Deep learning-based anticancer drug synergistic effect prediction method

Abstract

The invention discloses an anti-cancer drug synergistic effect prediction method based on deep learning, and the method can employ a network model which comprises a data input module, a drug sub-network module, a CrossNet module and a collaborative score prediction module. The data input is used for inputting drug pair-cell line characteristic data and cell line types, tissue types and cancer subtype data under the combined action of the two drugs, and the drug pair-cell line characteristic data comprises drug chemical characteristic data of the two drugs and cancer cell line gene expression data under the combined action of the two drugs; the input of the medicine sub-network module and the input of the CrossNet module are data output by the data input module; and the input of the collaborative score prediction module is the output of the drug sub-network module and the output of the CrossNet module. The method provided by the invention can more accurately and reliably predict whether the anticancer drugs have a synergistic effect.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and specifically relates to a method for predicting the synergistic effects of anticancer drugs based on deep learning. Background Art

[0002] In complex diseases, multiple cellular mechanisms of cells are often altered; therefore, treatment with a single drug and focusing on a single target is often ineffective and can lead to drug resistance with long-term use. Drug synergy is a promising solution, where the combination of multiple drugs is often more effective than the sum of the effects of the individual drugs. Especially in cancers where drug resistance is common, drug synergy has been an effective strategy for decades. Another advantage of drug synergy is that it reduces the risk of adverse reactions by lowering the required dose of each drug. Therefore, identifying the right drug combination with synergy is an important task. However, until recently, effective drug combinations were proposed based on clinical experience. Although it is possible to detect drug synergy by performing high-throughput screening experiments using cell lines, due to the huge number of drug combinations, it is not feasible to thoroughly test all drug combinations even with high-throughput screening. Therefore, it is necessary to use algorithms to predict the synergy of drug pairs to supplement and guide high-throughput synergy screening by prioritizing drug pairs. This approach enables the screening of large-scale drug combinations and reduces the time cost of experimental evaluation of drug synergy.

[0003] At present, many researchers have used machine learning and deep learning techniques to conduct research in the field of drug synergy prediction. For example, Preuer et al. used compound and genomic information as input in the paper "DeepSynergy: predicting anti-cancer drugsynergy with Deep Learning". By integrating genomic information, the model can distinguish different cancer cell lines and find specific drug combinations with the greatest efficacy for specific cell lines. The model combines information about cancer cell lines and drug combinations in its hidden layer, and ultimately produces accurate predictions of drug synergy. Kuru et al. proposed a drug synergy prediction algorithm based on deep neural networks in their paper "MatchMaker: a deep learning framework for drug synergy prediction". The model uses drug chemical structures and cell line gene expression profiles to predict the Loewe additive score of drug pairs. The model trains two parallel sub-networks for two drugs, one for each drug, to learn the specific performance of the drug on a specific cell line. The outputs of the two sub-networks are combined to become the input of the third sub-network, which is used to learn and predict the synergy of drug pairs.

[0004] Although the above studies have successfully predicted the synergistic effects between drugs through deep learning, some information about drug synergy is still ignored, such as the molecular subtype information of cancer. Heiser et al. proposed in the paper "Subtype and pathway specific responses to anticancer compounds in breast cancer" that in the trial of 77 compounds for the treatment of breast cancer, about one-third of the drugs showed subtype-specific responses. This shows that most drugs are specific to cancer subtypes. In addition, the feature interaction methods in existing deep learning methods are still insufficient. The feature interaction methods used by existing deep learning methods are all based on implicit feature interactions of neural networks, and cannot well mine the interactive information between features. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a network model for predicting drug synergy based on deep learning. The network model uses cancer cell line gene expression data, drug chemistry feature data and cancer subtype information, and can be used to predict whether there is a synergistic effect between two drugs used for anti-cancer treatment.

[0006] Another object of the present invention is to provide a method for predicting the synergistic effects of anticancer drugs based on deep learning. The method for predicting the synergistic effects of anticancer drugs introduces a combination of implicit feature interactions and explicit feature interactions, which can better explore the interactions between features.

[0007] The purpose of the present invention is achieved through the following technical solutions.

[0008] A network model for predicting drug synergy based on deep learning, comprising: a data input module, a drug sub-network module, a CrossNet module based on an explicit feature interaction method, and a synergy score prediction module, wherein:

[0009] The data input module includes: an Input layer, an Embedding layer, and a Concatenate layer. The Input layer is used to input drug pair-cell line feature data, which includes drug chemical feature data of two drugs and gene expression data of cancer cell lines on which the two drugs act together. The Embedding layer is used to input cell line types, tissue types, and cancer subtype data on which the two drugs act together. The Concatenate layer connects the data passing through the Input layer with the data passing through the Embedding layer.

[0010] The input of the drug sub-network module is the data output by the data input module; the drug sub-network module includes multiple Keras layers: the Dense layer, Dropout layer, Dense layer, Dropout layer and Dense layer through which the data passes in sequence;

[0011] The input of the CrossNet module is the data output by the data input module. The CrossNet module consists of multiple Cross layers, and each layer is calculated through the feature cross formula;

[0012] The input of the synergy score prediction module is the output of the drug sub-network module and the output of the CrossNet module. The output of the synergy score prediction module is: the predicted value of the synergy score value with the highest density distribution of drug pairs and the label value of whether the drug pair has synergistic effect.

[0013] In the above technical solution, the feature intersection formula is:

[0014] X l+1 =X0X l TW l +b l +X l ,

[0015] Among them, X l+1 , X lare the outputs of the l+1th Cross layer and the lth Cross layer, respectively. X0 is the input of the first Cross layer. W l and b l is the weight and bias of the lth Cross layer.

[0016] In the above technical solution, the collaborative score prediction module involves multiple Keras layers: the Concatenate layer, the Dense layer, the Dropout layer, the Dense layer, the Dropout layer and the Dense layer through which the data passes in sequence.

[0017] A method for predicting anticancer drug synergy based on deep learning, comprising the following steps:

[0018] S1, prepare data of multiple samples: the data of each sample includes: drug chemical characteristic data of two drugs and gene expression data of cancer cell lines on which the two drugs act together;

[0019] S2, data preprocessing, including S2-1 and S2-2;

[0020] S2-1, normalize each dimension of the sample data to obtain normalized data;

[0021] S2-2, perform data dimension reduction on the normalized data to obtain drug pair-cell line feature data;

[0022] S3, substituting the drug pair-cell line feature data as the training set and the cell line type, tissue type and cancer subtype data corresponding to the drug pair into the network model for training. When the loss value of the network model does not decrease within 50 training rounds, the trained network model is obtained;

[0023] S4, substituting the drug pair-cell line feature data as the prediction sample and the cell line type, tissue type and cancer subtype data corresponding to the drug pair in the prediction sample into the trained network model for prediction.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] The present invention proposes for the first time a deep learning model that introduces cancer subtype information and combines implicit feature interactions with explicit feature interactions, called a network model (MolCross). The performance of the network model is evaluated by cross-validation. The network model proposed in the present invention shows competitive results in various evaluation indicators, and can more accurately and reliably predict whether anticancer drugs have synergistic effects. After interpretable analysis, it is found that cancer subtype is one of the important factors affecting whether drugs have synergistic effects.

[0026] The network model proposed in the present invention includes a data input module, a drug sub-network module, a CrossNet module and a synergistic score prediction module. The synergistic score prediction module outputs the predicted value of the synergistic score value between drug pairs and the label value of whether the drug pairs have synergistic effects, thereby realizing the prediction of the synergistic score between drugs combining implicit feature interaction and explicit feature interaction; by dividing the synergistic scores of two drugs into two categories of whether they have synergistic effects according to the judgment value, the regression prediction method can be converted into a binary classification prediction method, and the effectiveness of the anticancer drug synergistic effect prediction method based on deep learning of the present invention can be evaluated from two aspects of regression and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of the deep learning-based method for predicting anticancer drug synergy;

[0028] Figure 2 It is a schematic diagram of the structure of the network model (MolCross);

[0029] Figure 3 It is a structural diagram of the data input module;

[0030] Figure 4 It is a schematic diagram of the structure of the drug sub-network module;

[0031] Figure 5 This is a schematic diagram of the structure of the CrossNet module;

[0032] Figure 6 It is a schematic diagram of the structure of the collaborative score prediction module;

[0033] Figure 7 It is a structural diagram of the network model;

[0034] Figure 8 It is a density distribution diagram of drug synergy score values;

[0035] Fig. 9 This is a comparison chart of the predicted results and actual results of the network model for predicting drug synergy based on deep learning for different tissue types in Example 3;

[0036] Fig.10 It is the MSE heat map of the prediction results of the network model for predicting drug synergy based on deep learning for different drug pairs in Example 3;

[0037] Fig.11 is a ROC curve diagram of the network model for predicting drug synergy based on deep learning in Example 3;

[0038] Fig.12 This is a PR curve diagram of the network model for predicting drug synergy based on deep learning in Example 3. DETAILED DESCRIPTION

[0039] The technical solution of the present invention is further described below in conjunction with specific embodiments.

[0040] In the following example, the synergy score of two drugs in the sample is performed according to the Synergy drug synergy score to obtain the density distribution of the synergy score value of the sample (the density distribution is visualized as shown in Figure 8 As shown in the figure, the label value of the sample is obtained according to the density distribution: the synergy score value with the largest density distribution is selected as the judgment value. When the judgment value is not higher than 0, there is no synergy between the two drugs in the sample (marked as a negative class); when the judgment value is higher than 0, there is a synergy between the two drugs in the sample (marked as a positive class).

[0041] Drug pair: two drugs whose synergistic effect needs to be predicted.

[0042] Example 1

[0043] A network model (MolCross) based on deep learning for predicting drug synergy, such as Figure 2 As shown, the network model includes: a data input module, a drug sub-network module, a CrossNet module based on an explicit feature interaction method, and a synergy score prediction module, where:

[0044] like Figure 3 As shown, the data input module includes: an Input layer, an Embedding layer and a Concatenate layer. The Input layer is used to input drug pair-cell line feature data, which includes drug chemistry feature data of two drugs (drug pair) and gene expression data of cancer cell lines on which the two drugs act together. The Embedding layer is used to input cell line types, tissue types, and cancer subtype data on which the two drugs act together, to facilitate feature extraction during training. The Concatenate layer connects the data passing through the Input layer with the data passing through the Embedding layer.

[0045] The input of the drug sub-network module is the data output by the data input module; Figure 4As shown in the figure, the drug sub-network module includes multiple Keras layers: Dense layer, Dropout layer, Dense layer, Dropout layer and Dense layer, through which the data passes in sequence. In the first two Dense layers and the first two Dropout layers, the relationship between features is mined by implicit feature interaction, where the first two Dense layers use the relu activation function for the activated neurons. The third Dense layer converts the elements into one dimension through the linear activation function, so that the positions of the elements in the subsequent synergy score prediction module correspond to each other. In the entire drug sub-network module, the drug chemical features and cell line types, tissue types, and molecular subtypes of cancer are mined through the implicit feature interaction of traditional neural networks to mine the relationship between features, and different weights are assigned to each feature during prediction. The drug sub-network module constructs a drug sub-network for each drug in the drug pair, which can better mine the feature relationship between a single drug and the cell line on which the drug acts, and helps to better mine the feature relationship between drugs in the following modules.

[0046] The input of the CrossNet module is the data output by the data input module. The CrossNet module consists of six Cross layers, and each layer is calculated through the feature cross formula:

[0047] X l+1 =X0X l T W l +b l +X l

[0048] Among them, X l+1 , X l are the outputs of the l+1th Cross layer and the lth Cross layer, respectively. X0 is the input of the first Cross layer. W l and b l are the weights and biases of the lth Cross layer. The output of each Cross layer depends on the output of the previous Cross layer and the data output from the data input module. In the training process of the CrossNet module, the idea of ​​Residual Learning is also introduced, which can effectively avoid the problem of gradient disappearance, and can help build deeper networks and enhance the network's expressiveness. The CrossNet module uses explicit feature interaction to increase the degree of features as the depth of the network increases. In a parameter-sharing manner, by controlling the number of superimposed layers, it can efficiently learn low-dimensional feature cross combinations and avoid artificial feature engineering.

[0049] The input of the synergy score prediction module is the output of the drug sub-network module and the output of the CrossNet module. The output of the synergy score prediction module is: the predicted value of the synergy score value with the highest density distribution of drug pairs and the label value of whether the drug pair has synergy. (When the predicted value of the synergy score value is not higher than 0, there is no synergy between the two drugs in the sample; when the predicted value of the synergy score value is higher than 0, there is synergy between the two drugs in the sample; the feature space of the label value of each sample is 1*1). Figure 6 As shown in the figure, the synergy score prediction module involves multiple Keras layers: the data passes through the Concatenate layer, Dense layer, Dropout layer, Dense layer, Dropout layer and Dense layer in sequence. The first Concatenate layer combines the data processed by the drug sub-network module and the data processed by the CrossNet module. The relu activation function is used for the activated neurons in the first two Dense layers. The two Dropout layers are used to prevent the overfitting of the model. The third Dense layer converts the elements into one dimension and then outputs them through the linear activation function. The synergy score prediction module can more effectively mine the interaction relationship between drug pairs and cell lines, which helps to predict the synergy score more accurately. The third Dense layer combines the prediction of the synergy score between drugs with implicit feature interactions and explicit feature interactions.

[0050] The prediction of drug synergy is a regression prediction problem, which can also be converted into a binary classification problem. According to the predicted value of the synergy score, a classification can be made as to whether the drug pairs have synergy. The present invention makes full use of the gene expression data of cancer cell lines and the pathological information of drugs to construct a network model (MolCross) for predicting drug synergy based on deep learning. The drug pair-cell line feature data is input into the network model (MolCross) to predict the synergy scores of different drug pair combinations.

[0051] Example 2

[0052] like Figure 1 As shown, a method for predicting the synergistic effect of anticancer drugs based on deep learning comprises the following steps:

[0053] S1, prepare data of multiple samples: the data of each sample includes: drug chemical characteristic data of two drugs and gene expression data of cancer cell lines on which the two drugs act together;

[0054] S2, data preprocessing, including S2-1 and S2-2;

[0055] S2-1, normalize each dimension of the sample data: use the maximum-minimum normalization method to perform [0,1] normalization according to formula (1) to obtain normalized data;

[0056]

[0057] In formula (1), X is the value before normalization, X max and X min are the maximum and minimum boundaries of the values ​​before normalization, X scale is the normalized value; through data normalization, the original data (the value before normalization) is converted to the range of [0,1] through a linear method. The value of the standardized data (the value after normalization) is more concentrated and suitable for comprehensive comparative evaluation.

[0058] S2-2, data dimensionality reduction: A deep autoencoder is used to perform feature selection on the normalized data to achieve data dimensionality reduction and obtain drug pair-cell line feature data; the deep autoencoder consists of two parts, the encoder and the decoder. During the deep autoencoder training process, the encoder is used to map high-dimensional input features to a low-dimensional space, and then the decoder is used to reconstruct the input features. Initialization is performed using the Xavier initializer, cross entropy is used as the loss function, and the Adam learning algorithm is used for training. Finally, the encoder part of the trained DAE is used to generate a low-dimensional potential representation of the high-dimensional input features.

[0059] S3, substituting the drug pair-cell line feature data as a training set and the cell line type, tissue type and cancer subtype data corresponding to the drug pair into the network model in Example 1 for training. When the loss value of the network model no longer decreases within 50 training rounds, a trained network model is obtained.

[0060] S4, substituting the drug pair-cell line feature data as the prediction sample and the cell line type, tissue type and cancer subtype data corresponding to the drug pair in the prediction sample into the trained network model for prediction.

[0061] Example 3

[0062] Based on Example 2, the drug chemical characteristic data from

[0063] http: / / www.bioinf.jku.at / software / DeepSynergy / (Merck & Co. released large-scale cancer screening). In this embodiment, the drug chemistry feature data includes 21,303 drug combination data from 36 cell lines of 6 different tissue types, wherein each drug combination data includes 1,309 ECFP_6 features, 802 physicochemical features, and 2,276 toxicological features.

[0064] The gene expression data of cancer cell lines and the cell line type, tissue type, and cancer subtype data in CCLE were downloaded from https: / / depmap.org / portal / download / all / . Specifically, first select the gene expression data of the mRNA transcriptome from Deepmap Protal, that is, select the ExpressionPublic data in Custom Downloads under Downloads, and repeat the following operations for each cell line: retrieve the corresponding cancer type, tissue type, and corresponding cancer molecular subtype according to the cell line name and record them. Taking the cell line MDAMB231 as an example, search the cell line MDAMB231 in Deepmap Protal and get the tissue type as Breast and the cancer molecular subtype as Basal.

[0065] The cancer cell line gene expression data are divided into six tissues through cell line type, tissue type, and cancer subtype data: lung tissue, colon tissue, breast tissue, skin tissue, prostate tissue, and ovarian tissue. Among them, lung tissue contains 8 cancer cell lines, colon tissue contains 8 cancer cell lines, breast tissue contains 5 cancer cell lines, skin tissue contains 6 cancer cell lines, prostate tissue contains 2 cancer cell lines, and ovarian tissue contains 7 cancer cell lines. Each tissue includes 19221-dimensional gene features.

[0066] The dimensions of the drug chemistry feature data of the two drugs obtained in this example and the gene expression data of the cancer cell line on which the two drugs act together are 21303×27995, that is, a total of 21303 samples. The feature space dimension of the normalized data is 21303×27995. The dimension of the drug pair-cell line feature data is 21303×6000.

[0067] In order to evaluate the performance of the network model, the drug pair-cell line feature data obtained above were divided into training set and prediction set according to K-fold cross validation, and the drug pair-cell line feature data were subjected to five-fold cross validation: the samples in the drug pair-cell line feature data with a dimension of 21303×6000 were divided into five parts, four of which were selected as training sets, and the remaining one was selected as the prediction set. The predicted value of the synergistic score value of each predicted sample in the prediction set was obtained, and the label value for judging whether there was synergistic effect between the drug pairs in the predicted sample was obtained.

[0068] According to the five-fold cross validation, a total of 5 trained network models were obtained, and the 5 trained network models were used to predict their corresponding prediction sets. The performance indicators of the evaluation network models were averaged, as shown in Tables 1 and 2.

[0069] The performance indicators for evaluating the network model include mean square error (MSE), absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), Pearson correlation coefficient (Pearson), Spearman correlation coefficient (Spearman), R square (R2) based on collaborative score prediction and accuracy score (ACC), recall rate (Recall), F1 score (F1), sensitivity (Precision), Kappa score (Kappa), area under the receiver operating characteristic curve (ROC-AUC) and area under the precision-recall curve (PR-AUC) based on binary classification prediction;

[0070] Experimental data show that for the prediction of drug synergy scores, the MSE reached 240.37. For the classification prediction of whether drugs have synergy, the ACC reached 78.66%. As shown in Tables 1 and 2, the drug synergy prediction method based on deep learning has high prediction performance for each prediction type.

[0071] Table 1 Evaluation results of regression prediction model

[0072]

[0073] Table 2 Classification model prediction results

[0074]

[0075] Fig. 9The prediction set was divided into 6 categories according to the cell line type. The predicted values ​​of drug synergy scores were obtained after prediction respectively, and compared with the original drug synergy scores. The results showed that the drug synergy prediction method based on deep learning has good prediction performance for any cell line in the prediction set. In addition, the heat map of the mean square error of the predicted values ​​of drug synergy scores between all drug pairs in the prediction set is shown in Figure 2. Fig.10 As shown in the figure, for most drug pair combinations, the drug synergy prediction method based on deep learning has good performance.

[0076] Fig.11 and Fig.12 The precision-recall curve and receiver operating characteristic curve of the drug synergy prediction method based on deep learning are shown respectively, where the area under the precision-recall curve (PR-AUC) is 0.8325 and the area under the receiver operating characteristic curve (ROC-AUC) is 0.8422, indicating that the drug synergy prediction method based on deep learning also has good classification performance.

[0077] The present invention is described above by way of example. It should be noted that, without departing from the core of the present invention, any simple deformation, modification or other equivalent replacement that can be made by those skilled in the art without inventive effort falls within the protection scope of the present invention.

Claims

1. A network model for predicting drug synergy based on deep learning, characterized in that: include: Data input module, drug sub-network module, CrossNet module based on explicit feature interaction and synergy score prediction module, among which, The data input module includes: an Input layer, an Embedding layer, and a Concatenate layer. The Input layer is used to input drug pair-cell line feature data, which includes drug chemical feature data of two drugs and gene expression data of cancer cell lines on which the two drugs act together. The Embedding layer is used to input cell line types, tissue types, and cancer subtype data on which the two drugs act together. The Concatenate layer connects the data passing through the Input layer with the data passing through the Embedding layer. The input of the drug sub-network module is the data output by the data input module; the drug sub-network module includes multiple Keras layers: the Dense layer, Dropout layer, Dense layer, Dropout layer and Dense layer through which the data passes in sequence; The input of the CrossNet module is the data output by the data input module. The CrossNet module consists of multiple Cross layers, and each layer is calculated through the feature cross formula; The input of the synergy score prediction module is the output of the drug sub-network module and the output of the CrossNet module. The output of the synergy score prediction module is: the predicted value of the synergy score value with the highest density distribution of drug pairs and the label value of whether the drug pair has synergistic effect.

2. The network model according to claim 1, characterized in that: The feature crossover formula is: X l+1 =X0X l T W l +b l +X l , Among them, X l+1 , X l are the outputs of the l+1th Cross layer and the lth Cross layer, respectively. X0 is the input of the first Cross layer. W l and b l is the weight and bias of the lth Cross layer.

3. The network model according to claim 1, characterized in that: The collaborative score prediction module involves multiple Keras layers: the Concatenate layer, Dense layer, Dropout layer, Dense layer, Dropout layer, and Dense layer, through which the data passes in sequence.

4. The network model according to claim 2, characterized in that: The collaborative score prediction module involves multiple Keras layers: the Concatenate layer, Dense layer, Dropout layer, Dense layer, Dropout layer, and Dense layer, through which the data passes in sequence.

5. A method for predicting anticancer drug synergy based on deep learning, characterized in that: The following steps are involved: S1, prepare data of multiple samples: the data of each sample includes: drug chemical characteristic data of two drugs and gene expression data of cancer cell lines on which the two drugs act together; S2, data preprocessing, including S2-1 and S2-2; S2-1, normalize each dimension of the sample data to obtain normalized data; S2-2, perform data dimension reduction on the normalized data to obtain drug pair-cell line feature data; S3, substituting the drug pair-cell line feature data as a training set and the cell line type, tissue type and cancer subtype data corresponding to the drug pair into the network model described in any one of claims 1 to 4 for training, and obtaining a trained network model when the loss value of the network model no longer decreases within 50 training rounds; S4, substituting the drug pair-cell line feature data as the prediction sample and the cell line type, tissue type and cancer subtype data corresponding to the drug pair in the prediction sample into the trained network model for prediction.

6. A method for predicting anticancer drug synergy based on deep learning, characterized in that: The network model is trained using the following data: drug chemical feature data of the two drugs, gene expression data of cancer cell lines on which the two drugs act together, and data on cell line types, tissue types, and cancer subtypes on which the two drugs act together.

Images