Lung adenocarcinoma prognosis prediction method based on machine learning

Abstract

The invention discloses a lung adenocarcinoma prognosis prediction method based on machine learning. RNA transcription data and clinical information of a lung adenocarcinoma patient are collected from a database, and a data set is constructed; performing gene expression data preprocessing and difference analysis to obtain differential expression genes; sequencing the risk coefficients of the differential genes by using single-factor COX analysis; carrying out survival related feature recognition on the screened genes by virtue of eight machine learning algorithms related to survival analysis to obtain related genes with prognosis values; a machine learning combinatorial algorithm is utilized to train multiple prognosis models, C-index average values of the prognosis models are calculated and compared, and an optimal StepCox-RSF algorithm is selected to carry out patient layering and prognosis prediction model construction according to feature genes; according to the risk score of the patient, predicting the total survival rate and the risk grouping condition of the patient; according to the method, the reflecting capacity of the model to the immune microenvironment is remarkably improved, the prediction precision is optimized, and a scientific basis is provided for personalized treatment of lung adenocarcinoma patients.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of biomedical detection, and particularly relates to a lung adenocarcinoma prognosis condition prediction method based on machine learning. BACKGROUND

[0002] Lung cancer is one of the most malignant tumors that threatens human life and health in the world, and non-small cell lung cancer is the main type of lung cancer, accounting for 85% of the total incidence rate of the disease. Lung adenocarcinoma is one of the main subtypes of non-small cell lung cancer. At present, the main method for judging the clinical prognosis of lung adenocarcinoma is the TNM staging system proposed by the American Joint Committee on Cancer, but the staging system only considers the anatomical factors of the tumor and does not consider non-anatomical factors, so it cannot accurately predict the prognosis of LUAD. Through the study of high-throughput gene expression profiles, biomarkers constructed through the combination of a variety of biomarkers are more promising. Due to the genetic heterogeneity of LUAD and the great difference in the prognosis of patients, the survival of lung adenocarcinoma patients cannot be accurately predicted by a single parameter or a single gene. Therefore, the prognosis gene feature based on the combination of multiple genes plays an important role in guiding and assisting clinicians to choose appropriate treatment methods, highlighting cancer progression and detecting possible new treatment targets, and it is very important to establish an expression-based gene feature to predict the prognosis and progression of LUAD patients. SUMMARY

[0003] The purpose of the present application is to provide a lung adenocarcinoma prognosis prediction method based on machine learning; to solve the problems raised in the background art.

[0004] The technical scheme adopted by the present application to achieve the purpose of the present application is as follows:

[0005] A lung adenocarcinoma prognosis prediction method based on machine learning, the specific steps comprising:

[0006] (1) Collecting RNA transcription data and clinical information of LUAD patients from TCGA and GEO databases, and constructing a data set;

[0007] (2) Grouping lung adenocarcinoma patients and normal people and performing data preprocessing, and obtaining the corresponding differential expression genes through differential expression analysis;

[0008] (3) Screening related genes with prognostic value in combination with a variety of machine learning algorithms related to prognosis;

[0009] (4) Training a variety of prognosis models based on core genes using machine learning combination algorithms, comparing the average C-index of a variety of prognosis models, and selecting the combination algorithm model with the highest average value;

[0010] (5) Constructing a lung adenocarcinoma prognosis model and calculating the risk score value of each patient;

[0011] (6) According to the risk score value of each patient, the overall survival rate and risk grouping of lung adenocarcinoma patients in the GSE31210 and GSE80051 data sets are predicted.

[0012] Preferably, in the step (1),

[0013] S11. The data used is derived from the Cancer Genome Atlas and Gene Expression Omnibus, which includes RNA transcription data and clinical information of LUAD patients, and the expression of each gene in the sample is corrected by background, quantile normalized and logarithm transformed to obtain clean microarray data to construct the data set.

[0014] Preferably, in the step (2),

[0015] S21. The corresponding differentially expressed genes are obtained by differential expression analysis, and the differentially expressed genes of different groups are compared to screen out significantly differentially expressed genes.

[0016] Preferably, in the step (3),

[0017] S31. Based on the overall survival time and survival status of LUAD patients in the TCGA data set, a single factor Cox regression analysis is performed on the significantly differentially expressed genes to construct a new calculation framework composed of eight machine learning algorithms related to prognosis, which is used to screen out core prognostic variables according to the provided gene expression and corresponding survival information.

[0018] Preferably, in the step (4),

[0019] S41. Based on the core genes, 117 prognostic models are trained using machine learning combination algorithms, and the C-index and average C-index of each model in the TCGA, GSE31210 and GSE80051 data sets are calculated.

[0020] S42. For each model, the C-index is calculated on the test set of each data set according to the following formula:

[0021]

[0022] Wherein, C-index is the consistency index, f(x i ) and f(x j ) are the risk score values of the i th and j th samples of the model, T i and T j are the survival times of the i th and j th samples, and 1 is an indicator function, which is 1 when the condition in the parentheses is true, and 0 otherwise.

[0023] As preferred, in the step (5),

[0024] By comparing the average C-index values of all models, the model of the combination of StepCox and RSF algorithm has the highest average C-index value, then the combination of StepCox and RSF algorithm is selected to establish the final prognostic model, and the construction process of the prognostic model is as follows:

[0025] S51.StepCox is used to select important features, and the formula of StepCox is as follows:

[0026] h(t|X)=h0(t)×exp(β1X1+β2X2+···+β p X p )

[0027] In the formula:

[0028] h(t|X):is the risk function given a set of covariates X=(X1,X2,…,X p )。

[0029] h0(t):is called the baseline risk function, which is independent of the covariates X, and reflects the overall basic risk with time t.

[0030] β1,β2,…,β p ,:these are the regression coefficients that need to be estimated, which correspond to each covariate X1,X2,…,X p , respectively, and the size and sign of the regression coefficient are of great significance.

[0031] Using the features selected by StepCox, RSF is applied for modeling;

[0032] S52.For a single survival tree, the cumulative risk function of individual i at the terminal node h is:

[0033]

[0034] Where t i,h is a certain time point of the individual, d i,h is the number of deaths at t i,h , S i,h is the number of people at risk at t i,h , and the cumulative risk of the final random survival forest is

[0035]

[0036] Where H n (t|X i ) is the cumulative risk of the n th survival tree.

[0037] As preferred, in the step (6),

[0038] S61. The survival prediction performance of the constructed model is evaluated using KM survival curves and AUC over time. According to the risk score value of each patient, the overall survival rate and risk grouping of lung adenocarcinoma patients in the GSE31210 and GSE80051 datasets are predicted.

[0039] Compared with the prior art, the present application has the beneficial effects of:

[0040] The present application constructs a prognostic model by using the RNA transcription and clinical data of LUAD patients in the TCGA and GEO databases. First, 1294 differentially expressed genes are screened out by the limma package, 575 related genes are obtained by single factor COX analysis, 60 prognostic related genes are screened out by combining eight machine learning algorithms, and then a variety of prognostic models are trained based on these genes using machine learning combination algorithms, the C-index and C-index average value of a variety of prognostic models are calculated and compared, the best model is selected to construct a lung adenocarcinoma prognostic model, and according to the lung adenocarcinoma prognostic model, the risk score value of each patient is calculated, and according to the risk score of each patient, the overall survival rate and risk stratification of lung adenocarcinoma patients are predicted. Therefore, by the above method, the reflection ability of the model to the immune microenvironment is significantly improved, the prediction accuracy of the model is optimized, the survival rate and risk stratification prediction of lung adenocarcinoma patients can be more accurately performed, which is helpful for clinical decision-making and the development of individualized treatment plan. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a flowchart of the present application;

[0042] Figure 2 is a volcano plot of differential gene expression;

[0043] Figure 3 is a single factor COX analysis result graph;

[0044] Figure 4 is a repeated histogram of prognostic related genes selected by different machine learning algorithms;

[0045] Figure 5 is the relationship between the risk score calculated by the StepCox[forward]-RSF model and the prognosis of patients in different cohorts;

[0046] Figure 6 is the 1-year, 3-year and 5-year AUC and C-index value of the StepCox[forward]-RSF model in different cohorts. DETAILED DESCRIPTION

[0047] In order to make the purposes, technical methods and advantages of the present application more clear, the present application is further described in detail below in combination with specific embodiments.

[0048] The present application aims to provide a machine learning-based lung adenocarcinoma prognosis prediction method; to solve the problems raised in the above background art.

[0049] Embodiments:

[0050] Please refer to Figures 1 to 6 The present application provides a technical solution:

[0051] A machine learning-based lung adenocarcinoma prognosis prediction method, as shown in the specific steps: Figure 1

[0052] S1. Collect RNA transcription data and clinical information of LUAD patients from TCGA and GEO databases to construct a data set;

[0053] S2. Group lung adenocarcinoma patients and normal people and perform data preprocessing, and obtain the corresponding differential expression genes through differential expression analysis;

[0054] S3. Combine various machine learning algorithms related to prognosis to screen out relevant genes with prognostic value;

[0055] S4. Based on the core genes, use machine learning combination algorithm to train multiple prognosis models, compare the average C-index of multiple prognosis models, and select the combination algorithm model with the highest average value;

[0056] S5. Construct a lung adenocarcinoma prognosis model and calculate the risk score value of each patient;

[0057] S6. According to the risk score value of each patient, predict the overall survival rate and risk grouping of lung adenocarcinoma patients in GSE31210 and GSE80051 data sets.

[0058] In this embodiment, S1 includes the following steps:

[0059] S11. The data used is derived from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO), which includes RNA transcription data and clinical information of LUAD patients, and the expression of each gene in the sample is corrected by background, quantile normalized and logarithmically transformed to obtain clean microarray data to construct a data set.

[0060] In this embodiment, S2 includes the following steps:

[0061] S21. As​Figures 2-3 As shown, the corresponding differentially expressed genes were obtained by differential expression analysis, the differentially expressed genes of different groups were compared, the significantly differentially expressed genes were screened, the core genes with prognostic value were obtained by performing single factor Cox regression analysis on the total survival time and survival state of the patients in the TCGA data set based on the significantly differentially expressed genes, and the process was as follows:

[0062] S22. 1294 differentially expressed genes were identified in the TCGA data set with |1og, FC|>2 and p<0.05 as the screening criteria, FC is the expression level change value, and p is the gene expression difference confidence, among which, the mRNA expression levels of 572 genes were significantly up-regulated, and the mRNA expression levels of 722 genes were significantly down-regulated;

[0063] S23. The effect of 1294 differentially expressed genes on the survival prognosis of lung adenocarcinoma patients was analyzed in depth, and single factor Cox regression analysis was performed based on the total survival time and survival state of lung adenocarcinoma patients in the TCGA data set, and the analysis result showed that the expression levels of 575 genes were significantly correlated with the prognosis of lung adenocarcinoma patients, that is, p<0.05, which means that the expression confidence of the screened genes reaches 95%.

[0064] In this embodiment, S3 includes the following steps:

[0065] S31. As shown in Figure 4 As shown, the screened genes were analyzed in depth, and a new type of computing framework composed of eight machine learning algorithms was constructed for screening core prognostic variables according to the provided transcriptome atlas and corresponding survival information. The implementation of the framework includes three steps:

[0066] S32. First, determine the prognostic factors by single variable COX regression analysis;

[0067] S33. Second, evaluate eight machine learning algorithms related to survival analysis to determine the most critical features, execute the LASSO algorithm 1000 times using different seeds, and determine those variables that are selected more than 10 times in 1000 executions as core features;

[0068] S34. Third, select the most frequently screened variables as core features according to the frequency of variable screening, and the key genes captured by different algorithms when screening genes are quite different, and their screening mechanisms and focuses are different;

[0069] S35. Due to the differences in the number of genes shared between different algorithms, a single algorithm may not be able to fully cover all the important genes related to the research. In actual gene screening research, in order to identify key genes more comprehensively and accurately, it is necessary to use multiple algorithms in combination and integrate the gene sets screened by different algorithms to make up for the limitations of a single algorithm.

[0070] In this embodiment, S4 includes the following steps:

[0071] S41. Based on the core genes, 117 prognostic models were trained using machine learning combinatorial algorithms, and the C-index and average C-index of each model in the TCGA, GSE31210 and GSE80051 datasets were calculated.

[0072] S42. For each model, calculate the C-index on the test set of each dataset, using the following formula:

[0073]

[0074] Wherein, C-index is the consistency index, f(x) i ) and f(x j T represents the risk score of the model for the i-th and j-th samples. i With T j The survival time of the i-th and j-th samples is given by 1, which is an indicator function. The value is 1 when the condition in parentheses is true, and 0 otherwise.

[0075] The C-index value ranges from 0 to 1, and the closer the value is to 1, the better the model's predictive performance.

[0076] When C-index = 1, it means that the model makes a perfect prediction, that is, all predicted risk scores are completely consistent with the actual survival time;

[0077] When C-index = 0.5, it means that the model's predictive performance is equivalent to random guessing;

[0078] When C-index < 0.5, it means that the model's predictive performance is worse than random guessing;

[0079] S43. For each model, calculate its average C-index across all datasets using the following formula:

[0080]

[0081] in C is the average C-index for each model. KC-index for the kth dataset, N is the number of datasets, k ranges from 1, 2, …, N is the total number of datasets.

[0082] In this embodiment, S5 includes the following steps:

[0083] By comparing the average C-index values of all models, the model of the combination of StepCox and RSF algorithm has the highest average C-index value, so the combination of StepCox and RSF algorithm is selected to establish the final prognostic model, and the construction process of the prognostic model is as follows:

[0084] S51.StepCox is used to select important features, and the formula of StepCox is as follows:

[0085] h(t|X) = h0(t) x exp(β1X1 + β2X2 + ··· + β p X p )

[0086] In the formula:

[0087] h(t|X): is the risk function under the condition of a set of covariates X = (X1, X2, …, X p ), where X1, X2, …, X p is the specific covariate, and p is the number of covariates. The hazard function describes the instantaneous probability of corresponding death in survival analysis.

[0088] h0(t): called baseline hazard function, which is independent of covariates X, reflects the change of the overall basic risk with time t. That is, it embodies the trend of the risk of events over time when all covariates take the value of 0.

[0089] β1, β2, …, β p : These are regression coefficients that need to be estimated, which correspond to each covariate X1, X2, …, X p , and the size and sign of the regression coefficient are of great significance. The absolute value of the coefficient reflects the degree of influence of the corresponding covariate on the risk function. If β i (i = 1, 2, …, p) is positive, it means that as the value of the covariate X i increases, the value of the risk function h(t|X) will increase, that is, the risk of the event will increase, and vice versa. If the coefficient is negative, it means that as the value of the covariate X i increases, the risk of the event decreases.

[0090] S52. Using the features selected by Step Cox, apply RSF to modeling;

[0091] The main steps of random survival forest are:

[0092] S521. Randomly sample the data in the whole data set using bootstrap method, about 37% of the data in the whole data set is taken as out-of-bag (OOB) data, which is not involved in the modeling;

[0093] S522. Set the number of survival trees to be generated as N, then perform N times of step (1) to generate N survival trees, randomly select mtry variables as candidate variables, select the variable that makes the difference between the child nodes the largest, and the difference is evaluated using logrank or logrank score to compare the survival curves between the survival curves.

[0094] S523. Estimate the cumulative hazard function of each survival tree using Nelson-Aalen estimator, and then take the average as the cumulative hazard of random survival forest. For a single survival tree, the cumulative hazard function of individual i at the terminal node h is:

[0095]

[0096] Where t i,h is a certain time point at which the individual is at, d i,h is the number of deaths at t i,h , S i,h is the number of individuals at risk at t i,h , and the final cumulative hazard of random survival forest is

[0097]

[0098] Where H n (t|X i ) is the cumulative hazard of the n th survival tree;

[0099] S524. Use OOB data to evaluate the prediction performance of the constructed model.

[0100] In the embodiment, S6 includes the following steps:

[0101] S61. Estimate the baseline survival function according to the Cox model, estimate the baseline survival function using the Kaplan-Meier method, and the formula is as follows:

[0102]

[0103] Where S0(t) is the baseline survival function, which is the survival probability when all feature variables X are zero at time t, h0(u) is the baseline risk function, which is the risk value when all feature variables X are zero at time u, and t is time.

[0104] S62. Based on the patient's risk score RS τ Given the baseline survival function S0(t), the individual's survival function is calculated using the following formula:

[0105]

[0106] Wherein, S(t|X τ Let be the individual survival function of the τ-th patient, representing the eigenvector X. τ The survival probability of an individual at time t;

[0107] S63. The overall survival rate is obtained by averaging the individual survival functions of all patients, using the following formula:

[0108]

[0109] Among them, S a (t) represents the overall survival rate for all patients.

[0110] S64. For example Figures 5-6 As shown, based on the above embodiments, the risk grouping process for predicting the risk of lung adenocarcinoma patients in the GSE31210 and GSE80051 datasets according to each patient's risk score is as follows:

[0111] S641. The optimal cutoff value is the point at which continuous risk scores are divided into two groups (high risk and low risk) that maximizes the survival difference between the two groups. By iterating through all possible cutoff values, calculating the survival curves at each cutoff value and comparing their differences, the cutoff value that maximizes the difference between the groups is finally selected as the optimal cutoff value M.

[0112] When a patient's risk score is greater than the optimal cutoff value, i.e., RS τ If the result is >M, then the patient belongs to the high-risk group;

[0113] When a patient's risk score is less than or equal to the optimal cutoff value, i.e., RS τ If the value is ≤M, then the patient belongs to the low-risk group.

[0114] S642. After patients were divided into high-risk and low-risk groups, survival curves for these two groups were plotted using the Kaplan-Meier method. The Kaplan-Meier survival curves show the survival probability over time.

[0115] By comparing the Kaplan-Meier survival curves of the high-risk group and the low-risk group, it is observed whether there is a significant difference in the survival of the patients in the two groups. If the Kaplan-Meier curve shows that the survival rate of the high-risk group is significantly lower than that of the low-risk group, it indicates that the risk score is effective in distinguishing the prognosis of the patients.

[0116] The above formulas are dimensionless values calculated, and the formulas are obtained by collecting a large amount of data to simulate a formula of the latest real situation. The preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0117] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1.A method of prognosis prediction for lung adenocarcinoma based on machine learning, characterized in that, The specific steps include: S1. Collecting RNA transcription data and clinical information of LUAD patients from TCGA and GEO databases to construct a dataset; S2. Grouping lung adenocarcinoma patients and normal people and performing data preprocessing, and obtaining the corresponding differential expression genes through differential expression analysis; S3. Screening relevant genes with prognostic value by combining various machine learning algorithms related to prognosis; S4. Training various prognostic models based on core genes using machine learning combination algorithm, comparing the average C-index of various prognostic models, and selecting the combination algorithm model with the highest average value; S5. Constructing a lung adenocarcinoma prognostic model and calculating the risk score value of each patient; S6. Predicting the overall survival rate and risk grouping of lung adenocarcinoma patients in GSE31210 and GSE80051 datasets according to the risk score value of each patient. 2.The machine learning-based lung adenocarcinoma prognosis prediction method according to claim 1, characterized in that The S1 includes the following steps: S11. The data used is derived from the Cancer Genome Atlas and Gene Expression Omnibus, which includes RNA transcription data and clinical information of LUAD patients, and the expression of each gene in the sample is corrected by background, quantile normalized and logarithmically transformed to obtain clean microarray data to construct a dataset. 3.The machine learning-based lung adenocarcinoma prognosis prediction method of claim 1, wherein, The S2 includes the following steps: S21. Obtain the corresponding differential expression genes through differential expression analysis, compare the differential expression genes of different groups, and screen out the significantly differentially expressed genes. 4.The method of claim 1, wherein, The S3 includes the following steps: S31. Perform single-factor Cox regression analysis on the significantly differentially expressed genes based on the overall survival time and survival status of LUAD patients in the TCGA dataset, and construct a new calculation framework composed of eight machine learning algorithms related to prognosis, which is used to screen out core prognostic variables according to the provided gene expression and corresponding survival information. 5.The machine learning based prognosis prediction method for lung adenocarcinoma according to claim 1, wherein, The S4 includes the following steps: S41. Based on the core genes, 117 prognostic models are trained using machine learning combination algorithm, and the C-index and its average value of each model in TCGA, GSE31210 and GSE80051 datasets are calculated; S42. For each model, calculate the C-index on the test set of each dataset according to the following formula: where C-index is the concordance index, f(x i ) and f(x j ) are the risk score values of the model for the i-th and j-th samples, T i and T j are the survival times of the i-th and j-th samples, and 1 is an indicator function that takes the value 1 when the condition in the parentheses is true and 0 otherwise. 6.The method of claim 1, wherein, The S5 includes the following steps: By comparing the average C-index values of all models, the StepCox and RSF combination algorithm has the highest C-index average value, so the combination of StepCox and RSF algorithm is selected to establish the final prognostic model, and the construction process of the prognostic model is as follows: S51. StepCox is used to select important features, and the formula of StepCox is as follows: h(t|X) = h0(t) x exp(β1X1+ β2X2+ ··· + βpXp) p X p ) In the formula: h(t|X): is the risk function given a set of covariates X = (X1, X2,..., Xp)T. p ) the risk function given a set of covariates X = (X1, X2,..., Xp)T. h0(t): called the baseline risk function, which is independent of the covariate X, reflecting the overall basic risk with time t. β1, β2, …, β p , These are the regression coefficients that need to be estimated, which correspond to each of the covariates X1, X2, …, X p , The size and sign of the regression coefficients are of great importance; Using the features selected by StepCox, RSF is applied for modeling; S52. For a single survival tree, the cumulative risk function of individual i at the terminal node h is: where t i,h is the time point at which the individual is, d i,h is the number of deaths at time t i,h , S i,h is the number of individuals at risk at time t i,h , and the cumulative risk of the final random survival forest is where H n (t|X i ) is the cumulative hazard of the nth surviving tree. 7.The machine learning based lung adenocarcinoma prognosis prediction method of claim 1, wherein, The S6 includes the following steps: S61. The survival prediction performance of the constructed model is evaluated using KM survival curves and AUC over time.

Images