Anticancer drug combination prediction method and system based on multi-source data

Abstract

The invention discloses an anti-cancer drug combination prediction method and system based on multi-source data. The method comprises the following steps: capturing structural characteristics and chemical characteristics of drug molecules by constructing an isomeric graph of the drug molecules; a view-based graph Transform framework is adopted, and drug molecule characteristics are aggregated from different angles by utilizing a graph attention mechanism. Constructing comprehensive characteristics of the cell line by using gene expression data, gene mutation data and protein interaction network information; drug characteristics and cell line characteristics are integrated, and a multi-layer perceptron is used for drug combination prediction. Characteristics containing pharmacophore information are obtained through a heterogeneous graph model, PPI network information, cell line related gene expression and other information are designed and fused based on a graph neural network, a strategy for fusing and updating cell line characteristics is also designed, finally, drug characteristics and cell line characteristics are embedded and integrated, and the drug combination synergistic effect prediction system is completed.

Description

Technical Field

[0001] The present invention relates to the field of biomedicine and computer science technology, and in particular to a method and system for predicting anticancer combined drug use based on multi-source data. Background Art

[0002] Compared with single drug therapy, combination therapy has significant advantages in improving cancer treatment efficacy, reducing drug dose dependence, reducing drug toxicity and preventing drug resistance. Identifying drug combinations is crucial for treating complex diseases while avoiding serious drug side effects. Traditional drug combination prediction methods have problems such as low efficiency and low accuracy, and as the number of drugs increases, the combinatorial explosion phenomenon has become a major challenge for discovering new drug combinations. For example, early drug combination studies relied on in vivo experiments, which were time-consuming, labor-intensive and costly, and could cause unnecessary harm to trial patients. Statistical-based computational methods rely on manual feature engineering by establishing statistical models of biological systems. They are difficult to effectively learn the interaction information between drug pairs, difficult to adapt to complex nonlinear biological processes, and easily lead to relatively low performance.

[0003] With the development of artificial intelligence and deep learning, the use of these technologies for combination prediction has gradually become a trend, has been widely developed, and has demonstrated its ability to solve complex biological problems. Machine learning can simulate complex nonlinear processes. The typical research method for drug combination is to first build features for cell lines and drugs, and then use machine learning or deep learning models to predict drug combinations. According to statistics, studies before 2018 used various machine learning models, such as random forests, XGBoost, etc. In addition, the publication of multiple large-scale data sets provides a reliable data foundation for deep learning. Techniques such as ensemble learning, transfer learning, and tensor decomposition have also been used in the field of drug combination prediction for drug combination. However, although the existing machine learning methods have improved the accuracy of prediction to a certain extent, there is still room for improvement.

[0004] Among them, a drug combination validated on one cell line may not be effective on other cell lines, so cell line characteristics are crucial in discovering drug combinations. Although multi-omics data are often used to construct cell line characteristics in machine learning and deep learning technologies, protein networks are rarely considered. The proposed propagation strategies based on drug targets and gene interaction networks, although they can simulate cell line characteristics after treatment, only consider the direct association between target genes and non-target genes, simplifying the use of gene-to-gene interaction information. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for capturing drug pharmacophore information and combining cell line characteristics with PPI networks to predict drug combination in view of the fact that the existing technology does not fully consider the problem of multi-source information integration when predicting synergistic drug use. It obtains features containing pharmacophore information through a heterogeneous graph model, and designs a strategy for fusing PPI network information and cell line-related gene expression information and fusing and updating cell line characteristics based on a graph neural network. Finally, the drug characteristics and cell line characteristics are embedded and integrated to complete the drug combination synergistic effect prediction system.

[0006] The technical solution of the present invention is as follows:

[0007] The method for predicting anticancer combination drugs based on multi-source data includes the following steps:

[0008] Drug molecule processing: Capturing the structural and chemical characteristics of drug molecules by constructing isomerized graphs of drug molecules;

[0009] Drug feature extraction: A view-based graph Transformer framework is used to aggregate drug molecule features from different angles using the graph attention mechanism. The features of nodes and edges are propagated and aggregated in heterogeneous graphs through the message passing mechanism.

[0010] Construction and acquisition of cell line gene features: Use gene expression data, gene mutation data, and protein-protein interaction (PPI) network information to construct comprehensive features of cell lines;

[0011] Prediction of combined drug effects: Integrate drug features with cell line features and use a multi-layer perceptron (MLP) for prediction.

[0012] Furthermore, the drug feature extraction includes the following steps:

[0013] The heterogeneous graph is represented as a feature matrix. A multi-view graph transformer framework based on message passing mechanism is used to pass messages among the atom view, pharmacophore view, and the connection view between atoms and pharmacophores, respectively, to obtain the corresponding graph feature matrix.

[0014] The information of edges and nodes in the molecular isomerization graph is learned by integrating the reaction information at the atomic level, the pharmacophore level, and the connection fragment level, and the characteristics of drug molecules are aggregated from multiple perspectives.

[0015] The message passing mechanism is used to update the features of nodes and edges, and the multi-head attention mechanism is used to enhance the accuracy and stability of feature extraction;

[0016] The feature vector is read through the GRU module to generate a global feature representation of the drug molecule.

[0017] Furthermore, the drug feature extraction specifically includes the following steps:

[0018] The calculation of multi-head attention includes:

[0019]

[0020] head i =Attention(Q i ,K i ,V i ),

[0021] Head i =Concat(head 1 ,head 2 ,…,head n )W o ,

[0022] Among them, Q, K, V represent the query matrix, key matrix, and value matrix respectively, σ is the activation function, P is the set of view types, p∈P represents the specific view type, Ω p The weight matrix K represents the learnable view types. [T represents the transposed matrix of the key matrix of view p, d k is the dimension of the query matrix and key matrix; W o is the weight matrix of each head;

[0023] Using the graph attention message passing structure based on the multi-head attention mechanism, features from different angles are learned from multiple views. The features of nodes and edges are updated and propagated in each iteration. The depth of the current message passing is represented by t. The operation formula of the message passing is:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, represents the hidden state of the edge, represents the hidden state of the node, and Represents the message information of nodes and edges respectively, Indicates that the search points to node v i The function of all edges of ;

[0030] Construct a simple residual block:

[0031]

[0032]

[0033] Use GRU to extract feature vectors Z from the three views of the molecule α ,Z γ ,Z β , where Z α The vector representing the atomic view, Z γ The vector representing the pharmacophore view, Z β The vector representing the connected view aggregates the features of these three views into the final molecular representation by utilizing the attention layer.

[0034] Furthermore, the construction and acquisition of the cell line gene characteristics specifically includes the following steps:

[0035] Using gene expression data, gene mutation data, and protein interaction network information, we construct a comprehensive characterization of the cell line; by combining nodes in the protein interaction network with gene expression data and mutation data through gene identifiers and symbols;

[0036] The graph attention network (GAT) model is used to construct the hidden state of genes based on the protein interaction network, in which proteins are nodes and the interactions between proteins are edges.

[0037] Randomly initialize the embedding matrix Where M represents the number of cell lines, and e cell represents the embedding dimension of the cell line state, the gene expression level of the cell line Exp and mutation result O Mut Represents the explicit state of the gene in the cell line, by mapping the hidden state matrix Z of the gene to the cell line space and combining it with the cell line state vector c j Performing a dot product operation, we can obtain the predicted clear state of the i-th gene in the cell line The calculation formula is:

[0038]

[0039] Among them, t∈Exp,Mut, Indicates the explicit status of a gene. j and Z represent the cell line state vector and the gene hidden state matrix, respectively.

[0040] Furthermore, the construction and acquisition of the cell line gene characteristics also includes the following steps:

[0041] Set the loss function so that the gene expression data and gene mutation data are completely matched with all nodes. The function is:

[0042]

[0043] Among them, N t represents the number of genes corresponding to the nodes in the PPI network;

[0044] By calculating the loss and back-propagating the gradient, the hidden state matrix Z of the gene and the state matrix C of the cell line are optimized at the same time. The embedding vectors obtained for the same cell line in different states are connected and normalized by Z-score as the final cell line embedding representation.

[0045] Furthermore, the prediction of the effect of combined medication specifically includes the following steps:

[0046] Connect the cell line feature vector with the cell line feature and embed it through a multi-layer perceptron;

[0047] Integrate the embedded cell line features with the drug features to finally form a drug pair feature vector;

[0048] The drug pair feature vector is classified and predicted by a multi-layer perceptron, and the probability of combined drug use is output; training is performed by minimizing the cross entropy loss function, and the specific loss function is:

[0049]

[0050] The performance was evaluated using a variety of evaluation metrics, including the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPR), accuracy (ACC), balanced accuracy (BACC), precision rate (PREC), true positive rate (TPR), and Cohen's Kappa value (KAPPA).

[0051] Furthermore, the drug molecule processing specifically includes the following steps:

[0052] The BRICS algorithm is used to decompose drug molecules into multiple fragments containing pharmacophores while retaining the chemical reaction information between the fragments;

[0053] The characteristic information of atoms and pharmacophores is combined to construct a heterogeneous graph, which is represented by G = {V, E}, where V is the node set and E is the edge set; the node type mapping function φ: V→O and the edge type mapping function ψ: E→P define the types of nodes and edges in the graph, respectively.

[0054] Furthermore, the heterogeneous graph includes an atomic view, a pharmacophore view including pharmacophore and reaction information, and a connection view that comprehensively represents the association between molecular atoms and pharmacophores.

[0055] Furthermore, the atomic view is represented as: α =(V α ,E α ), where for each atom Both exist Among them, 1≤i≤|N α |,|N α | represents the number of atoms, for each edge Where 1≤i,j≤|N α |;

[0056] The pharmacophore view is shown as: Figure G β =(V β ,E β ), for each pharmacophore fragment Exist Among them, 1≤i≤|N β |,|N β | represents the number of pharmacophore fragments, for each BRICS reaction type Where 1≤i,j≤|N β |;

[0057] The connection view can be represented as: γ =(Vγ,E γ ), for each node Exist 1≤i≤|Nγ|, |Nγ| is the total number of atoms and pharmacophores, for each edge Where 1≤i,j≤|Nγ|.

[0058] This application also includes an anti-cancer drug combination prediction system based on multi-source data, including:

[0059] The drug molecule processing module captures the structural and chemical characteristics of drug molecules by constructing heterogeneous graphs;

[0060] The drug feature acquisition module adopts a view-based graph Transformer framework and uses the graph attention mechanism to aggregate drug molecular features from different angles;

[0061] The cell line feature acquisition module uses gene expression data, gene mutation data, and protein interaction network information to construct comprehensive features of cell lines;

[0062] The drug combination prediction module, including feature integration and model prediction, embeds and integrates the features of two drugs with the features of cell lines to predict the effects of drug combination.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The method and system for predicting anticancer combination drugs based on multi-source data can capture the characteristics of drug molecules from a more comprehensive perspective through the comprehensive use of heterogeneous graphs and multi-view information, thus improving the prediction accuracy of the model;

[0065] 2. The method and system for predicting anticancer combination drugs based on multi-source data provide a more representative method for expressing cell line characteristics through comprehensive analysis of gene expression, gene mutation and PPI network information, thus improving the accuracy of prediction of combination drug effects;

[0066] 3. The anti-cancer combination drug prediction method and system based on multi-source data introduced a variety of optimization strategies in the model training and feature extraction process to ensure the stability and effectiveness of the model when processing complex drug and cell line data. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a model diagram of the anti-cancer combination drug prediction method based on multi-source data. DETAILED DESCRIPTION

[0068] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0069] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0070] See also Figure 1 , the anticancer combination drug prediction method based on multi-source data includes the following steps:

[0071] Drug molecule processing: Capturing the structural and chemical characteristics of drug molecules by constructing isomerized graphs of drug molecules;

[0072] Drug molecule processing specifically includes the following steps:

[0073] The BRICS algorithm is used to decompose drug molecules into multiple fragments containing pharmacophores while retaining the chemical reaction information between the fragments;

[0074] The characteristic information of atoms and pharmacophores is combined to construct a heterogeneous graph, which is represented by G = {V, E}, where V is a node set and E is an edge set. The heterogeneous graph includes an atomic view, a pharmacophore view containing pharmacophore and reaction information, and a connection view that comprehensively represents the relationship between molecular atoms and pharmacophores.

[0075] The atomic view is represented as: α =(V α ,E α ), where for each atom Both exist Among them, 1≤i≤|N α |,|N α | represents the number of atoms, for each edge Where 1≤i,j≤|N α |;

[0076] The pharmacophore view is shown as: Figure G β =(V β ,E β ), for each pharmacophore fragment Exist Among them, 1≤i≤|N β |,|N β | represents the number of pharmacophore fragments, for each BRICS reaction type Where 1≤i,j≤|N β |;

[0077] The connection view can be represented as: γ =(V γ ,E γ ), for each node Exist 1≤i≤|N γ |,|N γ | is the total number of atoms and pharmacophores, for each edge Where 1≤i,j≤|N γ |.

[0078] Drug feature extraction: A view-based graph Transformer framework is adopted, and the graph attention mechanism is used to aggregate drug molecular features from different angles. Through the message passing mechanism, the features of nodes and edges are propagated and aggregated in the heterogeneous graph, thereby capturing more structural information.

[0079] The heterogeneous molecular graph is represented as a feature matrix. A multi-view graph Transformer framework based on the message passing mechanism is used to pass messages between the atomic view, the pharmacophore view, and the connection view between atoms and pharmacophores to obtain their corresponding graph feature matrices. The information of edges and nodes in the molecular heterogeneous graph is learned by integrating the reaction information at the atomic level, the pharmacophore level, and the connection fragment level, and the features of drug molecules are aggregated from multiple perspectives. The message passing mechanism is used to update the features of nodes and edges, and the multi-head attention mechanism is used to enhance the accuracy and stability of feature extraction. Finally, the feature vector is read through the GRU module to generate a global feature representation of the drug molecule.

[0080] Drug feature extraction includes the following steps:

[0081] The heterogeneous graph is represented as a feature matrix. A multi-view graph transformer framework based on message passing mechanism is used to pass messages among the atom view, pharmacophore view, and the connection view between atoms and pharmacophores, respectively, to obtain the corresponding graph feature matrix.

[0082] The information of edges and nodes in the molecular isomerization graph is learned by integrating the reaction information at the atomic level, the pharmacophore level, and the connection fragment level, and the characteristics of drug molecules are aggregated from multiple perspectives.

[0083] The message passing mechanism is used to update the features of nodes and edges, and the multi-head attention mechanism is used to enhance the accuracy and stability of feature extraction;

[0084] The feature vector is read through the GRU module to generate a global feature representation of the drug molecule.

[0085] Drug feature extraction specifically includes the following steps:

[0086] The calculation of multi-head attention includes:

[0087]

[0088] head i =Attention(Q i ,K i ,V i ),

[0089] Head i =Concat(head 1 ,head 2 ,…,head n )W o ,

[0090] Among them, Q, K, V represent the query matrix, key matrix, and value matrix respectively, σ is the activation function, P is the set of view types, p∈P represents the specific view type, Ω p The weight matrix representing the learnable view types, represents the transposed matrix of the key matrix of view p, d k is the dimension of query matrix and key matrix; W o It is the weight matrix of each head. By splicing the outputs of each head, a richer feature expression is obtained;

[0091] In order to ensure that the magnitude of the inner product does not become too large when the input vector dimension increases, and to avoid numerical instability during calculation, the present invention introduces d k The normalization factor is obtained by dividing by This ensures that the output of attention calculation is smoother and more stable.

[0092] Using the graph attention message passing structure based on the multi-head attention mechanism, features from different angles are learned from multiple views. The features of nodes and edges are updated and propagated in each iteration. The depth of the current message passing is represented by t. The operation formula of the message passing is:

[0093]

[0094]

[0095]

[0096]

[0097]

[0098] in, represents the hidden state of the edge, represents the hidden state of the node, and Represents the message information of nodes and edges respectively, Indicates that the search points to node v i The function of all edges of ;

[0099] In order to avoid the gradient vanishing problem, a simple residual block is constructed to ensure that the training of the module is more stable during the multi-view message passing process:

[0100]

[0101]

[0102] Use GRU to extract feature vectors Z from the three views of the molecule α ,Zγ ,Z β , where Z α The vector representing the atomic view, Z γ The vector representing the pharmacophore view, Z β The vector representing the connected view is used to aggregate the features of these three views into the final molecular representation by using the attention layer. The attention aggregation function is defined as:

[0103]

[0104] The features of the pharmacophore and the connection view are aggregated to capture the reaction information between atoms and pharmacophores and between pharmacophores. The aggregated features are further fused with the feature aggregation of the atomic view to obtain the global feature representation of the molecule.

[0105] The multi-head attention mechanism is used to capture multiple levels of structural information in drug molecules. Each attention head focuses on different structural dimensions of the molecule, making the feature extraction process more comprehensive and accurate. To ensure the effective propagation of information, a residual block is introduced in the message passing process to enhance the stability of multi-view message passing. Finally, the GRU module is used to read out and aggregate features of the atomic view, pharmacophore view, and connection view to generate a global feature representation of the drug molecule.

[0106] Construction and acquisition of cell line gene features: Use gene expression data, gene mutation data, and protein interaction network information to construct comprehensive features of cell lines;

[0107] Gene expression data were obtained from the ArrayExpress database, summarized by the Robust MicroarraySummarization method, and normalized by z-score. Gene mutation data were obtained from the COSMIC cell line project, and after deleting invalid data, a 12333-dimensional binary gene mutation vector was formed.

[0108] In addition, PPI network data was obtained from the STRING database, and interaction information with a comprehensive score greater than 0.7 was retained. This information was combined with gene expression data and gene mutation data to construct feature embeddings for cell lines. The GAT model was used to capture node attributes and topological structure information in the PPI network and generate a hidden state matrix for genes.

[0109] In order to further represent the state of the cell line, the StateEncoder module is constructed. First, the embedding matrix is ​​randomly initialized to map the hidden state of the gene to the cell line space and interact with the cell line state vector to predict the displayed state of the gene. This displayed state combines the gene expression level and mutation results to form a characteristic representation of the cell line.

[0110] Through gene identifiers and symbols, the nodes in the PPI network are combined with gene expression data and mutation data to achieve unified data integration. By filtering key genes and using encoders to reduce the dimension of omics data, instead of directly using raw omics data, the number of model parameters can be reduced, the computational efficiency can be improved, and the performance of the model can be improved. The information of the PPI network is integrated into the cell line embedding. Considering the expression differences of the same gene in different cell lines, it is assumed that the displayed state of the gene is generated by the interaction between the hidden state of the gene independent of the cell line and the state of the cell line. Through this method, a comprehensive feature representation of the cell line is constructed.

[0111] The construction and acquisition of cell line gene characteristics specifically include the following steps:

[0112] Using gene expression data, gene mutation data, and protein interaction network information, we construct a comprehensive characterization of the cell line; by combining nodes in the protein interaction network with gene expression data and mutation data through gene identifiers and symbols;

[0113] A graph attention network model is used to construct the hidden state of genes based on the protein interaction network, in which proteins are nodes and the interactions between proteins are edges.

[0114] Randomly initialize the embedding matrix Where M represents the number of cell lines, and e cell represents the embedding dimension of the cell line state, the gene expression level of the cell line Exp and mutation result O Mut Represents the explicit state of the gene in the cell line, by mapping the hidden state matrix Z of the gene to the cell line space and combining it with the cell line state vector c j Performing a dot product operation, we can obtain the predicted clear state of the i-th gene in the cell line The calculation formula is:

[0115]

[0116] Among them, t∈Exp,Mut, Indicates the explicit status of a gene. j and Z represent the cell line state vector and the gene hidden state matrix, respectively.

[0117] Set the loss function so that the gene expression data and gene mutation data are completely matched with all nodes. The function is:

[0118]

[0119] Among them, N trepresents the number of genes corresponding to the nodes in the PPI network;

[0120] By calculating the loss and back-propagating the gradient, the hidden state matrix Z of the gene and the state matrix C of the cell line are optimized at the same time. The embedding vectors obtained for the same cell line in different states are connected and normalized by Z-score as the final cell line embedding representation.

[0121] Gene expression data, gene mutation data and PPI network information are combined and analyzed comprehensively through the GAT model to obtain a high-dimensional representation of cell line characteristics.

[0122] Prediction of combined drug effects: Integrate drug features with cell line features and use a multi-layer perceptron for prediction.

[0123] The prediction of the effect of combined medication specifically includes the following steps:

[0124] Connect the cell line feature vector with the cell line feature and embed it through a multi-layer perceptron;

[0125] Integrate the embedded cell line features with the drug features to finally form a drug pair feature vector;

[0126] The drug pair feature vector is classified and predicted by a multi-layer perceptron, and the probability of combined drug use is output; training is performed by minimizing the cross entropy loss function, and the specific loss function is:

[0127]

[0128] The performance was evaluated using a variety of evaluation metrics, including the area under the receiver operating characteristic curve, the area under the precision-recall curve, accuracy, balanced accuracy, precision, true positive rate, and Cohen's Kappa value.

[0129] The efficient prediction of the combined drug effect is achieved through feature integration and MLP model. The present invention particularly optimizes the MLP architecture and combines multiple evaluation indicators to ensure the generalization ability and practical application value of the model.

[0130] This application also includes an anti-cancer drug combination prediction system based on multi-source data, including:

[0131] The drug molecule processing module captures the structural and chemical characteristics of drug molecules by constructing heterogeneous graphs;

[0132] The drug feature acquisition module adopts a view-based graph Transformer framework and uses the graph attention mechanism to aggregate drug molecular features from different angles;

[0133] The cell line feature acquisition module uses gene expression data, gene mutation data, and protein interaction network information to construct comprehensive features of cell lines;

[0134] The drug combination prediction module, including feature integration and model prediction, embeds and integrates the features of two drugs with the features of cell lines to predict the effects of drug combination.

[0135] In another specific embodiment, the O'Neil drug combination data set is trained to predict drug combination. The experimental execution of drug combination using machine learning provided by the present invention is based on Python 3.11.5 and PyTorch 2.1.0. The experimental parameter settings are shown in Table 1.

[0136] Table 1

[0137] parameter set up d_dim 300 batch_size 2000 lr 0.0001 dropout 0.2

[0138] The provided experiment of predicting drug combination based on machine learning was trained on the O'Neil dataset. In this set of experiments, the evaluation indexes of the generated model performance and the prediction results are shown in Table 2.

[0139] Table 2

[0140] index result AUROC 0.94 AUPR 0.93 ACC 0.86 BACC 0.86 PREC 0.84 TPR 0.85 KAPPA 0.72

[0141] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A method for predicting anticancer drug combinations based on multi-source data, characterized in that: The following steps are involved: Drug molecule processing: Capturing the structural and chemical characteristics of drug molecules by constructing isomerized graphs of drug molecules; Drug feature extraction: A view-based graph Transformer framework is used to aggregate drug molecule features from different angles using the graph attention mechanism. The features of nodes and edges are propagated and aggregated in heterogeneous graphs through the message passing mechanism. Construction and acquisition of cell line gene features: Use gene expression data, gene mutation data, and protein interaction network information to construct comprehensive features of cell lines; Prediction of combined drug effects: Integrate drug features with cell line features and use a multi-layer perceptron for prediction.

2. The method for predicting anticancer drug combinations based on multi-source data according to claim 1, characterized in that: The drug feature extraction comprises the following steps: The heterogeneous graph is represented as a feature matrix. A multi-view graph transformer framework based on message passing is used to pass messages between the atom view, pharmacophore view, and the connection view between atoms and pharmacophores to obtain the corresponding graph feature matrix. The information of edges and nodes in the molecular isomerization graph is learned by integrating the reaction information at the atomic level, the pharmacophore level, and the connection fragment level, and the characteristics of drug molecules are aggregated from multiple perspectives. The message passing mechanism is used to update the features of nodes and edges, and the multi-head attention mechanism is used to enhance the accuracy and stability of feature extraction; The feature vector is read through the GRU module to generate a global feature representation of the drug molecule.

3. The method for predicting anticancer drug combination based on multi-source data according to claim 2, characterized in that: The drug feature extraction specifically comprises the following steps: The calculation of multi-head attention includes: head i =Attention(Q i ,K i ,V i ), Head i =Concat(head1,head2,…,head n )W o , Among them, Q, K, V represent the query matrix, key matrix, and value matrix respectively, σ is the activation function, P is the set of view types, p∈P represents the specific view type, Ω p The weight matrix representing the learnable view types, represents the transposed matrix of the key matrix of view p, d k is the dimension of query matrix and key matrix; W o is the weight matrix of each head; Using the graph attention message passing structure based on the multi-head attention mechanism, features from different angles are learned from multiple views. The features of nodes and edges are updated and propagated in each iteration. The depth of the current message passing is represented by t. The operation formula of the message passing is: in, represents the hidden state of the edge, represents the hidden state of the node, and Represents the message information of nodes and edges respectively, Indicates that the search points to node v i The function of all edges of ; Construct a simple residual block: Use GRU to extract the feature vector Z of each view from the three views of the molecule α ,Z γ ,Z β , where Z α The vector representing the atomic view, Z γ The vector representing the pharmacophore view, Z β Represents the directions of connected views, integrated to form a comprehensive representation of drug molecules.

4. The method for predicting anticancer drug combinations based on multi-source data according to claim 1, characterized in that: The construction and acquisition of the cell line gene characteristics specifically include the following steps: Use gene expression data, gene mutation data, and protein interaction network information to construct a comprehensive characterization of the cell line; combine nodes in the protein interaction network with gene expression data and mutation data through gene identifiers and symbols; A graph attention network model is used to construct the hidden state of genes based on the protein interaction network, in which proteins are nodes and the interactions between proteins are edges. Randomly initialize the embedding matrix Where M represents the number of cell lines, and e cell represents the embedding dimension of the cell line state, the gene expression level of the cell line Exp and mutation result O Mut Represents the explicit state of the gene in the cell line, by mapping the hidden state matrix Z of the gene to the cell line space and combining it with the cell line state vector c j Performing a dot product operation, we can obtain the predicted clear state of the i-th gene in the cell line The calculation formula is: Among them, t∈Exp,Mut, Indicates the explicit state of the gene, c j and Z represent the cell line state vector and the gene hidden state matrix, respectively.

5. The method for predicting anticancer drug combinations based on multi-source data according to claim 4, characterized in that: The cell line gene feature construction and acquisition further comprises the following steps: Set the loss function so that the gene expression data and gene mutation data are completely matched with all nodes. The function is: Among them, N t represents the number of genes corresponding to the nodes in the PPI network; By calculating the loss and back-propagating the gradient, the hidden state matrix Z of the gene and the state matrix C of the cell line are optimized at the same time. The embedding vectors obtained for the same cell line in different states are connected and normalized by Z-score as the final cell line embedding representation.

6. The method for predicting anticancer drug combination based on multi-source data according to claim 1, characterized in that: The prediction of the drug combination effect specifically comprises the following steps: Connect the cell line feature vector with the cell line feature and embed it through a multi-layer perceptron; Integrate the embedded cell line features with the drug features to finally form a drug pair feature vector; The drug pair feature vector is classified and predicted by a multi-layer perceptron, and the probability of combined drug use is output; training is performed by minimizing the cross entropy loss function, and the specific loss function is: The performance was evaluated using a variety of metrics, including the area under the receiver operating characteristic curve, the area under the precision-recall curve, accuracy, balanced accuracy, precision, true positive rate, and Cohen's Kappa value.

7. The method for predicting anticancer drug combinations based on multi-source data according to claim 1, characterized in that: The drug molecule processing specifically comprises the following steps: The BRICS algorithm is used to decompose drug molecules into multiple fragments containing pharmacophores while retaining the chemical reaction information between the fragments; The characteristic information of atoms and pharmacophores is combined to construct a heterogeneous graph, which is represented by G = {V, E}, where V is the node set and E is the edge set.

8. The method for predicting anticancer drug combinations based on multi-source data according to claim 7, characterized in that: The heterogeneous graph includes an atomic view, a pharmacophore view including pharmacophore and reaction information, and a connection view which comprehensively represents the association relationship between molecular atoms and pharmacophores.

9. The method for predicting anticancer drug combinations based on multi-source data according to claim 8, characterized in that: The atomic view is represented as: α =(V α ,E α ), where for each atom Both exist Among them, 1≤i≤|N α |,|N α | represents the number of atoms, for each edge Where 1≤i,j≤|N α |; The pharmacophore view is shown as: Figure G β =(V β ,E β ), for each pharmacophore fragment Exist Among them, 1≤i≤|N β |,|N β | represents the number of pharmacophore fragments, for each BRICS reaction type Where 1≤i,j≤|N β |; The connection view can be represented as: γ =(V γ ,E γ ), for each node Exist 1≤i≤|N γ |,|N γ | is the total number of atoms and pharmacophores, for each edge Where 1≤i,j≤|N γ |. 10.Anti-cancer drug combination prediction system based on multi-source data, characterized in that: include: The drug molecule processing module captures the structural and chemical characteristics of drug molecules by constructing heterogeneous graphs; The drug feature acquisition module adopts a view-based graph Transformer framework and uses the graph attention mechanism to aggregate drug molecular features from different angles; The cell line feature acquisition module uses gene expression data, gene mutation data, and protein interaction network information to construct comprehensive features of cell lines; The drug combination prediction module, including feature integration and model prediction, embeds and integrates the features of two drugs with the features of cell lines to predict the effects of drug combination.

Images