Solid tumor taxane / immunotherapy response evaluation system based on integrated machine learning and working method thereof

Abstract

The invention discloses a solid tumor taxane / immunotherapy response evaluation system based on integrated machine learning and a working method thereof, and belongs to the technical field of bioinformatics and artificial intelligence crossing. The method comprises the following steps of: 1, preprocessing acquired solid tumor taxane chemotherapy / immunotherapy response transcriptome data or a source data set, and dividing a training set and a verification set; step 2, extracting overlapped genes as candidate genes from the gene expression level based on the preprocessed data; step 3, carrying out reprocessing based on the candidate genes; 4, constructing a consensus prediction model; 5, training the constructed model by adopting the training set in the step 3; and 6, verifying and screening the trained model to obtain an optimal scheme. According to the method, a stable and generalizable prediction tool for solving the problem of curative effect heterogeneity of complex combined treatment and how to accurately screen potential benefit people and avoid excessive treatment is filled.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bioinformatics, tumor precision medicine and artificial intelligence, and particularly relates to an entity tumor taxane / immunotherapy response evaluation system based on integrated machine learning and a working method thereof. BACKGROUND

[0002] As a representative of broad-spectrum chemotherapy drugs, taxanes block mitosis of tumor cells by stabilizing microtubule structure, and occupy a core position in the treatment of various solid tumors such as breast cancer, ovarian cancer and non-small cell lung cancer. However, its clinical efficacy is limited by drug resistance and side effects, and only part of the patients can benefit from it in the long term. At the same time, the immune checkpoint inhibitor (ICI) represented by PD-1 / PD-L1 inhibitor brings hope for patients with advanced tumors by activating anti-tumor immune response, but the bottleneck of less than 30% response rate still needs to be broken through. In recent years, based on the synergistic mechanism of taxane chemotherapy and immunotherapy (such as induction of tumor antigen release and remodeling of immune microenvironment), the combination therapy strategy has gradually become a clinical research hotspot, and many phase II / III clinical trials have confirmed that it can significantly improve the objective remission rate and survival benefit in some tumors. However, the heterogeneity of the efficacy of combination therapy is more complex, and how to accurately screen the potential benefit population and avoid over-treatment has become the core challenge of clinical decision-making. Therefore, the development of a stable and generalizable prediction tool has become an important demand for precision treatment. SUMMARY

[0003] The application provides an entity tumor taxane / immunotherapy response evaluation system based on integrated machine learning and a working method thereof, which fills the gap in the stable and generalizable prediction tool for the complex heterogeneity of combination therapy and how to accurately screen the potential benefit population and avoid over-treatment.

[0004] The application is implemented by the following technical solutions: A working method of an entity tumor taxane / immunotherapy response evaluation system based on integrated machine learning, the method comprising the following steps: Step 1: preprocessing the acquired entity tumor taxane chemotherapy / immunotherapy response transcriptome data or source data set, and dividing the preprocessed entity tumor taxane chemotherapy / immunotherapy response transcriptome data or source data set into a training set and a validation set; Step 2: extracting overlapping genes as candidate genes from the gene expression level based on the data preprocessed in step 1; Step 3: reprocessing the candidate genes in step 2; Step 4: constructing a consensus prediction model; Step 5: training the model constructed in step 4 using the training set in step 3; Step 6: Verify the trained model and select the optimal solution.

[0005] Further, the step 1 of screening, including and processing the transcriptome data of solid tumor taxane chemotherapy / immunotherapy response is specifically as follows: The limitation conditions are specifically as follows: 1. solid tumor tissue sample; 2. containing clinical outcome information; 3. baseline pretreatment specimen; 4. Tax cohort receiving any pure chemotherapy regimen including taxane, ICI cohort receiving any one or more immune checkpoint inhibitors, and samples receiving other types of treatment regimen are excluded; 5. If there are different treatment regimens in the same research source dataset, split it into multiple single treatment regimen datasets; 6. The total sample size of each dataset is ≥10, and the sample size of R group and NR group is ≥3; 7. The number of genes is ≥10,000.

[0006] Further, the preprocessing of the source dataset in step 1 is specifically as follows: If the original FASTQ file of RNA-seq data and the original CEL, TXT file of microarray data are available, first, data quality control is performed by TrimGalore (0.6.10) to remove reads with base quality score less than 25 and sequence length less than 36 bp, and to remove and trim adapter sequences; Then, HISAT2 (2.2.1) and SAMtools (1.13) are used to align the reads to the human reference genome GNECODE version 36 and generate BAM files; The count data is then calculated by FeatureCounts in the subread (2.0.1) toolkit, and the original count reads are converted to transcripts per kilobase; For raw microarray data from Affymetrix, Illumina and Agilent, use the affy, lumi and limma packages for processing, respectively, and use the mean when multiple probes map to a single gene symbol; all gene expression values are transformed to log2 scale; the expression profiles without providing raw data are directly downloaded from the original literature or public database; all data are re-annotated based on GENCODE Release36.

[0007] Further, the step 2 is specifically as follows: Based on the gene expression level, the transcriptomic feature categories are obtained, including pathways, transcription factors, cytokines, key immune interactions, immune cell composition and cell-cell communication. The Meta analysis is used to extract overlapping genes from the transcriptomic feature categories as candidate genes.

[0008] Further, the step 3 is specifically KNN imputation on the training queue composed of candidate genes to fill in missing values, while deleting genes with missing rates exceeding 50%; in the training queue, the ComBat method is further used to correct batch effects; Finally, 70% of the training queue is divided into a training set, and 30% is divided into a validation set.

[0009] Further, the step 4 specifically includes 12 feature selection methods, 20 machine learning models, 8 signature scoring algorithms, 12 feature selection methods, and 28 classification methods.

[0010] Further, the step 5 is specifically based on the model of step 4 to remove genes with correlation greater than 0.85 and highest average absolute correlation in the training set using the Spearman correlation coefficient; The 12 feature selection methods in the model of step 4 are used to further screen features that can distinguish between treatment-responsive and non-responsive patients at the expression or co-expression level; The 20 machine learning models or 8 signature scoring algorithms in the model of step 4; Therefore, each of the 336 models or signatures starts with a "route" composed of one feature selection method and one classification method, and all possible combinations of the 12 feature selection methods and 28 classification methods in the model of step 4 are used; grid search and 5-fold 10-time repeated cross-validation are used to find the best model parameters, and the area under the receiver operating characteristic curve is used as the evaluation standard for model parameter optimization.

[0011] Further, the step 6 is specifically to compare all models and scoring methods in the validation set, and to screen the model with the highest AUC value and the difference between the training set and the validation set AUC value not more than 0.15 as the optimal scheme, i.e. the best model is the combination of Lasso and GSVA and the combination of RandomForest and ssGSEA.

[0012] An entity tumor taxane / immunotherapy response evaluation system based on integrated machine learning, the evaluation system is to construct a robust consensus prediction model, the performance of 336 possible models composed of 12 different feature selection routes and 28 classification methods is used to check the classification ability of the treatment response transcriptomic data.

[0013] Further, the evaluation system performs KNN interpolation on the training queue to fill in missing values, while deleting genes with missing rates exceeding 50%; in the training queue, further ComBat method is used to correct batch effect, so as to reduce systematic bias caused by experimental conditions or data source differences; The training queue is divided into a training set and a validation set, the correlation of all gene pairs is calculated by using a Spearman correlation coefficient, and genes with a correlation greater than 0.85 and the highest average absolute correlation are deleted; Twelve feature selection methods are used to further screen features that can distinguish between treatment-responsive and non-responsive patients in terms of expression or co-expression levels; Twenty machine learning models or eight signature scoring algorithms are used for each selected feature set; therefore, each of the 336 models or signatures starts with a "route" consisting of one feature selection method and one classification method, and all possible combinations of 12 feature selection methods and 28 classification methods are used; in order to evaluate the performance of the machine learning model, the platform uses grid search and 5-fold 10-time repeated cross-validation to find the best model parameters, and uses the area under the receiver operating characteristic curve as the evaluation standard for model parameter optimization; finally, all models and scoring methods are compared in the validation set, and the model with the highest AUC value and a difference in AUC values between the training set and the validation set not exceeding 0.15 is selected as the optimal scheme; The best model in Tax is the combination of Lasso and GSVA; the best model in ICI is the combination of RandomForest and ssGSEA.

[0014] The application of the above method is particularly related to the application of multi-omics data integration and machine learning model in tumor treatment response prediction.

[0015] The beneficial effects of the application are: The two models of the present application exhibit robust prediction performance and generalization ability. The chemotherapy test cohort and the immunotherapy test cohort are assembled from 8 / 9 different datasets (Materials and Methods), respectively. The model PredictTax shows high performance (AUC = 0.88, 0.86, 0.86, 0.86, 0.86, 0.86, 0.86, 0.88, 0.86) on 9 independent test cohorts. It is also observed that PredictICI achieves excellent prediction ability with an average AUC = 0.71 on 8 independent test cohorts. 96 / 102 published features associated with various biological features are assembled, respectively, to compare the performance of PredictTax and PredictICI with other features in predicting Tax / ICI treatment response. In comparing the overall performance, the two models outperform other features. In addition, three other comprehensive evaluation indicators, F1 score, Matthews Correlation Coefficient (MCC) and Balanced.Accuracy, are also calculated. All these indicators show that PredictTax and PredictICI have stable and robust performance in multiple independent cohorts, ranking first among hundreds of signatures. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a method flowchart of the present application. DETAILED DESCRIPTION

[0017] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, technologies, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0018] It is to be understood that the terminology "includes", "has", "holds", "contains" and / or "comprising", when used in this specification, including the accompanying claims, includes the presence of stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] It is also to be understood that the terminology used in the description of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in this specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0020] With reference to the drawings accompanying the present application specification, the technical solutions in the embodiments of the present application are described clearly and completely, obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative efforts fall within the scope of protection of the present application.

[0021] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be practiced without other different ways from those described herein, and those skilled in the art can make similar extensions without departing from the connotation of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0022] Embodiment one The present embodiment provides a working method of an entity tumor taxane / immunotherapy response evaluation system based on integrated machine learning, the method comprising the following steps: Step 1: preprocessing the acquired entity tumor taxane chemotherapy / immunotherapy response transcriptome data or source data set, and dividing the preprocessed entity tumor taxane chemotherapy / immunotherapy response transcriptome data or source data set into training set and validation set; Step 2: extracting overlapping genes as candidate genes from gene expression levels based on the data preprocessed in step 1; Step 3: reprocessing the candidate genes in step 2; Step 4: constructing a consensus prediction model; Step 5: training the model constructed in step 4 using the training set in step 3; Step 6: verifying the trained model and selecting the optimal scheme.

[0023] Further, the step 1 screens, includes and processes the entity tumor taxane chemotherapy / immunotherapy response transcriptome data under the following restrictions: The restrictions are specifically: 1. Entity tumor tissue sample; 2. Contains clinical outcome information; 3. Baseline pretreatment specimen; 4. Tax queue accepts any pure chemotherapy regimen including taxane, ICI queue accepts any one or more immune checkpoint inhibitors, samples receiving other types of treatment regimen (such as radiotherapy, targeted therapy) are excluded; 5. If there are different treatment regimens in the same research source data set, split it into multiple single treatment regimen data sets; 6. Total sample size of each data set ≥10, R group and NR group sample size ≥3; 7. The number of genes is ≥10000.

[0024] Further, the preprocessing of the step 1 source data set is specifically, if the original FASTQ file of the RNA-seq data and the original CEL, TXT file of the microarray data are available, first, data quality control is performed by TrimGalore (0.6.10) to remove reads with base quality score less than 25 and sequence length less than 36 bp, and to remove and trim the adapter sequence; Then, HISAT2 (2.2.1) and SAMtools (1.13) are used to align the reads to the human reference genome GNECODE version 36 and generate a BAM file; The subsequent count data is calculated by FeatureCounts in the subread (2.0.1) toolkit, and the original count reads are converted into transcripts per kilobase transcript (TPM); For the original microarray data from Affymetrix, Illumina and Agilent, the affy, lumi and limma packages are used for processing, respectively, and the mean value is used when multiple probes are mapped to a single gene symbol; all gene expression values are transformed to log2 scale; the expression profiles not providing original data are directly downloaded from the original literature or public database; all data are re-annotated based on GENCODE Release36.

[0025] Further, the step 2 is specifically, based on the gene expression level, a transcriptomic feature category is obtained, and the transcriptomic feature category includes a pathway, a transcription factor, a cytokine, a key immune interaction, an immune cell composition and cell-cell communication; Using Meta analysis, overlapping genes are extracted from the transcriptomic feature category genes and multi-dimensional features such as pathways, transcription factors, cytokines, key immune interactions, immune cell compositions and cell-cell communication as candidate genes. Each gene obtained in this way is not only confirmed by meta analysis to be a robust treatment response-related biomarker, but also is explicitly involved in biological functions such as pathways, cytokines, transcription factors, immune cell infiltration, etc.

[0026] Further, the step 3 is specifically, KNN imputation is performed on the training queue composed of candidate genes to fill in missing values, and genes with a missing rate exceeding 50% are deleted to ensure the integrity and reliability of the data; in the training queue, the ComBat method is further used to correct the batch effect to reduce systematic bias caused by differences in experimental conditions or data sources; Finally, 70% of the training queue is divided into a training set and 30% is divided into a validation set, so as to carry out subsequent model training and performance evaluation.

[0027] Further, the step 4 specifically includes 12 feature selection methods, 20 machine learning models, 8 signature scoring algorithms, 12 feature selection and 28 classification methods.

[0028] Further, the step 5 is specifically based on the model of step 4, using the Spearman correlation coefficient to remove genes with correlation greater than 0.85 and the highest average absolute correlation in the training set, in order to avoid the problem of multicollinearity; Specifically, the findCorrelation function in the caret package of R language is used to calculate the correlation of all gene pairs, and the gene with the highest average absolute correlation and the correlation greater than 0.85 is deleted, so as to ensure that the correlation between each feature is less than 0.85.

[0029] In order to reduce noise or redundant genes, 12 feature selection methods in step 4 are used to further screen the features that can distinguish between treatment response and non-response patients at the expression or co-expression level, and reduce the number of features; 20 machine learning models or 8 signature scoring algorithms in step 4 are used for each selected feature set; Therefore, each of the 336 models or signatures starts with a "route" composed of 1 feature selection method and 1 classification method, and all possible combinations of 12 feature selection methods and 28 classification methods in step 4 are used; In order to evaluate the performance of the machine learning model, the platform uses grid search and 5-fold 10-time repeated cross-validation to find the best model parameters, and uses the area under the receiver operating characteristic curve (AUC-ROC) as the evaluation standard for model parameter optimization.

[0030] Further, the step 6 is specifically comparing all models and scoring methods in the validation set, screening the model with the highest AUC value and the difference between the training set and the validation set AUC value not more than 0.15 as the optimal scheme, that is, the best model is the combination of Lasso and GSVA and the combination of RandomForest and ssGSEA.

[0031] A total of 200 integrated models are obtained in Tax, and the best model is the combination of Lasso and GSVA (PredictTax), with the highest AUC (0.696) in the validation set. A total of 211 integrated models are retained in ICI, and the best model is the combination of RandomForest and ssGSEA (PredictICI), with the highest AUC (0.711) in the validation set.

[0032] PredictTax for taxane chemotherapy prediction and PredictICI for immunotherapy prediction.

[0033] In use, the user inputs the pre-treatment RNA-seq or microarray expression matrix, the system automatically completes data standardization and preprocessing, and performs response scoring through the PredictTax and PredictICI models to evaluate the patient's response to taxane treatment and immunotherapy.

[0034] Specifically, data processing: in the independent test queue and the training verification queue, KNN interpolation is performed respectively, and genes with more than 50% missing values are deleted before interpolation. In the training verification queue, ComBat is used to remove batch effects, and 70% of the samples are divided into the training set, and the remaining 30% of the samples are divided into the training set.

[0035] Feature selection: for the genes obtained by Meta analysis, feature selection is performed in the training set. First, remove genes with high correlation between features, use spearman correlation coefficient to find gene pairs with correlation greater than 0.85, compare their average absolute correlation with each gene, delete the gene with the largest average absolute correlation to ensure that the correlation between each feature is less than 0.85. Second, WGCNA, Fair, HC, ExpHC, SIS, Boruta, xgboost, gbm, abess, rpart, randomForest, LASSO are used to select features in the training set. WGCNA: select the module related to patient response. Each module is used as a feature gene set. Fair, HC, ExpHC: SelectV function from HiDimDA package (from three literatures). SIS: SIS package. Boruta: Boruta package, only select Confirmed features. Xgboost, gbm: mlr3 package, select features with feature importance not equal to 0. Abess, rpart: mlr3 package, direct feature selection. randomForest: mlr3 package, select the top 100 genes with feature importance. LASSO: mlr3 package, use lambda.1se as the threshold. In order to ensure the availability of subsequent training models, the screening algorithm with feature number greater than 10% of the training set sample number is deleted.

[0036] Model construction: based on the mlr3 package. 20 different classification algorithms are applied to the results of the above feature selection algorithms, grid search and 5-fold 10-time repeated cross-validation are used to find the best model parameters, and AUC value is used as the evaluation standard for model parameter tuning.

[0037] 20 kinds of classification algorithms and adjusted parameters: ranger: random forest. min.node.size, num.trees, xgboost: extreme gradient boosting. alpha, eta, max_depth lightgbm: light gradient boosting machine. max_depth, num_leaves catboost: gradient boosting on decision trees for classification. learning_rate, iterations svm: support vector machine. cost, gamma nnet: neural network classifier. size, decay, abstol naive_bayes: naive bayes. eps, laplace rpart: decision tree. cp, maxdepth, minsplit log_reg: logistic regression. epsilon, maxit glmnet: generalized linear models with elastic net regularization. alpha, nlambda, eps kknn: K-nearest neighbors classification. k, distance abess: adaptive best subset selection. lambda, early.stop, nfolds multinom: multinomial logistic linear classifier with neural network. decay, abstol C50: C50 decision tree. trials, minCases, CF ctree: classification tree. alpha, maxdepth, minbucket glmboost: generalized linear classification with boosting. nu, mstop lda: linear discriminant analysis. method qda: quadratic discriminant analysis. method earth: multivariate adaptive regression splines. newvar.penalty, nfold LMT: classification tree with logistic regression models at leaves. W, R, I Build scoring algorithm: apply 8 different scoring algorithms to the results of the above feature selection algorithms, according to the results of the Meta analysis, divide the feature genes into up-regulated genes and down-regulated genes, and use the score of the up-regulated genes minus the score of the down-regulated genes as the final result. 8 scoring algorithms: mean (mean), CRDscore, PCA, plage, ssGSEA, GSVA, singscore, TCellSI.

[0038] All models and scoring algorithms are compared in the validation set, and the algorithm with the highest AUC value in the validation set and the AUC difference between the training set and the validation set is not greater than 0.15 is selected as the optimal algorithm.

[0039] Language integration: In the present application, machine learning models and scoring algorithms for chemotherapy and immunotherapy response are developed based on gene expression data, and the optimal model and scoring system are selected through systematic comparison, in order to realize precise patient stratification and treatment response prediction. The research process includes data preprocessing, feature selection, model construction, scoring system construction and final model selection.

[0040] First, in the data preprocessing, KNN imputation is performed on the independent test queue and the training and validation queue respectively to fill in the missing values, and genes with missing rate exceeding 50% are deleted to ensure the integrity and reliability of the data. In the training and validation queue, the ComBat method is further used to correct the batch effect to reduce the systematic bias caused by differences in experimental conditions or data sources. Then, the training and validation queue is divided into training set (70%) and validation set (30%) for subsequent model training and performance evaluation.

[0041] In the feature selection stage, for the gene set screened by Meta analysis, first, the Spearman correlation coefficient is used to remove highly correlated genes to avoid multicollinearity problems. Specifically, the correlation of all gene pairs is calculated, and the genes with correlation greater than 0.85 and the highest average absolute correlation are deleted to ensure that the correlation between each feature is less than 0.85. Subsequently, in the training set, a variety of feature selection methods are used, including WGCNA, Fair, HC, ExpHC, SIS, Boruta, XGBoost, GBM, Abess, Rpart, RandomForest and LASSO, etc., to select the most biologically and statistically significant features. WGCNA is used to identify gene modules related to patient response, and each module is used as a feature gene set; Fair, HC, ExpHC are based on the SelectV function in the HiDimDA package; SIS method is derived from the SIS package; Boruta selects Confirmed features through the Boruta package; XGBoost, GBM extract genes with non-zero feature importance based on the mlr3 package; Abess, Rpart use the mlr3 package for direct feature selection; RandomForest selects the top 100 genes with feature importance; LASSO uses the mlr3 package to select features with lambda.1se as the threshold. To ensure the robustness of the model, further remove the feature selection algorithms with the number of features exceeding 10% of the training set sample size to reduce the impact of dimension disaster.

[0042] In the model construction phase, based on the mlr3 framework, the feature sets selected by different feature selection methods were applied to 20 different classification algorithms, including random forest (ranger), extreme gradient boosting (XGBoost), light gradient boosting (LightGBM), CatBoost, support vector machine (SVM), neural network (nnet), Naive Bayes, decision tree (Rpart), logistic regression (log_reg), elastic net regularized generalized linear model (GLMNet), K nearest neighbor classification (kknn), adaptive optimal subset selection (Abess), multinomial logistic model (Multinom), C50 decision tree, classification conditional inference tree (CTree), boosted generalized linear classification (GLMBoost), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), multivariate adaptive regression splines (Earth), and classification tree with logistic regression in leaf nodes (LMT). Grid search and 5-fold 10-time repeated cross-validation were used for hyperparameter optimization, and AUC value was used as the model performance evaluation index to select the optimal classification algorithm.

[0043] In the scoring system construction aspect, 8 different scoring methods were applied to the gene sets selected by different feature selection methods, including Mean, CRDscore, PCA, Plage, ssGSEA, GSVA, Singscore and TCellSI. According to the results of Meta analysis, the feature genes were divided into up-regulated genes and down-regulated genes, and the score of up-regulated genes minus the score of down-regulated genes was used as the final scoring standard to improve the biological interpretability and clinical applicability of the model.

[0044] Finally, in the final selection process of the model and the scoring system, all models and scoring methods were compared in the validation set, and the model with the highest AUC value and the difference between the training set and the validation set AUC value not more than 0.15 was selected as the optimal scheme. This process ensures the balance of the selected model in terms of generalization performance and stability, and provides a powerful tool for precise stratification and individualized treatment of tumor patients.

[0045] Embodiment Two The present embodiment provides an integrated machine learning-based solid tumor taxane / immunotherapy response evaluation system, which is to construct a robust consensus prediction model, and the performance of 336 possible models composed of 12 different feature selection routes and 28 classification methods is used to check the classification ability of the treatment response transcriptome data.

[0046] Further, the evaluation system performs KNN imputation on the training cohort to fill in missing values, while deleting genes with missing rates exceeding 50% to ensure the integrity and reliability of the data; in the training cohort, the ComBat method is further used to correct batch effects to reduce systematic bias caused by experimental conditions or differences in data sources; The training cohort is divided into a training set (70%) and a validation set (30%) for subsequent model training and performance evaluation; the Spearman correlation coefficient is used to remove highly correlated genes to avoid multicollinearity problems. Specifically, the Spearman correlation coefficient is used to calculate the correlation of all gene pairs, and the genes with a correlation greater than 0.85 and the highest average absolute correlation are deleted, so that the correlation between each feature is less than 0.85; Third, in order to reduce noise or redundant genes, a set of 12 feature selection methods are used to further screen features that can distinguish between treatment-responsive and non-responsive patients at the expression or co-expression level, reducing the number of features; Fourth, 20 machine learning models or 8 signature scoring algorithms are used for each selected feature set; therefore, each of the 336 models or signatures starts with a "route" consisting of one feature selection method and one classification method, and all possible combinations of 12 feature selection and 28 classification methods are used; in order to evaluate the performance of the machine learning model, the platform uses grid search and 5-fold 10-time repeated cross-validation to find the best model parameters, and uses the area under the receiver operating characteristic curve (AUC-ROC) as the evaluation standard for model parameter optimization; finally, all models and scoring methods are compared in the validation set, and the model with the highest AUC value and a difference in AUC values between the training set and the validation set not exceeding 0.15 is selected as the optimal solution; Through the unified classification integration platform, Tax has a total of 200 integrated models, and the best model is the combination of Lasso and GSVA (PredictTax), with the highest AUC (0.696) in the validation set. ICI has a total of 211 integrated models, and the best model is the combination of RandomForest and ssGSEA (PredictICI), with the highest AUC (0.711) in the validation set.

[0047] Embodiment Three The embodiment discloses a method as in Embodiment One, especially relates to the application of multi-omics data integration and machine learning model in tumor treatment response prediction.

Claims

1. A working method for a taxane / immunotherapy response assessment system for solid tumors based on ensemble machine learning, characterized in that, The method includes the following steps: Step 1: Preprocess the obtained solid tumor taxane chemotherapy / immunotherapy response transcriptome data or source dataset, and divide the preprocessed solid tumor taxane chemotherapy / immunotherapy response transcriptome data or source dataset into training set and validation set; Step 2: Based on the preprocessed data from Step 1, extract overlapping genes as candidate genes from the gene expression level; Step 3: Process the candidate genes from Step 2 again; Step 4: Build a consensus prediction model; Step 5: Train the model built in Step 4 using the training set from Step 3; Step 6: Validate the trained model and select the optimal solution.

2. The evaluation method according to claim 1, characterized in that, Step 1 involves screening, including, and processing the transcriptome data on the response to taxane chemotherapy / immunotherapy in solid tumors under the following restrictive conditions: The specific restrictions are as follows: Solid tumor tissue samples; Includes clinical outcome information; Baseline pretreatment of specimens; The Tax cohort accepted any chemotherapy regimen that included taxanes, the ICI cohort accepted any one or more immune checkpoint inhibitors, and samples that received other types of treatment were excluded. If a dataset from the same research source has different treatment options, it should be split into multiple datasets for each treatment option. The total sample size of each dataset is ≥10, and the sample size of the R group and NR group is ≥3; Number of genes ≥ 10,000.

3. The evaluation method according to claim 1, characterized in that, The preprocessing of the source dataset in step 1 is as follows: if the original FASTQ file of RNA-seq data and the original CEL and TXT files of microarray data are available, data quality control is first performed using TrimGalore to remove reads with a base mass fraction of less than 25 and a sequence length of less than 36 bp, and adapter sequences are removed and trimmed. Then use HISAT2 and SAMtools to align the readings to the human reference genome GNECODEv36 version and generate a BAM file; Subsequent counting data were calculated using FeatureCounts in the subread toolkit, and the raw count readings were converted to transcripts per kilobase. Raw microarray data from Affymetrix, Illumina, and Agilent were processed using the affy, lumi, and limma packages, respectively. When multiple probes were mapped to a single gene symbol, the mean was used. All gene expression values ​​were converted to a log2 scale. Expression profiles for which original data were not provided were downloaded directly from the original literature or public databases; all data were re-annotated based on GENCODERelease36.

4. The evaluation method according to claim 1, characterized in that, Step 2 specifically involves obtaining transcriptomics feature categories based on gene expression levels. These transcriptomics feature categories include pathways, transcription factors, cytokines, key immune interactions, immune cell composition, and intercellular communication. Meta-analysis was used to extract overlapping genes as candidate genes from transcriptomic feature categories.

5. The evaluation method according to claim 4, characterized in that, Step 3 specifically involves performing KNN imputation on the training queue composed of candidate genes to fill in missing values, while deleting genes with a missing rate exceeding 50%; and further using the ComBat method to correct batch effects in the training queue. Finally, 70% of the training queue was divided into the training set and 30% into the validation set.

6. The evaluation method according to claim 1, characterized in that, Step 4 specifically includes 12 feature selection methods, 20 machine learning models, 8 signature scoring algorithms, 12 feature selection methods, and 28 classification methods.

7. The evaluation method according to claim 5 or 6, characterized in that, Specifically, step 5 involves using the Spearman correlation coefficient based on the model from step 4 to remove genes in the training set with a correlation greater than 0.85 and the highest average absolute correlation. The 12 feature selection methods in the model from step 4 were used to further screen features that distinguish between treatment-responsive and non-responsive patients at the expression or co-expression level; The model in step 4 includes 20 machine learning models or 8 signature scoring algorithms; Therefore, each of the 336 models or schemes begins with a "route" consisting of one feature selection method and one classification method, using all possible combinations of the 12 feature selection methods and 28 classification methods from the model in step 4; the optimal model parameters are found using grid search and 5-fold 10-times repeated cross-validation, and the area under the receiver operating characteristic curve is used as the evaluation criterion for model parameter tuning.

8. The evaluation method according to claim 7, characterized in that, Step 6 specifically involves comparing all models and scoring methods in the validation set, and selecting the model with the highest AUC value and a difference of no more than 0.15 between the AUC values ​​of the training set and the validation set as the optimal solution. That is, the best model is the combination of Lasso and GSVA and the best model is the combination of RandomForest and ssGSEA.

9. A system for evaluating the response to taxane / immunotherapy in solid tumors based on ensemble machine learning, characterized in that, The evaluation system was designed to examine the classification capabilities of treatment response transcriptome data by constructing robust consensus prediction models, consisting of 336 possible models comprising 12 different feature selection routes and 28 classification methods.

10. An application of the method as described in any one of claims 1-8, particularly relating to the integration of multi-omics data and machine learning models in predicting tumor treatment response.

Images