A high-dimensional gene expression profile cancer subtype classification method and system based on double-layer co-evolution and reinforcement learning

Abstract

The application discloses a high-dimensional gene expression profile cancer subtype classification method and system based on double-layer cooperative evolution and reinforcement learning, and belongs to the fields of bioinformatics and intelligent computing technology.The method comprises the following steps: obtaining gene expression profile data; constructing a double-layer evolution framework of main population and elite population cooperative evolution; dynamically selecting a guide vector to drive the evolution of the main population by using a reinforcement learning mechanism; screening three types of elites to inject into the elite population by using an adaptive multi-type elite management mechanism; performing differential evolution on the elite population by using a type-aware elite evolution strategy; iteratively executing until a termination condition is met, and outputting an optimal feature subset; and performing classification based on the optimal feature subset to obtain a cancer subtype classification result.The application solves the problems of easy loss of excellent combinations, easy search for local optimal and difficult balance between diversity and convergence in high-dimensional feature selection, and improves the stability and classification precision of gene subset screening.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and intelligent computing, and in particular to a method and system for classifying cancer subtypes based on high-dimensional gene expression profiles using two-layer co-evolution and reinforcement learning. Background Technology

[0002] In cancer genomics research, accurately classifying cancer subtypes by analyzing patients' gene expression profiles is a crucial step in achieving personalized precision medicine. For example, in the clinical diagnosis of diseases such as glioma, leukemia, and lung cancer, identifying key disease-related gene biomarkers can provide crucial evidence for targeted therapy and prognostic assessment.

[0003] However, gene expression profiling data are characterized by high dimensionality and small sample size: the number of genes typically reaches thousands or even tens of thousands, while the number of usable samples is often only tens to hundreds. Furthermore, the expression levels of a large number of genes are irrelevant to disease classification or exhibit high redundancy. Screening out a subset of key genes highly relevant to diseases from massive gene datasets while maintaining classification accuracy is a core technical challenge currently facing the field of bioinformatics.

[0004] Existing high-dimensional feature selection methods mainly suffer from the following technical drawbacks: 1. Excellent gene combinations are easily lost: During the evolutionary search process, key gene combinations that are highly related to the disease are easily destroyed, resulting in poor solution set stability and difficulty in continuous optimization.

[0005] 2. The search process is prone to getting trapped in local optima: Traditional algorithms have limited exploration capabilities in high-dimensional sparse spaces and are prone to prematurely converging to suboptimal solutions, making it difficult to discover better gene combinations.

[0006] 3. Difficulty in balancing diversity and convergence: Single-layer evolutionary structures cannot simultaneously explore diverse gene combinations and develop high-quality gene combinations, resulting in the final output gene subset failing to meet the dual requirements of feature quantity and classification accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a high-dimensional gene expression profile cancer subtype classification method and system based on two-layer co-evolution and reinforcement learning. It aims to solve the technical problems of existing feature selection methods under the condition of high-dimensional and small sample gene expression profile data, such as the easy loss of excellent gene combinations, the easy getting trapped in local optima, and the difficulty in balancing diversity and convergence.

[0008] To achieve the above objectives, this invention provides a high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning, comprising the following steps: S1. Obtain gene expression profile data to be classified. The gene expression profile data includes multiple samples and multiple gene features corresponding to each sample. S2. Construct a two-layer evolutionary framework in which the main population and the elite population iterate in parallel and cooperate with each other. The main population is used to maintain search diversity, and the elite population is used to retain and develop high-quality feature subsets. S3. Employ reinforcement learning to dynamically select the guiding vector, and drive the evolutionary process of the main population based on the guiding vector; S4. An adaptive multi-type elite management mechanism is adopted to select elite individuals from the main population and inject them into the elite population. The elite individuals are divided into three categories: performance-type elites, error-oriented elites, and convergence-balance-type elites. S5. Employ a type-aware elite evolution strategy to evolve the elite population, setting differentiated parent selection methods, crossover strategies, and mutation intensities for different types of elite individuals. S6. Iterate through steps S3 to S5 until a preset termination condition is reached. The preset termination condition is reaching a preset maximum number of evaluations. Then, output the final elite population as the optimal feature subset. S7. Classify the gene expression profile data based on the optimal feature subset to obtain the cancer subtype classification results.

[0009] Preferably, the guiding vector in step S3 includes a prior guiding vector, a sparse distribution guiding vector, and an exploratory guiding vector; The reinforcement learning mechanism uses a deep Q-network (DQN). The state space of the deep Q-network includes the average feature selection ratio of the current population, the dispersion of the feature selection ratio of each individual in the population, and the evaluation count consumption ratio. The reward signal of deep Q-networks is calculated based on the absolute gain of the hypervolume index between two adjacent generations.

[0010] Preferably, the prior guiding vector estimates the potential contribution of features based on the statistical information of the initial main population. For the th The contribution of each feature is defined as: ; in, This indicates that the main population contains the first A set of individuals with certain characteristics , Indicates the first Individual, Indicates the first The classification error rate of each individual; The sparsity distribution guide vector is constructed by statistically analyzing the frequency of each feature being selected in the main population, and is defined as: ; in, Indicates the first The individual in the first The values ​​that can be taken on each feature For indicator functions, Population size; The exploratory guide vector is generated by adding a Gaussian perturbation to the standardized prior guide vector, and the perturbation strength decays exponentially with the number of generations.

[0011] Preferably, the adaptive multi-type elite management mechanism in step S4 specifically includes: S41. Extract individuals from the current main population that are the first non-dominant frontier as performance elites; S42. In the remaining population after removing the selected individuals, select the individual with the smallest classification error rate as the error-oriented elite. S43. Among the individuals that have not yet been selected, based on the positional relationship in the normalized target space, select the individuals located in the lower left corner of the target space as convergent equilibrium elites.

[0012] Preferably, the type-aware elite evolution strategy in step S5 specifically includes: The elite population is divided into several subgroups according to type, and each type of elite only undergoes crossover and mutation operations within its own subgroup. For error-oriented elites, a binary tournament based on classification error rate is used to select the parent, without performing crossover operations, only performing local mutation; For performance-oriented elites and convergent equilibrium elites, the parent generation is randomly selected, and a uniform crossover operation is performed when the diversity of the elite population is higher than a preset threshold. Elite population diversity is measured using the average normalized Hamming distance, and is defined as: ; in, This indicates the diversity of the current elite population. Indicates the size of the elite population. Representing feature dimension, Represents an individual With individuals Hamming distance between them; Different upper bound constraints are set for the mutation intensity of different types of elites.

[0013] Preferably, in step S3, the selection of the guiding vector is decided by the deep Q network based on the current state, and the deep Q network is trained and updated using the hypervolume gain reward signal.

[0014] Preferably, in step S4, performance-oriented elites are used to maintain the frontier representativeness of the elite pool in the multi-objective space, error-oriented elites are used to provide directions for performance improvement, and convergence-balancing elites are used to promote the convergence of the population to the ideal solution region.

[0015] Preferably, the diversity threshold of the elite population in step S5 is used to control the start and stop of crossover operations. When the diversity of the elite population is lower than the diversity threshold, the crossover operation between performance-oriented elites and convergent equilibrium-oriented elites is disabled.

[0016] This invention also provides a high-dimensional gene expression profile cancer subtype classification system based on two-layer co-evolution and reinforcement learning, comprising: The data acquisition module is used to acquire gene expression profile data to be classified. The feature selection module is connected to the data acquisition module and includes: a framework construction unit, a reinforcement learning guidance unit, an elite management unit, an elite evolution unit, and an iterative control unit. The classification module, connected to the feature selection module, is used to classify gene expression profile data based on the optimal feature subset to obtain cancer subtype classification results.

[0017] Preferably, the feature selection module specifically includes: The framework building unit is used to construct a two-layer evolutionary framework in which the main population and the elite population co-evolve. The main population is used to maintain search diversity, and the elite population is used to retain and develop high-quality feature subsets. A reinforcement learning guidance unit, connected to the framework building unit, is used to dynamically select guidance vectors using a reinforcement learning mechanism, and drive the evolutionary process of the main population based on the guidance vectors. The elite management unit, connected to the framework construction unit, is used to select elite individuals from the main population and inject them into the elite population using an adaptive multi-type elite management mechanism, and to classify elite individuals into three categories: performance-oriented elites, error-oriented elites, and convergence-balanced elites. The elite evolution unit, connected to the elite management unit, is used to evolve the elite population using a type-aware elite evolution strategy, setting differentiated parent selection methods, crossover strategies, and mutation intensities for different types of elite individuals. The iterative control unit is connected to the reinforcement learning guidance unit, the elite management unit, and the elite evolution unit, respectively. It is used to iteratively execute reinforcement learning guidance, elite management, and elite evolution until a preset termination condition is reached. The preset termination condition is reaching a preset maximum number of evaluations. After that, the final elite population is output as the optimal feature subset.

[0018] Therefore, the present invention employs the above-mentioned high-dimensional gene expression profile cancer subtype classification method and system based on two-layer co-evolution and reinforcement learning, and achieves the following technical effects: (1) The elite population serves as an independent container for excellent solutions. Combined with the sequential screening and maintenance of the three types of elites, the key gene combinations are not easily destroyed during the evolution process, and the stability of the solution set is significantly improved, which directly solves the problem of the easy loss of excellent combinations.

[0019] (2) The deep Q network dynamically selects the exploratory guiding vector and the perturbation intensity decreases with the algebra. Combined with the local mutation of the error-oriented elite, it achieves an adaptive balance between exploration and development, effectively avoiding the search from getting trapped in local optima.

[0020] (3) The two-layer division of labor between the main population exploration and the elite population development, combined with the cross operation of diversity threshold control, enables the algorithm to maintain both population diversity and convergence accuracy at the same time. The output gene subset achieves a better balance between the number of features and the classification accuracy, meeting the requirement of "few but excellent" feature subsets for cancer subtype classification.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart of a high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning according to the present invention; Figure 2 This is an architecture diagram of a high-dimensional gene expression profile cancer subtype classification system based on two-layer co-evolution and reinforcement learning, as proposed in this invention. Figure 3 This is a flowchart of the RL-TECE algorithm, an embodiment of a high-dimensional gene expression profile cancer subtype classification method and system based on two-layer co-evolution and reinforcement learning, according to the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example 1 This embodiment provides a high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning. The overall process is as follows: Figure 1 As shown, it includes the following steps: Step S1: Obtain gene expression profile data to be classified.

[0026] Obtain gene expression profile data to be classified. This data includes multiple samples and multiple gene features corresponding to each sample. Let the number of samples be... The number of gene features is , Much larger The data is characterized by high dimensionality and small sample size. Each sample corresponds to a cancer subtype category label.

[0027] Step S2: Construct a two-layer evolutionary framework in which the main population and the elite population iterate in parallel and cooperate with each other.

[0028] Initialize the main population Population size is Each individual is represented by a binary encoded vector, with a length equal to the feature dimension. , of which A value of 1 indicates that the gene feature is selected, while a value of 0 indicates that it is not selected. The objective function for each individual includes: the proportion of selected features. (The smaller the better) and classification error rate (The smaller the better).

[0029] Initialize elite population The upper limit of the scale is Initially, elites are selected from the main population through an adaptive multi-type elite management mechanism (see step S4 for details). The main population is responsible for global exploration and maintaining population diversity; the elite population is responsible for preserving and developing high-quality feature subsets. The two achieve co-evolution through information exchange in each generation.

[0030] The feature selection process described in steps S2 to S6 of this invention constitutes a reinforcement learning-assisted two-layer elite co-evolutionary multi-objective feature selection algorithm (RL-TECE). The detailed flow of this algorithm is as follows: Figure 3 As shown below. Figure 3 This paper provides a detailed explanation of each core step of the RL-TECE algorithm.

[0031] Step S3: Employ reinforcement learning to dynamically select a guiding vector, and drive the evolution of the main population based on the guiding vector.

[0032] This step further includes the following sub-steps: S31: Construct three types of guiding vectors.

[0033] The three types of guiding vectors, as the set of optional actions for a deep Q-network (DQN), are defined as follows: (1) Prior guidance vector: Used to provide performance-oriented information for the search. The potential contribution of each feature is estimated based on the statistics of the initial main population. For the ... Each feature has a contribution defined as: ; in, This indicates that the main population contains the first A set of individuals with certain characteristics , Indicates the first Individual, Indicates the first The classification error rate of each individual; (2) Sparsity distribution guiding vector: used to characterize the overall structural information of the current main population in the feature selection space. It is constructed by statistically analyzing the frequency of each feature being selected in the main population and is defined as: ; in, Indicates the first The individual in the first The values ​​that can be taken on each feature For indicator functions, Population size; (3) Exploratory guiding vector: used to enhance the randomness and exploration ability of the search process while utilizing prior information. First, the prior guiding vector is standardized: ; in, This represents the standardized prior guiding vector. This represents the prior guiding vector. Represents the prior guiding vector The mean, Represents the prior guiding vector standard deviation Represent a very small positive number (e.g.) This is used to prevent the calculation result from being infinite or undefined due to a denominator of zero. Then, a Gaussian perturbation is introduced based on the standardized result to construct a noisy exploration guidance vector: ; Among them, noise intensity With each generation of evolution Exponential decay enhances exploration capabilities in the early stages of a search and gradually strengthens the convergence trend in later stages. This represents a random noise vector that follows a standard multivariate normal distribution. .

[0034] S32: Establish the state space, action space, and reward signal of the Deep Q Network (DQN).

[0035] state space It consists of multidimensional statistical information, including: the average feature selection ratio of the current population, the dispersion (standard deviation) of the feature selection ratio of individuals in the population, and the evaluation count consumption ratio (current evaluation count / maximum evaluation count).

[0036] The action space is for the selection of three types of guiding vectors, corresponding to the selection of prior guiding vectors, sparse distribution guiding vectors, or exploratory guiding vectors, respectively.

[0037] The reward signal is based on the improvement in multi-objective performance between two adjacent generations of the main population, and the absolute gain of hypervolume (HV) is used as the immediate reward: ; in, Indicates an immediate reward. Indicates contemporary (i.e., the first) HV value of the main population (generation) This represents the HV value of the previous generation of the main population.

[0038] S33: DQN Training and Guided Vector Selection.

[0039] Initialize the DQN network parameters. In each generation of evolution, DQN adjusts the parameters based on the current state. The action is selected, and the guiding vector to be used is determined. Then, based on the selected guiding vector, the AGSEA (Adaptive Guided Subspace Evolution Algorithm) method is used to generate the offspring population: First, the guiding vector is used to perform 2-means clustering on the feature dimensions, dividing the features into two groups; then, two parent individuals are randomly selected from the main population, and one set of features is randomly selected, and the gene loci corresponding to that set of features are swapped to generate two offspring; subsequently, position flipping mutation is performed according to the mutation probability. The mutation probability is adaptively adjusted with the number of generations.

[0040] After generating the offspring population, the offspring are merged with the parent generation, and the SPEA2 environmental selection mechanism is used for preservation. Individuals constitute the new generation of the main population. Simultaneously, the current generation's state transition storage is used to create an experience replay pool, and the DQN network is updated using gradient descent at regular intervals. The training objective is to minimize the temporal difference error.

[0041] Step S4: Adaptive multi-type elite management mechanism is used to select elite individuals from the main population and inject them into the elite population. The elite individuals are divided into three categories: performance-oriented elites, error-oriented elites, and convergence-balanced elites.

[0042] This step executes the following filtering sub-steps in a preset order: S41: Extract the individuals at the first non-dominated front from the current main population as performance elites. Perform non-dominated ranking on the individuals in the main population; individuals at the first front represent the solution that achieves the optimal trade-off between multiple objectives (classification error rate and the proportion of selected features), and label them as type 1 and add them to the elite population.

[0043] S42: In the remaining population after removing the selected individuals, select the individual with the lowest classification error rate as an error-oriented elite. Exclude the individuals selected in step S41 from the main population, find the individual with the lowest classification error rate among the remaining individuals, mark it as type 2, and add it to the elite population. Although this type of elite may select more features, it provides a clear direction for performance improvement in the search.

[0044] S43: Among the unselected individuals, based on the positional relationships in the normalized target space, select individuals located in the lower left corner of the target space as convergent equilibrium elites. Normalize the two objective function values ​​(classification error rate, feature ratio) of the unselected individuals in the main population, then divide the target space into regions, focusing on the lower left corner region (i.e., the region where both objective values ​​are low). Select the individual closest to the origin within this region, label it as type 3, and add it to the elite population. This type of elite represents the potential search direction within the ideal solution region.

[0045] The selection order for the three types of elites is fixed, but the actual number injected is adaptively determined by the distribution of the current main population, without the need for preset ratio parameters. If the elite population size exceeds the upper limit... Then, the environment selection mechanism will prune redundant individuals and update the type markers of each elite simultaneously.

[0046] Step S5: Employ a type-aware elite evolution strategy to evolve the elite population, setting differentiated parent selection methods, crossover strategies, and mutation intensities for different types of elite individuals.

[0047] This step further includes the following sub-steps: S51: Type Division and Subgroup Isolation. The elite population is divided into three subgroups according to type (1, 2, 3). Elites of each type only undergo crossover and mutation operations within their respective subgroups. There is no direct genetic interaction between different types, thus avoiding mutual interference of search preferences.

[0048] S52: Differentiated parent selection.

[0049] For Type 2 (error-oriented elites): a binary tournament selection of parents based on classification error rate is used to reinforce the genetic advantage of individuals with high classification performance.

[0050] For Type 1 (performance-oriented elites) and Type 3 (convergent equilibrium elites): random selection of parents is used to reduce selection pressure and maintain diversity within the subgroup.

[0051] S53: Differentiated cross-selling strategy.

[0052] First, the diversity of the elite population is calculated. The average normalized Hamming distance is used as a diversity measure, defined as: ; in, This indicates the diversity of the current elite population. Indicates the size of the elite population. Representing feature dimension, Represents an individual With individuals Hamming distance between them; when When the diversity threshold is greater than the preset threshold: Subgroups of type 1 and type 3 perform uniform crossover operations to recombine different structural patterns.

[0053] Type 2 does not perform crossover operations.

[0054] when When less than or equal to the preset diversity threshold: Temporarily disable type 1 and type 3 cross operations to protect the established good structure.

[0055] S54: Differentiated mutation strategy.

[0056] First, define the basic variation strength: ; Then, different upper limits for mutation intensity are set for different types of elites: Type 1 (Performance Elite): Allowed mutation strength upper limit is 2 This is to enhance the exploration of non-dominated solution regions.

[0057] Type 2 (Error-oriented Elite): Allowable upper bound for mutation strength is This allows for localized fine-tuning.

[0058] Type 3 (Convergent Equilibrium Elite): Allows an upper bound of 0.5 for mutation intensity. (Round down) to promote smooth convergence.

[0059] In the specific implementation, for each elite individual, the maximum number of mutation bits is determined according to its type. Then randomly generate a number not exceeding integers Randomly select from the individual's code Each gene position is flipped (i.e., 0 becomes 1, 1 becomes 0).

[0060] S55: Elite Population Update and Pruning. The evolved elite population is merged with the original elite population; if the size exceeds the upper limit... Then, the SPEA2 environment selection mechanism will be used to cut the limits and update the type tags of each elite simultaneously.

[0061] Step S6: Iterate through steps S3 to S5 until a preset termination condition is met. The preset termination condition is reaching a preset maximum number of evaluations (function evaluation count). Then, output the final elite population as the optimal feature subset. This subset corresponds to a set of selected gene features.

[0062] Each assessment of the classification error rate for individuals in the main population or elite population is included in the total number of assessments. The iteration terminates when the number of assessments reaches a preset maximum value.

[0063] Step S7: Classify the gene expression profile data based on the optimal feature subset to obtain the cancer subtype classification results.

[0064] Gene feature columns corresponding to the optimal feature subset are extracted from the gene expression profile data to form a dimensionality-reduced dataset. A classifier (e.g., support vector machine, k-nearest neighbors, or random forest) is used to train and classify this dataset, outputting the cancer subtype category for each sample. Classifier parameters are set using default settings or determined through cross-validation.

[0065] Example 2 This embodiment provides a high-dimensional gene expression profile cancer subtype classification system based on two-layer co-evolution and reinforcement learning, used to execute the method of Embodiment 1. Figure 2 As shown, the system includes: The data acquisition module is used to acquire gene expression profile data to be classified. The gene expression profile data includes multiple samples and multiple gene features corresponding to each sample.

[0066] A feature selection module is connected to the data acquisition module. This feature selection module further includes the following units: Framework Building Unit: This unit is used to construct a two-layer evolutionary framework in which the main population and elite population co-evolve. The main population is used to maintain search diversity, while the elite population is used to preserve and develop high-quality feature subsets. This unit is responsible for basic settings such as population initialization, individual encoding, and objective function definition.

[0067] The reinforcement learning guidance unit, connected to the framework building unit, is used to dynamically select guidance vectors using reinforcement learning mechanisms, driving the evolutionary process of the main population based on these guidance vectors. This unit implements the calculation of three types of guidance vectors, state awareness of the deep Q-network, action selection, reward calculation, and network training and updates.

[0068] Elite Management Unit: Connected to the framework construction unit, this unit employs an adaptive multi-type elite management mechanism to select elite individuals from the main population and inject them into the elite population. Elite individuals are categorized into three types: performance-oriented elites, error-oriented elites, and convergence-equilibrium elites. This unit executes the selection according to sequential rules and maintains the size and type labels of the elite population.

[0069] Elite Evolution Unit: Connected to the elite management unit, this unit employs a type-aware elite evolution strategy to evolve the elite population, setting differentiated parent selection methods, crossover strategies, and mutation intensities for different types of elite individuals. This unit implements subpopulation partitioning, diversity calculation, conditional crossover, and mutation operations with upper bounds on the intensity constraints.

[0070] Iterative Control Unit: Connected to the reinforcement learning guidance unit, elite management unit, and elite evolution unit, this unit iteratively executes reinforcement learning guidance, elite management, and elite evolution until a preset termination condition (maximum number of evaluations) is reached, outputting the final elite population as the optimal feature subset. This unit is responsible for accumulating the total number of evaluations, determining termination, and extracting the optimal solution.

[0071] The classification module, connected to the feature selection module, is used to classify gene expression profile data based on the optimal feature subset to obtain cancer subtype classification results. This module uses the selected gene features to reduce the dimensionality of the original data and calls the classifier to output the final category label.

[0072] Therefore, this invention employs the aforementioned high-dimensional gene expression profile cancer subtype classification method and system based on bilayer co-evolution and reinforcement learning. Addressing the challenges of high-dimensional, small-sample gene expression profile data, easy loss of key gene combinations, susceptibility to local optima in search, and the difficulty of balancing diversity and convergence in single-layer structures, this invention utilizes a bilayer elite co-evolution framework to continuously retain superior solutions, employs reinforcement learning to dynamically select guiding vectors to achieve an adaptive balance between exploration and development, and maintains population diversity and convergence accuracy through multi-type elite sequential screening and type-aware differential evolution strategies. This provides a stable and efficient feature screening scheme for cancer subtype classification.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning, characterized in that, Includes the following steps: S1. Obtain gene expression profile data to be classified. The gene expression profile data includes multiple samples and multiple gene features corresponding to each sample. S2. Construct a two-layer evolutionary framework in which the main population and the elite population iterate in parallel and cooperate with each other. The main population is used to maintain search diversity, and the elite population is used to retain and develop high-quality feature subsets. S3. Employ reinforcement learning to dynamically select the guiding vector, and drive the evolutionary process of the main population based on the guiding vector; S4. An adaptive multi-type elite management mechanism is adopted to select elite individuals from the main population and inject them into the elite population. The elite individuals are divided into three categories: performance-type elites, error-oriented elites, and convergence-balance-type elites. S5. Employ a type-aware elite evolution strategy to evolve the elite population, setting differentiated parent selection methods, crossover strategies, and mutation intensities for different types of elite individuals. S6. Iterate through steps S3 to S5 until a preset termination condition is reached. The preset termination condition is reaching a preset maximum number of evaluations. Then, output the final elite population as the optimal feature subset. S7. Classify gene expression profile data based on the optimal feature subset to obtain cancer subtype classification results; In step S3, the guiding vectors include prior guiding vectors, sparse distribution guiding vectors, and exploratory guiding vectors; The reinforcement learning mechanism uses a deep Q-network. The state space of the deep Q-network includes the average feature selection ratio of the current population, the dispersion of the feature selection ratio of each individual in the population, and the evaluation count consumption ratio. The reward signal of the deep Q network is calculated based on the absolute gain of the hypervolume index between two adjacent generations. The prior guiding vector estimates the latent contribution of features based on the statistical information of the initial main population. For the th... The contribution of each feature is defined as: ； in, This indicates that the main population contains the first The number of individuals with each characteristic , Indicates the first Individual, Indicates the first The classification error rate of each individual; The sparsity distribution guide vector is constructed by statistically analyzing the frequency of each feature being selected in the main population, and is defined as: ； in, Indicates the first The individual in the first The values ​​of each feature For indicator functions, Population size; The exploratory guide vector is generated by adding a Gaussian perturbation to the standardized prior guide vector, and the perturbation strength decays exponentially with the number of generations. The adaptive multi-type elite management mechanism in step S4 specifically includes: S41. Extract individuals from the current main population that are the first non-dominant frontier as performance elites; S42. In the remaining population after removing the selected individuals, select the individual with the smallest classification error rate as the error-oriented elite. S43. Among the individuals that have not yet been selected, based on the positional relationship in the normalized target space, select the individuals located in the lower left corner of the target space as convergent equilibrium elites.

2. The high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning according to claim 1, characterized in that, The type-aware elite evolution strategy in step S5 specifically includes: The elite population is divided into several subgroups according to type, and each type of elite only undergoes crossover and mutation operations within its own subgroup. For error-oriented elites, a binary tournament based on classification error rate is used to select the parent, without performing crossover operations, only performing local mutation; For performance-oriented elites and convergent equilibrium elites, the parent generation is randomly selected, and a uniform crossover operation is performed when the diversity of the elite population is higher than a preset threshold. Elite population diversity is measured using the average normalized Hamming distance, and is defined as: ； in, This indicates the diversity of the current elite population. Indicates the size of the elite population. Representing feature dimension, Represents an individual With individuals Hamming distance between them; Different upper bound constraints are set for the mutation intensity of different types of elites.

3. The high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning according to claim 1, characterized in that, In step S3, the selection of the guiding vector is decided by the deep Q network based on the current state. The deep Q network is trained and updated using the hypervolume gain reward signal.

4. The high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning according to claim 1, characterized in that, In step S4, performance-oriented elites are used to maintain the frontier representativeness of the elite pool in the multi-objective space, error-oriented elites are used to provide directions for improving classification performance, and convergence-balancing elites are used to promote the convergence of the population to the ideal solution region.

5. The high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning according to claim 1, characterized in that, In step S5, the diversity threshold of the elite population is used to control the start and stop of crossover operations. When the diversity of the elite population is lower than the diversity threshold, the crossover operation between performance-oriented elites and convergent equilibrium-oriented elites is disabled.

6. A high-dimensional gene expression profile cancer subtype classification system based on two-layer co-evolution and reinforcement learning, used to execute the high-dimensional gene expression profile cancer subtype classification method based on two-layer co-evolution and reinforcement learning as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire gene expression profile data to be classified. The feature selection module is connected to the data acquisition module and includes: a framework construction unit, a reinforcement learning guidance unit, an elite management unit, an elite evolution unit, and an iterative control unit. The classification module, connected to the feature selection module, is used to classify gene expression profile data based on the optimal feature subset to obtain cancer subtype classification results.

7. A high-dimensional gene expression profile cancer subtype classification system based on two-layer co-evolution and reinforcement learning according to claim 6, characterized in that, The feature selection module specifically includes: The framework building unit is used to construct a two-layer evolutionary framework for the co-evolution of the main population and the elite population. The main population is used to maintain search diversity, and the elite population is used to retain and develop high-quality feature subsets. A reinforcement learning guidance unit, connected to the framework building unit, is used to dynamically select guidance vectors using a reinforcement learning mechanism, and drive the evolutionary process of the main population based on the guidance vectors; The elite management unit, connected to the framework construction unit, is used to select elite individuals from the main population and inject them into the elite population using an adaptive multi-type elite management mechanism, and to classify elite individuals into three categories: performance-oriented elites, error-oriented elites, and convergence-balanced elites. The elite evolution unit, connected to the elite management unit, is used to evolve the elite population using a type-aware elite evolution strategy, setting differentiated parent selection methods, crossover strategies, and mutation intensities for different types of elite individuals. The iterative control unit is connected to the reinforcement learning guidance unit, the elite management unit, and the elite evolution unit, respectively. It is used to iteratively execute reinforcement learning guidance, elite management, and elite evolution until a preset termination condition is reached. The preset termination condition is reaching a preset maximum number of evaluations. After that, the final elite population is output as the optimal feature subset.

Images