combination of mutated gene characteristics; and These are two combinations of genetic traits randomly selected from the union of the parent population and the elite archives. As a reference vector; This is an adaptive scaling factor; and These are the gene feature combinations that perform optimally on the first and second preset optimization objectives, respectively, and are used to guide the second generation population to converge toward the boundary and middle of the Pareto front.

[0050] S421. Combining the characteristics of mutated genes Crossing with the parent population generates a second offspring population: ; in, It is the first in the second generation population A combination of gene characteristics; express Random numbers between; For crossover probability, For the parent population.

[0051] The adaptive scaling factor is implemented by establishing a historical success memory pool to record the adaptive scaling factor and crossover probability corresponding to the successful offspring of the extreme vector fusion operator. Before generating the second offspring population, based on the prior knowledge in the historical success memory pool, the corresponding adaptive scaling factor and crossover probability are generated for each gene feature combination in the second offspring population. ; in and These are the optimal historical scaling factor and the optimal crossover probability, randomly selected from the historical successful memory pool, respectively. rand It is a random number in the range [0.1]. randni Indicates that it follows the mean. variance is It follows a normal distribution.

[0052] For example, a third-generation population is generated by performing a random perturbation operation on selected combinations of gene traits using a random mutation operator. This includes: For example, after merging the parent population with the three offspring populations, the next generation is selected, and the subset of genes obtained from the Pareto front is updated to the elite archive; the fitness (classification error rate and number of features) of the newly generated offspring is calculated. The parent and offspring are then merged, and fast non-dominated sorting and crowding distance calculation are performed to select the next generation. N Individuals enter the next generation; simultaneously, all gene subsets obtained from the Pareto frontier are updated to the elite archive. E middle.

[0053] For example, Figure 4 This is the third flowchart illustrating a gene feature selection method provided in this specification, as shown below. Figure 4 As shown, the method may include: Step S1: Obtain the medical gene expression profile data to be analyzed, divide it into training set and test set, and generate an initial gene feature subset population and an initial elite archive.

[0054] Step S2: Calculate feature importance and construct a probability model by statistically analyzing the difference between the feature frequencies of solutions in the elite archive and those in non-elite solutions.

[0055] Step S3: Gene feature subsets are generated from the gene combination guide solution of the elite archive and the current population parent solution through differential inheritance, consistency perturbation and global mutation.

[0056] Step S4: Merge the parent and offspring populations, select the next generation population, and update the elite archive with the subset of genes obtained from the Pareto front.

[0057] Step S5: Calculate the stagnation generation. When the stagnation generation exceeds a preset threshold, trigger a local search operation to simplify or swap redundant genes in the elite archive.

[0058] Step S6: Determine whether the maximum number of evaluations has been reached. If so, output the optimal subset of gene features in the elite archive as candidates; otherwise, iterate until the end.

[0059] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0060] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A gene feature selection method, characterized in that, include: The high-dimensional medical gene expression data to be analyzed is used as the search space for gene feature selection. A population containing multiple gene feature combinations is generated within the search space, and an elite archive is constructed to store Pareto optimal solutions. The Pareto optimal solution is a high-quality gene feature combination that has no other gene feature combination that can simultaneously outperform it on all preset optimization objectives. Obtain the Pareto optimal solution for the current population, store the Pareto optimal solution for the current population into the elite archive to update the elite archive; after updating the elite archive, perform population update and stagnation monitoring. Population update includes: using the difference between the frequency of each gene feature in the updated elite archive and its frequency in the current population as information gain, selecting superior gene features in the updated elite archive to replace non-superior gene features in the current population based on the information gain, so as to iteratively update the current population; Stagnation monitoring includes: recording the number of stagnant iterations in the elite archive where no substantial changes have occurred; when the number of stagnant iterations exceeds a preset stagnation threshold, performing local redundant gene feature elimination and local gene feature replacement operations on the elite archive. When the number of population updates reaches the preset maximum value or the population update difference is less than the preset convergence threshold, the update stops and the optimal gene feature combination in the current elite archive is output.

2. The gene feature selection method according to claim 1, characterized in that, The step of selecting superior gene trait combinations from the elite archive and replacing inferior gene trait combinations in the current population based on information gain includes: The information gain is converted into gene feature selection probability using the Sigmoid function: ； ； in, It is the first in the elite archive. The frequency of occurrence of each genetic trait in elite archives It is the first in the elite archive. The frequency of occurrence of a gene trait in the current population; It is the first in the elite archive. Information gain of individual gene features; It is the first in the elite archive. The probability of gene feature selection corresponding to each gene feature; It is a natural constant; Based on the gene feature selection probability, select high-quality gene features from the elite archive and replace the low-quality gene features in the current population with the selected high-quality gene features. The low-quality gene features are the gene features in the gene feature combination with the smallest crowding distance and the lowest non-dominant level in the current population.

3. The gene feature selection method according to claim 1, characterized in that, The iterative update of the current population also includes: The Pareto-guided binary operation operator selects the best high-quality gene feature combination that satisfies a preset optimization objective from the elite archive as the guide template, and performs differential inheritance, consistency perturbation and global mutation on the guide template to generate the first generation population. The extreme vector fusion operator randomly selects gene feature combinations from the current population as the parent population, and then mutates the parent population to generate the second offspring population. A third-generation population is generated by performing random perturbation operations on selected gene feature combinations using a random mutation operator. The first, second, and third offspring populations and the parent population are merged. The merged population is then subjected to non-dominated sorting and crowding distance assessment. Based on the sorting and assessment results, the current population is iteratively updated.

4. The gene feature selection method according to claim 3, characterized in that, The selection probabilities of the Pareto-guided binary operator, the extreme vector fusion operator, and the random mutation operator in the current iteration period are dynamically adjusted based on the ratio of successful offspring generated by the Pareto-guided binary operator, the extreme vector fusion operator, and the random mutation operator in the previous iteration period. Successful offspring are offspring populations that possess an optimal combination of gene characteristics that satisfies at least one preset optimization objective. ； ； in, Indicates the first k The operator at the th ... t The percentage of successfully generated offspring within the iteration cycle. ; Indicates the first k The operator at the th ... t The number of times a successful offspring is generated within the iteration cycle; Indicates the first k The operator at the th ... t Total number of uses within the iteration cycle; For memory factors; Indicates the first t Operator probability vector within the iteration period, This represents the Pareto-guided binary operator selection probability. This represents the selection probability of the extreme vector fusion operator. This represents the selection probability of a random mutation operator; This is represented as a vector showing the ratio of successfully generated offspring. This represents the percentage of successful offspring generated by Pareto-guided binary operators. This represents the ratio of successful offspring generated by the extreme vector fusion operator. This represents the percentage of offspring successfully generated by the random mutation operator. This is represented as L1 norm normalization.

5. The gene feature selection method according to claim 3, characterized in that, The generation of the first offspring population includes: The first initial offspring population is generated by combining high-quality gene features from a pre-defined probability differential inheritance guide template. ； in, It is the first in the first initial offspring population A combination of gene characteristics; It is the first in the guide template A combination of gene characteristics; It is the first in the current population One parent population; It is obedience The first uniform probability distribution 1 independent random variable; Indicates the first initial offspring population. The genetic trait combination was selected from the paternal population. A combination of gene characteristics; For gene feature combinations in the first initial offspring population with the same dimension as those in the parent population, a random bit flipping perturbation is introduced to generate a second initial offspring population. ； in, It is the first in the second initial offspring population A combination of gene characteristics; It is obedience The second uniform probability distribution 1 independent random variable; Perform a global bit-flip mutation on the second initial offspring population to generate the first offspring population: ； in, It is the first generation of the first offspring population. A combination of gene characteristics; It is obedience The third uniform probability distribution in the 1 independent random variable; It is the dimension of the combination of gene characteristics in the second initial offspring population.

6. The gene feature selection method according to claim 3, characterized in that, The generation of the second generation population includes: The optimal combination of gene features in the elite archive that performs best on a single preset optimization target is taken as the extreme solution, and the difference vector formed by the extreme solution guides the mutation of the baseline vector. ； in, It is a combination of mutated gene characteristics; and These are two combinations of genetic traits randomly selected from the union of the parent population and the elite archives. As a reference vector; This is an adaptive scaling factor; and The optimal gene feature combinations, which perform best under the first and second preset optimization objectives, are used to guide the second generation population to converge toward the boundary and middle of the Pareto front. Combining variant gene characteristics Crossing with the parent population generates a second offspring population: ； in, It is the first in the second generation population A combination of gene characteristics; express Random numbers between; For crossover probability, For the first One parent population.

7. The gene feature selection method according to claim 6, characterized in that, The adaptive scaling factor is implemented in the following ways: Establish an adaptive scaling factor and crossover probability historical success memory pool to record the corresponding successful offspring of the extreme vector fusion operator; Before generating the second generation population, based on the prior knowledge in the historical success memory pool, corresponding adaptive scaling factors and crossover probabilities are generated for each gene feature combination in the second generation population. ； in, and These are the optimal historical scaling factor and the optimal crossover probability, randomly selected from the historical successful memory pool, respectively. rand It is a random number in the range [0.1]. randni Indicates that it follows the mean. variance is It follows a normal distribution.

8. The gene feature selection method according to claim 1, characterized in that, The local redundant gene feature elimination operation includes: Randomly remove a gene trait from the elite save file. If the corresponding preset optimization target does not decrease after removal, the simplified combination is retained. If the preset optimization target decreases after removal and causes the overall preset optimization target to deteriorate, the removal operation is abandoned and the original gene trait combination remains unchanged.

9. The gene feature selection method according to claim 1, characterized in that, The local gene feature replacement operation includes: Randomly remove a gene from the elite archive and add a new gene to form a new gene feature combination. If the preset optimization goal corresponding to the new gene feature combination is better, the new gene feature combination is accepted; if the preset optimization goal corresponding to the new gene feature combination is lower, the replacement operation is rejected and the original gene feature combination is retained.

10. The gene feature selection method according to claim 1, characterized in that, The multiple preset optimization objectives include minimizing the classification error rate and minimizing the number of gene features. The classification error rate is obtained by inputting gene feature combinations into a preset classification model and calculating the deviation between the model's prediction result and the sample's true label. The number of gene features is obtained by counting the number of genes corresponding to the gene feature combinations.

Images