Gastric cancer gene mutation map construction method based on machine learning

Abstract

The invention relates to the technical field of gastric cancer gene mutation map construction methods, and particularly discloses a gastric cancer gene mutation map construction method based on machine learning, which comprises the following steps: step S1, collecting a public data set TCGA (Thecancerous Genomeatlas) of various sequencing data of various human cancers; s2, selecting a TCGA-STAD item, and excluding gene data of residual adenoma and adenocarcinoma patients except cystic, mucus and serous tumors; s3, performing a whole transcriptome sequencing (RNA-Seq) experiment to obtain sequencing data of gene expression quantification, and further obtaining clinical data of the patient; a step; according to the method, gastric cancer gene expression data is deeply analyzed and researched, firstly, a data set is explored and screened, then key gene mutation points are determined through statistical data analysis methods such as difference analysis and enrichment analysis, and then classification of gastric cancer subtypes and construction of a gastric cancer gene mutation map are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of methods for constructing a gastric cancer gene mutation map, and in particular to a method for constructing a gastric cancer gene mutation map based on machine learning. Background Art

[0002] Gastric cancer is one of the malignant tumors with extremely high morbidity and mortality rates worldwide. Its complex molecular characteristics and multi-stage development process make its early diagnosis and precise treatment extremely challenging. According to the Global Cancer Statistics (GLOBOCAN), new cases of gastric cancer each year account for approximately 5.6% of all malignant tumors, and it ranks among the top five causes of cancer-related deaths. Despite the continuous advancement of surgery, radiotherapy, chemotherapy, and targeted therapy techniques in recent years, the overall five-year survival rate of gastric cancer patients is still low, especially in advanced cases, where the treatment effect is often unsatisfactory. Despite the continuous advancement of gastric cancer treatment methods, early diagnosis and effective treatment strategies remain the key to improving patient survival rates;

[0003] The occurrence and development of gastric cancer is a complex, multi-stage process involving mutations and expression changes of multiple genes. In recent years, many gene mutations associated with gastric cancer have been discovered, such as TP53, KRAS, EGFR, etc. These gene mutations may affect cell proliferation, apoptosis, differentiation and other processes, leading to the occurrence and development of gastric cancer. However, the gene mutation map of gastric cancer is still not completely clear and further research is needed.

[0004] However, how to find the gene mutation that is most closely related to gastric cancer from a large amount of gene mutation data, and how to use this data to build a model that can predict gastric cancer gene mutations, are still major defects in existing technologies. Summary of the invention

[0005] In view of the technical problem that there is a lack of a model capable of predicting gastric cancer gene mutations in the prior art, the present invention provides a method for constructing a gastric cancer gene mutation map based on machine learning.

[0006] The technical solution adopted by the present invention is: a method for constructing a gastric cancer gene mutation map based on machine learning, which specifically includes the following steps:

[0007] Step S1, collecting the public dataset TCGA (The cancer genome atlas) of various sequencing data of various human cancers;

[0008] Step S2, select the genetic data of the remaining adenoma and adenocarcinoma patients in the TCGA-STAD project, excluding cystic, mucinous and serous tumors;

[0009] Step S3, then through the whole transcriptome sequencing (RNA-Seq) experiment, the sequencing data of gene expression quantification is obtained, and the clinical data of the patient is also obtained;

[0010] Step S4, based on the Count sequencing data, use R language to clean and integrate the data, remove low-expression data, remove duplicate genes, retain the results of the maximum expression of each gene, use the gene name as the row name and the sample name as the column name to obtain the gene expression matrix;

[0011] Step S5, performing variance stabilizing transformation (VST) on the count sequencing data, so that the variance of the count sequencing data is independent of the average expression level of the gene;

[0012] Step S6, perform PCA analysis and visualization on the grouped gene data, and identify group aggregation, the presence of outliers, and potential data structures by observing the distribution of samples in the principal component space;

[0013] Step S7, select the unsupervised clustering algorithm KMeans to perform layer-by-layer classification, the specific method is as follows: first randomly select K sample data points as K cluster centers of the initial clustering, calculate the distance between each data point and the K cluster centers, and assign it to the cluster center with the closest distance, each time after a sample is assigned, recalculate the cluster center to adapt to the existing cluster objects, and repeat until the preset stop condition is met;

[0014] Step S8, selecting the gene expression levels of CDKN2A, MLH1, and PIK3CA as classification criteria for k-means clustering;

[0015] Step S9, split all sample data into training set and test set at a ratio of 4:1, use cross entropy as the splitting criterion, and use the training set to build a random forest model;

[0016] Step S10, split all sample data into training set and test set at a ratio of 2:1, and use the training set to build a parameter-free XGBoost model;

[0017] Step S11, combining the results of difference analysis, random forest and XGBoost to obtain a gene mutation map that distinguishes the normal gastric cancer group from the tumor group.

[0018] The present invention is further configured such that the clinical data of the patient includes the patient's age, gender, race, analysis, and prognosis data.

[0019] The present invention is further configured such that, in step S5, the steps of the VST standardization model are as follows:

[0020] First, for each gene, calculate the mean and variance in each sample

[0021] E(Y)=u=α+m η μ (2-1)

[0022] Var(Y)=v=s η 2 μ 2 +σ ε 2 (2-2)

[0023] The functional relationship between the estimated mean and variance is as follows:

[0024] v(u)=(s η / m η ) 2 (u-α) 2 +σ ε 2 = (c 1 u+c 2 ) 2 +c 3 (2-3)

[0025] Among them, v represents variance, u represents mean, and v(u) represents the variance v as a function of mean u. Then according to mathematical derivation, we can get the h(y) function as:

[0026]

[0027] Convert the matrix Y before normalization to the normalized matrix Make an identity transformation of formula (2-3):

[0028]

[0029] A linear model was fitted using the variance and mean of each gene to obtain c 1 , c 2 and c 3；

[0030] c 1 , c 2 and c 3 Substitute it back into formula (2-4) and solve the integral to get:

[0031]

[0032] That is, the VST standardized conversion formula h(Y) is obtained.

[0033] The present invention is further configured to also include, during the transfer learning process, using the MAE model for pre-training on the ImageNet-1K dataset, with the masked image block ratio set to 75% and the training rounds being 800 times; then, fine-tuning the pre-trained model on the snake image dataset, setting the learning rate to 0.001 and the batch size to 16 to optimize the classification performance.

[0034] The present invention is further configured such that the output form of the classification result includes: the name of the snake species, the regional distribution probability of the snake, and the confidence score of the model prediction.

[0035] The beneficial effects of the present invention are as follows: In the present invention, gastric cancer gene data collected in the Cancer Genome Atlas (TCGA) project are used, and multiple machine learning models, including KMeans, random forests, and XGBoost, are used to conduct in-depth analysis of gastric cancer gene expression data. The study first explores and screens the data set, and then uses statistical data analysis methods such as differential analysis and enrichment analysis to determine key gene mutation points, thereby achieving the classification of gastric cancer subtypes and the construction of gastric cancer gene mutation maps. The research results reveal the differences in gene expression and mutation patterns between gastric cancer subtypes, providing new strategies for the molecular classification and precision treatment of gastric cancer. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a sample barcode display diagram of the present invention;

[0037] Figure 2 is a principal component analysis diagram of the gastric adenocarcinoma grouping data in the present invention;

[0038] Figure 3 This is a volcano plot of DESeq2 differential analysis of gastric adenocarcinoma gene expression data using R language in the present invention;

[0039] Figure 4 This is the difference analysis-MA diagram in the present invention;

[0040] Figure 5 This is a diagram of the difference analysis results in the present invention;

[0041] Figure 6 This is the result diagram of GO enrichment analysis in the present invention;

[0042] Figure 7 This is the KEGG enrichment analysis result diagram in the present invention;

[0043] Figure 8 This is the K-means clustering result diagram of EBV subtypes in the present invention;

[0044] Fig. 9 This is the K-means clustering result diagram of MSI subtypes in the present invention;

[0045] Fig.10 It is the main component scree map in the present invention;

[0046] Fig.11 This is the decision tree diagram for gastric cancer subtype classification in the present invention

[0047] Fig.12 The important gene map selected by random forest in the present invention;

[0048] Fig.13 It is the important gene map selected by XGBoost in the present invention;

[0049] Fig.14 is a gastric cancer gene mutation spectrum diagram in the present invention;

[0050] Fig.15 It is the optimal range diagram of random forest parameters in the present invention;

[0051] Fig.16 The model score diagram corresponding to different numbers of base classifiers in the present invention, (a) the number of coarse-grained base classifiers; (b) the number of fine-grained base classifiers;

[0052] Fig.17 The gene map of the top 20 genes in terms of feature importance after optimization in the present invention;

[0053] Fig.18 This is a gene mutation spectrum of gastric cancer subtypes in the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] In order to solve the problems existing in the background technology, this application proposes the following technical solutions:

[0056] A method for constructing a gastric cancer gene mutation map based on machine learning, specifically comprising the following steps:

[0057] Step S1, collecting the public dataset TCGA (The cancer genome atlas) of various sequencing data of various human cancers;

[0058] Step S2, select the genetic data of the remaining adenoma and adenocarcinoma patients in the TCGA-STAD project, excluding cystic, mucinous and serous tumors;

[0059] Step S3, then through the whole transcriptome sequencing (RNA-Seq) experiment, the sequencing data of gene expression quantification is obtained, and the clinical data of the patient is also obtained;

[0060] Step S4, based on the Count sequencing data, use R language to clean and integrate the data, remove low-expression data, remove duplicate genes, retain the results of the maximum expression of each gene, use the gene name as the row name and the sample name as the column name to obtain the gene expression matrix;

[0061] Step S5, performing variance stabilizing transformation (VST) on the count sequencing data, so that the variance of the count sequencing data is independent of the average expression level of the gene;

[0062] Step S6, perform PCA analysis and visualization on the grouped gene data, and identify group aggregation, the presence of outliers, and potential data structures by observing the distribution of samples in the principal component space;

[0063] Step S7, select the unsupervised clustering algorithm KMeans to perform layer-by-layer classification. The specific method is as follows: first randomly select K sample data points as the K class centers of the initial clustering, calculate the distance between each data point and the K cluster centers, and assign it to the cluster center with the closest distance. After each sample is assigned, recalculate the cluster center to adapt to the existing clustering objects. This process is iterated continuously until the preset stop condition is met (such as the change in the class center is less than a certain threshold or the maximum number of iterations is reached). Through layer-by-layer optimization, this method can gradually converge to a stable clustering result and effectively reveal the distribution characteristics of the data;

[0064] Step S8, selecting the gene expression levels of CDKN2A, MLH1, and PIK3CA as classification criteria for k-means clustering;

[0065] Step S9, split all sample data into training set and test set at a ratio of 4:1, use cross entropy as the splitting criterion, and use the training set to build a random forest model;

[0066] Step S10, split all sample data into training set and test set at a ratio of 2:1, and use the training set to build a parameter-free XGBoost model;

[0067] Step S11, combining the results of difference analysis, random forest and XGBoost to obtain a gene mutation map that distinguishes the normal gastric cancer group from the tumor group.

[0068] It also includes the use of the MAE model for pre-training on the ImageNet-1K dataset during the transfer learning process, with the masked image block ratio set to 75% and the training rounds to 800 times; then, the pre-trained model is fine-tuned on the snake image dataset, with the learning rate set to 0.001 and the batch size set to 16 to optimize the classification performance.

[0069] The output forms of classification results include: snake species name, regional distribution probability of snakes, and confidence score of model prediction.

[0070] The patient's clinical data includes the patient's age, gender, race, analysis, and prognosis data.

[0071] In step S5, the steps of the VST standardization model are as follows:

[0072] First, for each gene, calculate the mean and variance in each sample

[0073] E(Y)=u=α+m η μ (2-1)

[0074] Var(Y)=v=s η 2 μ 2 +σ ε 2 (2-2)

[0075] The functional relationship between the estimated mean and variance is as follows:

[0076] v(u)=(s η / m η ) 2 (u-α) 2 +σ ε 2 = (c 1 u+c 2 ) 2 +c 3 (2-3)

[0077] Among them, v represents variance, u represents mean, and v(u) represents the variance v as a function of mean u. Then according to mathematical derivation, we can get the h(y) function as:

[0078]

[0079] Convert the matrix Y before normalization to the normalized matrix Make an identity transformation of formula (2-3):

[0080]

[0081] A linear model was fitted using the variance and mean of each gene to obtain c 1 , c 2 and c 3；

[0082] c 1 , c 2 and c 3 Substitute it back into formula (2-4) and solve the integral to get:

[0083]

[0084] That is, the VST standardized conversion formula h(Y) is obtained.

[0085] The above technical solution is further explained as follows:

[0086] Gastric cancer genomes and proteomes are widely used in molecular typing studies in the hope of finding specifically expressed pathways and potential therapeutic targets;

[0087] A comprehensive analysis of genomic and proteomic data for gastric cancer revealed four molecular subtypes:

[0088] EBV: Epstein-Barr virus subtype with extreme DNA hypermethylation.

[0089] MSI: microsatellite unstable subtype with extremely high mutation rate and high methylation.

[0090] GS: Genome-stable subtype with less obvious chromosomal structural changes.

[0091] CIN: A chromosomally unstable subtype that exhibits high chromosomal instability and increased expression of receptor tyrosine kinases (RTKs), including VEGFR2.

[0092] The subsequent construction of the gene map and prediction model of the present invention needs to be based on this molecular typing. Therefore, selecting a suitable classification model to accurately divide the gastric cancer data into these four subtypes is the first step of this project.

[0093] The classification method is as follows:

[0094] Because the TCGA gastric cancer dataset is an unlabeled dataset, the present invention selects the unsupervised clustering algorithm KMeans to perform layer-by-layer classification.

[0095] The core idea of ​​the K-Means algorithm is: first randomly select K sample data points as the K cluster centers of the initial clustering, then calculate the distance between each data point and the K cluster centers, and assign it to the cluster center closest to it. After each sample is assigned, the cluster center is recalculated to adapt to the existing clustering objects, and this process is repeated until the preset stopping condition is met.

[0096] The selection of K in KMeans can be assisted by the silhouette coefficient, which combines the two factors of cohesion and separation. The closer the silhouette coefficient is to 1, the more reasonable the sample clustering is. The present invention uses the method of making a silhouette analysis graph to assist in judging the selection of K.

[0097] The idea of ​​this project for gastric cancer gene expression data typing is to use the gene expression matrix and the grouping matrix to extract the tumor groups, first classified by EBV positivity, then classified by MSI high status, and the remaining tumors were divided into genomically stable (GS) or showing chromosomal instability (CIN) according to the degree of aneuploidy.

[0098] Acquisition of EBV subtypes: All EBV-positive tumors tested showed CDKN2A (p16INK4A) promoter hypermethylation but lacked the MLH1 hypermethylation characteristic of MSI-associated CIMP, and a strong tendency for PIK3CA mutations in EBV-positive gastric cancer was also observed.

[0099] In addition, studies on EBV infection have shown that hypermethylation of tumor suppressor genes is significantly more frequent in EBV-positive gastric cancer than in EBV-negative gastric cancer. Therefore, the present invention selects the gene expression levels of CDKN2A, MLH1, and PIK3CA as classification criteria for k-means clustering.

[0100] First, read the gene expression matrix and grouping matrix, and obtain 379 tumor group data through innerjoin for cluster analysis of gene expression. Before clustering, first output extreme values, means, variances and other information to determine whether standardization is required. The specific statistical information is shown in Table 1 below;

[0101] Table 1 Statistical information of EBV classification-related genes

[0102] CDKN2A MLH1 PIK3CA min 6 68 304 max 37338 5297 19933 mean 1993.477573 1645.712401 2259.401055 std 3900.386893 895.019969 1482.086384

[0103] In summary, it is obvious that there are obvious differences in the order of magnitude of the expression levels of these three genes. If the distance is directly calculated for clustering, the results will be seriously inaccurate. Therefore, it is necessary to standardize the data before using the KMeans algorithm.

[0104] After standardization, set k = 3, the clustering effect is as follows Figure 1As shown, the blue markers are EBV subtypes, which are particularly expressed in CDKN2A and have low expression levels of MLH1. A total of 23 groups of data were divided into EBV subtypes, accounting for 6.07% of tumor samples.

[0105] Acquisition of MSI subtypes: MSI subtypes have obvious MSI enrichment and MLH1 promoter hypermethylation characteristics, so the present invention selects the gene expression levels of MSI1, MSI3, and MLH1 as classification features for k-means clustering.

[0106] According to the guidance of the TCGA gastric cancer molecular subtype classification decision tree, before clustering, the 23 data sets with EBV label 1 should be filtered out, and then the remaining 356 data sets should be classified into MSI and non-MSI groups.

[0107] After standardization, set k = 2, the clustering effect is as follows Figure 2 As shown in the figure, the black dots are MSI subtypes, which have a prominent expression in MLH1 and a high expression in MSI2. A total of 102 groups of data were divided into MSI subtypes, accounting for 26.91%.

[0108] Obtaining GS and CIN subtypes: The remaining two groups are distinguished by the presence or absence of widespread somatic copy number aberrations (SCNA). Based on genomic data, TCGA uses iCluster to integrate clustering of multiple data types as an alternative method to define different gastric cancer subtypes.

[0109] Referring to the iCluster method used by TCGA, the present invention reads the gene expression matrix and the grouping matrix, filters out the data of EBV type and MSI type, and obtains the standardized gene expression matrix GEB (254*59427) of the remaining 254 tumor groups. However, if the expression of 59427 genes is directly used as a feature for clustering,

[0110] The computational cost is high and the operation time is slow. Therefore, the present invention first performs principal component analysis on nearly 60,000 genes, and the scree plot of the first 80 principal components is as follows Figure 3 shown.

[0111] like Figure 3 , observe the scree plot, the inflection point is at 13, so we choose n_components to be 13 for dimensionality reduction, and get the principal component matrix A of 13*59427. Next, GEB·A T (254*13) was used as the data set for binary classification, and finally 76 samples were GS type, accounting for 20.05% of the tumor group; 178 samples were CIN type, accounting for 46.96% of the tumor group.

[0112] In summary, the present invention uses TCGA gastric cancer sequencing data, according to Figure 4The decision tree was clustered layer by layer using KMeans, and the 379 tumor samples were divided into four subtypes: EBV, MSI, GS, and CIN.

[0113] Some subtype classification results are shown below:

[0114] Table 2 Classification results of gastric cancer molecular subtypes

[0115]

[0116] In summary, based on the unsupervised clustering KMeans algorithm and PCA dimensionality reduction, the gastric cancer dataset was divided into four molecular subtypes. For the classification of each subtype, different gene expression features were used:

[0117] (1) EBV subtype: based on the expression levels of CDKN2A, MLH1 and PIK3CA.

[0118] (2) MSI subtype: based on the expression levels of MSI1, MSI3, and MLH1.

[0119] (3) GS and CIN subtypes: Data dimensionality reduction was performed through somatic copy number aberration and multidimensional data integration cluster analysis combined with principal component analysis.

[0120] Finally, EBV-positive subtypes accounting for 6.07% of tumor samples, MSI subtype samples accounting for 26.91%, GS subtype samples accounting for 20.05% and CIN subtype samples accounting for 46.96% were successfully classified.

[0121] The construction of the gastric cancer gene mutation map is explained as follows:

[0122] First, using DESeq2 differential analysis, 389 significantly differentially expressed genes were identified; GO and KEGG analysis results highlighted the importance of specific biological processes and pathways (such as apical plasma membrane, skin development, digestive ability, and keratinocyte differentiation) in the pathogenesis of gastric adenocarcinoma.

[0123] Secondly, KMeans clustering was used to divide the tumor group data into four subtypes.

[0124] Finally, for each subtype, an integrated learning model is used to identify abnormally expressed genes and construct a subtype-specific gene mutation map;

[0125] Based on the random forest method:

[0126] First, all sample data are split into training set and test set at a ratio of 4:1. The training set has 330 groups of samples, including 28 normal samples and 302 tumor samples; the test set has 83 groups of samples, including 6 normal samples and 77 tumor samples. Then, the number of base classifiers is set to 100, and cross entropy is used as the splitting criterion. A random forest model is constructed using the training set. The specific results are shown in Table 3.

[0127] Table 3 Random Forest Binary Classification Effect

[0128] Accuracy Recall F1-Score support normal 1.00 0.67 0.80 6 tumor 0.97 1.00 0.99 77 Accuracy 0.97 83 Macroavg 0.99 0.83 0.89 83 Weightedavg 0.98 0.98 0.97 83

[0129] The mean square error of the random forest classification model is 0.0241, and the accuracy rate is 97.59%, which shows that the prediction effect is very good, and the output is

[0130] Based on the XGBoost method: First, all sample data are split into training set and test set at a ratio of 2:1. The training set has 276 groups of samples, 21 normal samples, and 255 tumor samples; the test set has 137 groups of samples, 13 normal samples, and 124 tumor samples. Then, a parameter-free XGBoost model is constructed using the training set. The specific results are shown in Table 4.

[0131] Table 4 Binary classification XGBoost effect

[0132]

[0133] It can be seen that compared with the random forest model, the XGBoost model has a higher accuracy rate, reaching 98%. Therefore, it is reasonable to use the XGBoost model to fit the sample data and realize the classification of gastric cancer subtypes. In the parameter-free XGBoost model, the 13 gene features ESM1, AC016727.3, AJ009632.2, RPSAP52, STRA6, HOXC-AS2, AC068643.1, HOTAIR, CST1, PACC1, WNT7B, GPX3, and CLEC3B are ranked high in importance, which plays a key role in using the XGBoost model to determine whether the sample is a tumor group. The feature importance bar chart is as follows Fig.13 Shown

[0134] Gastric cancer gene mutation map: Based on the results of differential analysis, random forest and XGBoost, this paper obtained the gene mutation map that distinguishes the normal group and tumor group of gastric cancer, with a total of 930 characteristic genes, such as Fig.14As shown in the figure, the horizontal axis represents the sequence number of the 930 genes, the vertical axis represents the log2FoldChange value, the size of the point represents the feature importance of the machine learning classification model, and the color of the point represents the p value.

[0135] Random forest method and optimization: First, the 379 tumor sample data were split into a training set and a test set at a ratio of 4:1. The training set had 303 samples; the test set had 76 samples, with EBV:MSI:GS:CIN=6:31:27:62. Then, a parameter-free random forest model was constructed using the training set. The specific results are shown in Table 5.

[0136] Table 5. Parameter-free random forest subtype classification effect

[0137] Accuracy Recall F1-Score support EBV 1.00 1.00 1.00 3 MSI 0.70 0.74 0.72 20 GS 0.85 0.81 0.83 17 CIN 0.79 0.77 0.78 36 Accuracy 0.76 76 Macroavg 0.59 0.59 0.59 76 Weightedavg 0.74 0.76 0.75 76

[0138] As shown in Table 5, the parameter-free random forest model performs poorly in gastric cancer subtype classification, with an accuracy of only 76.32% and a mean square error of 0.9211. Therefore, the random forest model needs to be further optimized. First, the parameters of the random forest model are explained as follows:

[0139] Table 6 Random forest classifier parameter description

[0140] Parameter name significance default value n_estimators The number of trees in the forest 100 criterion Criteria for optimal feature selection gini max_depth Maximum depth of the tree None min_samples_split Minimum number of samples required to split an internal node 2 min_samples_leaf Minimum number of samples required for a leaf node 1 max_features The number of features considered when finding the best split auto

[0141] Use cross-validation to gradually adjust parameters. First, determine the optimal range of values ​​for the number of decision trees in the forest (n_estimators parameter) and the maximum depth of the tree (max_depth parameter).

[0142] Depend on Fig.14 It can be seen that the optimal number of decision trees is 100 or 250, and the optimal range of the maximum depth of the tree is 5, 9, or 11. Because tuning only one parameter may fall into a local optimal solution, it is necessary to combine multiple parameters for further tuning. The tuning process is shown in Table 7 below.

[0143] Table 7 Joint tuning of random forest parameters

[0144] Number of decision trees n_estimators Maximum tree depth max_depth Score 100 3 0.7237 100 5 0.7895 100 9 0.7763 100 11 0.7763 250 3 0.7237 250 5 0.7632 250 9 0.8026 250 11 0.7763

[0145] However, the adjustment of criterion, min_samples_split, min_samples_leaf and max_features did not have a significant effect. Therefore, the parameters of random forest were finally selected as follows: the number of decision trees in the forest was 250, the maximum depth of the tree was 9, and the classification standard criterion was the Gini coefficient. The final model classification accuracy was 80.26%, and 6978 characteristic genes that played a key role in random forest classification were obtained.

[0146] XGBoost method and optimization: First, the 379 tumor sample data were split into training set and test set at a ratio of 2:1. The training set had 276 samples; the test set had 126 samples, with EBV:MSI:GS:CIN=6:31:27:62. Then, a parameter-free XGBoost model was constructed using the training set. The specific results are shown in Table 8.

[0147] Table 8 Parameter-free multi-classification XGBoost effect

[0148] Accuracy Recall F1-Score support EBV 1.00 1.00 1.00 6 MSI 0.70 0.74 0.72 31 GS 0.85 0.81 0.83 27 CIN 0.79 0.77 0.78 62 Accuracy 0.79 126 Macroavg 0.83 0.83 0.83 126 Weightedavg 0.79 0.79 0.79 126

[0149] As can be seen from Table 8, the parameter-free XGBoost model performs poorly for gastric cancer subtype classification, with an accuracy rate of only 79%, so it is necessary to fine-tune the parameters of the XGBoost model. First, the parameters of XGBoost are explained as follows:

[0150] Table 9 XGBoost parameter description

[0151] Parameter name significance n_estimators Number of Gradient Boosted Trees max_depth Maximum depth of the base classifier tree max_leaves Maximum number of leaf nodes min_child_weight Minimum weight of leaf node learning_rate The step size of each iteration subsample The proportion of subsamples used to train the model to the entire sample set colsample_bytree The ratio of features sampled when building a tree reg_alpha Regularization parameter gamma Loss threshold, the minimum loss reduction required for a leaf node to further split objective Specify binary classification / multi-classification / regression

[0152] Next, we use cross-validation to gradually adjust the parameters. First, we adjust the number of base classifiers, that is, the optimal parameter range of the n_estimators parameter. The figure shows the model scores corresponding to the number of base classifiers with a step size of 10 and coarse-grained changes, that is, n_estimators = 10k + 1, k = 0, 1, ..., 9. It can be seen that when the number of base classifiers is about 11, the model effect is best. Therefore, the number of classifiers is limited to 1 to 20, and the number of base classifiers with a step size of 1 and fine-grained is tested, that is, n_estimators = k, k = 1, 2, ..., 20. It can be seen that when the number of base classifiers is 9, the XGBoost model works best.

[0153] Next, 5-fold cross validation is used to gradually adjust the remaining parameters. The parameter selection range, specific adjustment steps, and final adjustment results are shown in the following table:

[0154] Table 10 Multi-classification XGBoost parameter adjustment results

[0155] Parameter name Select range Best value Score n_estimators (1,90,1) 9 0.81746 max_depth (2,10,1) 3 0.84126 min_child_weight (1,5,1) 1 0.84126 gamma (0,0.5,0.1) 0.1 0.84126 reg_alpha [1e-5,0.01,0.1,1,100] 0.01 0.84126 subsample (0.6,0.9,0.05) 0.8 0.83333 colsample_bytree (0.6,0.9,0.05) 0.75 0.83333 learning_rate [0.01,0.05,0.1,0.2] 0.1 0.84126

[0156] The optimized results are shown in Table 10 and Figure 16-17 shown.

[0157] Table 11 Multi-classification XGBoost effect after optimization

[0158] Accuracy Recall F1-Score support EBV 1.00 1.00 1.00 6 MSI 0.82 0.74 0.78 31 GS 0.92 0.81 0.86 27 CIN 0.81 0.89 0.85 62 Accuracy 0.84 126 Macroavg 0.89 0.86 0.87 126 Weightedavg 0.84 0.84 0.84 126

[0159] Figure 16-17 The top 20 genes in terms of feature importance in the XGBoost multi-classification model are shown. Among them, MLH1, CDKN2A, RCOR3, MSI2, WDR33, TLK2, RICTOR, POGZ, GARS1-DT, and C12orf65 performed particularly well. In fact, the tuned XGBoost model screened out 143 characteristic genes as characteristic genes for distinguishing gastric cancer subtypes. The model accuracy reached 84.13%, so the gastric cancer gene mutation map for subtypes can be improved based on the tuned XGBoost model.

[0160] Gastric cancer gene mutation maps for different subtypes: Combining the results of random forest and XGBoost, the present invention obtains gene mutation maps for differentiating four subtypes of gastric cancer, with a total of 17,242 characteristic genes, such as Fig.18 As shown in the figure, the horizontal axis represents the serial number of 17242 genes, the vertical axis is the log2FoldChange value, the size of the point is the feature importance of the machine learning classification model, and the color depth of the point represents the p value.

[0161] In summary, in the present invention, random forest and XGBoost integrated learning models are used for tumor group vs normal group and four subtypes of gastric cancer respectively. For tumor group vs normal group, based on the random forest method, 558 genes that play an important role in identifying tumor samples and normal samples were successfully identified. Based on the parameter-free XGBoost model, the accuracy rate reached 98%, and 13 genes played a key role in sample classification. Based on the results of comprehensive difference analysis, random forest and XGBoost, the gene mutation map of the normal group and tumor group of gastric cancer was constructed, which contained a total of 930 characteristic genes. Similar work was done for the four subtypes of gastric cancer, and the model was finely adjusted, and finally a subtype gastric cancer gene mutation map containing 17242 characteristic genes was constructed. These maps can help further understand the pathogenesis of gastric adenocarcinoma and provide guidance for clinical treatment.

[0162] Therefore, the present invention uses a large amount of gastric cancer gene data in the TCGA database to analyze and predict the association between gene mutations in gastric cancer and different subtypes through machine learning methods such as KMeans clustering, random forest and XGBoost, achieving effective classification of gastric cancer subtypes and constructing a gastric cancer gene mutation map for subtypes. Finally, a prediction model was established and optimized to improve the accuracy of early diagnosis of gastric cancer and the effect of personalized treatment.

[0163] The main creative works of this paper are summarized as follows:

[0164] 1. A gastric cancer molecular subtype classification model based on machine learning was established. Since the molecular classification of gastric cancer in the TCGA study is based on a highly complex comprehensive analysis of multiple genomic and proteomic data sets, it is unlikely to be easily translated into clinical practice. The KMeans layer-by-layer clustering provided by the present invention can achieve molecular subtype classification based only on gene expression data.

[0165] 2. Construct gastric cancer gene mutation maps for four subtypes. The key gene mutations commonly found in gastric adenocarcinoma were revealed, and clear gene mutation maps for different gastric cancer subtypes were constructed, guiding biomedical workers to pay attention to many potential gastric cancer-related genes that cannot be detected by existing biotechnology.

[0166] 3. This paper uses gastric cancer gene data collected in the Cancer Genome Atlas (TCGA) project and adopts a variety of machine learning models, including KMeans, random forests, and XGBoost, to conduct in-depth analysis of gastric cancer gene expression data. The study first explores and screens the data set, and then uses statistical data analysis methods such as differential analysis and enrichment analysis to determine key gene mutation points, thereby achieving the classification of gastric cancer subtypes and the construction of gastric cancer gene mutation maps. The research results reveal the differences in gene expression and mutation patterns between gastric cancer subtypes, providing new strategies for the molecular classification and precision treatment of gastric cancer.

[0167] 4. In response to the multi-data processing requirements of gastric cancer classification, this study developed an automated analysis framework, including full-process automation of data screening, feature extraction, model training, subtype classification, and map construction. This framework greatly reduces the steps of manual intervention, improves analysis efficiency, and supports larger-scale gastric cancer multi-omics data analysis in the future.

[0168] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a gastric cancer gene mutation map based on machine learning, characterized in that: The specific steps include: Step S1, collecting the public dataset TCGA (The cancer genome atlas) of various sequencing data of various human cancers; Step S2, select the genetic data of the remaining adenoma and adenocarcinoma patients in the TCGA-STAD project, excluding cystic, mucinous and serous tumors; Step S3, then through the whole transcriptome sequencing (RNA-Seq) experiment, the sequencing data of gene expression quantification is obtained, and the clinical data of the patient is also obtained; Step S4, based on the Count sequencing data, use R language to clean and integrate the data, remove low-expression data, remove duplicate genes, retain the results of the maximum expression of each gene, use the gene name as the row name and the sample name as the column name to obtain the gene expression matrix; Step S5, performing variance stabilizing transformation (VST) on the count sequencing data, so that the variance of the count sequencing data is independent of the average expression level of the gene; Step S6, perform PCA analysis and visualization on the grouped gene data, and identify group aggregation, the presence of outliers, and potential data structures by observing the distribution of samples in the principal component space; Step S7, select the unsupervised clustering algorithm KMeans to perform layer-by-layer classification, the specific method is as follows: first randomly select K sample data points as K cluster centers of the initial clustering, calculate the distance between each data point and the K cluster centers, and assign it to the cluster center with the closest distance, each time after a sample is assigned, recalculate the cluster center to adapt to the existing cluster objects, and repeat until the preset stop condition is met; Step S8, selecting the gene expression levels of CDKN2A, MLH1, and PIK3CA as classification criteria for k-means clustering; Step S9, split all sample data into training set and test set at a ratio of 4:1, use cross entropy as the splitting criterion, and use the training set to build a random forest model; Step S10, split all sample data into training set and test set at a ratio of 2:1, and use the training set to build a parameter-free XGBoost model; Step S11, combining the results of difference analysis, random forest and XGBoost to obtain a gene mutation map that distinguishes the normal gastric cancer group from the tumor group.

2. The method for constructing a gastric cancer gene mutation map based on machine learning according to claim 1, characterized in that: The patient's clinical data includes the patient's age, gender, race, analysis, and prognosis data.

3. The method for constructing a gastric cancer gene mutation map based on machine learning according to claim 2, characterized in that: In step S5, the steps of the VST standardization model are as follows: First, for each gene, calculate the mean and variance in each sample E(Y)=u=α+m η m (2-1) Var(Y)=V=s η 2 m 2 +s ε 2 (2-2) The functional relationship between the estimated mean and variance is as follows: v(u)=(s η / m η ) 2 (u-α) 2 +σ ε 2 =(c1u+c2) 2 +c3 (2-3) Among them, v represents variance, u represents mean, and v(u) represents the variance v as a function of mean u. Then according to mathematical derivation, we can get the h(y) function as: Convert the matrix Y before normalization to the normalized matrix Make an identity transformation of formula (2-3): Fit a linear model using the variance and mean of each gene to obtain c1, c2, and c3; Substitute c1, c2 and c3 into formula (2-4) and solve the integral to obtain: That is, the VST standardized conversion formula h(Y) is obtained.

4. The method according to claim 1, characterized in that: First, in the process of transfer learning, the MAE model was pre-trained on the ImageNet-1K dataset, the masked image block ratio was set to 75%, and the training rounds were 800 times; Then, we fine-tune the pre-trained model on the snake image dataset, setting the learning rate to 0.001 and the batch size to 16 to optimize the classification performance.

Images