Method and system for predicting synergistic effects of drug combinations based on knowledge graph

Abstract

The present invention discloses a method and system for predicting the synergistic effect of drug combinations based on a knowledge graph. The method uses the relationship between proteins, drugs, cell lines and tissues to construct a large-scale knowledge graph. The neighbor node embedding of drugs and cell lines is combined with the self-embedding to obtain a global representation of drugs and cell lines. RDkit is used to calculate the fingerprint of each drug and the descriptor based on SMILES expression to obtain a local representation of the drug. The gene expression data of the cell line is reduced in dimension using an autoencoder to obtain a local representation of the cell line. The local representations of drugs and cell lines are fused, and a deep neural network is used to predict the synergistic score of the drug combination. The present invention fully mines the complex semantic relationships in the knowledge graph to enhance the model's ability to understand potential drug combinations; at the same time, local molecular characteristics and global network topology information are introduced to provide a more comprehensive representation of drugs and cell lines, thereby enhancing the robustness of the prediction model.

Description

Technical Field

[0001] The present invention belongs to the field of medical informatics, and in particular relates to a method and system for predicting the synergistic effect of drug combinations based on a knowledge graph. Background Art

[0002] Drug synergy refers to the phenomenon that the effect of two or more drugs used simultaneously is better than that of using them alone. In complex diseases such as cancer, HIV and cardiovascular disease, many cellular mechanisms are commonly varied in different tissues and organ systems. Compared with monotherapy, drug combination therapy has great advantages, which can enhance the efficiency of cancer treatment, overcome drug resistance and reduce dose-dependent toxicity.

[0003] In the past few decades, many computational methods for drug synergy prediction have been developed to narrow the search space of drug combinations, thereby reducing costs and accelerating the identification process of new potential combination therapies. In general, methods for drug synergy prediction can be divided into three categories: structure-based methods, tensor decomposition-based methods, and network-based methods. Structure-based methods mainly select the chemical structure information of drugs as features for prediction. Tensor decomposition-based methods decompose multi-directional data constructed from the original data set to capture the potential relationships between different variables, and then use the decomposition results as features for downstream analysis. Network-based methods use various networks as side information to improve prediction performance, such as protein-protein interactions, drug-drug interactions, and drug-target interactions.

[0004] In recent years, network-based drug synergy prediction has received a lot of attention. Network data such as protein-protein interactions, drug-drug interactions, and drug-target interactions are integrated with omics data to predict anticancer synergistic drug combinations. Network-based methods can obtain more comprehensive biological information by integrating multiple network data and omics data, which helps to predict drug synergy more accurately; network-based methods can analyze the mechanism of drug action from the perspective of systems biology, taking into account multiple biological molecules and their interactions, which helps to understand the global effects of drugs. However, network-based drug synergy prediction also faces challenges such as high data dependence, high computational cost, and insufficient understanding of biological mechanisms.

[0005] At present, drug synergy prediction based on knowledge graphs constructs other entities including drugs, cell lines and proteins into knowledge graphs, and uses global information to predict drug synergy through information propagation and aggregation of knowledge graphs, such as the Chinese patent with publication number CN119495443A, entitled Anticancer Drug Synergy Prediction Method Based on Knowledge Graph Embedding, GNN and Bi-LSTM. However, the existing methods still have shortcomings. The information in the knowledge graph is not effectively extracted, and only the global information is taken into account, without considering the local information of drugs and cell lines. Summary of the invention

[0006] The technical solution of the present invention provides a method and system for predicting the synergistic effect of drug combinations based on knowledge graphs. First, a large-scale knowledge graph is constructed using drug-protein, cell line-protein, protein-protein, and cell line-tissue interactions. The hierarchical propagation of the knowledge graph is applied to find the neighbors of drugs and cell lines. Then, an attention network is used to capture the complex semantic information of various entities in the knowledge graph, and an aggregator is used to integrate all the embeddings of the attention layer to obtain the neighbor node embeddings of drugs and cell lines. The neighbor node embeddings of drugs and cell lines are combined with their own embeddings to enrich drugs and cell lines to obtain a global representation of drugs and cell lines. RDkit is used to calculate the fingerprint of each drug and the descriptor based on SMILES expressions, and the two are spliced ​​and normalized to obtain the local information of the drug. The gene expression data of the cell line is reduced in dimension using an autoencoder to obtain the local information of the cell line. Based on the global representation of drugs and cell lines learned from the knowledge graph, the local representation of drugs and cell lines is integrated, and a deep neural network is used to predict the synergistic score of the drug combination. The present invention proposes to fuse local and global information of drugs and cell lines, thereby obtaining a comprehensive and specific prediction model, thereby enhancing the robustness of the prediction model.

[0007] In order to achieve the above technical purpose, the technical solution of the present invention is:

[0008] In a first aspect, a method for predicting drug combination synergy based on a knowledge graph comprises the following steps:

[0009] Step S1: Using the associations between drugs and proteins, proteins and proteins, cell lines and proteins, and cell lines and tissues, a large-scale knowledge graph containing drugs, cell lines, proteins, and tissues is constructed;

[0010] Step S2: utilizing neighborhood sampling and information propagation of the large-scale knowledge graph to obtain the drug and cell line feature representations at each layer of the knowledge graph, and concatenating them with the initial embedding representations of the drugs and cell lines to obtain the final global features of the drugs and cell lines;

[0011] The initial embedding of drugs and cell lines uses the nn.Embedding module in the python library, through the entity embedding matrix E and the relationship embedding matrix Initialized to random values ​​and optimized during neighborhood sampling and information propagation in the knowledge graph;

[0012] Step S3: RDkit is used to calculate the fingerprint and descriptor of each drug based on SMILES expression, and the two are concatenated and normalized to obtain the local features of the drug;

[0013] Step S4: using an autoencoder to reduce the dimension of the gene expression data of the cell line to obtain the local features of the cell line;

[0014] Step S5: Based on the global features of drugs and cell lines learned from the knowledge graph, the local features of drugs and cell lines are integrated and the deep neural network is used to predict the synergy score of the drug combination.

[0015] RDKit is an open source chemical informatics toolkit used for tasks such as molecular modeling, drug design, and chemical data analysis; SMILES is the "Simplified Molecular Input Line Entry System", a linear symbol system for representing molecular structures.

[0016] This technical solution uses the hierarchical propagation of knowledge graphs to find the neighbors of drugs and cell lines, uses an attention network to capture the complex semantic information of various entities in the knowledge graph, and uses a deep neural network to predict the synergy score of drug combinations based on the global representation of drugs and cell lines learned from the knowledge graph and the local representation of fused drugs and cell lines. By analyzing the characteristic association relationship between drugs and cell lines, the complex semantic relationship in the knowledge graph is fully mined, and then the local and global information of the fused drugs and cell lines proposed by the present invention is used, while the local molecular characteristics and global network topology information are introduced to provide a more comprehensive representation of drugs and cell lines, and a prediction model that is both comprehensive and specific is obtained.

[0017] Furthermore, the step S1 includes the following specific steps:

[0018] 1) Drug-protein, cell line-protein, protein-protein, and cell line-tissue associations extracted from bioresource databases;

[0019] 2) Building a large-scale knowledge graph based on the associations between drugs, cell lines, proteins, and tissues;

[0020] The large-scale knowledge graph is represented as ,pass The triples of the form are constructed, each triple represents a relationship , and the head entity With tail entity Connected; among them, Represents a collection of entities, Represents a collection of relationships; entities include drugs, proteins, cell lines, and tissues. The relationships in large-scale knowledge graphs include the associations between drugs and proteins, cell lines and proteins, proteins and proteins, and cell lines and tissues.

[0021] Furthermore, step S2 includes the following specific steps:

[0022] 1) Let e ​​represent the set of drugs and cell lines, and the number of propagation layers of the knowledge graph is , In the The set of neighbor entities of a layer is defined as: ;

[0023] 2) The initial embedding representation of entities and relations is and , through the entity embedding matrix and the relation embedding matrix Initialized to random values ​​and optimized during neighborhood sampling and information propagation in the knowledge graph;

[0024] in, and , and Represent the number of entities and relationships respectively, and Respectively and The respective embedding dimensions, represents the field of real numbers;

[0025] 3) Use a two-layer neural network to build a weighted function ;

[0026] ;

[0027] in, is the intermediate state vector: , is the rectified linear unit activation function, and are trainable weights and biases, Represents a vector concatenation operation; is the Sigmoid activation function, with parameters , and , They are the trainable weight matrices and biases in the first and second layers of the neural network, respectively;

[0028] The first layer of the neural network corresponds to , that is, the activation function used by the first layer of the neural network is the rectified linear unit activation function, and the second layer of the neural network corresponds to , that is, the second layer of the neural network uses the Sigmoid activation function;

[0029] The weights are defined by the neural network, which are updated through back propagation and gradient descent throughout the process of acquiring global features of drugs and cell lines. Specifically, the weights of each layer in the neural network (e.g. , , , and the bias term , , ) will be adjusted continuously according to the feedback of the MSE loss function as the training progresses.

[0030] 4) Apply the softmax function to the weighting function Perform normalization;

[0031] ;

[0032] in, represents a weighting function with softmax normalization, Representing Entities Neighborhood set of Other triples in , excluding specific triples associated with entity h , is the head entity of the relationships associated with entity h excluding r, It is the tail entity associated with entity h excluding t;

[0033] 5) For triples Calculate the weighted sum to calculate the entity The initial aggregated neighbor embedding representation of :

[0034] ;

[0035] in, Representing Entities The neighborhood set of Represents the neighbor entity of entity h The initial embedding of

[0036] 6) Use multi-head attention mechanism to obtain entity The neighbors of are finally embedded in the output feature representation:

[0037] ;

[0038] in, is an activation function, represents the concatenation operation, with a total of K attention heads, Indicates The normalized attention coefficient calculated by the attention heads;

[0039] 7) Initial embedding of entity h The final embedding with its neighbors The final embedding of each entity is obtained by using a two-way interactive aggregator. as follows:

[0040] ;

[0041] in, and is a trainable weight matrix, Represents element-wise multiplication;

[0042] Knowledge Graph Layer Entity The final embedding Obtained by the following equation:

[0043] ;

[0044] in, Indicates The entity h of the layer is initially embedded, Indicates The neighbors of entity h at the layer are finally embedded;

[0045] ;

[0046] in, Represents the knowledge graph Layer Entity Neighbor entity The final embedding of

[0047] 8) The embeddings of all layers are concatenated, and the final embedding representation of drugs and cell lines, i.e., the global features, are defined as follows:

[0048] ;

[0049] ;

[0050] in, Drugs after splicing all layers The final embedding of Cell lineage after splicing all layers The final embedding of Represents the knowledge graph Layer of Drugs The aggregation embedding Represents the knowledge graph Cell lines The aggregation embedding Represents a concatenation operation.

[0051] Furthermore, step S3 includes the following specific steps:

[0052] 1) Generate a Morgan fingerprint for each drug's SMILES representation using RDkit and represent the Morgan fingerprint as a binary vector;

[0053] 2) Use RDkit to process the drug SMILES representation to obtain the descriptor of each drug, and form the descriptors of all drugs into a real-valued vector;

[0054] 3) Concatenate the binary vector and the real-valued vector of the drug's Morgan fingerprint, and screen the concatenated feature vectors to remove the feature vectors with zero variance;

[0055] 4) Use the z-score normalization method to preprocess the drug splicing features after screening to obtain the local characteristics of the drug.

[0056] Furthermore, step S4 includes the following specific steps:

[0057] 1) The original gene expression features of the cell line are input into the encoder in the autoencoder module for encoding. The encoding process is:

[0058] ;

[0059] Among them, the input data is x, and the output of the encoder is feature h1. The mapping function of the encoder is a parameterized neural network. The encoder of the autoencoder module has three fully connected layers and uses the ReLU activation function.

[0060] 2) Input the encoded feature h1 into the decoder of the autoencoder for decoding. The decoding process is:

[0061] ;

[0062] The output of the decoder is , Represents the mapping function of the decoder. The decoder and encoder of the autoencoder module adopt a symmetrical structure and use the ReLU activation function;

[0063] 3) During the training process of the autoencoder module, the parameters of the autoencoder module are learned on the dataset by minimizing the mean square loss function. The loss function is defined as follows:

[0064] ;

[0065] Among them, n is the dimension of the input data. After training, the intermediate variable h1 is extracted as the local feature of the cell line after dimensionality reduction.

[0066] Furthermore, step S5 includes the following specific steps:

[0067] 1) Take the medicine and medications The local and global features are expressed as , local and global features of the cell line are represented as and , perform feature fusion based on weighting and obtain the fusion vector:

[0068] ;

[0069] in, represents the weighted fusion vector, is the weight of the global feature, is the weight of the local feature, and its initial value is set to 0.5;

[0070] 2) Input the fusion vector of drugs and cell lines into the deep neural network to predict the synergistic effect score of drug combination;

[0071] During training, the deep neural network uses the known drug combination synergy scores of the drug and cell line as the output of the deep neural network, and uses the fusion vector of the drug and cell line as the input. , taking the minimum error between the predicted value and the known value of the drug combination synergistic effect score as the training goal, the parameters of the deep neural network are continuously adjusted during iterative training, and finally a drug combination synergistic effect score prediction model based on the deep neural network is obtained.

[0072] In a second aspect, a prediction system using the above-mentioned knowledge graph-based drug combination synergy prediction method comprises:

[0073] Knowledge graph construction module: Utilize the associations between drugs and proteins, proteins and proteins, cell lines and proteins, and cell lines and tissues to build a large-scale knowledge graph containing drugs, cell lines, proteins, and tissues;

[0074] Drug and cell line global feature extraction module: utilizes neighborhood sampling and information propagation of the large-scale knowledge graph to obtain the feature representation of drugs and cell lines at each layer of the knowledge graph, and concatenates it with the initial embedding representation of drugs and cell lines to obtain the final global features of drugs and cell lines;

[0075] Drug local feature calculation module: RDkit is used to calculate the fingerprint and descriptor of each drug based on SMILES expression, and the two are concatenated and normalized to obtain the local features of the drug;

[0076] Cell line local feature calculation module: Use autoencoder to reduce the dimension of gene expression data of cell lines to obtain local features of cell lines;

[0077] Synergy score prediction module: Based on the global features of drugs and cell lines learned from the knowledge graph, the local features of drugs and cell lines are integrated and a deep neural network is used to predict the synergy score of the drug combination.

[0078] In a third aspect, a computer storage medium stores a computer program, wherein the computer program is called by a processor to implement:

[0079] The steps of the above-mentioned knowledge graph-based drug combination synergistic effect prediction method.

[0080] In a fourth aspect, an electronic terminal at least comprises:

[0081] one or more processors;

[0082] and a memory storing one or more computer programs;

[0083] Wherein, the processor calls the computer program to execute: the steps of the above-mentioned knowledge graph-based drug combination synergistic effect prediction method.

[0084] The technical effect of the present invention is to construct a large-scale knowledge graph using drug-protein, cell line-protein, protein-protein, and cell line-tissue interactions. Hierarchical propagation of knowledge graphs is applied to find neighbors of drugs and cell lines. An attention network is used to capture the complex semantic information of various entities in the knowledge graph, and an aggregator is used to integrate all embeddings of the attention layer to obtain neighbor node embeddings of drugs and cell lines. The attention mechanism can learn different weights for different neighbors. The neighbor node embeddings of drugs and cell lines are combined with their own embeddings, and the original feature representation is used to enrich drugs and cell lines to obtain global representations of drugs and cell lines. RDkit is used to calculate the fingerprint of each drug and the descriptor based on SMILES expression, and the two are spliced ​​and normalized to obtain local information of the drug. The gene expression data of the cell line is reduced by an autoencoder to obtain local information of the cell line. Based on the global representation of drugs and cell lines learned from the knowledge graph, the local representation of drugs and cell lines is integrated, and a deep neural network is used to predict the synergy score of the drug combination. By analyzing the characteristic associations between drugs and cell lines, fully mining the complex semantic relationships in the knowledge graph, and then utilizing the local and global information of the fusion drugs and cell lines proposed in the present invention, while introducing local molecular characteristics and global network topology information, a more comprehensive drug and cell line characterization is provided, and a comprehensive and specific prediction model is obtained, which enhances the robustness of the model and improves the model's ability to understand potential drug combinations, so that the synergistic effects between drug combinations can be obtained more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a flowchart of the overall process of predicting drug synergy according to an example of the present invention;

[0086] Figure 2 It is the overall framework diagram of predicting drug synergy according to the example of the present invention;

[0087] Figure 3 A flowchart of the steps for obtaining local characteristics of a drug according to an example of the present invention;

[0088] Figure 4 A flowchart of the steps for obtaining local characteristics of a cell line according to an example of the present invention;

[0089] Figure 5 A flowchart of obtaining global features of drugs and cell lines at each layer during the knowledge graph information propagation process described in an example of the present invention;

[0090] Figure 6 Schematic diagram of the comparison between the example of the present invention and other methods. DETAILED DESCRIPTION

[0091] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0092] Example 1

[0093] like Figure 1 and Figure 2 As shown, a drug combination synergy prediction method based on knowledge graph adopted in the technical solution of this embodiment is implemented in Python programming language, using the drug synergy dataset proposed by O'Neil et al., which includes 38 drugs and 39 cell lines from 7 different tissues, a total of 583 drug combinations, and 23052 samples. The experiment adopts five-fold cross validation. The specific steps are as follows:

[0094] Step 1: Utilize drug-protein, protein-protein, cell line-protein, and cell line-tissue interactions to build a large-scale knowledge graph containing drugs, cell lines, proteins, and tissues.

[0095] 1) The biological resources are the associated data of drugs, cell lines, proteins and tissues involved in the benchmark datasets extracted from databases such as DrugBank, BindingDB, ChEMBL, PharmGKB and CCLE, totaling 229,782 associated data.

[0096] 2) The large-scale knowledge graph is represented as ,in Represents a collection of entities, Represents a collection of relationships. In this graph, entities include drugs, proteins, cell lines, and tissues. The relationships in the knowledge graph include the associations between drugs and proteins, cell lines and proteins, proteins and proteins, and cell lines and tissues. The triples are constructed in the form of , the head entity With tail entity Connect it up.

[0097] Step 2: Use neighborhood sampling and information propagation of the knowledge graph to obtain the drug and cell line representations at each layer of the knowledge graph, and then concatenate them with their original representations to obtain the final global features of the drugs and cell lines, such as Figure 5 shown.

[0098] 1) Let e ​​represent the drug and cell line, and the number of propagation layers is , In the The set of neighbor entities of a layer is defined as: ,in represents a knowledge graph triple, Represents the knowledge graph. The size of the neighborhood set of each entity is limited to a maximum of 5 neighbors. Experiments were conducted by setting different numbers of neighbors, and the results showed that the best effect was achieved when the number of neighbors was set to 5.

[0099] 2) The initial embedding representation of entities and relations is and , through the entity embedding matrix and the relation embedding matrix is initialized to random values ​​and optimized during training. Specifically, and ,in and Represent the number of entities and relationships respectively, and denote their respective embedding dimensions.

[0100] 3) Define the weighting function , first calculate the intermediate state vector:

[0101] ;in, is the rectified linear unit activation function, which ensures that the output value is in the range [0, 1]. and are trainable weights and biases, Represents a vector concatenation operation. and Represent entities and relationship Embedding.

[0102] Use a two-layer neural network to define the weighting function:

[0103] ;

[0104] in, is the Sigmoid activation function, with parameters , , and are the trainable weight matrices and biases in the network.

[0105] 4) Apply the softmax function to normalize the attention coefficient. The weighting function with softmax normalization is defined as follows:

[0106] ;

[0107] in, Represents a neighborhood set Other triples in , excluding specific triples , that is, the triples related to entity h.

[0108] Pair of triples Calculate the weighted sum to calculate the entity Aggregate neighborhood embedding representation of :

[0109] ;

[0110] in, Representing Entities The neighborhood set of Represents a neighbor entity Embedding.

[0111] 5) Use a multi-head attention mechanism. The final output feature representation is:

[0112] ;

[0113] in, is an activation function, represents the concatenation operation, with a total of K attention heads, Indicates The normalized attention coefficient calculated by the attention heads.

[0114] 6) Embed each entity Embedding with its neighbors The bi-interaction aggregator is used for fusion. The aggregation function as follows:

[0115] ;

[0116] in, and is a trainable weight matrix, Represents element-wise multiplication.

[0117] In the Aggregate neighbor embeddings during layer propagation The calculation of is as follows:

[0118] ;

[0119] in, Indicates Layer Neighbor Entity The embedding of contains information from its first-order neighbors. The entity embedding of each layer is updated by the following equation:

[0120] ;

[0121] in, Indicates Entity embedding of the layer.

[0122] 7) Concatenate the embeddings of all layers. The final representation is defined as follows:

[0123] ;

[0124] ;

[0125] in, Drug embedding after all layers are stitched together, For all layers of cell line embedding, Indicates Layer of Drugs Embedded, Indicates Cell lines The embedding Represents a concatenation operation.

[0126] L={1, 2, 3, 4}. When L=4, the constructed knowledge graph achieves the best performance.

[0127] Step 3: Use RDkit to calculate the fingerprint and descriptor of each drug based on SMILES expression, concatenate the two and normalize them to obtain the local features of the drug, such as Figure 3 As shown;

[0128] 1) Generate a Morgan fingerprint for each drug's SMILES representation using RDkit, and represent the Morgan fingerprint as a 256-dimensional binary vector;

[0129] 2) Use RDkit to process the drug SMILES representation to obtain 200 descriptors for each drug, and combine the descriptors of all drugs into a real-valued vector;

[0130] 3) Concatenate the binary vector and the real-valued vector of the drug's Morgan fingerprint, and screen the concatenated feature vectors to remove the feature vectors with zero variance;

[0131] 4) The z-score normalization method was used to preprocess the screened drug splicing features to obtain 416-dimensional drug local features.

[0132] Step 4: Use the autoencoder to reduce the dimension of the gene expression data of the cell line and obtain the local features of the cell line, such as Figure 4 As shown;

[0133] 1) The original gene expression features of the cell line are input into the encoder in the autoencoder module for encoding. The encoding process is:

[0134] ;

[0135] Among them, the input data is x, and the output of the encoder is feature h1. The mapping function of the encoder is a parameterized neural network. The encoder of the autoencoder module has three fully connected layers, each of which uses 1024, 512 and 300 neuron nodes respectively, and uses the ReLU activation function;

[0136] 2) Input the encoded feature h1 into the decoder of the autoencoder for decoding. The decoding process is:

[0137] ;

[0138] The output of the decoder is , Represents the mapping function of the decoder. The decoder and encoder of the autoencoder module adopt a symmetrical structure. Each layer uses 300, 512, and 1024 neuron nodes respectively, and uses the ReLU activation function;

[0139] 3) During the training process of the autoencoder module, the parameters of the autoencoder module are learned on the dataset by minimizing the mean square loss function. The loss function is defined as follows:

[0140] ;

[0141] Among them, n is the dimension of the input data. After training, the intermediate variable h1 is extracted as the local feature of the cell line after dimensionality reduction.

[0142] Step 5: Based on the global features of drugs and cell lines learned from the knowledge graph, the local features of drugs and cell lines are integrated and the deep neural network is used to predict the synergy score of the drug combination.

[0143] 1) Take the medicine and medications The local and global features are expressed as , local and global features of the cell line are represented as and , perform feature fusion based on weighting and obtain the fusion vector:

[0144] ;

[0145] in, represents the weighted fusion vector, is the weight of the global feature, is the weight of the local feature, and its initial value is set to 0.5;

[0146] In this example, a splicing-based method is also used to fuse feature vectors: all feature vectors are directly spliced ​​into a large feature vector, and the spliced ​​fusion vector is Defined as:

[0147] ;

[0148] In addition, an average-based approach is used to perform feature vector fusion: the local and global feature vectors of drugs and cell lines are element-wise averaged, and the average-based fusion vector Defined as:

[0149] ;

[0150] Experiments show that assigning learnable weights to local and global feature vectors allows feature fusion to be performed in a weighted manner, so that the drug combination synergy effect score prediction model can learn the best contribution of each feature type during training;

[0151] 2) Input the fusion vector of drugs and cell lines into the deep neural network to predict the drug combination synergy score. The deep neural network model receives an input vector with a dimension of 1088, which is a weighted fusion of the local and global features of drug A, drug B and cell lines. The deep neural network includes two hidden layers, the first layer has a dimension of 4096 and the second layer has a dimension of 2048. Both use the ReLU activation function for nonlinear transformation. Finally, the output layer generates a real value with a dimension of 1;

[0152] During training, the deep neural network uses the known drug combination synergy scores of the drug and cell line as the output of the deep neural network, and uses the fusion vector of the drug and cell line as the input. , taking the minimum error between the predicted value and the known value of the drug combination synergistic effect score as the training goal, the parameters of the deep neural network are continuously adjusted during iterative training, and finally a drug combination synergistic effect score prediction model based on the deep neural network is obtained.

[0153] like Figure 6As shown, the horizontal axis in the figure is the cell line involved in the present invention, and the vertical axis is the root mean square error (RMSE) and the Pearson correlation coefficient (PCC), both of which are performance indicators for evaluating the method of predicting drug synergy. The smaller the RMSE, the closer the predicted value is to the true value, and the better the prediction effect. The larger the PCC, the stronger the correlation between the predicted value and the true value, and the better the prediction ability. In the final deep neural network training, five-fold cross validation is used. The mean of each fold in the figure is represented by a solid line, and the error bar of the standard deviation is represented by the shaded area. LGSyn-con (based on splicing), LGSyn-ave (based on average) and LGSyn-wei (based on weighting) are the names of the methods based on different fusion methods of the present invention, and the remaining methods (PRODeepSyn, MGAE-DC, CGMS, SNRMPACDC, DeepDDS) are comparative methods. The figure shows that compared with the comparative method, the technical solution LGSyn-ave proposed in the present invention shows lower RMSE and higher PCC on most cell lines, proving that it is superior to existing methods in terms of prediction accuracy and correlation.

[0154] Example 2

[0155] A prediction system using the above-mentioned knowledge graph-based drug combination synergy prediction method comprises:

[0156] Knowledge graph construction module: Utilize the associations between drugs and proteins, proteins and proteins, cell lines and proteins, and cell lines and tissues to build a large-scale knowledge graph containing drugs, cell lines, proteins, and tissues;

[0157] Drug and cell line global feature extraction module: utilizes neighborhood sampling and information propagation of the large-scale knowledge graph to obtain the feature representation of drugs and cell lines at each layer of the knowledge graph, and concatenates it with the initial embedding representation of drugs and cell lines to obtain the final global features of drugs and cell lines;

[0158] Drug local feature calculation module: RDkit is used to calculate the fingerprint and descriptor of each drug based on SMILES expression, and the two are concatenated and normalized to obtain the local features of the drug;

[0159] Cell line local feature calculation module: Use autoencoder to reduce the dimension of gene expression data of cell lines to obtain local features of cell lines;

[0160] Synergy score prediction module: Based on the global features of drugs and cell lines learned from the knowledge graph, the local features of drugs and cell lines are integrated and a deep neural network is used to predict the synergy score of the drug combination.

[0161] It should be understood that the specific implementation process of each module please refer to the above method content, the present invention will not be repeated here, and the division of the above functional modules is only for example illustration. In some embodiments, some functional modules can be merged, and some functional modules can be split. Each functional module can be implemented in software or hardware or a combination of software and hardware. Among them, the hardware and software equipment includes but is not limited to general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0162] Example 3

[0163] A computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to implement:

[0164] The above steps of a method for predicting the synergistic effect of drug combinations based on knowledge graph.

[0165] For the specific implementation process of each step, please refer to the description of the above method.

[0166] It should be understood that in the embodiments of the present invention, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the type of device.

[0167] The readable storage medium is a computer-readable storage medium, which may be an internal storage unit of the software and hardware device described in any of the aforementioned embodiments, such as a hard disk or memory of a controller. The readable storage medium may also be an external storage device of the controller, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the controller. Furthermore, the readable storage medium may also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium may also be used to temporarily store data that has been output or is to be output.

[0168] Example 4

[0169] An electronic terminal comprises at least: one or more processors; and a memory storing one or more computer programs; wherein the processor calls the computer program to execute:

[0170] Steps of a method for predicting synergistic effects of drug combinations based on knowledge graph.

[0171] For the specific implementation process of each step, please refer to the description of the above method.

[0172] Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0173] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is a device for realizing the function specified in one flow or multiple flows of the flow chart and / or one or multiple blocks of the block diagram with reference to the instructions executed by the processor according to the method, device (system) and computer program product of the embodiment of the present application. These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which realizes the function specified in one flow or multiple flows of the flow chart and / or one or multiple blocks of the block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0174] It should be emphasized that the examples described in the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. A method for predicting drug combination synergy based on knowledge graph, characterized in that: The following steps are involved: Step S1: Using the associations between drugs and proteins, proteins and proteins, cell lines and proteins, and cell lines and tissues, a large-scale knowledge graph containing drugs, cell lines, proteins, and tissues is constructed; Step S2: utilizing neighborhood sampling and information propagation of the large-scale knowledge graph to obtain the drug and cell line feature representations at each layer of the knowledge graph, and concatenating them with the initial embedding representations of the drugs and cell lines to obtain the final global features of the drugs and cell lines; Step S3: RDkit is used to calculate the fingerprint and descriptor of each drug based on SMILES expression, and the two are concatenated and normalized to obtain the local features of the drug; Step S4: using an autoencoder to reduce the dimension of the gene expression data of the cell line to obtain the local features of the cell line; Step S5: Based on the global features of drugs and cell lines learned from the knowledge graph, the local features of drugs and cell lines are integrated and the synergy score of the drug combination is predicted using a deep neural network; The step S3 comprises the following specific steps: 1) Generate a Morgan fingerprint for each drug's SMILES representation using RDkit and represent the Morgan fingerprint as a binary vector; 2) Use RDkit to process the drug SMILES representation to obtain the descriptor of each drug, and form the descriptors of all drugs into a real-valued vector; 3) Concatenate the binary vector and the real-valued vector of the drug's Morgan fingerprint, and screen the concatenated feature vectors to remove the feature vectors with zero variance; 4) Use the z-score normalization method to preprocess the drug splicing features after screening to obtain the local features of the drug; The step S5 comprises the following specific steps: 1) Take the medicine and medications The local and global features are expressed as , local and global features of the cell line are represented as and , perform feature fusion based on weighting and obtain the fusion vector: ； in, represents the weighted fusion vector, is the weight of the global feature, is the weight of the local feature, and its initial value is set to 0.5; 2) Input the fusion vector of drugs and cell lines into the deep neural network to predict the synergistic effect score of drug combination; During training, the deep neural network uses the known drug combination synergy scores of the drug and cell line as the output of the deep neural network, and uses the fusion vector of the drug and cell line as the input. , taking the minimum error between the predicted value and the known value of the drug combination synergistic effect score as the training goal, the parameters of the deep neural network are continuously adjusted during iterative training, and finally a drug combination synergistic effect score prediction model based on the deep neural network is obtained.

2. The method according to claim 1, characterized in that The step S1 comprises the following specific steps: 1) Drug-protein, cell line-protein, protein-protein, and cell line-tissue associations extracted from bioresource databases; 2) Building a large-scale knowledge graph based on the associations between drugs, cell lines, proteins, and tissues; The large-scale knowledge graph is represented as ,pass The triples of the form are constructed, each triple represents a relationship , and the head entity With tail entity Connected; among them, Represents a collection of entities, Represents a collection of relationships; entities include drugs, proteins, cell lines, and tissues. The relationships in large-scale knowledge graphs include the associations between drugs and proteins, cell lines and proteins, proteins and proteins, and cell lines and tissues.

3. The method according to claim 2, characterized in that The step S2 includes the following specific steps: 1) Let e ​​represent the set of drugs and cell lines, and the number of propagation layers of the knowledge graph is , In the The set of neighbor entities of a layer is defined as: ; 2) The initial embedding representation of entities and relations is and , through the entity embedding matrix and the relation embedding matrix Initialized to random values ​​and optimized during neighborhood sampling and information propagation in the knowledge graph; in, and , and Represent the number of entities and relationships respectively, and Respectively and The respective embedding dimensions, represents the field of real numbers; 3) Use a two-layer neural network to build a weighted function ; ； in, is the intermediate state vector: , is the rectified linear unit activation function, and are trainable weights and biases, Represents a vector concatenation operation; is the Sigmoid activation function, with parameters , and , They are the trainable weight matrices and biases in the first and second layers of the neural network, respectively; 4) Apply the softmax function to the weighting function Perform normalization; ； in, represents a weighting function with softmax normalization, Representing Entities Neighborhood set of Other triples in , excluding specific triples associated with entity h , is the head entity of the relationships associated with entity h excluding r, It is the tail entity associated with entity h excluding t; 5) For triples Calculate the weighted sum to calculate the entity The initial aggregated neighbor embedding representation of : ； in, Representing Entities The neighborhood set of Represents the neighbor entity of entity h The initial embedding of 6) Use multi-head attention mechanism to obtain entity The neighbors of are finally embedded in the output feature representation: ； in, is an activation function, represents the concatenation operation, with a total of K attention heads, Indicates The normalized attention coefficient calculated by the attention heads; 7) Initial embedding of entity h The final embedding with its neighbors The final embedding of each entity is obtained by using a two-way interactive aggregator. as follows: ； in, and is a trainable weight matrix, Represents element-wise multiplication; Knowledge Graph Layer Entity The final embedding Obtained by the following equation: ； in, Indicates The entity h of the layer is initially embedded, Indicates The neighbors of entity h at the layer are finally embedded; ； in, Represents the knowledge graph Layer Entity Neighbor entity The final embedding of 8) The embeddings of all layers are concatenated, and the final embedding representation of drugs and cell lines, i.e., the global features, are defined as follows: ； ； in, Drugs after splicing all layers The final embedding of Cell lineage after splicing all layers The final embedding of Represents the knowledge graph Layer of Drugs The aggregation embedding Represents the knowledge graph Cell lines The aggregation embedding Represents a concatenation operation.

4. The method according to claim 1, characterized in that: The step S4 comprises the following specific steps: 1) The original gene expression features of the cell line are input into the encoder in the autoencoder module for encoding. The encoding process is: ； Among them, the input data is x, and the output of the encoder is the feature h1. The mapping function of the encoder is a parameterized neural network. The encoder of the autoencoder module has three fully connected layers and uses the ReLU activation function. 2) Input the encoded feature h1 into the decoder of the autoencoder for decoding. The decoding process is: ； The output of the decoder is , Represents the mapping function of the decoder. The decoder and encoder of the autoencoder module adopt a symmetrical structure and use the ReLU activation function; 3) During the training process of the autoencoder module, the parameters of the autoencoder module are learned on the dataset by minimizing the mean square loss function. The loss function is defined as follows: ； Among them, n is the dimension of the input data. After training, the intermediate variable h1 is extracted as the local feature of the cell line after dimensionality reduction.

5. A prediction system using the drug combination synergistic effect prediction method based on knowledge graph according to any one of claims 1 to 4, characterized in that: include: Knowledge graph construction module: Utilize the associations between drugs and proteins, proteins and proteins, cell lines and proteins, and cell lines and tissues to build a large-scale knowledge graph containing drugs, cell lines, proteins, and tissues; Drug and cell line global feature extraction module: utilizes neighborhood sampling and information propagation of the large-scale knowledge graph to obtain the feature representation of drugs and cell lines at each layer of the knowledge graph, and concatenates it with the initial embedding representation of drugs and cell lines to obtain the final global features of drugs and cell lines; Drug local feature calculation module: RDkit is used to calculate the fingerprint and descriptor of each drug based on SMILES expression, and the two are concatenated and normalized to obtain the local features of the drug; Cell line local feature calculation module: Use autoencoder to reduce the dimension of gene expression data of cell lines to obtain local features of cell lines; Synergy score prediction module: Based on the global features of drugs and cell lines learned from the knowledge graph, the local features of drugs and cell lines are integrated and a deep neural network is used to predict the synergy score of the drug combination.

6. A computer storage medium, characterized in that: A computer program is stored, and the computer program is called by a processor to implement: the steps of the method according to any one of claims 1 to 4.

7. An electronic terminal, characterized in that: At least: one or more processors; and a memory storing one or more computer programs; Wherein, the processor calls the computer program to execute: the steps of the method described in any one of claims 1-4.

Images