Cancer classification and gene discovery algorithm based on adaptive learning search

Abstract

The application relates to a cancer classification and gene discovery algorithm based on adaptive learning search, and comprises the following steps: performing preliminary screening on genomic data by using F scores, calculating the F scores according to the correlation between genes and sample labels, and selecting genes with higher F scores as an initial feature gene set for model training; introducing a learning search algorithm, integrating historical experience and collective wisdom, and further optimizing the gene subset through two stages of global exploration and local development to finally construct an optimal gene feature subset. The gene selection algorithm Fs-LSA can quickly identify key genes for tumor classification and diagnosis from high-dimensional data by preliminarily screening genes by using F-scores and further optimizing gene selection by combining a learning search algorithm (LSA), so that targeted treatment intervention can be carried out in tumor management.

Description

Technical Field

[0001] This invention belongs to the field of biomedical data analysis technology, specifically relating to a cancer classification and gene discovery algorithm based on adaptive learning search. Background Technology

[0002] Cancer is one of the world's major health problems, with a large number of new cases each year. Despite advances in medical technology for cancer diagnosis and treatment, its complexity still presents numerous challenges. With the development of DNA microarray technology, researchers have been able to acquire vast amounts of gene expression data, providing a powerful tool for early cancer diagnosis and prognosis. However, the overfitting problem caused by high-dimensional data and the challenge of selecting genes closely related to cancer classification from massive datasets remain urgent problems to be solved. Existing gene selection methods (such as traditional machine learning algorithms, Cytohubba methods, and other metaheuristic algorithms) suffer from high computational costs and a tendency to get trapped in local optima. Therefore, a more efficient and accurate gene selection algorithm is needed to improve the performance of cancer classification. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a cancer classification and gene discovery algorithm based on adaptive learning search (i.e., Fs-LSA algorithm, F-score based Learning Search Algorithm). By combining F-score and learning search algorithm (LSA), it efficiently and accurately identifies key genes for cancer classification from high-dimensional gene expression data, thereby improving cancer diagnosis and prognosis capabilities.

[0004] The specific technical solution is: a cancer classification and gene discovery algorithm based on adaptive learning search, including the following steps:

[0005] Step 1: Use F-scores to perform preliminary screening of genes. Calculate F-scores based on the correlation between genes and sample labels, and select genes with higher F-scores as the initial feature gene set for model training. This step ensures that the genes in the initial feature subset have high correlation and discriminative power.

[0006] Step 2: Introduce the Learning Search Algorithm (LSA). Based on the principle of swarm intelligence, it integrates historical experience and collective wisdom. Through two phases—global exploration and local development—it dynamically balances exploration and utilization to further optimize the gene subset, ultimately constructing the optimal gene feature subset. This step effectively optimizes the size of the feature subset, ensuring that while reducing the number of features, it maintains high classification accuracy and good generalization ability. This not only improves the model's efficiency but also provides valuable gene selection results for biomedical research.

[0007] Furthermore, step 1 specifically includes the following steps:

[0008] A1. Input Data: Obtain a high-dimensional gene expression dataset from a public database {( ),( ),......,( ),......,( Data sets for diseases such as colon cancer, central nervous system cancer, and leukemia;

[0009] in, This represents a dim-dimensional feature vector, while The label represents the category label, and sc represents the number of samples.

[0010] A2. Data preprocessing: Standardize the data, remove noise and missing values, and ensure data quality;

[0011] A3. F-score calculation: The F-score is calculated based on the correlation between each gene feature and the sample label. The specific calculation formula is shown in equation (1):

[0012] (1)

[0013] Where c is the number of categories; It is the number of samples for the i-th gene in the k-th category; It is the j-th sample value of the ith gene in the k-th category; It is the average value of the i-th gene across all samples; It is the average value of the i-th gene in the k-th category;

[0014] A4. Feature Ranking and Selection: The features are ranked according to the F score, and the top n features are selected to form the initial feature gene set for model training (the higher the F score, the stronger the differential expression ability of the gene in different categories of samples, which means that the expression level of these genes in tumor samples is significantly different from that in normal samples, and therefore can be used as an effective biomarker to distinguish between tumor and non-tumor samples).

[0015] Furthermore, step 2 specifically includes the following steps:

[0016] B1. Initialize the population: Randomly generate an initial population, where each individual represents a subset of genes. The population size can be set according to the problem size and complexity.

[0017] B2. Global Exploration: Using historical experience and information from the current population to guide the algorithm for global search, avoiding local optima, specifically updating the individual position through formula (2):

[0018] (2)

[0019] in, Let i and j represent individuals i and j in the t-th generation of the current population, respectively. Let i represent individual i in the (t+1)th generation of the current population; the variable rand is a random number with dimension n and a value range of 0-1; Let x represent any individual in the historical experience record. In the initial stage, a historical population of the same size as x will be randomly generated. In each iteration, the historical record will be updated with a certain probability using formula (3), and then the individuals in the historical record will be randomly arranged using formula (4):

[0020] (3)

[0021] (4)

[0022] Where a and b represent random floating-point values ​​between 0 and 1, and permuting represents random reordering operations;

[0023] B3. Local Development: Enhance individual learning ability through exemplary popular science or teaching behaviors and proactive learning behaviors, and update individual position according to formula (5):

[0024] (5)

[0025] Where rand and r are random numbers, It is the best individual in the current period; It is the learning factor, determined by formula (6); The average value of the subjects taught in the example is calculated using formula (7):

[0026] (6)

[0027] (7)

[0028] Where sub represents the number of entities in the example professor. During the experimental testing, it was found that the algorithm achieved the best performance when sub=3; randi is a random integer in the range [1,sub].

[0029] B4. Dynamic Balance Control: By using a balance control factor δ to balance global exploration and local exploitation, the algorithm ensures that it performs global search in the early stages and local optimization in the later stages, thereby improving the overall performance of the algorithm. The specific formula is as follows:

[0030] (8)

[0031] (9)

[0032] in, and These represent the initial and final values ​​of δ, respectively. Indicates the maximum number of iterations; This represents an unordered sequence generated by a logical self-mapping function, used to introduce nonlinear changes;

[0033] This step introduces a dynamic balance control factor δ to dynamically balance global exploration and local development. The specific balancing mechanism is as follows:

[0034] In the early stages of the algorithm, the value of δ is relatively large, and the algorithm mainly performs global exploration, extensively searching the solution space through historical experience and swarm intelligence to avoid getting trapped in local optima. Subsequently, as the number of iterations increases, the value of δ gradually decreases, and the algorithm gradually shifts to local exploitation. Through demonstrative popular science (educational) behavior and active learning behavior, it refines and optimizes the current solution, improving its quality. Thus, by dynamically adjusting the value of δ, the Fs-LSA algorithm can achieve a smooth transition between global exploration and local exploitation, avoiding premature trapping in local optima while allowing for refined optimization of the solution in later stages, thereby improving the overall performance of the algorithm. This dynamic balancing mechanism enables the Fs-LSA algorithm to effectively identify genes related to cancer classification when processing high-dimensional gene expression data, while maintaining high classification accuracy and stability.

[0035] B5. Model Training and Validation: First, a selected subset of genes is used to train the machine learning model (e.g., using a discriminant analysis classifier). During training, the model learns how to distinguish different cancer types or sample categories based on the expression data of these genes. Then, a separate validation dataset is used to evaluate the model's performance. Starting from the initial feature subset, the number of features is gradually reduced. After each reduction, the model is retrained and its performance is evaluated. The performance of the new model is compared with the performance of the previous model to determine whether the model's performance has decreased after reducing the number of features.

[0036] B6. Determine the optimal feature subset size: By iterating and adjusting the feature subset size multiple times, find the number of features that maximizes the classification accuracy without affecting the model performance.

[0037] Furthermore, due to the nature of the gene recognition problem, which requires the selection of specific genes, each solution is represented as an n-dimensional binary vector, where 1 represents the selected feature and 1 represents the unselected feature; each dimension of the vector is denoted as... , corresponding to an input gene vector If genes If the j-th position is selected, then Set it to 1, otherwise set it to 0. Here, m represents the population size, and n represents the dimension (number of genes). Let represent the value of the j-th dimension of the i-th individual. To make the Latent Semantic Analysis (LSA) algorithm applicable to solving gene identification problems, the individual positions in step B3 are converted into discrete values ​​of 0 or 1 according to formulas (10) and (11) (e.g., ...). Figure 3 (As shown), indicating whether a gene has been selected:

[0038] (10)

[0039] (11)

[0040] Among them, rand ) is a random number.

[0041] Furthermore, step B5 evaluates the model's performance metrics, including classification accuracy, precision, recall, and F1 score. These metrics measure the model's performance on unseen data, ensuring that the model has good generalization ability.

[0042] Furthermore, step B6 determines the specific criteria for the optimal feature subset size as follows: while ensuring that the classification accuracy does not decrease significantly, precision, recall, and F1 score should also be considered. Under the premise that these indicators are all good, the number of features should be reduced as much as possible to simplify the model, improve computational efficiency, and enhance the interpretability of the model.

[0043] Furthermore, in order to achieve a balance between minimizing the number of features and maximizing classification accuracy, formula (12) is used as the objective function value calculation method for the fitness function:

[0044] (12)

[0045] Where accuracy is the precision of the discriminant analysis classifier; α and β are weight parameters, and rand β=1-α; n i is the number of features selected by individual i, and n is the total number of features.

[0046] The beneficial effects of this invention are as follows: The gene selection algorithm Fs-LSA described in this invention performs preliminary screening of genes using F-scores, and further optimizes gene selection by combining it with a learning search algorithm (LSA). This enables rapid identification of key genes for tumor classification and diagnosis from high-dimensional data, thereby facilitating targeted treatment interventions in tumor management. Furthermore, the algorithm employs a combination of multiple search strategies, significantly improving its flexibility and adaptability, allowing it to perform well in handling different types of cancer data. In particular, the learning search algorithm utilizes historical experience and other search individuals to guide the learning process (the indirect nature of human learning) to achieve a global search; on the other hand, it uses the current best individual (role model) to educate other individuals, while the algorithm itself actively learns from the best individual (the proactive nature of human learning), achieving a local search. Furthermore, it introduces a dynamic balance control factor δ to achieve a smooth transition between global exploration and local development. This mechanism allows the algorithm to broadly search the solution space in the early stages, avoiding getting trapped in local optima; and in the later stages, it refines and optimizes the solutions, improving their quality. This not only improves the convergence speed of the algorithm but also enhances its adaptability to complex datasets. Compared to traditional gene selection methods, Fs-LSA significantly reduces computation time while maintaining high accuracy. Fs-LSA not only identifies key genes but also generates a minimized feature subset, simplifying the model and reducing computational resource consumption. This is particularly important when dealing with high-dimensional data, effectively avoiding overfitting and noise, and improving the model's generalization ability, which is crucial for accurate cancer detection and classification. Furthermore, the driver genes identified by Fs-LSA can be used to build independent prognostic models, supporting personalized medicine. Attached Figure Description

[0047] Figure 1 This is an overall flowchart of an embodiment;

[0048] Figure 2 This is a progress graph of the learning search algorithm in the embodiment;

[0049] Figure 3 This is a schematic diagram illustrating the formation process of the new solution in the embodiment;

[0050] Figure 4 This is a schematic diagram illustrating the calculation process of fitness values. Detailed Implementation

[0051] To make the technical problems and solutions solved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0052] like Figure 1As shown, this embodiment provides a cancer classification and gene discovery algorithm based on adaptive learning search, including the following steps:

[0053] 1. Data Preparation

[0054] Using publicly available cancer gene expression datasets, this embodiment selects the GSE4290 and GSE68848 datasets from the GEO (Gene Expression Omnibus) database, containing 305 glioblastoma samples and 51 normal brain tissue samples. Then, the "ComBat" function is used to handle batch effects and ensure data consistency.

[0055] 2. Feature gene screening based on F-score

[0056] The F-score of each gene is calculated according to formula (1), and the top 100 genes with the highest F-scores are selected as the initial feature gene set. This number was determined experimentally to balance classification accuracy and computational efficiency.

[0057] 3. Learning Search Algorithm (LSA) Optimization Selection (e.g.) Figure 2 Figure 4 (As shown)

[0058] (1) Initialize the population: Randomly generate the initial population, each individual represents a subset of genes, and the population size is 50.

[0059] (2) During the iteration process, the LSA algorithm optimizes the gene subset through two stages: global exploration and local exploitation.

[0060] Global exploration: The algorithm is guided by historical experience and information in the current population to conduct a global search, avoiding local optima. Specifically, the individual position is updated by formula (2).

[0061] Local development: Improve individual learning ability through exemplary popular science or teaching behaviors and active learning behaviors, and update individual position according to formula (5).

[0062] (3) Dynamic balance control: By balancing the global exploration and local development through the balance control factor δ, the algorithm can ensure that it performs global search in the early stage and local optimization in the later stage, thereby improving the overall performance of the algorithm.

[0063] (4) After each iteration, update the population according to the fitness function (considering classification accuracy and number of feature genes) and retain the best individual.

[0064] Maximum number of iterations during LSA optimization =200; the initial and final values ​​of the balance control factor δ are δ init =1 and δ final =0.1.

[0065] 4. Model Training and Validation: The classifier was trained using an initial feature subset. The number of features was gradually reduced, and the model was retrained and its performance evaluated after each reduction. After approximately 150 iterations, the model converged to the globally optimal feature subset. At this point, the model's average classification accuracy reached 99.494%, the AUC value reached 0.92837, and an average of 8.525 key genes were identified (specifically, ANXA2, BTG1, DDIT4, IGFBP2, SOX2, PYGL, RBBP8, PLTP, and SOX9). The expression levels of these genes in glioblastoma samples were significantly higher than those in normal tissue samples.

[0066] 5. Biological Significance: These key genes likely play important roles in the biological processes of glioblastoma, involving cell cycle regulation, cell proliferation, and apoptosis. Enrichment analysis revealed that these genes are primarily enriched in the Wnt signaling pathway and negative cell cycle regulation. Furthermore, the risk scoring model constructed using these key genes demonstrated good performance in prognostic assessment, with AUC values ​​of 0.71 (1-year survival prediction) and 0.78 (3-year survival prediction), respectively. The risk scoring model accurately predicts patient survival, providing important reference for clinical treatment and prognostic assessment.

[0067] This example demonstrates that the Fs-LSA algorithm exhibits high classification accuracy, low feature count, and good stability in glioblastoma gene identification, significantly outperforming other algorithms. This algorithm not only improves the accuracy of cancer diagnosis but also provides potential biomarkers and therapeutic targets for personalized treatment.

[0068] The present invention has been described in detail above through specific and preferred embodiments. However, those skilled in the art should understand that the present invention is not limited to the embodiments described above. Any modifications, equivalent substitutions, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cancer classification and gene discovery algorithm based on adaptive learning search, characterized in that, Includes the following steps: Step 1: Use F-scores to perform preliminary screening of genes. Calculate F-scores based on the correlation between genes and sample labels, and select genes with higher F-scores as the initial gene set for model training. Step 2: Introduce a learning search algorithm, integrate historical experience and collective wisdom, and further optimize the gene subset through two stages: global exploration and local development, and finally construct the optimal gene feature subset.

2. The cancer classification and gene discovery algorithm based on adaptive learning search according to claim 1, characterized in that, Step 1 specifically includes the following steps: A1. Input Data: Obtain a high-dimensional gene expression dataset from a public database {( ),( ),......,( ),......,( )};in, This represents a dim-dimensional feature vector, while The label represents the category label, and sc represents the number of samples. A2. Data preprocessing: Standardize the data, remove noise and missing values, and ensure data quality; A3. F-score calculation: The F-score is calculated based on the correlation between each gene feature and the sample label. The specific calculation formula is shown in equation (1): （1） Where c is the number of categories; It is the number of samples for the i-th gene in the k-th category; It is the j-th sample value of the ith gene in the k-th category; It is the average value of the i-th gene across all samples; It is the average value of the i-th gene in the k-th category; A4. Feature Ranking and Selection: Sort the features according to the F-score and select the top n features to form the initial feature set for model training.

3. The cancer classification and gene discovery algorithm based on adaptive learning search according to claim 1 or 2, characterized in that, Step 2 specifically includes the following steps: B1. Initialize the population: Randomly generate an initial population, where each individual represents a subset of genes, and the population size is set according to the problem size and complexity. B2. Global Exploration: Using historical experience and information from the current population to guide the algorithm for global search, avoiding local optima, specifically updating the individual position through formula (2): （2） in, Let i and j represent individuals i and j in the t-th generation of the current population, respectively. Let i represent individual i in the (t+1)th generation of the current population; the variable rand is a random number with dimension n and a value range of 0-1; Let x represent any individual in the historical experience record. In the initial stage, a historical population of the same size as x will be randomly generated. In each iteration, the historical record will be updated with a certain probability using formula (3), and then the individuals in the historical record will be randomly arranged using formula (4): （3） （4） Where a and b represent random floating-point values ​​between 0 and 1, and permuting represents random reordering operations; B3. Local Development: Enhance individual learning ability through exemplary popular science or teaching behaviors and proactive learning behaviors, and update individual position according to formula (5): （5） Where rand and r are random numbers, It is the best individual in the current period; It is the learning factor, determined by formula (6); The average value of the subjects taught in the example is calculated using formula (7): （6） （7） Where sub represents the number of entities in the example professor. During the experimental testing, it was found that the algorithm achieved the best performance when sub=3; randi is a random integer in the range [1,sub]. B4. Dynamic Balance Control: By using a balance control factor δ to balance global exploration and local exploitation, the algorithm ensures that it performs global search in the early stages and local optimization in the later stages, thereby improving the overall performance of the algorithm. The specific formula is as follows: （8） （9） in, and These represent the initial and final values ​​of δ, respectively. Indicates the maximum number of iterations; This represents an unordered sequence generated by a logical self-mapping function, used to introduce nonlinear changes; B5. Model Training and Validation: First, a selected subset of genes is used to train the machine learning model. During training, the model learns how to distinguish different cancer types or sample categories based on the expression data of these genes. Then, a separate validation dataset is used to evaluate the model's performance. Starting from the initial feature subset, the number of features is gradually reduced. After each reduction, the model is retrained and its performance is evaluated. The performance of the new model is compared with the performance of the previous model to determine whether the model's performance has decreased after reducing the number of features. B6. Determine the optimal feature subset size: By iterating and adjusting the feature subset size multiple times, find the number of features that maximizes the classification accuracy without affecting the model performance.

4. The cancer classification and gene discovery algorithm based on adaptive learning search according to claim 3, characterized in that, The individual positions in step B3 are converted into discrete values ​​of 0 or 1 according to formulas (10) and (11), indicating whether the gene was selected: （10） （11） Among them, rand ) is a random number.

5. The cancer classification and gene discovery algorithm based on adaptive learning search according to claim 3, characterized in that, The performance metrics evaluated in step B5 include classification accuracy, precision, recall, and F1 score.

6. The cancer classification and gene discovery algorithm based on adaptive learning search according to claim 3, characterized in that, Step B6 determines the specific criteria for the optimal feature subset size: while ensuring that the classification accuracy does not decrease significantly, precision, recall and F1 score should also be considered. Under the premise that these indicators are all good, the number of features should be reduced as much as possible.

7. The cancer classification and gene discovery algorithm based on adaptive learning search according to claim 6, characterized in that, To achieve a balance between minimizing the number of features and maximizing classification accuracy, formula (12) is used as the objective function value calculation method for the fitness function: （12） Where accuracy is the precision of the discriminant analysis classifier; α and β are weight parameters, and rand β=1-α; n i is the number of features selected by individual i, and n is the total number of features.

Images