Construction method and system of anticancer drug sensitivity prediction model

Abstract

The application provides a construction method and system of an anticancer drug response prediction model based on machine learning, the model proposes the concept of a drug genome space, uses matrix decomposition to project drug and cell line data to a high-dimensional latent space, so as to extract biological characteristics of the drug and the cell line, and has high biological interpretability. The model of recommendation learning is compared with multiple models including a traditional machine learning model to evaluate the ability of the model. In addition to correlation verification, the application first uses a weighted ranking indicator (NDCG) frequently used in a recommendation system to evaluate the ability of the model to recommend sensitive drugs for new cell lines, and the ability can assist doctors in selecting targeted drugs in the clinic. In the two verification evaluation modes, the model based on recommendation learning all shows good prediction ability, and can more accurately and reliably predict the sensitivity of anticancer drugs.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning research and computer development of calculation model, and particularly relates to a method for improving the performance of a machine learning model and predicting anticancer drug sensitivity based on a machine learning algorithm. BACKGROUND

[0002] Cancer is a highly complex and diverse disease, which is caused by abnormal cell proliferation and malignant transformation, leading to dysfunction of tissues and organs. Generally, normal cells will proliferate and divide within limits according to the needs of the body, and orderly die after completing their functions. However, cancer cells will proliferate and spread unlimitedly, forming malignant tumors. Cancer can occur in any part of the body tissue and organ, seriously affecting the quality of life of patients, and can endanger life. The basic goal of the era of precision medicine is to design targeted drugs and determine individual chemotherapy to maximize drug efficacy while minimizing adverse side effects. How to accurately calculate and predict the response of cancer patients to various targeted drug treatments based on the molecular and clinical characteristics of patients has become a crucial basic problem in the era of precision medicine.

[0003] In recent years, the number of patients with molecular data and tumor biological information in cancer information has been increasing. At the same time, large-scale researches such as Cancer Genome Atlas (TCGA) and International Cancer Genome Consortium (ICGC) have integrated data information, and cancer-driven gene identification research has also benefited greatly. However, these data sources only contain gene information of cancer groups, and usually lack of drug response information, so they are not suitable for identifying biomarkers of drug response. On the other hand, many research works including Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) project have screened drugs on multiple cancer cell lines. These cell line datasets enable research to utilize the information characteristics of the genome, apply mathematical and statistical methods to decipher the functional relationship between drug response and genomic information, and construct models that can predict precise patient-specific drug response.

[0004] Researchers have proposed various types of models that fully utilize genomic features to predict drug responses. Among these, the most widely used model is the drug-specific model. This model is trained independently for each drug, based on the genetic information of cell lines and drug response data tested individually. Models in this category generally use structurally similar datasets, often based on baseline gene expression or a combination of gene expression and other genomic information (such as copy number alterations and DNA methylation). Model architectures vary, including linear regression models, neural networks, random forests, support vector machines and kernel regression based on multiple genomic information, and neural network models incorporating drug attribute information. Drug-specific models are typically limited by the number of cell lines tested with a given drug. To increase the number of data points and obtain a more robust and general drug sensitivity model, Costello et al. proposed the Bayesian Multitask Multikernel Learning (BMTMKL) method, which demonstrated state-of-the-art performance in the DREAM challenge for drug response prediction. This work highlights the importance of sharing information across drugs for improving the accuracy of drug response prediction.

[0005] Although the aforementioned studies have successfully predicted the accuracy of anticancer drugs using various methods, the latest version of the "Guidelines for Clinical Application of New Antitumor Drugs (2020 Edition)" proposes that the use of antitumor drugs should not only be based on the results of pathological diagnosis, but also follow the principle that targeted drug therapy should only be used after gene testing; that is, targeted therapy, testing first. Precision cancer diagnosis and treatment cannot be separated from the development of testing technologies and their widespread clinical application. Improving the accuracy and sensitivity of testing technologies is crucial for precision diagnosis and treatment. Machine learning has the ability to learn from data and experience, providing important technical support for many interdisciplinary fields. Bioinformatics research includes data acquisition, data management, data analysis, and simulation experiments; machine learning can be used to extract necessary knowledge from large amounts of biological data for learning and prediction.

[0006] Cancer drug sensitivity testing applies computer technology to the field of biology to improve the prediction of cancer drug sensitivity and the identification of biomarkers related to drug response, thereby accelerating drug discovery and saving the lives of cancer patients. This research aims to improve the performance of machine learning models based on current research, thereby enhancing the efficiency and accuracy of cancer drug sensitivity prediction. Summary of the Invention

[0007] The purpose of this invention is to address the technical deficiencies in the existing technology for predicting drug sensitivity by providing a performance-based method for predicting drug sensitivity based on a machine learning model.

[0008] The technical solution adopted to achieve the purpose of this invention is:

[0009] A method for constructing an anticancer drug sensitivity prediction model includes the following steps:

[0010] Step 1, Dataset Construction and Preprocessing:

[0011] Step 1-1: The dataset is from two studies, CCLE and GDSC.

[0012] Steps 1-2 involve recalculating the IC50 values ​​for the two datasets.

[0013] Steps 1-3: Compare the IC50 values ​​obtained from the two datasets to determine the final dataset;

[0014] Step 2, construct a model for drug response calculation:

[0015] Step 2-1: Based on the idea of ​​using latent factors for feature extraction in collaborative filtering recommendation algorithm, construct the drug response matrix W and decompose it into bias S, cell line matrix C and drug matrix M;

[0016] Step 2-2: Use the decomposed drug matrix M to learn the pharmacogenomics space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response.

[0017] Steps 2-3: Construct the relationship between the drug response matrix W, the cell line matrix C, and the drug matrix M to determine the potential association between cell lines and drug genomes;

[0018] Step 3: Train and evaluate the drug response calculation model;

[0019] The optimized sum of squared errors loss function was selected as the optimization algorithm for the drug response calculation model; the sum of squared errors loss function is:

[0020]

[0021]

[0022] in and W ij These are the sensitivity scores of cell line i observed and predicted using drug j, and the drug response matrix, respectively. It represents the number of drug response experiments in the training dataset;

[0023] The method for optimizing the sum of squared errors loss function is as follows: The sum of squared errors loss function is optimized using gradient descent; all parameters in the drug response matrix W are obtained based on the measured drug response values; parameters M P MQ The gradient function is derived as follows:

[0024]

[0025] In the drug response calculation model, initialization M is 0 P and M Q Using the minimum random number, perform batch gradient descent using the gradient calculation formula described above to optimize... M P and M Q .

[0026] Then, the evaluation method for the drug response calculation model was determined, specifically: five-fold cross-validation was used to compare the correlation between the drug response calculation model results and the known responses of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; mean normalized depreciation cumulative gain (NDCG) was also used to evaluate the drug response calculation model for each cell line, and the calculation formula is as follows:

[0027]

[0028] Where g is the predicted rank of the drug tested on the cell line, and w is the list of observed drug sensitivity scores. It is a known drug ranking calculated based on measured drug response values; the NDCG value ranges from 0 to 1, where 1 indicates that the drug response calculation model correctly predicted the drug's ranking.

[0029] Furthermore, in step 1-1, the GDSC and CCLE datasets are used as benchmark datasets.

[0030] Furthermore, in steps 1-2, for the CCLE and GDSC datasets, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method.

[0031] In CCLE, each cell line is tested using six different concentrations of the drug, and a cell activity value (AV) is measured at each concentration; in GDSC, a cell activity value (AV) is measured using seven different concentrations of the drug.

[0032]

[0033] Where T, P, and N represent the intensity values ​​measured in the test drug, positive control, and negative control, respectively;

[0034] The JAGS method was used to fit the Bayesian sigmoid curves to obtain the dose-response curves for each drug and cell line pair:

[0035]

[0036] Define a as cell line ID, b as drug ID, and x as... ab For logarithmic-scale dose, y ab As a result of the response to a given drug and cell line pair, This represents precision, which is the reciprocal of the variance. Represents the partial derivatives of the input variables; using the parameter γ 0b and γ 1b Two parameters are used to simulate the distribution of the control scale parameter θ; where γ 0b The center of the drug response curve, i.e., the IC50 value, γ 1b Used to capture the steepness of a curve.

[0037] Furthermore, in steps 1-3, the final dataset is determined by recalculating the IC50 value; after recalculating the IC50 value, a paired p-test is first used to measure the paired differences of the variable. Assuming that the difference between the two IC50 values ​​is 0, the obtained p-value is less than 0.01, and a decision can be made not to reject the null hypothesis.

[0038] Furthermore, in step 2-1, a collaborative filtering algorithm is used to analyze the relationships between cell lines and the dependencies between drugs to identify new drug response associations; suppose a user group consists of m users {u1, u2, ..., u...} m A list of n items {h1, h2, ..., h} and n items {h1, h2, ..., h} n The list is composed of items, and each user u1 has a corresponding list of items. This is used to record the user's ratings of these items, or to record the user's preferences inferred by the model from the user's behavior.

[0039] Furthermore, in step 2-2, matrix factorization is used to learn the pharmacogenomics space, projecting drug and cell line data into a high-dimensional latent space. The cell line-specific drug response is provided by the dot product between the cell line vector and the drug vector, and the drug response model is calculated according to the following formula:

[0040]

[0041] M P It is a transformation matrix that transforms cell features x i Projected onto the potential space, These are the bias terms for drug j and cell line i, respectively, and k j and d iM is the vector of drug j and cell line i in the n-dimensional latent space. P It is a transformation matrix that transforms cell features x i Projected onto the potential space.

[0042] Furthermore, in step 3, in order to better train the model, a drug sensitivity score coefficient w = -log(IC50) is defined. The higher the score, the more sensitive the cell line is to the drug. Through performance testing, the model can extract more biological features to predict the reaction results, making the results more accurate.

[0043] This invention also provides a system for constructing an anticancer drug sensitivity prediction model, comprising,

[0044] The dataset building module is used to build and preprocess datasets, including the following processes:

[0045] Step 1-1: The dataset is from two studies, CCLE and GDSC.

[0046] Steps 1-2 involve recalculating the IC50 values ​​for the two datasets.

[0047] Steps 1-3: Compare the IC50 values ​​obtained from the two datasets to determine the final dataset;

[0048] The model building module is used to build models for drug response calculations, and includes the following processes:

[0049] Step 2-1: Based on the idea of ​​using latent factors for feature extraction in collaborative filtering recommendation algorithm, construct the drug response matrix W and decompose it into bias S, cell line matrix C and drug matrix M;

[0050] Step 2-2: Use the decomposed drug matrix M to learn the pharmacogenomics space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response.

[0051] Steps 2-3: Construct the relationship between the drug response matrix W, the cell line matrix C, and the drug matrix M to determine the potential association between cell lines and drug genomes;

[0052] The training module is used to train and evaluate the drug response calculation model;

[0053] The optimized sum of squared errors loss function was selected as the optimization algorithm for the drug response calculation model; the sum of squared errors loss function is:

[0054]

[0055]

[0056] in and W ij These are the sensitivity scores of cell line i observed and predicted using drug j, and the drug response matrix, respectively. It represents the number of drug response experiments in the training dataset;

[0057] The method for optimizing the sum of squared errors loss function is as follows: The sum of squared errors loss function is optimized using gradient descent; all parameters in the drug response matrix W are obtained based on the measured drug response values; parameters M P M Q The gradient function is derived as follows:

[0058]

[0059] In the drug response calculation model, initialization M is 0 P and M Q Using the minimum random number, perform batch gradient descent using the gradient calculation formula described above to optimize... M P and M Q .

[0060] Then, the evaluation method for the drug response calculation model was determined, specifically: five-fold cross-validation was used to compare the correlation between the drug response calculation model results and the known responses of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; mean normalized depreciation cumulative gain (NDCG) was also used to evaluate the drug response calculation model for each cell line, and the calculation formula is as follows:

[0061]

[0062] Where g is the predicted rank of the drug tested on the cell line, and w is the list of observed drug sensitivity scores. It is a known drug ranking calculated based on measured drug response values; the NDCG value ranges from 0 to 1, where 1 indicates that the drug response calculation model correctly predicted the drug's ranking.

[0063] Furthermore, in step 1-1, the GDSC and CCLE datasets are used as benchmark datasets.

[0064] Furthermore, in steps 1-2, for the CCLE and GDSC datasets, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method.

[0065] In CCLE, each cell line is tested using six different concentrations of the drug, and a cell activity value (AV) is measured at each concentration; in GDSC, a cell activity value (AV) is measured using seven different concentrations of the drug.

[0066]

[0067] Where T, P, and N represent the intensity values ​​measured in the test drug, positive control, and negative control, respectively;

[0068] The JAGS method was used to fit the Bayesian sigmoid curves to obtain the dose-response curves for each drug and cell line pair:

[0069]

[0070] Define a as cell line ID, b as drug ID, and x as... ab For logarithmic-scale dose, y ab As a result of the response to a given drug and cell line pair, This represents precision, which is the reciprocal of the variance. Represents the partial derivatives of the input variables; using γ 0b and γ 1b Two parameters are used to simulate the distribution of the control scale parameter θ; where γ 0b The center of the drug response curve, i.e., the IC50 value, γ 1b Used to capture the steepness of a curve.

[0071] Compared with the prior art, the beneficial effects of the present invention are:

[0072] 1. This invention not only uses the relevance of each cell line's prediction of drug response as an evaluation metric, but also employs the weighted ranking metric (NDCG) commonly used in recommender systems to evaluate the model's ability to correctly rank drugs for a given cell line. Under both evaluation metrics, the recommender system-based model consistently delivers good results, and performs well even on datasets the model has never encountered before.

[0073] 2. The recommendation system-based model of this invention has excellent interpretability in biological genomics. The trained model projects cell lines and drugs into the pharmacogenomics space, establishes feature vectors of drugs and cell lines in the latent space, and can be used to explore drug-drug, cell-cell-cell, and drug-cell-cell relationships, conduct visualization studies, explore genomic laws, and provide biological interpretations. The recommendation system-based model enables this embodiment to determine the strength and direction of drug-pathway associations, providing guidance for future research directions. Attached Figure Description

[0074] Figure 1 This is a flowchart of the present invention;

[0075] Figure 2 This is a schematic diagram illustrating the correlation between the recalculated IC50 values ​​and different datasets in the example.

[0076] Figure 3 Example Spearman correlation diagrams between the CCLE and GDSC datasets for the three drugs in this embodiment;

[0077] Figure 4 This is a schematic diagram illustrating the lack of specificity of drugs with a median IC50 < 1 μm in the examples; drugs with a median IC50 < 1 μm lack specificity, (A) CCLE dataset (B) GDSC dataset;

[0078] Figure 5 This is an example diagram of the potential space of the pharmacogenomics genome in the embodiments;

[0079] Figure 6 The example shows that 10 models trained based on different random starting points produced similar performance results.

[0080] Figure 7 This is a schematic diagram illustrating the dimension of the latent space and the model accuracy in the embodiment; Detailed Implementation

[0081] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0082] This embodiment provides a method for constructing an anticancer drug sensitivity prediction model, including the following steps:

[0083] Step 1: Constructing the dataset and preprocessing

[0084] Step 1-1: Prepare the dataset. The dataset comes from two large-scale studies: CCLE (Cancer Cell Line Encyclopedia) and GDSC (Genomics of Drug Sensitivity in Cancer).

[0085] In step 1-1, the CCLE project is a research project aimed at collecting and sharing genomic information and drug sensitivity data from a variety of cancer cell lines. It generated a large-scale genomic dataset for 947 human cancer cell lines and performed pharmacological analyses on 24 compounds from approximately 500 of these cell lines, generating pharmacological maps of 24 compounds (targeting and cytotoxic agents) across 479 cell lines, accelerating progress in cancer biology and drug discovery.

[0086] In step 1-1, the GDSC project, a project aimed at studying the relationship between cancer drug sensitivity and genomics, is a larger dataset developed in 2012 by Wanjuan Yang et al., containing drug sensitivity data from nearly 75,000 experiments, describing the responses of nearly 700 cancer cell lines to 138 anticancer drugs. To determine the molecular markers of drug response, the GDSC dataset is combined with a large genomic dataset obtained from the COSMIC (Catalogue of Somatic Mutations in Cancer) database, summarizing somatic mutation information, gene amplification and deletion, tissue type, and transcriptional data of cancer genes, etc.

[0087] Both the CCLE and GDSC datasets contain abundant genomic information and show considerable consistency in terms of drug response measures and predictive characteristics of these responses.

[0088] Step 1-2 recalculates the IC50 values ​​of the dataset from Step 1-1;

[0089] In steps 1-2, the IC50 (half maximal inhibitory concentration) value was selected as an indicator of drug sensitivity. IC50 is expressed as concentration; the higher the drug concentration required to kill a cell line, the higher the resistance of that cell line. If the IC50 is very low, then the drug concentration (or amount) required to kill the cell line is also very low, meaning the cell line is more sensitive. For both the CCLE and GDSC datasets, all cell lines with baseline gene expression data were retained. To make the values ​​between the datasets more comparable, the IC50 values ​​were recalculated using a Bayesian sigmoid curve fitting method.

[0090] (1) Observation of dose-response experimental data for CCLE and GDSC. In CCLE, each cell line was tested using 6 different concentrations of the drug, and a cell activity value (AV) was measured at each concentration. In GDSC, 7 different concentrations of the drug were used to measure a cell activity value (AV).

[0091]

[0092] Wherein, T, P, and N represent the intensity values ​​(cell counts) measured in the test drug, positive control, and negative control, respectively.

[0093] (2) The JAGS (Just another Gibbs sampler) method was used to fit Bayesian sigmoid curves to obtain the dose-response curve for each drug and cell line pair. Simply put, let a be the cell line ID, b be the drug ID, and x be the cell line ID. ab For logarithmic-scale dose, y ab This represents the response results for a given drug and cell line pair. The parameter γ is used. 0b and γ 1b Two parameters are used to simulate the s-distribution of students. Where γ... 0b The center of the drug response curve (i.e., the IC50 value), γ 1b This is used to capture the steepness of the curve. The resulting fitting formula is as follows:

[0094]

[0095] Furthermore, assuming the drug response curve γ 0b and γ 1b The two parameters follow Normal and Gamma distributions, respectively.

[0096]

[0097] For the remaining parameters, the theory described by Kruschke is adopted, using non-informative priors. After the fitting calculation is completed, the IC50 value of the cells is recalculated using the fitted curve.

[0098] Steps 1-3 compare the IC50 values ​​obtained in Steps 1-2 to determine the final dataset: After recalculating the IC50 values, the study compared the two IC50 values ​​to better facilitate subsequent work. First, a paired p-test was used to assess paired differences in the measured variables. Assuming the difference between the two IC50 values ​​was 0, the resulting p-value was less than 0.01, allowing for a decision not to reject the null hypothesis. Furthermore, the correlation between the IC50 values ​​of the two datasets was calculated. (See Appendix...) Figure 2 It can be seen that when the IC50 values ​​are recalculated from the original dose-response data, the Spearman correlation between each drug is higher than the provided IC50 values. (Appendix) Figure 3 An example of Spearman correlation between three drugs on the CCLE and GDSC datasets is given. The processed IC50 values ​​increase the correlation between the data in the two datasets. As described above, the recalculated IC50 values ​​provide better comparability between different datasets.

[0099] All cell lines showed very high drug sensitivity to drugs with low median IC50, suggesting that these drugs may lack specificity (low rank entropy) and are cytotoxic. (See attached image) Figure 4As shown, statistical analysis of data from the (A)CCLE and (B)GDSC datasets revealed that drugs with a median IC50 < 1 μm tend to be cytotoxic and consistently exhibit high toxicity across all cell lines. Their presence simplifies drug response prediction; therefore, this embodiment excludes them to focus on predicting responses to targeted anticancer drugs.

[0100] In an embodiment of the present invention, after screening, 492 cell lines of CCLE and 983 cell lines of GDSC were obtained.

[0101] Step 2: Construct a model for drug response calculation (the model for drug response calculation adopts a collaborative filtering recommendation algorithm, hereinafter referred to as the model).

[0102] The main idea of ​​the model is that the drug response matrix W is decomposed into deviation S, cell line matrix C, and drug matrix M.

[0103] Step 2-1 constructs the drug response matrix and decomposes it into a cell line matrix and a drug matrix. This embodiment is based on the idea of ​​using latent factors for feature extraction in the collaborative filtering recommendation algorithm.

[0104] Recommendation algorithms can be simply summarized as algorithms that recommend the most suitable items, products, or services to a specific user. These include methods based on collaborative filtering, content-based methods, knowledge-based methods, and hybrid methods. As one of the most successful methods for building recommendation systems, collaborative filtering (CF) is a widely used technique that uses the known preferences of a set of users to recommend or predict the unknown preferences of other users.

[0105] To date, collaborative filtering algorithms can be roughly divided into three types: ① memory-based collaborative filtering algorithms, among which neighborhood-based collaborative filtering algorithms are the most prevalent; ② model-based collaborative filtering algorithms; and ③ collaborative filtering techniques based on matrix factorization models of cell line and drug similarity.

[0106] This embodiment combines the above three collaborative filtering algorithms to analyze the relationships between cell lines and the dependencies between drugs in order to identify new drug response associations.

[0107] Step 2-2: Constructing a model for calculating drug response: Matrix factorization is used to learn the "pharmacogenomic space" (pharmacogenomic space refers to a high-dimensional space where each dimension represents a biological characteristic related to drug response, such as drug properties (e.g., chemical structure, drug mechanism) and genetic and phenotypic characteristics of cell lines (e.g., gene expression, mutation state)). Drug and cell line data are projected into this high-dimensional latent space. The dot product between the cell line vector and the drug vector provides the cell line-specific drug response. The drug response model is calculated using the following formula:

[0108]

[0109] k j =y j M Q (6)

[0110] d i =x i M P (7)

[0111] Substituting (5), (6), and (7) into the equation, we obtain the simplified result as shown in formula (8).

[0112]

[0113] W ij is the sensitivity score of cell line i to drug j predicted by the model, and μ is the overall average drug response. These are the bias terms for drug j and cell line i, respectively, and k j and d i M is the vector of drug j and cell line i in the n-dimensional latent space. Q M is a matrix containing drug vectors, where y is the drug indicator vector. P It is a transformation matrix that transforms cell features x i Projected into the latent space. Based on the performance of cross-validation, the n value for the CCLE and GDSC datasets was set to 12.

[0114] Step 2-3 simplifies formula (8) to obtain the following formula:

[0115] W = S + CM T (9)

[0116] The cell line matrix C can be viewed as a vector of cell line carriers in the latent space, and the drug matrix M can be viewed as a vector of drug carriers in the latent space. The latent pharmacogenomic space can capture the interaction between the drug and the genomic background of the cell line; therefore, the dot product (k, d) between the cell line carrier and the drug carrier can represent the interaction between the drug and the cell line, as shown in the appendix. Figure 5As shown, using vectors to represent the characteristics of cell lines and drugs, it is easy to see that cell line i1 is sensitive to drugs j1 and j2, but insensitive to drug j3. Similarly, cell line i2 differs from cell line i1; it does not respond to either drugs j1 or j2. This representation can predict drug responses in unseen cell lines or patient samples.

[0117] Step 3: Train the drug response calculation model and evaluate it.

[0118] Step 3-1: Training the model for drug response calculation

[0119] This embodiment uses the "sum of squared errors" loss function described below for optimization.

[0120]

[0121] in and W ij These are the sensitivity scores of cell line i observed and predicted using drug j, respectively. It represents the number of drug response experiments in the training dataset.

[0122] Step 3-2 Optimize the loss function

[0123] The loss function in step 3-2 is optimized using gradient descent, and all parameters in W are obtained based on the measured drug response values. M P M Q The gradient function is derived as follows:

[0124]

[0125]

[0126] During model training, initialization M is 0 P and M Q Using the minimum random number, perform batch gradient descent using the gradient calculation formula described above to optimize... M P and M Q .

[0127] Step 3-3 Determine the model evaluation method

[0128] Five-fold cross-validation was performed on the dataset from step 2-2, comparing the model results with the correlation (or squared difference) of known responses for each drug (across cell lines). Pearson correlation was calculated to evaluate the cell line ranking for each drug, and the average correlation for each drug was obtained. In addition, this embodiment also used Normalized Discounted Cumulative Gain (NDCG) and Discounted Cumulative Gain (DCG) to evaluate the model for each cell line. The calculation formulas are as follows:

[0129]

[0130] Where g is the predicted rank of the drug tested on the cell line, and w is the list of observed drug sensitivity scores. The known drug ranking is calculated based on measured drug response values. NDCG values ​​range from 0 to 1, where 1 indicates that the model correctly predicted the drug's ranking. The numerator in DCG is designed to assign greater weight to drugs with higher sensitivity scores, while the denominator prioritizes drugs predicted to have higher rankings.

[0131] Step 3-4 Predicting Model Performance

[0132] To better train the model, the study defined a drug sensitivity score coefficient w = -log(IC50). The higher the score, the more sensitive the cell line is to the drug.

[0133] This embodiment trains the model on the complete dataset to obtain unique features of drugs and cell lines, thereby deriving the potential pharmacogenomic space of drug-drug, cell-cell-cell, and drug-cell-cell associations in the CCLE and GDSC datasets.

[0134] To test the robustness of the model, this embodiment uses the same dataset and constructs 10 different models from different random starting points (S). These models are optimized using gradient descent, and the normalized discounted cumulative returns are used to evaluate the models. The results are represented by box plots as shown in the attached figure. Figure 6 As shown in the figure, the mean NDCG of these models is stable at around 0.94, demonstrating very similar performance and proving the reproducibility of the models. To obtain better training results, the dimension of the latent space was changed, and multiple sets of comparative experiments were conducted using five-fold cross-validation to explore the impact of the number of dimensions on model accuracy. The Pearson correlation coefficient between the IC50 values ​​obtained by the models and the true values ​​was calculated, and the results are represented by scatter plots as shown in the attached figure. Figure 7As shown in the figure, as the dimension of the potential space increases, the correlation between the predicted result and the true value continuously increases. The model can extract more biological features to predict the reaction result, making the result more accurate.

Claims

1. A method for constructing an anticancer drug sensitivity prediction model, characterized in that, Includes the following steps, Step 1, Dataset Construction and Preprocessing: Step 1-1: The dataset is from two studies, CCLE and GDSC. Steps 1-2 involve recalculating the IC50 values ​​for the two datasets. Steps 1-3: Compare the IC50 values ​​obtained from the two datasets to determine the final dataset; Step 2, construct a model for drug response calculation: Step 2-1: Based on the idea of ​​using latent factors for feature extraction in collaborative filtering recommendation algorithm, construct the drug response matrix W and decompose it into bias S, cell line matrix C and drug matrix M; Step 2-2: Use the decomposed drug matrix M to learn the pharmacogenomics space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response. Steps 2-3: Construct the relationship between the drug response matrix W, the cell line matrix C, and the drug matrix M to determine the potential association between cell lines and drug genomes; Step 3: Train and evaluate the drug response calculation model; The optimized sum of squared errors loss function was selected as the optimization algorithm for the drug response calculation model; the sum of squared errors loss function is: R （4） （5） in and These are the sensitivity scores of cell line i observed and predicted using drug j, and the drug response matrix, respectively. = , It represents the number of drug response experiments in the training dataset; The optimization method for the sum of squared errors loss function is as follows: the sum of squared errors loss function is optimized by using gradient descent; the drug response matrix is ​​obtained based on the measured drug response values. All parameters in; parameters The gradient function is derived as follows: （6） （7） （8） （9） in, and These are the bias terms for drug j and cell line i, respectively. and It is the vector of drug j and cell line i in the n-dimensional potential space; It is a matrix containing drug vectors, where y is the drug indicator vector. j The indicator vector of drug j. It is a transformation matrix that transforms cell features Projecting onto the latent space; initialization in the drug response computational model. and =0, and Using the minimum random number, perform batch gradient descent using the gradient calculation formula described above to optimize... , , and ; Then, the evaluation method for the drug response calculation model was determined, specifically: five-fold cross-validation was used to compare the correlation between the drug response calculation model results and the known responses of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; mean normalized depreciation cumulative gain (NDCG) was also used to evaluate the drug response calculation model for each cell line, and the calculation formula is as follows: （10） （11）； in, is the predicted rank of the drug tested on cell lines, and w is the list of observed drug sensitivity scores. It is a known drug ranking calculated based on measured drug response values; the NDCG value ranges from 0 to 1, where 1 indicates that the drug response calculation model correctly predicted the drug's ranking.

2. The method for constructing the anticancer drug sensitivity prediction model as described in claim 1, characterized in that, In step 1-1, the GDSC and CCLE datasets are used as the baseline datasets.

3. The method for constructing the anticancer drug sensitivity prediction model as described in claim 2, characterized in that, In steps 1-2, for the CCLE and GDSC datasets, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method. In CCLE, each cell line is tested using six different concentrations of the drug, and a cell activity value (AV) is measured at each concentration; in GDSC, a cell activity value (AV) is measured using seven different concentrations of the drug. OFF= (1) Where T, P, and N represent the intensity values ​​measured in the test drug, positive control, and negative control, respectively; The JAGS method was used to fit the Bayesian sigmoid curves to obtain the dose-response curves for each drug and cell line pair: （2） Let a be the cell line ID and b be the drug ID. Logarithmic scale dose, As a result of the response to a given drug and cell line pair, This represents precision, which is the reciprocal of the variance. Represents the partial derivatives of the input variables; using parameters and Two parameters are used to simulate the control scale parameter. Distribution; among which The IC50 value represents the center of the drug response curve. Used to capture the steepness of a curve.

4. The method for constructing the anticancer drug sensitivity prediction model as described in claim 1, characterized in that, In steps 1-3, the final dataset is determined by recalculating the IC50 value. After recalculating the IC50 value, a paired p-test is first used to measure the paired differences of the measured variables. Assuming that the difference between the two IC50 values ​​is 0, the obtained p-value is less than 0.01, and a decision can be made not to reject the null hypothesis.

5. The method for constructing the anticancer drug sensitivity prediction model as described in claim 1, characterized in that, In step 2-1, a collaborative filtering algorithm is used to analyze the relationships between cell lines and the dependencies between drugs to identify new drug response associations; assuming a user base of m... A list of n items The list consists of each user Each has a corresponding item list. This is used to record the user's ratings of these items, or to record the user's preferences inferred by the model from the user's behavior.

6. The method for constructing the anticancer drug sensitivity prediction model as described in claim 1, characterized in that, In step 2-2, matrix factorization is used to learn the pharmacogenomics space, projecting drug and cell line data into a high-dimensional latent space. The cell line-specific drug response is provided by the dot product between the cell line vector and the drug vector. The drug response model is calculated according to the following formula: =μ+ (3) It is a transformation matrix that transforms cell features Projected onto the potential space, and These are the bias terms for drug j and cell line i, respectively. and Let be the vectors of drug j and cell line i in the n-dimensional latent space. It is a transformation matrix that transforms cell features Projected onto the potential space.

7. The method for constructing the anticancer drug sensitivity prediction model as described in claim 1, characterized in that, In step 3, to better train the model, a drug sensitivity score coefficient w = -log(IC50) is defined. The higher the score, the more sensitive the cell line is to the drug. Through performance testing, the model can extract more biological features to predict the reaction results, making the results more accurate.

8. A system for constructing an anticancer drug sensitivity prediction model, characterized in that, include, The dataset building module is used to build and preprocess datasets, including the following processes: Step 1-1: The dataset is from two studies, CCLE and GDSC. Steps 1-2 involve recalculating the IC50 values ​​for the two datasets. Steps 1-3: Compare the IC50 values ​​obtained from the two datasets to determine the final dataset; The model building module is used to build models for drug response calculations, and includes the following processes: Step 2-1: Based on the idea of ​​using latent factors for feature extraction in collaborative filtering recommendation algorithm, construct the drug response matrix W and decompose it into bias S, cell line matrix C and drug matrix M; Step 2-2: Use the decomposed drug matrix M to learn the pharmacogenomics space, project the drug and cell line data into a high-dimensional latent space, and use the dot product between the cell line vector and the drug vector to provide the cell line-specific drug response. Steps 2-3: Construct the relationship between the drug response matrix W, the cell line matrix C, and the drug matrix M to determine the potential association between cell lines and drug genomes; The training module is used to train and evaluate the drug response calculation model; The optimized sum of squared errors loss function was selected as the optimization algorithm for the drug response calculation model; the sum of squared errors loss function is: R （4） （5） in and These are the sensitivity scores of cell line i observed and predicted using drug j, and the drug response matrix, respectively. = , It represents the number of drug response experiments in the training dataset; The optimization method for the sum of squared errors loss function is as follows: the sum of squared errors loss function is optimized by using gradient descent; the drug response matrix is ​​obtained based on the measured drug response values. All parameters in; parameters The gradient function is derived as follows: （6） （7） （8） （9） in, and These are the bias terms for drug j and cell line i, respectively. and It is the vector of drug j and cell line i in the n-dimensional potential space; It is a matrix containing drug vectors, where y is the drug indicator vector. j The indicator vector of drug j; It is a transformation matrix that transforms cell features Projecting onto the latent space; initialization in the drug response computational model. and =0, and Using the minimum random number, perform batch gradient descent using the gradient calculation formula described above to optimize... , , and ; Then, the evaluation method for the drug response calculation model was determined, specifically: five-fold cross-validation was used to compare the correlation between the drug response calculation model results and the known responses of each drug; Pearson correlation was calculated to evaluate the cell line ranking of each drug, and the average correlation of each drug was obtained; mean normalized depreciation cumulative gain (NDCG) was also used to evaluate the drug response calculation model for each cell line, and the calculation formula is as follows: （10） （11）； in, is the predicted rank of the drug tested on cell lines, and w is the list of observed drug sensitivity scores. It is a known drug ranking calculated based on measured drug response values; the NDCG value ranges from 0 to 1, where 1 indicates that the drug response calculation model correctly predicted the drug's ranking.

9. The system for constructing the anticancer drug sensitivity prediction model as described in claim 8, characterized in that, In step 1-1, the GDSC and CCLE datasets are used as the baseline datasets.

10. The system for constructing the anticancer drug sensitivity prediction model as described in claim 9, characterized in that, In steps 1-2, for the CCLE and GDSC datasets, the IC50 value was selected as the indicator of drug sensitivity, and the IC50 value was recalculated using the Bayesian sigmoid curve fitting method. In CCLE, each cell line is tested using six different concentrations of the drug, and a cell activity value (AV) is measured at each concentration; in GDSC, a cell activity value (AV) is measured using seven different concentrations of the drug. OFF= (1) Where T, P, and N represent the intensity values ​​measured in the test drug, positive control, and negative control, respectively; The JAGS method was used to fit the Bayesian sigmoid curves to obtain the dose-response curves for each drug and cell line pair: （2） Let a be the cell line ID and b be the drug ID. Logarithmic scale dose, As a result of the response to a given drug and cell line pair, This represents precision, which is the reciprocal of the variance. Represents the partial derivatives of the input variables; using and Two parameters are used to simulate the control scale parameter. Distribution; among which The IC50 value represents the center of the drug response curve. Used to capture the steepness of a curve.

Images