Cancer drug sensitivity prediction method based on pathway activity and elastic net
By constructing a cancer drug sensitivity prediction method based on pathway activity and elastic network, the problem of lack of consideration of intergenic interaction relationships in the pathway in the prior art is solved, and higher-precision drug sensitivity prediction and better biological explanation are achieved.
Patent Information
- Application Number
- CN202011539984.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-23
- Publication Date
- 2025-04-08
- Estimated Expiration
- 2040-12-23
AI Technical Summary
The prior art lacks consideration of intergenic interaction relationships and key genes in the pathways when predicting cancer drug sensitivity, resulting in low repeatability of the prediction model.
Using a prediction method based on pathway activity and elastic network, the gene expression characteristic data of cancer cell lines and chemical characteristic data of anti-cancer drugs are obtained, and the pathway activity vector is calculated, and combined with the medicinal chemical characteristic matrix is combined to construct an elastic network model for prediction.
Improves the accuracy of drug sensitivity prediction, enables learning the potential relationships of anti-cancer drugs from multi-dimensional characteristics, has better biological explanations, and reduces the time and cost of clinical experiments.
Smart Images
Figure CN112599207B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of bioinformatics and machine learning, and specifically relates to a method for predicting cancer drug sensitivity based on pathway activity and elastic net regression, and more specifically to a method for predicting the sensitivity of 24 anticancer drugs to cancer cell lines. Background Art
[0002] Determining whether a patient will respond to an anticancer drug usually takes months, with repeated trials and the possibility of misuse. Developing effective methods to predict anticancer drug response has become an important topic in oncology research.
[0003] Currently, most methods predict drug sensitivity by establishing prediction models for single genes, without taking into account the interactions between genes in the pathway and the key genes in the pathway. Due to biodiversity and the difficulty in explaining the biological relevance of new gene-drug associations, the reproducibility of the prediction models is very low. Summary of the invention
[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for predicting cancer drug sensitivity based on pathway activity and elastic net. The model has high prediction accuracy and can effectively predict drug sensitivity. The method proposed in the present invention, which selects the expression level of key genes in the pathway to represent the pathway activity level, has a better prediction effect and better biological explanation for cancer drug sensitivity.
[0005] To achieve the above object, the solution of the present invention is:
[0006] A method for predicting cancer drug sensitivity based on pathway activity and elastic net, comprising the following steps:
[0007] S1: Obtain gene expression signature data of cancer cell lines and chemical signature data of anticancer drugs;
[0008] S2: Download and integrate the gene interaction network table in each pathway (gene set) from the KEGG pathway database, and select key genes that are closely connected to the gene expression signature data genes of cancer cell lines from the network table;
[0009] S3: Calculate the pathway activity vector of each pathway and merge all pathway activity vectors to obtain the pathway activity feature matrix;
[0010] S4: integrating the chemical feature data of the anticancer drugs in step S1 into a drug chemistry feature matrix, connecting and merging the drug chemistry feature matrix and the pathway activity feature matrix in step S3 to obtain a new matrix including cancer cell line features and drug chemistry features;
[0011] S5: Construct an elastic net model for predicting anticancer drug sensitivity based on steps S1-S4, input the new matrix containing cancer cell line characteristics and drug chemistry characteristics into the model as a feature tensor for training and prediction, and verify the reproducibility of the model.
[0012] As a preferred solution, in step S1, the chemical characteristics of anticancer drugs and the gene expression characteristics of cancer cell lines are collected and integrated from the public database on the Internet, corresponding to the known sensitivity levels of anticancer drugs, to form a two-dimensional tensor of chemical characteristics of drugs and a two-dimensional tensor of gene expression characteristics. At the same time, the pathway gene interaction network data is downloaded from the pathway database.
[0013] As a preferred solution, in step S2, the two-dimensional tensor of gene expression features of cancer cell lines and the two-dimensional tensor of chemical features of anticancer drugs are standardized, and then genes with higher degrees are selected from the pathway gene interaction network as key genes, and the mean / variance of these genes in gene expression is taken. Drug response is a synergistic behavior of multiple genes, and applying the key genes of the gene interaction network in the pathway to drug sensitivity will improve the prediction effect.
[0014] As a preferred solution, in step S3, based on all pathway data in step S1, the mean / variance activity vectors of the expression of all key genes in each pathway in step S2 in the cancer cell line sample are merged into a pathway activity vector. At this time, the feature is the pathway and the sample is the cancer cell line. Among them, each pathway activity vector is the mean / variance of the expression of the key gene in the pathway interaction network table in the gene expression matrix, and the feature in the pathway activity matrix after merging all pathways becomes the pathway.
[0015] As a preferred solution, in step S4, the pathway activity feature matrix and the drug chemistry feature matrix are merged into a new feature matrix, that is, a new matrix containing cancer cell line features and drug chemistry features, and used as an input value of the elastic net model for regression prediction.
[0016] As a preferred solution, in step S5, an elastic net model is constructed for the new matrix and drug sensitivity level obtained in step S4, and the new matrix containing cancer cell line characteristics and drug chemical characteristics is used as the model input value, and the drug sensitivity level is used as the model output value. The parameters are debugged and optimized according to the mean square error (MSE) to verify the model.
[0017] Due to the adoption of the above scheme, the beneficial effects of the present invention are:
[0018] First, the method of the present invention using pathway activity and elastic net to predict cancer drug sensitivity has higher accuracy for most drugs and can help clinical trials reduce time and money costs, thereby effectively predicting drug responses.
[0019] Second, the method proposed in the present invention combines the chemical characteristics of anticancer drugs and the pathway activity characteristics of cancer cell lines, which can learn the potential relationship of anticancer drugs from multi-dimensional characteristics and has a better biological explanation. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 Schematic diagram of pathway activity inference in an embodiment of the present invention.
[0021] Figure 2 This is the prediction performance of 24 drugs in the examples of the present invention on the validation set.
[0022] Figure 3 This is a pathway activity diagram of the drug PF2341066 between sensitive cell lines and resistant cell lines in an example of the present invention.
[0023] Figure 4 This is a diagram of the pathway activity of the drug 17-AAG between sensitive cell lines and resistant cell lines in an example of the present invention. DETAILED DESCRIPTION
[0024] The present invention provides a method for predicting cancer drug sensitivity based on pathway activity and elastic net.
[0025] The present invention will be further described below with reference to the embodiments.
[0026] Example:
[0027] The cancer drug sensitivity prediction method based on pathway activity and elastic net of this embodiment includes:
[0028] 1 Materials and analysis
[0029] 1.1 Data Source
[0030] The data used for model training and prediction in the present invention are derived from data sets publicly available on the Internet. Specifically, the anticancer drug chemical feature data set is derived from the results calculated by the alvaDesc software, the gene expression data and anticancer drug sensitivity values of cancer cell lines are derived from the database CCLE (Cancer Cell Line Encyclopedia https: / / portals.broadinstitute.org / ccle / ), and the pathway database is derived from the KEGG database (Kyoto Encyclopedia of Genes and Genomes https: / / www.genome.jp / kegg / ).
[0031] 1.2 Model Method
[0032] Figure 1This is the inferred pathway activity flow chart of the present invention. The present invention first calculates the expression of high-connectivity genes (key genes) in a pathway gene interaction network to obtain a pathway activity vector, then merges all pathway activity vectors to obtain a pathway activity feature matrix, and then merges the pathway activity feature matrix and the drug chemistry feature matrix into a new matrix, which is used as the input value of the elastic net model to finally predict drug sensitivity.
[0033] 1.2.1 Pathway activity inference
[0034] Inferring pathway activity involves the following steps:
[0035] Step 1: Standardize the gene expression data. Standardization is to subtract the mean of each dimension of the data tensor from all elements in that dimension, and then divide it by the standard deviation of each dimension. The new data tensor obtained is the standardized data.
[0036] Step 2: The mean / variance of the expression levels of genes with high connectivity in each pathway gene interaction network is used as the activity vector of the pathway.
[0037] Step 3: Merge all pathway activity vectors into a pathway activity feature matrix.
[0038] 1.2.2 Model Building
[0039] The pathway activity feature matrix and the drug chemistry feature matrix were merged into a new feature matrix, that is, a new matrix containing cancer cell line features and drug chemistry features, which was used as the input value of the elastic net model, corresponding to the cancer drug sensitivity IC50 value. A grid search algorithm was used to select the optimal parameters to predict the drug sensitivity level. At the same time, the top 10% of the pathway features were selected for analysis based on the ranking of the elastic net model coefficients.
[0040] 2 Results and analysis
[0041] 2.1 Pathway activity inference
[0042] Pathway activity inference takes into account the interaction between genes in the pathway, selects the genes in the top 10% of the degree in the pathway, and uses the gene relationship in the pathway to convert the gene features in the gene expression matrix into pathway features in the pathway activity matrix. The more than 20,000 gene features in the original gene expression matrix are converted into 388 pathway features, greatly reducing the dimensionality.
[0043] 2.2 Prediction of cancer drug sensitivity based on pathway activity and elastic net
[0044] The elastic net was used to construct a regression model. The input data was a new matrix containing cancer cell line characteristics and drug chemical characteristics. The output data was the drug sensitivity IC50 value. A prediction model of pathway activity for drug sensitivity was constructed, and the glmnet package in R language was used to simulate the prediction model.
[0045] The model is trained and tested and predicted on the test set. The experimental results on the validation set are as follows: Figure 2 As shown, it can be seen that the cancer drug sensitivity prediction method has a good prediction effect on most drugs. The experimental results obtained on the test set are shown in Table 1.
[0046] Table 1 Prediction performance of drug sensitivity prediction models based on pathway activity on the test set
[0047] drug MSE drug MSE drug MSE drug MSE 17-AAG 3.730 L-68545 3.464 Panobinostat 0.006 RAF265 8.241 AEW541 6.566 Lapatinib 7.265 PD-0325901 9.112 Sorafenib 2.753 AZD0530 7.287 LBW242 3.107 PD-0332991 3.601 TAE684 7.465 AZD6244 10.381 Nilotinib 2.437 PF2341066 4.488 TKI258 5.766 Erlotinib 5.765 Nutin-3 1.614 PHA-665752 1.814 Topotecan 4.256 lrinotecan 1.516 Paclitaxel 4.945 PLX4720 2.412 ZD-6474 6.595
[0048] 2.2 Biological explanations of pathways and drug sensitivity
[0049] Figure 3 For PF2341066, the 30 cell lines with the lowest and highest IC50 were selected as the sensitive and resistant groups, respectively, and the p-value of the t-test provided the difference in pathway activity. The TGF-beta signaling pathway ranked high in the drug PF2341066, and in fact, studies have found that activation of this pathway can lead to resistance to the drug PF2341066.
[0050] Figure 4 The data also showed the difference in pathway activity between sensitive and resistant cell lines of the drug 17-AAG, which is consistent with the facts.
[0051] The above description of the embodiments is to facilitate the understanding and use of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art based on the principles of the present invention without departing from the scope of the present invention should be within the protection scope of the present invention.
Claims
1. A method for predicting cancer drug sensitivity based on pathway activity and elastic net, characterized by: It includes the following steps: S1: Obtain gene expression signature data of cancer cell lines and chemical signature data of anticancer drugs; S2: Use the pathway database to obtain a network table of gene interactions in each pathway, and select key genes that are closely connected to the gene expression signature data of cancer cell lines from the network table; S3: Calculate the pathway activity vector of each pathway, and merge all pathway activity vectors to obtain the pathway activity feature matrix; S4: integrating the chemical feature data of the anticancer drugs in step S1 into a drug chemistry feature matrix, and connecting and merging the drug chemistry feature matrix and the pathway activity feature matrix in step S3 to obtain a new matrix including cancer cell line features and drug chemistry features; S5: Construct an elastic net model for predicting anticancer drug sensitivity based on steps S1-S4, input a new matrix containing cancer cell line characteristics and drug chemical characteristics into the model as a feature tensor for training and prediction, and verify the reproducibility of the model; In step S1, gene expression feature data of cancer cell lines and chemical feature data of anticancer drugs are collected and integrated from the Internet, the sensitivity level of anticancer drugs is learned from the multi-dimensional features of cancer cell lines, a two-dimensional tensor of chemical features of drugs and a two-dimensional tensor of gene expression features are formed, and pathway gene interaction network data are downloaded from the pathway database; In step S2, the two-dimensional tensor of gene expression characteristics of cancer cell lines and the two-dimensional tensor of chemical characteristics of anticancer drugs are standardized, genes with higher gene interaction network degrees in the pathway are selected as key genes, and the mean / variance of gene expression of these genes is taken; In step S3, the pathway activity vectors are merged according to the mean / variance activity vectors of the gene expression to obtain a pathway activity feature matrix, specifically: according to all pathway data in step S1, the mean / variance activity vectors of the expression of all key genes in each pathway in the cancer cell line sample in step S2 are merged into a pathway activity vector, wherein each pathway activity vector is the mean / variance of the expression of the key gene in the gene expression matrix in the pathway interaction network table, and the features in the pathway activity matrix after merging all pathways become pathways; In step S5, the elastic net model constructed in steps S2-S4 is integrated, a new matrix containing cancer cell line characteristics and drug chemistry characteristics is used as a model input value, and the drug sensitivity level is used as a model output value. The parameters of the model are debugged and optimized according to the mean square error, and training and verification are performed.