KR102653969B1 - A system of predicting drug response with convolutional neural networks based on similarity matrices of drugs and cell lines - Google Patents

Abstract

Human cancer cell lines are frequently used in research to study the biology of cancer and test the efficacy of cancer treatments. Accurately predicting drug response using pharmacogenomics data is an essential issue in precision medicine in oncology. Based on the fact that similar cell lines react similarly to similar drugs, an ensemble deep learning model that applies the cross product between the column vectors of the drug-drug similarity matrix and the cell line-cell line similarity matrix to a similarity-based convolutional neural network. provides. This shows that it can be useful in precision medicine by linking a patient's genetic characteristics to drug sensitivity.

Description

A system for predicting drug response using a convolutional neural network based on the similarity matrix of drugs and cell lines {A SYSTEM OF PREDICTING DRUG RESPONSE WITH CONVOLUTIONAL NEURAL NETWORKS BASED ON SIMILARITY MATRICES OF DRUGS AND CELL LINES}

Precision medicine aims to precisely select cancer treatments based on each patient’s genetic information. One of the most important problems in real medicine is predicting anticancer drug response for each patient. Because of tumor heterogeneity, patients with the same type of cancer may respond differently to similar drugs. Therefore, it is very important to provide a predictive method that reveals the relationship between genomic information and drug response, which can be helpful in precision medicine.

Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) are two projects that have provided molecular profiles and drug response values for hundreds of cancer cell lines treated with multiple anticancer drugs. These large-scale datasets allow us to develop methods to predict drug response for each patient. Generally, methods for predicting drug response are classified into two types. The first is a classification approach that predicts sensitive drug-cell line pairs. The second is a regression analysis approach that predicts reference values for measuring a cell line's response to a drug. One embodiment of the present invention is a cell line for a drug. Quantified through (half-maximal inhibitory concentration) value A regression analysis approach to predict is disclosed. Various regression analysis approaches have been proposed to predict drug response using gene expression profiles or other molecular information of a cell line. Some prediction methods have improved drug response prediction by integrating drug information such as the drug's chemical substructure and cell line information. Additionally, numerous machine learning methods have been applied to the drug response prediction problem. For example, regression analysis methods such as lasso (least absolute shrinkage and selection operator), elastic net, random forest, kernel-based methods, neural networks, and deep learning. were applied. Ali and Aittokallio provide a comprehensive recent review (Ali, M. & Aittokallio, T. Machine learning and feature selection for drug response prediction in precision oncology applications. Biophys. Rev. 11, 31-39 (2019)] ). Recent advances in deep learning have opened a new field for finding regression models that predict drug response, and may ultimately provide more accurate tools for treatment response.

Wang et al. proposed a similarity-regularized matrix factorization (SRMF) method for predicting drug response, which simultaneously includes similarity of gene expression profiles of cell lines and similarity of chemical substructure of drugs (Reference [Wang, L., Li, X., Zhang, L. & Gao, Q. Improved anticancer drug response prediction in cell lines using matrix factorization with similarity regularization. BMC Cancer 17, 513 (2017)]. Patients with similar genetic characteristics have been shown to have similar responses to similar drugs. Suphavilai et al. designed a matrix decomposition-based recommender system called “CaDRReS” that can predict drug responses to new drugs and new cell lines through learning predictions of drugs and cell lines in the latent space (Suphavilai, C., Bertrand, D. & Nagarajan, N. Predicting cancer drug response using a recommender system. Bioinformatics 34, 3907-3914 (2018)]). This showed that the properties of the latent space were correlated with the drug path. Chang et al. proposed “CDRscan,” an ensemble model including five convolutional neural networks (CNNs). It used the mutation profile of the cell line and the chemical substructure of the drug as input features for CNNs. Drug response values were measured as the average of the output values of five CNNs. However, “CDRscan” tends to be poor at predicting drug responses to new drugs and new cell lines. Wei et al. recently designed a cell line-drug complex network (CDCN), which predicts drug response by inferring information from a simple network composed of cell lines and drugs (Wei et al. , D., Liu, C., Zheng, X. & Li, Y. Comprehensive anticancer drug response prediction based on a simple cell line-drug complex network model. BMC Bioinf. 20, 44 (2019)] CDCN provides satisfactory results for imputing missing drug information. Moughari and Eslahchi designed a model for Anticancer Drug Response Prediction using Manifold Learning (ADRML) (document [Moughari, F. A. & Eslahchi, C. ADRML: anticancer drug response prediction using manifold learning. Sci. Rep. 10, 14245 (2020)]). ADRML maps drug response values to a low-dimensional latent space and computes drug response values for new cell line-drug pairs from the latent space. It takes into account the similarity and drug similarity of various types of cell lines and utilizes them in the manifold learning procedure. ADRML has been shown to provide accurate and robust predictions.

Even patients with the same type of cancer may react differently to similar drugs. Therefore, it is very important to accurately predict drug response for individual patients.

Additionally, drug response prediction tends to be poor at predicting drug responses to new drugs and new cell lines. Therefore, a new approach is needed to overcome this.

Additionally, a new method is required that can provide higher accuracy drug response prediction while simplifying the computational process.

One embodiment of the present invention is a method of predicting drug response of a drug and a cell line using a convolutional neural network model, comprising: preparing a first similarity matrix between drugs and drugs; preparing a second similarity matrix between cell lines; As a step of calculating a cross product between the first similarity matrix and the second similarity matrix, the ith column vector of the first similarity matrix (i = 1, 2, ... m: m is an integer) and the second similarity matrix Calculating a cross product between the j-th column vectors (j = 1, 2, ..., n: n is an integer); Learning the convolutional neural network model using the cross product as an input value and the drug response value of the i-th drug and the j-th cell line as output values; and predicting and outputting drug response values of a new drug and cell line using the learned convolutional neural network model.

In addition, the first similarity matrix is an m x m drug-drug similarity matrix based on the Tanimoto coefficient, and the second similarity matrix is an n x n cell line-cell line similarity matrix that is a radial basis function (RBF) kernel matrix. , where m and n represent the number of drugs and cell lines in the learning dataset, respectively, providing a drug response prediction method.

Additionally, the i-th column vector of the first similarity matrix is the Tanimoto similarity between the i-th drug and other drugs including the i-th drug, and the j-th column vector of the second similarity matrix is the j-th cell. A method for predicting drug response is provided, which is gene expression similarity between a line and other cell lines including the j-th cell line.

In addition, the drug response value is quantified through the IC ₅₀ (half-maximal inhibitory concentration) value of the cell line for the drug. It could be a value. In addition, the second similarity matrix includes a first submatrix and a second submatrix, and calculating a cross product between the first similarity matrix and the second similarity includes the first similarity matrix and the second similarity. Providing a method for predicting drug response, comprising calculating a first cross product between a first submatrix of a matrix, and calculating a second cross product between the first similarity matrix and a second submatrix of the second similarity matrix. do.

In addition, the convolutional neural network model includes a first convolutional neural network model and a second convolutional neural network model, and the step of training the convolutional neural network model includes combining the first cross product and the second cross product with the first synthesis. A drug response prediction method using input values of a product neural network model and the second convolutional neural network model is provided.

Additionally, the convolutional neural network model may be a 2-dimensional model.

Additionally, the convolutional neural network model may include two 2-dimensional convolutional layers, two max pooling layers, one flattening layer, one dropout layer, and two fully connected layers.

Another embodiment of the present invention is a system for predicting drug responses of drugs and cell lines using a convolutional neural network model, comprising: a control unit for controlling the convolutional neural network model; a communication unit for communication with an external server; memory unit; display unit; and an input unit that receives a user's input, wherein the memory unit includes a first similarity matrix between drugs and drugs and a second similarity matrix between cell lines and cell lines, and the control unit stores the first similarity matrix. Calculate the cross product between the ith column vector (i = 1, 2, ... m: m is an integer) and the jth column vector (j = 1, 2, ..., n: n is an integer) of the second similarity matrix, and , the convolutional neural network model is trained using the cross product as an input value and the drug response value of the ith drug and jth cell line as an output value, and a new drug and cell line are developed using the learned convolutional neural network model. Provides a drug response prediction system that predicts drug response values.

In addition, the first similarity matrix is an m x m drug-drug similarity matrix based on the Tanimoto coefficient, and the second similarity matrix is an n x n cell line-cell line similarity matrix that is a radial basis function (RBF) kernel matrix. , where m and n represent the number of drugs and cell lines in the learning dataset, respectively, providing a drug response prediction system.

Additionally, the i-th column vector of the first similarity matrix is the Tanimoto similarity between the i-th drug and other drugs including the i-th drug, and the j-th column vector of the second similarity matrix is the j-th cell. A drug response prediction system is provided, which is gene expression similarity between a line and other cell lines including the j-th cell line.

In addition, the drug response value is quantified through the IC ₅₀ (half-maximal inhibitory concentration) value of the cell line for the drug. It could be a value.

In addition, the second similarity matrix includes a first submatrix and a second submatrix, and calculating a cross product between the first similarity matrix and the second similarity includes the first similarity matrix and the second similarity. Providing a drug response prediction system comprising calculating a first cross product between a first submatrix of a matrix, and calculating a second cross product between the first similarity matrix and a second submatrix of the second similarity matrix. do.

Additionally, the convolutional neural network model includes a first convolutional neural network model and a second convolutional neural network model, and learning of the convolutional neural network model involves combining the first cross product and the second cross product with the first convolutional neural network. A drug response prediction system is provided, which uses input values of each model and the second convolutional neural network model.

Additionally, the convolutional neural network model may be a 2-dimensional model.

Additionally, the convolutional neural network model may include two 2-dimensional convolutional layers, two max pooling layers, one flattening layer, one dropout layer, and two fully connected layers.

One embodiment discloses the Dr.CNN model to solve the problem of predicting drug response values. Dr.CNN simply divides the RBF kernel matrix associated with a cell line into two submatrices and calculates the cross product of the column vector of the Tanimoto similarity matrix and the column vector of each RBF kernel submatrix. Then, this is used as the input value for each of the two CNN models. This is to obtain faster computation and improved prediction performance through ensemble learning. The RBF kernel matrix may be divided randomly and may be divided into two or more submatrices depending on the number of cell lines. Dr.CNN is the first nonlinear method that applies 2D CNN to the cross product between the column vector of the drug's Tanimoto similarity matrix and the column vector of the cell line's RBF kernel matrix to predict drug response.

Experimental results show that Dr.CNN surpasses the performance of existing models such as elastic net, RF, SVR, and 1D CNN ensemble. Dr. CNN can be further improved by adjusting the CNN architecture according to the data structure. The main idea behind Dr.CNN is to integrate two modalities using a cross product and apply CNN to the resulting matrix. Dr.CNN is a very effective approach for drug response prediction and can play a big role in the drug development process.

Figure 1 shows the overall workflow of one embodiment proposed for predicting drug response values.
Figure 2 shows a summary of the GDSC1 and GDSC2 datasets used in one embodiment.
Figure 3 shows the measured values of GDSC2 and GDSC1 datasets. Value predicted by one example Shows a scatterplot of values.
Figure 4 shows the workflow of an ensemble 1D CNN model for prediction of drug response values.
Figure 5 shows the architecture of a similarity-based CNN submodel in one embodiment.
Figure 6 is a block diagram showing an example of a system for implementing a method for predicting drug response according to an embodiment.

The present invention will become clear by referring to the embodiments described in detail below along with the drawings attached to this specification. However, the present invention is not limited to the embodiments disclosed below and will be implemented in various different forms. The present embodiments only serve to ensure that the disclosure of the present invention is complete and that common knowledge in the technical field to which the present invention pertains is not limited. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Meanwhile, the terms used in this specification are for describing embodiments and are not intended to limit the present invention. As used herein, singular forms also include plural forms, unless specifically stated otherwise in the context. As used herein, “comprises” or “comprising” does not exclude the presence or addition of one or more other components or steps in addition to the mentioned components or steps. Terms such as first, second, etc. may be used to describe various components, but the components should not be limited by the terms. Terms are used only to distinguish one component from another.

One embodiment of the present invention uses a two-dimensional CNN approach for the outer product of the column vector of the Tanimoto similarity matrix of the drug and the Radial Basis Function (RBF) kernel matrix of the cell line. This provides a similarity-based ensemble deep learning model that predicts drug response values. In one embodiment, this model is called Dr.CNN.

Figure 1 shows the overall workflow of Dr.CNN proposed for predicting drug response values. Dr.CNN's workflow predicts the final drug response value based on the drug response value obtained from two sub-networks running inside the ensemble deep learning architecture, for example, as an average value. A 2D CNN is used to learn features by using the cross product of the drug similarity vector and the cell line similarity vector as input. Each CNN has, for example, 2 convolutional layers, 2 max-pooling layers, 1 flatten layer, 1 dropout layer, and It may consist of three fully connected layers (FC layer). A detailed description of the CNN structure is described later in Figure 5.

In one embodiment, in order to obtain more sophisticated prediction performance and faster calculation through ensemble learning, the RBF kernel matrix of the cell line is divided into two submatrices, and a Tanimoto similarity matrix and each RBF kernel submatrix are formed. The cross product of the column vector of is constructed, and this is used as the input value of CNN. The RBF kernel matrix of a cell line may be divided into two or more submatrices depending on the number of cell lines. The embodiment may consist of two steps. In the first step, the Tanimoto similarity matrix of the drug and the RBF kernel matrix of the cell line are computed, and then the cross product between the Tanimoto similarity matrix and the column vector of each RBF kernel submatrix is computed. It's a step. In the second step, a 2D CNN model is applied to extract features from the extrinsic product and predict drug response values for the two subnetworks. The final prediction of the drug response value can be obtained by computing the drug response values from the two learned subnetworks, for example, taking the average value.

Dr.CNN according to one embodiment has root mean squared error (RMSE), concordance index (CI), and modified squared correlation coefficient: ) is verified by comparing it with other machine learning and deep learning models. CI is the rank correlation between observed and predicted data.

Embodiments of the present invention can help users predict drug response values.

experimental dataset

In one embodiment, the input values for Dr. CNN are SMILES (Simplified Molecular-Input Line-Entry System) of cell line gene expression and anticancer compounds. GDSC, a publicly available database, is used for observed drug responses in all pairs of cell lines and drugs. The GDSC drug SMILES is obtained from PubChem. GDSC is GDSC1 (Iorio, F. et al. A landscape of pharmacogenomic interactions in cancer. Cell 166, 740-754 (2016)] and GDSC2 (Picco, G. et al. Functional linkage of gene fusions to cancer cell fitness assessed by pharmacological and CRISPR-Cas9 screening. Nat. Commun. 10, 2198 (2019)]) dataset. In one embodiment, the GDSC1 and GDSC2 datasets are used to evaluate drug response prediction. The GDSC1 dataset tested 681 cell lines across 234 compounds using the Resazurin or Syto60 assays. The GDSC2 dataset tested 588 cell lines across 147 compounds using the CellTitreGlo assay. The GDSC1 dataset is indeed observed for all pairs of 234 drugs and 681 cell lines. Contains 131,894 drug response values measured in values. On the other hand, the GDSC2 dataset observed for all pairs of 147 drugs and 588 cell lines, Contains 72,393 observed drug response values. Table 1 shows these two datasets in the form used in the actual experiment. In one embodiment Use the value converted to logspace through .

Dataset Drugs Cell line Interactions Density (%) GDSC1 234 681 131,894 82.77 GDSC2 147 588 72,393 83.75

Figure 2 shows a summary of the GDSC1 and GDSC2 datasets. Figures 2 (a) and (b) show the distribution of drug response values of the GDSC1 and GDSC2 datasets, respectively, and (c) and (d) show the distribution of the length of the SMILES string of the GDSC1 and GDSC2 datasets, respectively. More specifically, (a) and (b) in Figures 2 Shows the distribution of values, and (c) and (d) show the distribution of SMILES string lengths of drugs in the GDSC1 and GDSC2 datasets, respectively. of the GDSC1 dataset. For , the mean and standard deviation are -0.9032 and 1.1777, respectively. of the GDSC2 dataset. The mean and standard deviation for are -1.2472 and 1.2182, respectively. For drugs in the GDSC1 dataset, the maximum SMILES length is 133 and the average is 62. The maximum SMILES length for drugs in the GDSC2 dataset is 126, and the average is 62.

Expression of input and output values

In one embodiment of Dr.CNN, drug-drug and cell line-cell line similarity matrices for the GDSC1 and GDSC2 datasets are used. These two matrices are each , It is expressed as To create an ensemble model, one embodiment uses a matrix Divide into two submatrices. in other words, It is expressed as, where: , am. Therefore, the input value of Dr.CNN for each drug-cell line pair is class cross product of , where: is the similarity matrix is the ith column of is the similarity matrix for is the jth column of represents the outer product. subscript represents the transpose of the vector. cross product is actually defined as equation (1) below.

Formula (1)

This cross product yields two sets of information. , Therefore, it yields bimodal interactions and raw unimodal representations of the individual modalities. therefore, Is and Includes any combination of information. this is In predicting the drug response of the ith drug and the jth cell line, class This indicates that it can be a more effective input value than simple concatenation of .

Drug-drug similarity is Tanimoto coefficient It is calculated using , which is the most popular similarity measure for comparing chemical structures expressed as fingerprints. One embodiment uses RDKit's topological fingerprint. The Tanimoto similarity scale has values from 0 to 1 and can be interpreted as the percentage of characteristics shared by two drugs. On the other hand, cell line-cell line similarity can be calculated based on the gene expression vector using the RBF kernel described by the equation in equation (2). The RBF kernel is a popular kernel function used in various kernel learning algorithms.

for Formula (2)

here, is the gene expression vector of the ith cell line, is the bandwidth parameter for which an estimate is obtained as in He et al. (He, T. et al. SimBoost: A read-across approach for predicting drug-target binding affinities using gradient boosting machines. J. Cheminf. 9, 24 . https://doi.org/10.1186/s1332 1-017-0209-z (2017)]). The RBF kernel is a measure of similarity between vectors. The value of the RBF kernel is the distance decreases according to , from 0 (threshold) to 1 ( ) has a range of up to . That is, if the two vectors are close to each other, becomes smaller. After that, It gets bigger and bigger in the limit. Therefore, close vectors have larger RBF kernel values than distant vectors.

The output value for each drug-cell line pair is Corresponds to the value.

performance rating scale

Elastic net, random forest (RF), and support vector regression (SVR) are general regression analysis methods proposed for predicting drug response, so one embodiment considers them as basic methods. One embodiment compares the performance of elastic net, RF, SVR, 1D CNN ensemble, and DR.CNN using the above-described evaluation dataset. The 1D CNN ensemble here is similar to that of Park et al. It is similar to the one used in (Park, H. et al. Detection of chromosome structural variation by targeted next-generation sequencing and deep learning application. Sci. Rep. 9, 3644 (2019)]). The inputs to the Elastic net, RF, SVR, and 1D CNN ensembles are concatenated vectors of extended-connectivity fingerprints (ECFPs) of 2048 bit length and a gene expression vector consisting of 172 values selected from 19,144 values. In one example, the prediction performance of the GDSC2 dataset was evaluated through a 5-fold cross validation experiment. This technique randomly splits the dataset into 5 folds of approximately equal size. One fold is treated as the validation set, and learning is applied to the remaining 4 folds. This procedure is repeated five times, and in each procedure, a different group of instances is processed into the validation set. One embodiment also evaluated the prediction performance of the GDSC1 dataset after training five models using the GDSC2 dataset as a training dataset. In one embodiment, RMSE, CI, and Pearson correlation coefficient are used to evaluate the performance of a regression model. , The same four indicators were used.

RMSE is a commonly used indicator for error in continuous prediction. Because it uses regression techniques, one embodiment uses RMSE, a commonly used metric for error in continuous prediction. here, is the actual output value, corresponds to the prediction. means the number of samples.

Formula (3)

Pahikkala et al. As suggested in, CI can be used as an evaluation indicator for prediction accuracy (Pahikkala, T. et al. Toward more realistic drug-target interaction predictions. Brief. Bioinform. 16, 325-337 (2015)] ). The intuition about CI is as follows. The CI for a set of paired data is the probability that the predictions for two randomly drawn drug-cell line pairs with different label values are in the correct order, with the larger affinity value prediction for has a smaller affinity value prediction of It means bigger than.

Formula (4)

here, is the normalization constant, is a step function (Pahikkala, T. et al. Toward more realistic drug-target interaction predictions. Brief. Bioinform. 16, 325-337 (2015)]).

Formula (5)

CI ranges from 0.5 to 1.0, where 0.5 corresponds to a random prediction and 1.0 corresponds to perfect prediction accuracy.

To increase the predictability of the model, Roy and Roy used the modified squared correlation coefficient was introduced (document [Roy, P. & Roy, K. On some aspects of variable selection for partial least squares regression models. QSAR Comb. Sci. 27, 302-313 (2008)]).

Formula (6)

here, class are the squared correlation coefficients for the cases including and not including the intercept, respectively. About the test dataset The model is judged to be an acceptable model.

Learning and Assessment

Hereinafter, prediction of drug response values for the GDSC1 and GDSC2 datasets of Dr.CNN, an example, The prediction performance of is explained. One example evaluated the performance of one example model on two benchmark datasets. First, because the GDSC2 dataset is more recent than the GDSC1 dataset, the prediction performance of the GDSC2 dataset was evaluated through 5-fold cross-validation. Table 2 shows the performance results of five models through nested 5-fold cross-validation on the GDSC2 dataset. Values in bold represent the highest performance results. Standard errors are listed in parentheses.

Model RMSE C.I. RF 0.5386
(0.0026) 0.8433 (0.0002) 0.8970
(0.0011) 0.7997
(0.0018) SVR 0.7305
(0.0398) 0.7923
(0.0021) 0.8307
(0.0033) 0.6455
(0.0227) ElasticNet 0.7522
(0.0377) 0.7843
(0.0094) 0.8084
(0.0176) 0.5416
(0.0410) 1D CNN 0.5390
(0.0022) 0.8448
(0.0004) 0.8979
(0.0009) 0.7909
(0.0060) Dr.CNN 0.5085
(0.0042) 0.8536
(0.0007) 0.9098
(0.0011) 0.8162
(0.0043)

As can be seen in Table 2, Dr.CNN in one example shows the best performance in all indicators of the GDSC2 dataset. To statistically evaluate the significant improvement of the model of one embodiment, a one-sided t-test was performed. We compared Dr.CNN, which had the best performance results, with other models. Therefore, the null hypotheses associated with Table 2 are given as follows. , , , . All relevant aspects of the above hypothesis testing The value is calculated to be less than 0.01. Therefore, Dr.CNN shows significantly better performance than other models for all four indicators. In particular, Dr.CNN has significantly greater results than other models in 5-fold cross-validation on the GDSC2 dataset. It is the most acceptable model because it derives .

Next, five models were trained using the GDSC2 dataset as a training dataset, and then the prediction performance on the GDSC1 dataset was evaluated. Table 3 shows the performance results of five models on the GDSC1 dataset. Values in bold represent the highest performance results, and standard errors are listed in parentheses. Since only GDSC1 is used to calculate the evaluation index of each model, a statistically significant model cannot be specified through the evaluation index. Therefore, the bootstrap method was used to estimate the mean and standard error of each evaluation index (Efron, B. & Tibshirani, R. An Introduction to the Bootstrap (Chapman Hall, 1993)]. This allows to specify a statistically significant model based on the estimated mean and standard error for each indicator. The bootstrap method is known to be useful for estimating the sampling distribution of evaluation indicators without using normal theory. The bootstrap method involves repeatedly sampling the GDSC1 dataset using reconstruction extraction. When applying bootstrap sampling to the GDSC1 dataset, set the bootstrap sample size to 131,894 and repeat the sampling process 20 times.

Model RMSE C.I. RF 1.1992
(0.0006) 0.6180
(0.0002) 0.4210
(0.0008) 0.1420
(0.0007) SVR 1.0948
(0.0004) 0.6176
(0.0002) 0.4606
(0.0005) 0.1756
(0.0005) Elastic Net 1.0890
(0.0004) 0.6398
(0.0002) 0.4829
(0.0006) 0.2034
(0.0006) 1D CNN 1.0652
(0.0006) 0.6321
(0.0001) 0.5080
(0.0005) 0.2536
(0.0003) Dr.CNN 1.0524
(0.0005) 0.6531
(0.0002) 0.5597
(0.0005) 0.3024
(0.0007)

As shown in Table 3, Dr.CNN shows the best performance in all four indicators for the 20 bootstrap samples of the GDSC1 dataset. As above, a one-tailed t-test was run to statistically evaluate the significant improvement of Dr.CNN. The Dr.CNN model with the best results was compared with other models. Therefore, the null hypothesis related to Table 3 is given as follows. , , , . All relevant factors in the above hypothesis test The value is calculated to be less than 0.01. Therefore, Dr.CNN shows significantly better performance than other models for all four indicators. Although Dr.CNN is large enough , but this is the most acceptable model. This is larger than other models for the 20 bootstrap samples of the GDSC1 dataset. Because it shows the value.

To visually demonstrate the predictive ability, we plot predicted values versus measured values for several datasets. Figure 3 shows the measured values of GDSC2 and GDSC1 datasets. Predicted by Dr.CNN compared to the value Shows a scatterplot of values. In the case of the GDSC2 dataset, the 4-fold of the GDSC2 dataset is alternately used as a training dataset and the remaining fold is used as a test dataset to obtain the prediction value. In the case of the GDSC1 dataset, the predicted value is obtained by using the GDSC2 dataset as a training dataset and the GDSC1 dataset as a test dataset. The ideal regression model is the predicted value This measurement It will be the same as It is expected that it will be. In particular, for the GDSC2 dataset, a straight line It shows high density in the surrounding area.

reference model

For the baseline model, three existing machine learning models and one deep learning-based model can be considered. Existing machine learning models are elastic net, RF, and SVR. Elastic net first emerged as a result of criticism of Lasso (least absolute shrinkage and selection operator), which said that variable selection was too dependent on the data and could be unstable. The solution is to combine ridge regression and Lasso's penalty to be optimal for both domains. Lasso is a regression analysis method that performs both variable selection and normalization to increase the prediction accuracy and interpretability of statistical models. RF assigns sampled input data with replacement to many decision trees for learning, collects decision results of drug-cell line pairs, and determines drug response through averaging. When the tree grows, division is determined by considering only a subset of all the characteristics at each node. This algorithm is simple, fast, and does not cause overfitting. In general, it shows better performance than using a single good regression model. SVR provides the flexibility to define the allowable errors in the model and find a high-dimensional hyperplane suitable for the data. The goal function of SVR is not to minimize the error squared but to minimize the coefficients, especially the L2-norm of the coefficient vector. Instead, the error term is the maximum error This is done in a constraint that sets an absolute error that is less than or equal to a specified margin, called . In one embodiment, to obtain the required accuracy of the model, can be tuned. SVR has proven to be an effective tool in real-value function estimation. One of the advantages of SVR is that its computational complexity does not depend on the dimension of the input space. Additionally, it has excellent generalization capabilities with high prediction accuracy. The inputs of the Elastic net, RF and SVR are a concatenation vector of ECFPs with a length of 2048 bits and a gene expression vector consisting of 172 values selected from 1,444 values. One of the deep learning models is a prediction model based on an ensemble 1D CNN that uses a 2048-bit long ECFP vector and a gene expression vector consisting of 172 values selected from 19,144 values as input to each individual 1D CNN.

Figure 4 shows the workflow of an ensemble 1D CNN model for prediction of drug response values. In a 1D CNN, the kernel and pooling move along one dimension. The stride for all convolutional layers is set to 1. Set the stride to 2 for all max-pooling layers.

Dr.CNN is a 2D CNN-based prediction model that uses the cross product of the drug similarity vector and the cell line similarity vector as input. For input to the Dr.CNN model, first, the mxm drug-drug similarity matrix based on the Tanimoto coefficient and an nxn cell line-cell line similarity matrix based on gene expression scores via the RBF kernel function. Calculate . RDKit's 2048-bit long ECFP fingerprint is used to calculate the Tanimoto coefficient. Here, m and n represent the number of drugs and cell lines in the training dataset, respectively. Then, for every drug-cell line pair, mx1 drug similarity vector and nx1 cell line similarity vector cross product of Calculate . here, class are the similarity matrices, respectively. similarity matrix with the ith column of Indicates the jth column of . in other words, consists of the Tanimoto similarity between the ith drug and other drugs including itself, consists of the gene expression similarity between the jth cell line and other cell lines including itself.

Parameters related to Dr.CNN can be obtained by using the cross product as input and the drug response as output.

Figure 5 shows the architecture of Dr.CNN's similarity-based CNN submodel. As shown in Figure 5, the CNN submodel consists of two 2D convolution layers followed by max-pooling, one flatten layer, FC(128), FC(64), and FC(1) layers, respectively. It can be. The numbers in parentheses indicate the number of nodes. FC(128), FC(64), and FC(1) layers can use ReLU (Rectified linear unit) and linear functions, respectively. In one embodiment, to remove overfitting, a drop ratio of 0.1 is used between the flattening layer and the FC (128) layer, between the FC (128) layer and the FC (64) layer, and between the FC (64) layer and FC (1) layer. Can include out layers. The number of filters for each convolution layer is 18 and 24, respectively. In one embodiment, filters with kernel sizes of 5 x 5 and 3 x 3 may be used for the convolution layer, respectively. The max pooling layer is size 2 and stride 2. In one embodiment, the batch size and number of epochs for the learning algorithm may be set to 32 and 20, respectively. One embodiment may use the Adam optimizer with a learning rate of 0.001.

Figure 6 is a block diagram showing an example of a system for implementing a method for predicting drug response according to an embodiment, and conceptually shows parts related to the embodiment. Each component may be provided in one device and processed independently, but the present invention is not limited to this, and may also include that each component is performed on a separate device connected through a network.

The external server 20 may be connected to the prediction system 10 through a network and may provide information on drug chemical characteristics, cell line information, drug reaction information, etc. For example, information on the chemical characteristics of a drug may include Simplified Molecular-Input Line-Entry System (SMILES) information, and cell line information may include cell line gene expression information. Specifically, GDSC provides information on drug responses observed in all pairs of cell lines and drugs, which can also be received from an external server 20. For example, the external server 20 may be a database for drug response prediction processing of the prediction system 10 or a server that provides the same.

The prediction system 10 may include a control unit 11, a communication unit 12, an input/output interface unit 13, and a memory unit 14.

The control unit 11 is a component that controls the entire prediction system 10 and may include, for example, a processing unit such as a CPU or GPU. The control unit 11 can train models to be described later using information stored in the memory unit 14, and can also calculate a predicted value for a new input through the learned model. Specifically, the control unit 11 can control a model that predicts drug response. To this end, the control unit 11 may include an internal memory for storing control programs such as an operating system (OS), programs defining various processing sequences, and data. And, the control unit 11 can perform information processing to execute various processes using these programs, etc.

Additionally, the communication unit 12 may include an interface that can be connected to a communication device such as a router connected to a communication line, and can control communication between the prediction system 10 and the external server 20. .

The input/output interface unit 13 may be an interface connected to the input unit 15 and/or the display unit 16. The prediction system 10 and the user can communicate through the input/output interface unit 13. For example, the display unit 16 may be a display means (for example, a display made of liquid crystal or organic EL, a monitor, a touch panel, etc.) that displays a display screen such as an application. Additionally, the input unit 15 may be, for example, a key input unit, a touch panel, a control pad (eg, a touch pad, a game pad, etc.), a mouse, a keyboard, a microphone, etc.

Additionally, the memory unit 14 may be a device that stores various databases, tables, etc. For example, the memory unit may provide information on drug chemical characteristics, cell line information, drug reaction information, etc. For example, information on the chemical characteristics of a drug may include Simplified Molecular-Input Line-Entry System (SMILES) information, and cell line information may include cell line gene expression information. Additionally, processes for input and output of the prediction system 10 can be stored, and result values for process processing can also be stored.

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. A computer-readable recording medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on a computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable by those skilled in the computer software field. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specifically configured to store and perform program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the invention and vice versa.

Additionally, the embodiments according to the present invention described above may be a set of program instructions that can be executed through various computer components and a user application for executing them. Specifically, it may be a program itself that can be downloaded from a server or via a storage medium and installed on a client computer.

In the above, the present invention has been described with specific details such as specific components and limited embodiments and drawings, but this is only provided to facilitate a more general understanding of the present invention, and the present invention is not limited to the above embodiments. , a person skilled in the art to which the present invention pertains can make various modifications and variations from this description.

Accordingly, the spirit of the present invention should not be limited to the above-described embodiments, and the scope of the patent claims described below as well as all modifications equivalent to or equivalent to the scope of the patent claims fall within the scope of the spirit of the present invention. They will say they do it.

Additionally, the embodiments of the present invention are not mutually exclusive, and the configuration of one embodiment may be applied to another embodiment. The embodiments of the present invention provide examples of some of the various forms that can be derived from various combinations of components, and are not limited to the specific embodiments of the present invention itself.

10: Prediction system 20: External server
11: Control unit 12: Communication unit
13: input/output interface unit 14: memory unit
15: input unit 16: display unit

Claims (16)

A method of predicting the drug response of drugs and cell lines using a two-dimensional convolutional neural network model,
preparing a first drug-drug similarity matrix;
preparing a second similarity matrix between cell lines;
As a step of calculating a cross product between the first similarity matrix and the second similarity matrix, the ith column vector of the first similarity matrix (i = 1, 2, ... m: m is an integer) and the second similarity matrix Calculating a cross product between the j-th column vectors (j = 1, 2, ..., n: n is an integer);
Learning the two-dimensional convolutional neural network model using the two-dimensional matrix that is the calculated value of the cross product as an input value and the drug response value of the i-th drug and the j-th cell line as output values; and
A step of predicting and outputting drug response values of a new drug and cell line using the learned two-dimensional convolutional neural network model,
The i-th column vector of the first similarity matrix is the similarity between the i-th drug and other drugs including the i-th drug,
The j-th column vector of the second similarity matrix is the gene expression similarity between the j-th cell line and other cell lines including the j-th cell line. According to claim 1,
The first similarity matrix is an mxm drug-drug similarity matrix based on the Tanimoto coefficient,
The second similarity matrix is an nxn cell line-cell line similarity matrix that is a radial basis function (RBF) kernel matrix,
A drug response prediction method, wherein m and n represent the number of drugs and cell lines in the learning dataset, respectively. According to claim 1,
The i-th column vector of the first similarity matrix is the Tanimoto similarity between the i-th drug and other drugs including the i-th drug,
The two-dimensional convolutional neural network model includes two two-dimensional convolutional layers, two max pooling layers, one flattening layer, one dropout layer, and two fully connected layers. A drug response prediction method. The method according to any one of claims 1 to 3,
The drug response value is quantified through the IC ₅₀ (half-maximal inhibitory concentration) value of the cell line for the drug. Value, method for predicting drug response. According to claim 1,
The first similarity matrix is an mxm drug-drug similarity matrix,
The second similarity matrix is an nxn cell line-cell line similarity matrix,
The m and n represent the number of drugs and cell lines in the training dataset, respectively,
The second similarity matrix includes a first submatrix and a second submatrix,
The step of calculating a cross product between the first similarity matrix and the second similarity matrix includes:
Calculating a first cross product between the first similarity matrix and a first submatrix of the second similarity matrix, and calculating a second cross product between the first similarity matrix and a second submatrix of the second similarity matrix. Method for predicting drug response, including. According to claim 5,
The two-dimensional convolutional neural network model includes a first convolutional neural network model and a second convolutional neural network model,
The step of learning the two-dimensional convolutional neural network model is,
A method for predicting drug response, which is a step of training the calculated value of the first cross product and the calculated value of the second cross product as input values for each of the first convolutional neural network model and the second convolutional neural network model. As a learning method of a 2-dimensional Convolutional Neural Network model for predicting drug responses of drugs and cell lines,
preparing a first drug-drug similarity matrix;
preparing a second similarity matrix between cell lines;
A step of calculating a cross product between the first similarity matrix and the second similarity matrix, comprising the i-th column vector of the first similarity matrix (i = 1, 2, ... m: m is an integer) and the second similarity matrix Calculating a cross product between the j-th column vectors (j = 1, 2, ..., n: n is an integer); and
Learning the two-dimensional convolutional neural network model using the two-dimensional matrix that is the calculated value of the cross product as an input value and the drug response value of the i-th drug and the j-th cell line as output values; Including,
The i-th column vector of the first similarity matrix is the similarity between the i-th drug and other drugs including the i-th drug,
The j-th column vector of the second similarity matrix is a two-dimensional convolution that predicts the drug response of a drug and a cell line, which is the gene expression similarity between the j-th cell line and other cell lines including the j-th cell line. Learning methods for neural network models. According to claim 7,
The second similarity matrix includes a first submatrix and a second submatrix,
The step of calculating a cross product between the first similarity matrix and the second similarity matrix includes:
Comprising a step of calculating a first cross product between the first similarity matrix and the first submatrix, and calculating a second cross product between the first similarity matrix and the second submatrix,
The two-dimensional convolutional neural network model includes a first convolutional neural network model and a second convolutional neural network model,
The first convolutional neural network model is learned using the calculated value of the first cross product as an input value, and the second convolutional neural network model is learned using the calculated value of the second cross product as an input value. Drug and cell line Learning method of a 2D convolutional neural network model to predict drug response. A system that predicts the drug response of drugs and cell lines using a two-dimensional convolutional neural network model,
a control unit for controlling the two-dimensional convolutional neural network model;
A communication unit for communication with an external server;
memory unit;
display unit; and
Includes an input unit that receives user input,
The memory unit includes a first similarity matrix between drugs and drugs and a second similarity matrix between cell lines and cell lines,
The control unit controls the ith column vector of the first similarity matrix (i = 1, 2, ... m: m is an integer) and the jth column vector of the second similarity matrix (j = 1, 2, ..., n: n is an integer), the two-dimensional matrix that is the calculated value of the cross product is used as the input value, and the drug response value of the i-th drug and the j-th cell line is used as the output value to learn the two-dimensional convolutional neural network model. And
The drug response values of a new drug and a cell line are predicted using the learned two-dimensional convolutional neural network model, and the i-th column vector of the first similarity matrix is the i-th drug and other drugs including the i-th drug. is the similarity between them,
The j-th column vector of the second similarity matrix is the gene expression similarity between the j-th cell line and other cell lines including the j-th cell line, a drug response prediction system. According to clause 9,
The first similarity matrix is an mxm drug-drug similarity matrix based on the Tanimoto coefficient,
The second similarity matrix is an nxn cell line-cell line similarity matrix that is a radial basis function (RBF) kernel matrix,
A drug response prediction system, where m and n represent the number of drugs and cell lines in the learning dataset, respectively. According to clause 9,
The second similarity matrix includes a first submatrix and a second submatrix,
The calculated value of the cross product is,
A drug response prediction system comprising a calculated value of a first cross product between the first similarity matrix and the first submatrix, and a calculated value of a second cross product between the first similarity matrix and the second submatrix. The method according to any one of claims 9 to 11,
The drug response value is quantified through the IC ₅₀ (half-maximal inhibitory concentration) value of the cell line for the drug. Value-in, drug response prediction system. According to claim 11,
The two-dimensional convolutional neural network model includes a first convolutional neural network model and a second convolutional neural network model,
The first convolutional neural network model is learned using the calculated value of the first cross product as an input value, and the second convolutional neural network model is learned using the calculated value of the second cross product as an input value. A drug response prediction system . According to claim 13,
Predicting drug response values of new drugs and cell lines using the learned two-dimensional convolutional neural network model is as follows:
The calculated values of the first cross product and the calculated values of the second cross product respectively calculated between the column vector of the first similarity matrix including the new drug, the column vector of the first partial matrix, and the column vector of the second partial matrix. Each input to the first convolutional neural network model and the second convolutional neural network model, and a final drug response based on two drug response values that are output values of each of the first convolutional neural network model and the second convolutional neural network model. Drug response prediction system that predicts value. According to clause 9,
The second similarity matrix includes two or more submatrices,
Calculation of the cross product between the first similarity matrix and the second similarity matrix is:
Comprising two or more cross products between the first similarity matrix and each of the two or more partial matrices,
The two-dimensional convolutional neural network model includes two or more two-dimensional convolutional neural network submodels,
The control unit trains the two or more two-dimensional convolutional neural network submodels by using the calculated values of the two or more cross products as input values for each of the two or more two-dimensional convolutional neural network submodels and the drug response values as output values, respectively. Drug response prediction system. According to clause 9,
The two-dimensional convolutional neural network model includes two two-dimensional convolutional layers, two max pooling layers, one smoothing layer, one dropout layer, and two fully connected layers. A drug response prediction system.