Feature gene screening method based on NSGA-ANN and ceRNA network construction method

Abstract

The invention discloses a characteristic gene screening method based on NSGA-ANN and a ceRNA network construction method, and relates to the technical field of gastric cancer research. A public gene expression database is utilized, a non-dominated sorting genetic algorithm and a multi-layer perceptron are combined with a biostatistical method to screen key related genes of gastric cancer, a competitive endogenous RNA regulatory network is constructed, 12 key genes including SPP1, MMP13, ERBB2 and the like and a proto-oncogene regulatory pathway are found, and the screening of the key related genes of gastric cancer is realized. Wherein 11 genes can obviously influence the prognosis condition of a gastric cancer patient, and have the potential to be used as a target of target treatment or an index for early molecular diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of gastric cancer research, and specifically to a characteristic gene screening method based on NSGA-ANN and a ceRNA network construction method. Background Art

[0002] Gastric cancer is a malignant tumor originating from the gastric mucosal epithelium, generally manifesting as adenocarcinoma. The symptoms of early gastric cancer are not obvious. A few patients will have upper gastrointestinal symptoms such as nausea, vomiting or gastric ulcers, which are difficult to attract enough attention. However, the progression of advanced gastric cancer is very rapid and prone to distant metastasis. The 5-year survival rate of patients with advanced gastric cancer is generally only 3% to 5%. Therefore, early screening of gastric cancer is particularly important for gaining treatment time, delaying the development of gastric cancer and even reducing mortality. The current gold standard for early screening of gastric cancer is endoscopic and endoscopic biopsy. Although it is very accurate, it requires higher time and equipment, and patients also have to endure certain pain. With the advancement of molecular biotechnology, methods for diagnosing and treating diseases at the molecular level are developing, among which targeted therapy technology is the representative. This technology targets specific binding sites on malignant tumor cells, designs corresponding anti-tumor drugs and binds to tumor cells in a targeted manner, or interferes with the expression of proto-oncogenes in tumor cells to achieve the purpose of accurately killing cancer cells. However, this technology requires highly precise target locations to maximize the therapeutic effect. Considering the heterogeneous characteristics of gastric cancer and the fact that target detection technology is still not precise enough, personalized targeted therapy is still developing slowly.

[0003] The development of tumors is closely related to gene expression. The growth of cancer cells is often regulated by multiple genes. Although this regulatory relationship is complex and diverse, it is significantly different from normal cells and is not easily interfered by individual and environmental factors. Therefore, differentially expressed genes in tumors can be used as a clear biomarker to distinguish tumor tissues from normal tissues and help clarify the pathogenesis of gastric cancer, which can provide new ideas for molecular diagnosis and targeted treatment of gastric cancer. With the development of information technology and high-throughput sequencing technology, a considerable amount of gene transcription spectrum data has been measured. The gene transcription data collected by gene chips generally have tens of thousands of dimensions, and only a few genes play a key role in cell carcinogenesis. The expression of most genes will be affected by these key genes. In order to explore the essential mechanism and key control targets of gastric cancer development, it is necessary to remove most redundant genes.

[0004] Removing redundant genes is essentially a process of feature selection. Currently, a variety of computational methods have been developed for feature selection tasks, which can be roughly divided into filtering, winding and embedding methods. Most of these methods are based on scoring and ranking the relevance of target genes to clinical traits, or are combined with classification models for simultaneous selection and verification. For example, Signal et al. used the signal-to-noise ratio as an evaluation criterion and used the K-nearest neighbor algorithm (KNN) as a classifier. The selected feature genes achieved good classification results on the prostate cancer dataset. Li et al. used an improved signal-to-noise ratio evaluation function to screen genes and used support vector machines (SVMs) for classification, achieving an accuracy rate of nearly 100% on the leukemia dataset. Li et al. used an improved sorting filter combined with a hill climbing algorithm for gene selection, tested it on 18 tumor gene expression datasets, and achieved a maximum accuracy rate of 94.29%. In recent years, deep learning and evolutionary optimization algorithms have also begun to be applied to gene screening tasks. For example, Yuvaraj et al. combined convolutional neural networks (CNNs) with improved whale optimization algorithms (IWOA) to build a reinforcement learning model and applied it to gene expression data of lung cancer samples, with a classification accuracy of 97.5%. Soumen et al. drew a heat map of inter-gene correlation based on the gene expression matrix, and selected the regions most relevant to cancer as regions of interest (ROIs), which were input into a specific graph neural network for training and prediction. Finally, 73 key genes were screened out of 7129 original genes for leukemia, and 89 key genes were screened out of 12533 ​​original genes for lung cancer, achieving good performance in classification accuracy. Fu et al. used an improved multi-objective marine predator optimization algorithm to test 16 commonly used tumor gene expression datasets. The results showed that this method has considerable advantages in dealing with high-dimensional data. Xiong et al. proposed an interpretable heterogeneous graph transformer model, which combines the protein-coding interaction relationship of differentially expressed genes and the KEGG regulatory pathway information. Compared with traditional screening methods, it can focus on the impact of both proteins and pathways on diseases at the same time, and is ahead of other methods in identifying key genes.

[0005] Although there are many methods for selecting key genes, most studies only focus on gene expression and ignore the impact of non-coding genes and miRNA in gene regulation. Some algorithms consider the complex relationship between genes and genes and genes and proteins, resulting in a high overall model complexity and difficulty in data acquisition. Therefore, a new solution to the above problems is needed. Summary of the invention

[0006] The purpose of the present invention is to provide a characteristic gene screening method based on NSGA-ANN and a ceRNA network construction method, aiming to use a machine learning algorithm combined with a bioinformatics statistical method to create a simple and easy gene recognition framework, and apply it to three types of data: mRNA, miRNA and lncRNA, to construct a competitive endogenous RNA (ceRNA) regulatory network, and deeply explore the key genes and prognostic indicators closely related to the development of gastric cancer, so as to provide a reference for the goals of early molecular diagnosis and targeted treatment of gastric cancer, so as to solve the technical problems raised in the background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a characteristic gene screening method based on NSGA-ANN, comprising at least the following steps:

[0008] S1: Perform data collection, obtain cancer samples and normal control samples as data, and then determine that the entire data set is composed of three types of nucleotide sequences: mRNA, miRNA, and lncRNA;

[0009] S2: Perform differential analysis on the three nucleotide data to obtain the nucleotides with significant differential expression between gastric cancer samples and normal samples, and perform WGCNA module division on the mRNA data to obtain differential miRNA, target selection module and differential IncRNA;

[0010] S3: Data preprocessing was performed to remove errors and invalid RNAs according to the annotation files and to separate the three nucleotides;

[0011] S4: Design a feature gene screening algorithm based on NSGA-ANN. The feature gene screening algorithm based on NSGA-ANN is an adaptive multi-step cycle optimization iterative method. When the number of genes selected by a single NSGA no longer changes, individuals with current fitness higher than a certain threshold are automatically selected, and their selections are combined to start the next complete NSGA iteration process. This iterative training method can ensure that the algorithm remains in the optimization process.

[0012] S5: Set the evaluation method for gene selection and determine the feasibility of the NSGA-ANN-based feature gene screening algorithm;

[0013] S6: Perform corresponding screening through the characteristic gene screening algorithm of NSGA-ANN.

[0014] Furthermore, the S2 at least includes the following steps:

[0015] For transcriptome data, negative binomial distribution was used for difference test. When the test value of the gene met the following conditions, it could be considered to be significantly different;

[0016] |log2 fold change|gene ≥1

[0017] This means that the fold change in gene expression is at least 2-fold (or reverse to 0.5-fold);

[0018] log2 represents a logarithmic transformation with base 2, so that a gene with a change factor of 1 has a change value of 1 or -1 on the logarithmic scale. This condition helps to screen out genes with biologically significant changes;

[0019] Adjust P gene <0.05

[0020] It is the p-value after multiple testing correction, which is used to control the false positive rate. A p-value lower than 0.05 indicates that the change in gene expression is statistically significant. The corrected p-value is used to deal with multiple comparison problems and reduce false positive findings. This means that in the statistical test, the p-value of the gene is less than 0.05 after multiple testing correction, indicating that the result is statistically significant;

[0021] Weighted gene co-expression network analysis, i.e., WGCNA, was used to describe the gene association patterns between different samples;

[0022] Gene groups with coordinated changes in expression often have upstream and downstream relationships. If the gene group has a strong correlation with clinical traits, it means that the core gene of the gene group may be a driver gene closely related to the disease.

[0023] Based on the idea of ​​clustering, WGCNA divides different gene modules according to the coordinated changes between genes and calculates the correlation between each gene module and clinical traits;

[0024] This method can quickly narrow down the search scope of key genes;

[0025] Since WGCNA constructs a co-expression network by calculating the correlation coefficients between genes, all genes should be taken into account, otherwise the weights of the network may change.

[0026] Furthermore, the pseudo code of the NSGA-ANN-based feature gene screening algorithm is as follows:

[0027]

[0028]

[0029] The construction method of the ceRNA network is based on the three differential nucleotides obtained in the above-mentioned NSGA-ANN-based characteristic gene screening method;

[0030] Combining the miRTarbase and miRbase databases, a regulatory network consisting of 7 mRNAs, 5 lncRNAs and 6 miRNAs was obtained. According to the competitive endogenous RNA theory, the existence of the HOTAIR / HOTTIP-has-mir-143 / has-mir-125b / has-mir559-ERBB2 regulatory pathway in gastric cancer was determined. ERBB2 is a known oncogene, and this pathway is worthy of further study for artificially controlling ERBB2 expression.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention utilizes a public gene expression database, applies a non-dominated sorting genetic algorithm and a multi-layer perceptron in combination with a biostatistical method to screen key gastric cancer-related genes, and constructs a competitive endogenous RNA regulatory network. Twelve key genes including SPP1, MMP13, ERBB2, etc. and a proto-oncogene regulatory pathway are discovered, 11 of which significantly affect the prognosis of gastric cancer patients and have the potential to be used as targets for targeted therapy or indicators for early molecular diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0034] Figure 1 It is the correlation heat map of each module of the present invention and 5 kinds of sample traits and the MM-GS scatter plot of the pink module;

[0035] Figure 2 Schematic diagram of the iterative process of the first embodiment of the present invention;

[0036] Figure 3 This is the ceRNA network of Example 2 of the present invention;

[0037] Figure 4 This is a simple schematic diagram of the process steps of Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0039] Embodiment 1:

[0040] The characteristic gene screening method based on NSGA-ANN at least includes the following steps:

[0041] S1: Perform data collection, obtain cancer samples and normal control samples as data, and then determine that the entire data set is composed of three types of nucleotide sequences: mRNA, miRNA, and lncRNA;

[0042] S2: Perform differential analysis on the three nucleotide data respectively to obtain nucleotides with significant differential expression between gastric cancer samples and normal samples. The DESeq2 package is used in the embodiment, and different settings can be used in different embodiments; and perform WGCNA division module on the mRNA data to obtain differential miRNA, target selection module and differential IncRNA;

[0043] S3: Data preprocessing was performed to remove errors and invalid RNAs according to the annotation files and to separate the three nucleotides;

[0044] S4: Design a feature gene screening algorithm based on NSGA-ANN. The feature gene screening algorithm based on NSGA-ANN is an adaptive multi-step cycle optimization iterative method. When the number of genes selected by a single NSGA no longer changes, individuals with current fitness higher than a certain threshold are automatically selected, and their selections are combined to start the next complete NSGA iteration process. This iterative training method can ensure that the algorithm remains in the optimization process.

[0045] S5: Set the evaluation method for gene selection and determine the feasibility of the NSGA-ANN-based feature gene screening algorithm;

[0046] S6: Perform corresponding screening through the characteristic gene screening algorithm of NSGA-ANN.

[0047] S2 includes at least the following steps:

[0048] For transcriptome data, negative binomial distribution was used for difference test. When the test value of the gene met the following conditions, it could be considered to be significantly different;

[0049] |log2 fold change|g ene ≥1

[0050] This means that the fold change in gene expression is at least 2-fold (or reverse to 0.5-fold);

[0051] log2 represents a logarithmic transformation with base 2, so that a gene with a change factor of 1 has a change value of 1 or -1 on the logarithmic scale. This condition helps to screen out genes with biologically significant changes;

[0052] Adjust P gene <0.05

[0053] It is the p-value after multiple testing correction, which is used to control the false positive rate. A p-value lower than 0.05 indicates that the change in gene expression is statistically significant. The corrected p-value is used to deal with multiple comparison problems and reduce false positive findings. This means that in the statistical test, the p-value of the gene is less than 0.05 after multiple testing correction, indicating that the result is statistically significant;

[0054] Weighted gene co-expression network analysis, i.e., WGCNA, was used to describe the gene association patterns between different samples;

[0055] Gene groups with coordinated changes in expression often have upstream and downstream relationships. If the gene group has a strong correlation with clinical traits, it means that the core gene of the gene group may be a driver gene closely related to the disease.

[0056] Based on the idea of ​​clustering, WGCNA divides different gene modules according to the coordinated changes between genes and calculates the correlation between each gene module and clinical traits;

[0057] This method can quickly narrow down the search scope of key genes;

[0058] Since WGCNA constructs a co-expression network by calculating the correlation coefficients between genes, all genes should be taken into account, otherwise the weights of the network may change.

[0059] The pseudo code of the feature gene screening algorithm based on NSGA-ANN is as follows:

[0060]

[0061]

[0062] The characteristic gene screening algorithm based on NSGA-ANN in this embodiment can be verified as follows:

[0063] Biological Function Analysis Methods

[0064] In order to explore the biological significance of the selected key genes, it is necessary to conduct gene enrichment analysis. GO analysis and KEGG pathway analysis are common gene enrichment analysis methods. GO (Gene Ontology) aims to investigate gene functions at three levels, namely cellular components, molecular functions, and biological processes. KEGG (Kyoto Encyclopedia of Genes and Genomes) focuses on the metabolism and signal transduction pathways in which genes are involved, which helps to understand the gene regulation mechanism in gastric cancer. In addition, in this study, the Reactome database was also used to supplement pathway analysis. This database includes all gene regulatory pathways reported in the literature, and covers a wider range than the KEGG database.

[0065] In order to determine whether a gene affects the patient's prognosis, Kaplan-Meier survival analysis can be used. KM survival analysis draws a survival curve by statistically calculating the probability of the target event occurring as the recording time increases. In this study, the target event is that the patient did not die. The KM curve will be deformed when there are too few samples or too many deleted samples, resulting in inaccurate analysis. This study will be completed directly using the Kaplan-Meierplotter online platform because it has more sample data and improves the accuracy of the analysis.

[0066] And further results and analysis of the method:

[0067] Difference Analysis and WCGNA Module Division

[0068] For the raw data of all nucleotide types, we first used the R package DESeq2 to perform differential analysis, and removed genes with no significant differences in both the cancer group and the normal group. Table 1 shows the changes in the number of various types of nucleotides before and after the differential analysis.

[0069] Table 1 Changes in the number of mRNA, miRNA and lncRNA before and after differential analysis, completed using R package

[0070]

[0071] Then, WCGNA was used to divide all differentially expressed codable genes into modules. According to the pre-test, the soft threshold was selected as 4, and the division was completed using the R package. At the same time, the correlation between each module and the clinical trait was calculated. It was observed that the pink module had the highest correlation with the clinical trait. Further investigation of the correlation between module genes and the correlation between genes and traits (MM-GS relationship) showed that the pink module showed a significant positive correlation, indicating that the core genes in the module had a strong relationship with the clinical traits. The subsequent screening of key genes will be carried out in the pink module. Figure 1 The heat map of the correlation between each module and the five sample traits and the MM-GS scatter plot of the pink module are shown in Figure 2. Figure 1 The left picture is a heat map of the correlation between each module and five sample traits, where the first trait is whether it is a cancer sample; the right picture is the MM-GS scatter plot of the pink module, with a correlation coefficient of 0.39.

[0072] NSGA-ANN algorithm screening and biological function analysis

[0073] By organizing the expression matrix of the pink module genes and combining them with clinical traits as labels, a data set consisting of 98 genes and 407 samples can be obtained. Due to the imbalance in the number of positive and negative samples, smoothing and standardization operations are required. The data balancing method refers to the Borderline-SMOTE processing used by Zhang et al. The screening algorithm is programmed in Python, in which the multilayer perceptron (ANN) is set to 1 hidden layer based on the results of pre-training, containing 30 neurons, and the activation function uses Relu. The learning rate is set to 0.01 to increase the iteration speed. The NSGA initializes the population size to 200, and each individual is a binary code with a length equal to the number of genes to be selected, and 1 represents the selection of the gene. The number of complete iterations is set to 30, and the iteration process is as follows: Figure 2 As shown, Figure 2 From left to right, the screening iterative process of mRNA, miRNA, and lncRNA is shown. The horizontal axis is the number of iterations, and the vertical axis is the number of genes.

[0074] The encodable genes obtained by the algorithm were imported into the STRING database to draw a protein interaction network, and single genes were removed. Finally, nine characteristic genes, SPP1, COL10A1, COL11A1, COL8A1, ASPN, ADAMTS2, TREM2, MMP13 and EPYC, were obtained. Biological function analysis showed that these genes were related to the generation, degradation and fixation of the extracellular matrix, which explains why gastric cancer is prone to distant metastasis. SPP1 and MMP13 have opposite effects on the prognosis of patients. Consulting relevant data also confirmed that high expression of SPP1 will reduce the survival time of patients, while MMP13 will destroy the matrix environment of cancer cells.

[0075] Embodiment 2:

[0076] See also Figure 3 , a method for constructing a ceRNA network, based on the three differential nucleotides obtained in the characteristic gene screening method based on NSGA-ANN in the above-mentioned embodiment 1;

[0077] Combining the miRTarbase and miRbase databases, a regulatory network consisting of 7 mRNAs, 5 lncRNAs and 6 miRNAs was obtained. According to the competitive endogenous RNA theory, the existence of the HOTAIR / HOTTIP-has-mir-143 / has-mir-125b / has-mir559-ERBB2 regulatory pathway in gastric cancer was determined. ERBB2 is a known oncogene, and this pathway is worthy of further study for artificially controlling ERBB2 expression.

[0078] Embodiment three:

[0079] This embodiment is based on the embodiment, and further proposes another process based on the application of the optimized NSGA2-MLP gene selection algorithm, see Figure 4 ,The screening and identification process of the entire set of gene expression data includes differential analysis, WGCNA module division, NSGA2-MLP to obtain the optimal subset, protein interaction network construction and hub node identification. Enrichment analysis and KM survival analysis are used for the validation process, and the determination of ceRNA pathways relies on Spearman correlation test;

[0080] The NSGA2-MLP gene selection algorithm is as follows:

[0081] NSGA-2 is an improved genetic algorithm that differs from the original genetic algorithm only in the selection of operators. NSGA-2 stratifies individual solutions according to their fitness, and individuals with smaller fitness will dominate other individuals. Individuals with higher dominance relationships will be preferentially inherited into the next generation. In addition, in order to solve the redundancy problem, the crowding degree of the selected individuals is calculated, and individuals with lower crowding density are selected to ensure that the final solution is close to the Pareto optimal solution.

[0082] Here, assuming that the score of solution x1 is greater than the score of x2, and the number of genes selected by x1 is less than or equal to the number of genes selected by x2, then solution x1 is considered to dominate solution x2, expressed as x1≤x2. Function 1 describes how NSGA generates dominance relationships between individuals. Specifically, the function first traverses all solutions in the population and calculates the dominance number n for each solution. p and the dominated set S p . Then, for each solution with a dominance number of 0, traverse each member individual in its dominance set. If the dominance number of a member individual is equal to 1, it means that the member belongs to a higher level of non-dominated level, and so on until all solutions are sorted. P is a population composed of multiple binary coded individuals, where the length of the individual represents the size of the candidate gene set, 0 means that the corresponding gene is not selected, and 1 means that the gene is selected. The specific steps are as follows:

[0083]

[0084] The next step is to calculate the crowding degree between individuals belonging to the same dominance level. The smaller the crowding degree, that is, the individuals with a larger crowding distance, will be given priority in the next generation. The crowding distance D i It is defined as the difference in score between the two solutions adjacent to the ith solution. m represents the number of objective functions, which is equal to 2 here, i.e., the solution set size and fitness. and They represent the maximum and minimum values ​​of the mth objective function respectively. Function 2 describes the process of calculating the crowding distance.

[0085]

[0086] Since the gene dimension is large, a single optimization cannot obtain the optimal subset, so we propose a step-by-step selection iterative optimization process based on the Boosting idea. After the single optimization of NSGA-2 is completed, the set of individuals with the highest score is selected, and the genes selected by most individuals are added to the target gene set for the next optimization, and the NSGA-2 optimization process is restarted. Considering that the highest score will decrease as the number of genes decreases, we set a dynamic threshold to ensure that individuals with higher scores can be selected in each iteration, and the threshold will decrease as the number of iterations increases. Algorithm 1 defines the overall process of the NSGA2-MLP screening algorithm.

[0087]

[0088]

[0089] Based on this embodiment, an evaluation method for gene selection is proposed

[0090] In order to evaluate the stability of the gene screening results, we repeated the screening 10 times and used two methods to evaluate the similarity of the obtained characteristic gene sets. The first is the Jaccard similarity distance, which can be used to describe the similarity between two sets. The larger the Jaccard distance value, the lower the set similarity. It is defined as follows:

[0091]

[0092] ABJ(A,B) where and represent two sets, and when the set is empty, it is defined as 0. Another method is weight superposition, which comes from Cattelani et al. The genes selected in each solution are sorted according to the Pareto frontier and assigned weights ranging from 0 to 1. For any two groups of solution individuals, the sum of the minimum weights of the corresponding genes is calculated, and finally the average weight sum of all solution combinations is calculated. The calculation formula is as follows.

[0093]

[0094] ijw i,k iknn = 98 Here, and represent the numbers of two different solutions, represents the weight of the gene in the solution set, and when a gene is not selected, the weight is 0. is the number of genes, and in actual calculation. The average weight and value range are obtained. The higher the value, the higher the similarity of the set.

[0095] In order to verify the accuracy of the selected feature genes, support vector machines (SVM) were used to train and test on three GEO external data sets, and common classification evaluation indicators including accuracy, precision, recall and F1 index were used to observe the classification effect. The specific calculation formulas of the four indicators are as follows:

[0096]

[0097]

[0098] In addition, in order to test the generalization performance of the NSGA2-MLP screening algorithm, it was applied to the gene expression data of another 30 cancer types. The SVM was also used to classify the characteristic genes and compare the results with those of the random forest and Lasso regression methods.

[0099] Protein-protein interaction and enrichment analysis

[0100] For the obtained gastric cancer characteristic genes, although the expression data are related to the phenotype, the expression proteins of some genes may not be related to other genes, which means that the interaction between the gene and other genes is not strong and does not belong to the hub gene. The relationship between the expressed proteins is obtained using the STRING database. The protein interaction relationship between the characteristic genes will be represented in the form of a graph. The nodes are clustered and grouped, and the criticality of the characteristic genes is evaluated according to the degree index of the nodes. The nodes with a degree of 0 and the node groups with too few numbers are removed to obtain the appropriate hub gene group. There is a method to use the relationship between all genes in the STRING database as the original input. This multi-scale information may help the model learn more patterns between genes and phenotypes, but it puts higher requirements on storage space and model architecture.

[0101] Gene Ontology (GO) analysis is a commonly used gene enrichment analysis method, which focuses on the cellular components, molecular functions and biological processes involved in genes. Another enrichment method is KEGG pathway enrichment, which focuses on understanding the signal transduction pathways and metabolic pathways involved in genes. Through enrichment analysis, we can understand the main biological functions of the target gene set and speculate on its role in tumor development. We used the R software package to complete the enrichment analysis of the target gene set as the cutoff standard.

[0102] Construction of competing endogenous RNA networks

[0103] The competitive endogenous RNA (ceRNA) regulatory mechanism represents a new gene expression regulation mode. Compared with the miRNA regulatory network, the ceRNA regulatory network can be more refined because it contains more relationships between RNA molecules. ceRNA believes that miRNA can not only bind to the response unit MRE on mRNA to inhibit mRNA translation, but also bind to the MRE on long non-coding RNA (lncRNA). When lncRNA expression increases, it can attract more miRNAs to bind, which in turn relieves the inhibition of miRNA on mRNA translation, so that the gene expression level increases. Therefore, in a typical ceRNA regulatory pathway, mRNA is positively correlated with lncRNA in expression, but negatively correlated with miRNA. In order to further explore the ceRNA regulatory mechanism in gastric cancer, we screened mRNA, miRNA and lncRNA in gastric cancer data at the same time, combined with the miRcode database and multiMiR software package, obtained the mRNA-miRNA and miRNA-lncRNA relationship data, and constructed the ceRNA network after taking the intersection with the differentially expressed genes, and used the Spearman correlation coefficient to verify and confirm the most likely regulatory pathway.

[0104] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A characteristic gene screening method based on NSGA-ANN, characterized by: At least the following steps are included: S1: Perform data collection, obtain cancer samples and normal control samples as data, and then determine that the entire data set is composed of three types of nucleotide sequences: mRNA, miRNA, and lncRNA; S2: Perform differential analysis on the three nucleotide data to obtain the nucleotides with significant differential expression between gastric cancer samples and normal samples, and perform WGCNA module division on the mRNA data to obtain differential miRNA, target selection module and differential IncRNA; S3: Data preprocessing was performed to remove errors and invalid RNAs according to the annotation files and to separate the three nucleotides; S4: Design a feature gene screening algorithm based on NSGA-ANN. The feature gene screening algorithm based on NSGA-ANN is an adaptive multi-step cycle optimization iterative method. When the number of genes selected by a single NSGA no longer changes, individuals with current fitness higher than a certain threshold are automatically selected, and their selections are combined to start the next complete NSGA iteration process. This iterative training method can ensure that the algorithm remains in the optimization process. S5: Set the evaluation method for gene selection and determine the feasibility of the NSGA-ANN-based feature gene screening algorithm; S6: Perform corresponding screening through the characteristic gene screening algorithm of NSGA-ANN.

2. The characteristic gene screening method based on NSGA-ANN according to claim 1, characterized in that: The S3 at least includes the following steps: For transcriptome data, negative binomial distribution was used for difference test. When the test value of the gene met the following conditions, it could be considered to be significantly different; |log2 fold change| gene ≥1 Indicates that the fold change of gene expression is at least 2-fold; log2 represents a logarithmic transformation with base 2, so that a gene with a change factor of 1 has a change value of 1 or -1 on the logarithmic scale. This condition helps to screen out genes with biologically significant changes; Adjust P gene <0.05 It is the p-value after multiple testing correction, which is used to control the false positive rate. A p-value lower than 0.05 indicates that the change in gene expression is statistically significant. The corrected p-value is used to deal with multiple comparison problems and reduce false positive findings. This means that in the statistical test, the p-value of the gene is less than 0.05 after multiple testing correction, indicating that the result is statistically significant.

3. The method for screening characteristic genes based on NSGA-ANN and the method for constructing ceRNA network according to claim 1, characterized in that: The pseudo code of the NSGA-ANN-based feature gene screening algorithm is as follows:

4. A method for constructing a ceRNA network, characterized in that: Three differential nucleotides obtained in the characteristic gene screening method based on NSGA-ANN according to claim 1 above; Combining the miRTarbase and miRbase databases, a regulatory network consisting of 7 mRNAs, 5 lncRNAs and 6 miRNAs was obtained.

Images