Method for predicting drug response of tumor cell lines by integrating multi-omics and essential genes

Abstract

The application discloses a tumor cell line drug response prediction method integrating multiple omics and essential gene information, relates to the field of tumor drug response prediction, and is a DROEG prediction method (Drug Response based on Omics and Essential Genes). Four kinds of omics data, including gene expression data, copy number variation data, methylation data and somatic mutation data, and CRISPR gene effect data (CERES Score) are introduced to construct a drug response model, and the effect of the model is comprehensively evaluated by a qualitative evaluation method and a quantitative evaluation method. The application introduces CRISPR essential gene information, establishes a drug response prediction model, is more accurate than existing methods, has the functions of quantitative and qualitative prediction and evaluation, is more suitable, and is more flexible to operate.

Description

Technical Field

[0001] This invention relates to the field of tumor drug response prediction, and more particularly to a method for predicting tumor cell line drug response that integrates multi-omics and essential gene information. Background Technology

[0002] Drug responses in tumors exhibit significant individual differences, making effective prediction of tumor drug responses a challenging problem in precision medicine. Currently, several methods utilize omics data to build models for predicting drug sensitivity, primarily categorized into quantitative and qualitative methods. Quantitative methods predict continuous IC50 values ​​of drug responses and then calculate Pearson or Spearman correlation coefficients between predicted and actual values ​​to evaluate overall predictive effectiveness. Geelher et al. preprocessed cell line-level and clinical-level gene expression data, removed batch effects using Combat, and employed a ridge regression model for drug response prediction. Validation was performed on three independent clinical datasets, yielding a relatively good predictive model and identifying gene characteristic variables with good predictive performance. Wang et al., based on the fundamental understanding that similar drugs produce similar drug responses in similar cell lines, incorporated drug structure similarity networks and gene expression similarity information into their model. They then constructed a similarity-regularized matrix factorization (SRMF) by compressing these data into latent variables. Zhang et al. constructed a similarity-heterogeneity network by calculating Pearson correlations of gene expression, drug chemical structure, drug target information, and protein-protein interaction networks among different cell lines. They then used an information flow-based algorithm on this network to predict drug responses. The rapid development of deep learning has made it possible to obtain more accurate drug response prediction models. MOFGCN first constructed a heterogeneous network and calculated drug fingerprint similarity, while fusing multi-omics data (gene expression data, copy number variation data, and mutation data) from cell lines into a cell line similarity network. Then, MOFGCN used graph convolution for feature extraction on the heterogeneous network, and finally reconstructed the association matrix using a linear correlation coefficient decoder to obtain the predicted drug response values. This model considered the influence of drugs, cell lines, and drug-cell line associations on the model's predictions. DrugCGN first constructed a gene map containing a PPI network graph and graph signals based on gene expression, and then learned a graph convolutional network with local filters to obtain the final drug response model. The tumor microenvironment and other confounding factors may also affect the actual drug response results in patients. Therefore, a deep learning model based on the VAE model was validated in TCGA clinical samples by compressing genetic features into a low-dimensional latent variable and using an elastic network for drug prediction.

[0003] On the other hand, qualitative methods mainly distinguish the actual state of drug response into sensitivity or resistance, generally by constructing a classification model and using classification evaluation indicators for model evaluation. The Merida method uses prior knowledge from the literature to annotate and binarize gene expression data, mutation data, copy number variation data, and drug response data. Then, it uses the ILP formula to screen for sensitivity / resistance-related mutation information before predicting drug response. MOLI is a multi-omics fusion model based on deep neural networks. This model takes gene expression data, copy number variation, and mutation data as input and generates three independent subnetworks. The features of these three networks are then compressed to produce a final network, which serves as the input to the classification subnetwork, finally obtaining the predicted value of the drug response. From the above research on tumor drug response models, it can be seen that most current drug response prediction methods only perform either quantitative or qualitative assessments, and cannot simultaneously perform both types of assessments. Furthermore, the accuracy of these models needs further improvement.

[0004] Furthermore, current methods do not consider information on genes essential for tumor cell growth. Genome-wide CRISPR loss-of-function screening is crucial for understanding gene function in normal and disease states. Incorporating information on essential CRISPR genes holds promise for improving the effectiveness of drug response prediction models. The Depmap database provides the latest genome-wide loss-of-function screening data and drug dependence data for cancer cell lines. This database offers a list of essential genes selected based on CRISPR Cas9 screening, i.e., genes whose knockout significantly impacts tumor cell growth. We will utilize this essential gene information and integrate it with multi-omics to construct new and more accurate tumor drug response prediction models.

[0005] Existing technologies do not utilize the essential CRISPR gene information of tumors; they can only perform quantitative or qualitative predictions, and the accuracy needs to be improved; existing methods cannot discover prognostic biomarkers for specific drug treatments; and existing methods using multi-omics data do not systematically compare the contribution of data at various levels to the model's predictive performance.

[0006] Therefore, those skilled in the art are dedicated to developing a tumor cell line drug response prediction method based on the integration of multi-omics and CRISPR essential gene information, which can simultaneously predict and evaluate both continuous drug response values ​​and discrete values ​​of sensitivity or resistance. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to provide a tumor cell line drug response prediction method based on the integration of multi-omics and CRISPR essential gene information, which can simultaneously predict and evaluate both continuous values ​​of drug response and discrete values ​​of sensitivity or resistance.

[0008] To achieve the above objectives, this invention provides a method for predicting drug response in tumor cell lines, the DROEG prediction method, which is a drug response prediction method based on omics and essential genes. It constructs a drug response model by incorporating four types of omics data: gene expression data (mRNA), copy number variation data (CNV), methylation data, and somatic mutation data, as well as CRISPR gene effect data (CERES Score). The effectiveness of the model is comprehensively evaluated through qualitative and quantitative evaluation methods.

[0009] Furthermore, quantitative evaluation methods are based on predicting drug response values, and then evaluate the drug response by calculating the correlation index between the predicted and actual values; qualitative evaluation methods process drug response values ​​into categorical variables, then predict drug response by constructing a categorical model, and finally evaluate the drug response by categorical indicators.

[0010] Furthermore, the drug response value is the IC50 value or ActArea; the categorical variable is a binary variable of drug resistance or sensitivity; the categorical indicators are accuracy, precision, specificity, sensitivity, F1 score, and Matthew correlation coefficient.

[0011] Furthermore, the above method includes the following steps:

[0012] Step 1: Data preprocessing, processing of omics data and drug response data: The data provided by GDSC / CCLE has undergone preliminary preprocessing.

[0013] On the other hand, for drug sensitivity data, K-means clustering is used to cluster the actual drug sensitivity data IC50 / Actarea values ​​into two categories and obtain the corresponding clustering thresholds. Then, the predicted values ​​are also binarized according to the thresholds to convert them into sensitive or resistant drugs, and then compared with the actual response to evaluate the model.

[0014] Step 2: Model feature selection. The selection of model features is divided into two parts. The first part is feature selection based on prior knowledge from the literature. Prior knowledge is taken from important molecular feature information reported in the literature and essential gene information provided by various databases. The second part is further selection based on statistical methods. Among the features selected based on the literature in the first part, the feature variables with the highest correlation with drug sensitivity data are selected for inclusion in the model.

[0015] Step 3, Model Fitting and Parameter Selection: After completing the two parts of variable feature selection in Step 2, the selected feature variables are incorporated into the Support Vector Regression (SVR) model for parameter selection;

[0016] Step 4, Model Prediction: Under the parameters selected in Step 3, the final drug prediction uses the Leave-One-out-of-Vote (LOOCV) method to predict the drug sensitivity of each drug separately and obtain the predicted drug sensitivity value.

[0017] Step 5, Model Evaluation: Model evaluation includes quantitative evaluation methods and qualitative evaluation methods. The quantitative evaluation method calculates the Pearson correlation coefficient between the actual drug response value and the predicted response value. The qualitative evaluation method first evaluates the predicted drug sensitivity value obtained in Step 4 by binarizing it according to the clustering threshold in Step 1.

[0018] Furthermore, the omics data in step 1 includes CRISPR gene effect data, gene expression data, copy number variation data, gene mutation data, and methylation data. Gene expression data uses RMA-normalized baseline expression levels. Copy number variation data is processed into ternary variables of -1, 0, and 1, where 0 represents normal copy number, and -1 and +1 represent gene copy loss and increase, respectively. Gene mutation binary data can be directly obtained through matrix transformation to represent mutation occurrence and non-mutation, respectively. Methylation data (CpG island methylation) is obtained by preprocessing β values.

[0019] Furthermore, data preprocessing, specifically the processing of omics data (including CRISPR gene effect data) and drug response data: The data provided by GDSC / CCLE has already undergone preliminary preprocessing. Gene expression data uses RMA-normalized baseline expression levels. Copy number variation data is processed into ternary variables of -1, 0, and 1, where 0 represents normal copy number, and -1 and +1 represent gene copy loss and increase, respectively. Binary data on gene mutations can be directly obtained through matrix transformations, representing occurrence and absence of mutations, respectively. Methylation data (CpG island methylation) is obtained through preprocessing β values.

[0020] On the other hand, for drug sensitivity data (IC50 or ActArea), in order to perform qualitative analysis, K-means clustering is used to cluster the actual drug sensitivity data IC50 / Actarea values ​​into two classes and obtain the corresponding clustering thresholds. Then, based on the thresholds, the predicted values ​​are also binarized to convert them into sensitive or resistant values, and then compared with the actual responses for model evaluation. The definition of K-means is shown in Equation 1:

[0021]

[0022] Furthermore, in step 2, the gene expression data in the first part were screened using the IntOGen tumor driver gene catalog, the copy number variation data were referenced from the Sanchez-Vega copy number variation catalog, the somatic mutation data of gene mutations were annotated and screened using the OncoKB, CGI, and COSMIC database information annotation catalogs, the methylation data were screened using the catalog provided by the MethyCancer database, and the CRISPR data were screened using the essential gene catalog of the Depmap database. In the second part, the continuous variables of the characteristic variables were screened by calculating the Pearson correlation coefficient, while the continuous and discrete variables were screened by calculating the F-statistic using analysis of variance. The screening index was that variables greater than the mean ± two standard deviations were included in the model.

[0023] Furthermore, step 3 also includes selecting model parameters for each drug using five-fold cross-validation, with a Gaussian kernel as the kernel function, and model parameters including regularization coefficient C and gamma value.

[0024] Furthermore, in step 5, the qualitative evaluation method transforms the regression problem into a classification problem, and evaluates the model by calculating accuracy, precision, specificity, sensitivity, F1 score, and Matthew correlation coefficient classification index.

[0025] Furthermore, the above method also includes a comparative analysis of the contribution of five types of data—gene expression data (mRNA), copy number variation data (CNV), methylation data, somatic mutation data, and CRISPR gene effect data (CERES Score)—to the model's predictive performance. By removing one type of data from each of the five types and constructing a drug response model based on the remaining four types of data, the changes in the predictive results of the model are compared with those of the complete model, thereby measuring the importance of each level of data.

[0026] Furthermore, the importance of the five types of data in contributing to the model is ranked as follows: CRISPR essential gene information, mRNA expression data, SNV mutation data, DNA methylation data, and CNV copy number variation data.

[0027] In a preferred embodiment 1 of the present invention, the principle and process of a tumor cell line drug response prediction method integrating multi-omics and essential gene information are described in detail.

[0028] In another preferred embodiment 2 of the present invention, the qualitative and quantitative results evaluation of the prediction of DROEG drug sensitivity using the two datasets GDSC and CCLE are described in detail.

[0029] In another preferred embodiment of the present invention, the process of comparing and analyzing the contribution of five types of data to the model prediction effect is described in detail.

[0030] The beneficial technical effects of this invention are as follows:

[0031] 1. For the first time, essential CRISPR gene information is introduced and integrated with genomic, transcriptomic, and methylome data to establish a drug response prediction model that is more accurate than existing methods;

[0032] 2. A method that can simultaneously predict and evaluate both continuous values ​​of drug response and discrete values ​​of sensitivity or resistance, possessing both quantitative and qualitative prediction and evaluation functions, with wider applicability and more flexible operation.

[0033] 3. By comparing the contributions of omics data at different levels to the predictive performance of drug response models, it was revealed that the information on essential genes for tumor growth has the greatest impact on drug sensitivity, confirming that the information on essential genes for CRISPR has the most important influence on the predictive ability of the model, which has not been explored by existing methods.

[0034] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0035] Figure 1 This is a flowchart of a preferred embodiment 1 of the present invention, which is a method for predicting drug response in tumor cell lines that integrates multi-omics and essential gene information;

[0036] Figure 2 This is a graph showing the comparative analysis results of the contribution of five types of data to the model prediction effect in a preferred embodiment 2 of the present invention. Detailed Implementation

[0037] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0038] Example 1: Principle of a Tumor Cell Line Drug Response Prediction Method Integrating Multi-omics and Essential Gene Information

[0039] This paper introduces CRISPR-essential gene information and integrates it with genomic, transcriptomic, and methylome data to construct a novel method for predicting drug responses in tumor cell lines, DROEG (Drug Response based on Omics and Essential Genes), using support vector regression (SVR). This method possesses both quantitative and qualitative prediction and evaluation capabilities. The quantitative evaluation method is based on predicted drug response values ​​(usually IC50 values ​​or ActArea), followed by evaluation using correlation indices between predicted and actual drug response values. The qualitative evaluation method typically processes drug response values ​​into categorical variables (usually a binary variable of sensitivity / resistance), then constructs a classification model for drug response prediction, and finally evaluates the response using classification indices such as accuracy, precision, specificity, sensitivity, F1 score, and Matthew correlation coefficient. The workflow of this tumor cell line drug response prediction method integrating multi-omics and essential gene information is as follows: Figure 1 As shown, it mainly includes the following steps:

[0040] (1) Data preprocessing, i.e., processing of raw omics data (including CRISPR gene effect data) and drug response data: First, the cell lines corresponding to drug sensitivity data are intersected with the cell lines corresponding to the five omics data (i.e., copy number variation, somatic mutation, mRNA expression, methylation, and CRISPR essential genes) to obtain common cell lines for subsequent analysis. Drugs are screened based on these common cell lines; if the missing sensitivity data of a drug cell line exceeds 20%, that drug is not included.

[0041] On the other hand, for drug sensitivity data (IC50 or ActArea), K-means clustering is used to cluster the original drug sensitivity data into two classes and obtain the corresponding clustering thresholds. Then, the predicted sensitivity values ​​are binarized according to the thresholds and the model is evaluated.

[0042] (2) Model construction and feature selection: A support vector regression machine learning model was constructed, with copy number variation, somatic mutation, mRNA expression, methylation and CRISPR essential gene information as input. Feature selection is divided into two parts. The first part is based on prior knowledge (important molecular feature information reported in the literature and essential gene information provided by various databases). Specifically, gene expression data is selected through the IntOGen tumor driver gene catalog, copy number variation data is selected through the copy number variation catalog of Sanchez-Vega et al.

[76] , somatic mutation data is annotated and selected through the information annotation catalogs of databases such as OncoKB, CGI and COSMIC, methylation data is selected through the catalog provided by the MethyCancer database, and CRISPR is selected through the essential gene catalog of the Depmap database. The second part uses statistical methods for further screening. In the first part, which uses prior knowledge to screen features, we consider selecting the feature variables that are most correlated with drug sensitivity data to include in the model. Among them, continuous variables are screened by calculating the Pearson correlation coefficient, while continuous variables and discrete variables are screened by calculating the F-statistic through analysis of variance. The screening index is that variables greater than the mean ± two standard deviations are included in the model.

[0043] (3) Model Fitting and Parameter Selection: After completing the above two parts of variable feature screening, the screened feature variables are incorporated into the SVM model to predict drug sensitivity. Specifically, five-fold cross-validation is used for each drug to select model parameters for a single drug. The kernel function is a Gaussian kernel, and the model parameters include regularization coefficient C, gamma value, and other indicators.

[0044] (4) Model prediction: Under the parameters selected above, the final drug prediction adopts the leave-one-out method (LOOCV) to predict the drug sensitivity of each drug separately and obtain the drug sensitivity prediction value.

[0045] (5) Model Evaluation: Model evaluation includes quantitative evaluation methods and qualitative evaluation methods. The quantitative evaluation method is to calculate the Pearson correlation coefficient between the actual drug response value and the predicted response value. The qualitative evaluation method first evaluates the predicted drug sensitivity value obtained in (4) by binarizing it according to the clustering threshold in (1) above, that is, transforming the regression problem into a classification problem, and then evaluating the model by calculating classification indicators such as accuracy, precision, specificity, sensitivity, F1 score, and Matthew correlation coefficient.

[0046] Example 2: Qualitative and quantitative results evaluation of DROEG drug sensitivity prediction using two datasets, GDSC and CCLE.

[0047] For the GDSC dataset, 214 drugs were selected for evaluation based on the principle that the missing values ​​of cell line sensitivity data for a single drug should not exceed 20%. Furthermore, considering that only cell lines sharing five omics data points (including CRISPR-essential genes) with the GDSC drug response data were eligible for subsequent analysis, a final selection of 495 cell lines was obtained. For the CCLE dataset, the same method was used to select 24 drugs for drug sensitivity data and corresponding omics data in 264 cell lines.

[0048] We compared the prediction performance of DROEG with several major prediction methods, including MOFGCN, HNMDRP, SRMF, and VAEN. MOFGCN, HNMDRP, and SRMF can only perform qualitative classification predictions, while VAEN can only predict quantitative continuous response values. Our DROEG method, however, can perform both quantitative and qualitative predictions simultaneously. Table 1 shows that the DROEG model performs best overall on both the GDSC and CCLE datasets, especially on the CCLE dataset, where it significantly outperforms other methods in both qualitative and quantitative prediction evaluations. For classification evaluation metrics, DROEG outperforms other methods in accuracy, precision, F1 score, and Matthew correlation coefficient. For quantitative prediction, the Pearson correlation coefficient (PCC) is significantly higher than that of VAEN. In the GDSC dataset, 87.85% of the drug predictions have a correlation greater than 0.5, while in the CCLE dataset, the PCC for all drugs is greater than 0.5, ranging from 0.55 to 0.73.

[0049] Table 1 Comparison of the performance of the DROEG model with other models

[0050]

[0051] We further explored the effect of CRISPR-essential gene information on the predictive performance of the DROEG model. First, we evaluated the performance of the DROEG model on the GDSC2 and CCLE datasets with and without CRISPR-essential gene data. The model results show that when CRISPR-essential gene information was included in DROEG, the sensitivity, specificity, precision, and accuracy were all significantly improved (p<0.05). In the GDSC dataset with 173 drugs, the average sensitivity, specificity, precision, and accuracy of the model with CRISPR-essential genes included were 0.6538, 0.7273, 0.7316, and 0.7388, respectively, while those without were 0.6007, 0.6994, 0.6916, and 0.7057. In the CCLE dataset, when the essential CRISPR genes are included, these four metrics are 0.8775, 0.5271, 0.8215, and 0.8045, respectively; when they are not included, they are 0.8644, 0.3688, 0.7725, and 0.7558, respectively.

[0052] Example 3 compares and analyzes the contribution of five types of data to the model's predictive performance.

[0053] We constructed drug response models based on the remaining four data points by removing one data point at a time, and compared their predictive results with the complete model to measure the importance of each data level. A greater decrease in a particular indicator after removing a specific omics indicates a more significant improvement in the model's performance due to that omics. First, we compared the predictive performance of the model with the data from any omics removed to the complete DROEG model. The comparison results are as follows: Figure 2 As shown, the significance level indicates whether the four metrics (accuracy, precision, sensitivity, and specificity) of the model constructed after removing a certain omics data are significantly different from the four metrics of the complete DROEG model. * represents p < 0.05, ** represents p < 0.01, *** represents p < 0.001, and NS represents not significant. It can be seen that CRISPR essential gene information has the greatest impact on the model's predictive performance, followed by mRNA data and mutation data.

[0054] Furthermore, we compared models removing any one omics with models removing another, using paired t-tests to determine statistical significance. We found that removing CRISPR-essential gene information had the greatest impact on model performance. The model with CRISPR-essential information removed showed significantly lower performance across all four metrics compared to models removing any other omics (p<0.05), indicating that CRISPR-essential gene information most significantly improves model predictive performance. Additionally, in models removing transcriptome data (mRNA), precision and specificity were significantly lower (p<0.05) than those removing methylation data. The importance of the five data types in contributing to the model, ranked as follows: CRISPR-essential gene information, mRNA expression data, SNV mutation data, DNA methylation data, and CNV copy number variation data. Since CRISPR-essential genes are based on tumor cell lethality information obtained after gene knockout, they more accurately and comprehensively reflect cellular genetic background and gene importance compared to gene expression information, thus having the most significant impact on model improvement. This fully demonstrates the value of our new method in incorporating CRISPR-essential gene information.

[0055] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for predicting drug response in tumor cell lines, characterized in that, The prediction method is the DROEG prediction method (Drug Response based on Omics and Essential Genes), which constructs a drug response model by incorporating four types of omics data: gene expression data (mRNA), copy number variation data (CNV), methylation data, and somatic mutation data, as well as CRISPR gene effect data (CERES Score). The model's effectiveness is comprehensively evaluated using both qualitative and quantitative methods. The drug response value is the IC50 value or ActArea; the categorical variable is a binary variable of drug resistance or sensitivity; the categorical indicators are accuracy, precision, specificity, sensitivity, F1 score, and Matthew correlation coefficient. The method includes the following steps: Step 1: Data preprocessing. This involves processing the omics data and the drug response data, specifically performing preliminary preprocessing on the data provided by GDSC / CCLE. For drug sensitivity data, K-means clustering is used to cluster the actual drug sensitivity data (IC50 value / Actarea value) into two categories and obtain the corresponding clustering thresholds. Subsequently, the predicted drug sensitivity values ​​are also binarized and converted into sensitive or resistant values ​​based on the thresholds, and then compared with the actual response for model evaluation. Step 2: Model feature screening. The screening of model features is divided into two parts. The first part is based on prior knowledge from the literature, which is taken from important molecular feature information reported in the literature and essential gene information provided by various databases. The second part is based on statistical methods for further screening. Among the features screened in the first part based on the literature, the feature variables that are most correlated with the drug sensitivity data are selected and included in the model. Step 3, Model Fitting and Parameter Selection: After completing the two-part variable feature screening in Step 2, the screened feature variables are incorporated into the Support Vector Regression (SVR) model, and the parameters are selected to obtain the selected parameters. Step 4, Model Prediction: Under the parameters selected in Step 3, the final drug prediction uses the Leave-One-out-of-Vote (LOOCV) method to predict the drug sensitivity of each drug separately, and obtains the predicted drug sensitivity value. Step 5, Model Evaluation: Model evaluation includes the quantitative evaluation method and the qualitative evaluation method; wherein the quantitative evaluation method is to calculate the Pearson correlation coefficient between the actual drug response value and the predicted response value; and the qualitative evaluation method first evaluates the predicted drug sensitivity value obtained in Step 4 by binarizing it according to the clustering threshold in Step 1.

2. The prediction method as described in claim 1, characterized in that, The quantitative evaluation method is based on predicting drug response values, and then evaluates the drug response by calculating the correlation index between the predicted and actual values; the qualitative evaluation method processes the drug response values ​​into categorical variables, then constructs a classification model to predict the drug response, and finally evaluates it using classification indicators.

3. The prediction method as described in claim 1, characterized in that, In step 1, the gene expression data included in the omics data adopts the baseline expression level normalized by RMA; the copy number variation data is processed into a ternary variable of -1, 0, 1, where 0 represents the normal copy number, and -1 and +1 represent the loss and increase of gene copies, respectively; the binary data of gene mutation is directly obtained through matrix transformation to represent the occurrence and non-occurrence of mutation, respectively; and the methylation data is obtained by preprocessing β values.

4. The prediction method as described in claim 1, characterized in that, In step 2, the gene expression data in the first part is screened using the IntOGen tumor driver gene catalog, the copy number variation data is referenced from the Sanchez-Vega copy number variation catalog, the somatic mutation data of gene mutations is annotated and screened using the OncoKB, CGI, and COSMIC database information annotation catalogs, the methylation data is screened using the catalog provided by the MethyCancer database, and the CRISPR gene effect data is screened using the essential gene catalog of the Depmap database. In the second part, the continuous variables of the feature variables are screened by calculating the Pearson correlation coefficient, while the continuous variables and discrete variables are screened by calculating the F-statistic using analysis of variance. Variables greater than the mean ± two standard deviations are included in the model.

5. The prediction method as described in claim 1, characterized in that, Step 3 also includes selecting model parameters for each drug using five-fold cross-validation, with a Gaussian kernel as the kernel function, and the model parameters including the regularization coefficient C and the gamma value.

6. The prediction method as described in claim 1, characterized in that, The qualitative evaluation method in step 5 transforms the regression problem into a classification problem, and evaluates the model by calculating the accuracy, precision, specificity, sensitivity, F1 score, and Matthew correlation coefficient.

7. The prediction method as described in claim 1, characterized in that, The method further includes a comparative analysis of the contribution of five types of data—gene expression data (mRNA), copy number variation data (CNV), methylation data, somatic mutation data, and CRISPR gene effect data (CERES Score)—to the model's predictive performance. By removing one type of data from each of the five types and constructing a drug response model based on the remaining four types of data, the changes in the predictive results of the model are compared with those of the complete model, thereby measuring the importance of each level of data.

8. The prediction method as described in claim 7, characterized in that, The importance of the five types of data in contributing to the model is ranked as follows: CRISPR gene effect data, gene expression data, somatic mutation data, methylation data, and copy number variation data.

Images