Cancer drug synergistic prediction method based on multi-element data enhanced hypergraph neural network
Patent Information
- Application Number
- CN202210818968.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-12
- Publication Date
- 2026-08-28
- Estimated Expiration
- 2042-07-12
AI Technical Summary
当当前并没有相关方法应用超图神经网络对药物协同作用进行建模的方案
[0051] Compared with existing technologies, the advantages of this invention are: it provides a unified deep learning-based framework for predicting the synergistic effects of anticancer drugs, and for the first time models drug synergy data from a hypergraph perspective, combining the biochemical information of drugs and cell lines with the multi-directional correlations of drug synergy data. This invention can not only help clinical medicine discover novel synergistic drug combinations for cancer treatment, but also contribute to the discovery of potential mechanisms of synergistic effects of anticancer drugs.
Smart Images
Figure CN115295163B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biomedical data mining, specifically to a method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks. Background Technology
[0002] Drug combinations typically refer to the use of two or more drugs for treatment. Compared to monotherapy, synergistic drug combinations can enhance efficacy and reduce toxicity, playing a crucial role in treating various complex diseases, particularly cancer. However, in practical applications, some drug combinations may produce antagonistic effects, exacerbating the condition, thus requiring accurate screening. Although high-throughput experimental screening techniques are now well-developed, traditional wet experiments are time-consuming and labor-intensive, and the drug combination problem can be modeled as a combinatorial optimization problem, with its search space growing exponentially with the number of drugs considered. Identifying synergistic drug combinations remains challenging.
[0003] The problem of drug synergy prediction can be formulated as a classification task, dividing drug combinations into two categories: those with synergistic effects and those without. In recent years, numerous machine learning-based and deep learning-based methods have been proposed to predict drug synergies. These methods predict synergies by learning the complex nonlinear relationships between the input features (chemical structure, omics features, etc.) of drugs and cell lines. It is noteworthy that complex multidirectional relationships also exist between drug combinations and cell lines. For example… Figure 1 As shown, the combination of oxaliplatin and MK-8776 only exhibits a synergistic effect on the OCUB-M cell line; while the combination of Zolinza and erlotinib shows a synergistic effect on multiple cell lines (e.g., MSTO-211H, SW-837, SK-OV-3); and the combination of Dactolisib with BEZ-235 and MK-8669 both showed a synergistic effect on the T47D cell line. The limitation of existing methods is that they only consider the biochemical characteristics of drugs and cell lines, without fully considering the correlation information inherent in the multidirectional relationship between drug combinations and cell lines. Furthermore, while drug synergy data serves as a priori correlation information, it belongs to two different domains than the biochemical characteristics of drugs and cell lines. Fusion of information from these two domains can greatly improve prediction accuracy; however, how to effectively fuse information from these two domains is a problem that urgently needs to be solved.
[0004] A hypergraph is a graph structure that differs from a traditional graph in that a single edge can connect multiple vertices simultaneously. It has been used to model multi-directional relationships between real-world data, such as those found in social networks. Drug synergy data can be represented as a hypergraph, with drugs and cell lines considered as nodes, and synergistic drug-drug-cell line triples considered as hyperedges. Hypergraph neural networks (HGNNs), a graph neural network method specifically designed for handling hypergraphs, have been proposed to address problems such as co-author hypergraphs and protein-ligand complex-based hypergraphs. Compared to traditional classification methods, HGNNs can propagate more complex higher-order relationships and offer superior computational performance and efficiency. Currently, there are no established methods for modeling drug synergy using HGNNs. Summary of the Invention
[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method for predicting the synergistic effect of anticancer drugs based on multivariate data augmented hypergraph neural networks. This method can not only help clinical medicine discover new synergistic drug combinations for cancer treatment, but also help discover the potential mechanisms of synergistic effects of anticancer drugs.
[0006] To achieve the above objectives, the technical solution adopted by the present invention includes:
[0007] The dataset containing all drug-drug-cell line triplets is divided into a first set and a second set, and then the training and testing phases are performed.
[0008] The training phase is used to learn the drug structure data of the molecular map based on the first set, as well as the gene expression data of the cell line, to perform embedding learning and complete the training of the synergistic effect predictor.
[0009] The testing phase is used to learn the embeddings of drug structure data from molecular graphs obtained from the second set, as well as gene expression data from cell lines, to obtain node embeddings, and to make predictions based on the obtained node embeddings and the trained synergistic effect predictor.
[0010] Based on the above technical solution, the training phase specifically includes:
[0011] The first set was used as the training set. Each drug-drug-cell line triplet in the training set was preprocessed to obtain drug structure data based on molecular graphs and gene expression data of cell lines.
[0012] GCN was used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, and FCN was used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features.
[0013] All obtained drug structure features and cell line gene features are used as the initial features of nodes in the hypergraph. All drug-drug-cell line triples in the training set are organized to construct hyperedges. HGCN is used for embedding learning to obtain node embeddings.
[0014] Based on the obtained node embeddings, all drug-drug-cell line triples in the training set are fed into the synergistic effect predictor for training to generate a prediction model.
[0015] Based on the above technical solution, the testing phase specifically includes:
[0016] The second set was used as the test set. Each drug-drug-cell line triplet in the test set was preprocessed to obtain drug structure data based on molecular diagrams and gene expression data of cell lines.
[0017] GCN was used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, and FCN was used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features.
[0018] All obtained drug structure features and cell line gene features were used as initial features for nodes in the hypergraph. HGCN was used for embedding learning to obtain node embeddings.
[0019] The obtained node embeddings are fed into the constructed synergy effect predictor for prediction.
[0020] Based on the above technical solutions,
[0021] Each sample in the first and second sets contains biochemical information;
[0022] The sample is a drug-drug-cell line triplet, denoted as D. a D b C;
[0023] The biochemical information includes SMILES data of the drug and gene expression data of the cell line.
[0024] Based on the above technical solution, the preprocessing step includes:
[0025] Extraction of drug-drug-cell line ternary group D aand D b Molecular graph-based drug structure data;
[0026] Gene expression data of cell line C in the drug-drug-cell line ternary set were extracted.
[0027] Based on the above technical solutions,
[0028] The method employs GCN to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, specifically including:
[0029] The obtained drug structure data based on molecular graphs is input into a 2-layer GCN for deep structural feature learning to obtain drug structure features.
[0030] The method employs GCN to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structural features, specifically including:
[0031] The obtained drug structure data based on molecular graphs is input into a 2-layer GCN for deep structural feature learning to obtain drug structure features.
[0032] In GCN, the first layer has a hidden layer dimension of 128, the second layer has a hidden layer dimension of 100, each layer of GCN uses ReLU as the activation function and is normalized by a Batch Normalization layer, and the first layer of GCN also includes a dropout layer, and the second layer of GCN also includes a global max pooling layer.
[0033] Based on the above technical solutions,
[0034] The method employs FCN to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features, specifically including:
[0035] The obtained gene expression data of the cell line is input into a 2-layer FCN for deep gene feature learning to obtain the gene features of the cell line.
[0036] The method of using FCN to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features specifically includes:
[0037] The obtained gene expression data of the cell line is input into a 2-layer FCN for deep gene feature learning to obtain the gene features of the cell line.
[0038] In FCN, the first layer has a hidden layer dimension of 128, the second layer has a hidden layer dimension of 100, and after the first layer, FCN also includes a Tanh activation function, a BatchNormalization layer, and a dropout layer. The second layer of FCN uses ReLU as the activation function.
[0039] Based on the above technical solution, the method involves using all obtained drug structural features and cell line gene features as initial features for nodes in the hypergraph, organizing all drug-drug-cell line triples in the training set to construct hyperedges, and using HGCN for embedding learning to obtain node embeddings. Specific steps include:
[0040] All obtained drug structural features and cell line gene features are used as the initial features of nodes in the hypergraph, and the feature dimension of each node is 100.
[0041] For all drug-drug-cell line triples in the training set, use a hyperedge to connect drug-drug-cell line triples with a cooperability score higher than 30.
[0042] The constructed hypergraph is input into a 3-layer HGCN for embedding learning to obtain node embeddings.
[0043] Based on the above technical solution, the specific network propagation rules of HGCN are as follows:
[0044] H (l) =σ(O -1 YWB -1 Y T H (l-1) θ (l-1) )
[0045]
[0046] Where H represents the node embedding matrix, σ represents the nonlinear activation function, O represents the node diagonal matrix, B represents the hyperedge degree diagonal matrix, Y represents the hyperedge correlation matrix, W represents the hyperedge weight matrix, θ represents the learnable weight matrix, the superscript l represents the layer number in the network iteration process, the superscript T represents the matrix transpose operation, the subscript i represents any node in the node set, v represents the number of nodes in the node set, the subscript e represents any hyperedge in the hyperedge set, and ε represents the number of hyperedges in the hyperedge set.
[0047] Based on the above technical solution, the specific steps of embedding the obtained nodes into the constructed synergy effect predictor for prediction include:
[0048] A drug-cooperative hypergraph encoder is used to learn the node embeddings of the drug-drug-cell line triplet to be predicted, and the vectors are concatenated to obtain the input vector of the cooperative effect predictor.
[0049] A synergistic effect predictor is constructed using a 3-layer MLP. The first layer of the MLP has a hidden layer dimension of 384, the second layer has a hidden layer dimension of 192, and the third layer has a hidden layer dimension of 1. The first and second layers of the MLP are followed by a Tanh activation function, a Batch Normalization layer, and a dropout layer. The third layer of the MLP is followed by a Sigmoid activation function.
[0050] The node embeddings learned by the drug co-synergistic hypergraph encoder are categorized into drugs and cell lines, and then fed into the constructed co-synergistic effect predictor for prediction.
[0051] Compared with existing technologies, the advantages of this invention are: it provides a unified deep learning-based framework for predicting the synergistic effects of anticancer drugs, and for the first time models drug synergy data from a hypergraph perspective, combining the biochemical information of drugs and cell lines with the multi-directional correlations of drug synergy data. This invention can not only help clinical medicine discover novel synergistic drug combinations for cancer treatment, but also contribute to the discovery of potential mechanisms of synergistic effects of anticancer drugs. Attached Figure Description
[0052] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0053] Figure 1 This is a schematic diagram illustrating the relationship between drug combinations and cell lines.
[0054] Figure 2 This is a flowchart of the training phase in an embodiment of the present invention;
[0055] Figure 3 This is a flowchart of the testing phase in an embodiment of the present invention;
[0056] Figure 4 The image shows the ablation analysis results for initial node features, auxiliary tasks for similarity network reconstruction, drug graph neural networks, and hypergraph neural networks. Detailed Implementation
[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.
[0058] This invention provides a method for synergistic prediction of anticancer drugs based on multivariate data-augmented hypergraph neural networks, to help clinical medicine discover novel synergistic drug combinations for cancer treatment, and improve the accuracy and efficiency of prediction. The method specifically includes:
[0059] The dataset containing all drug-drug-cell line triplets is divided into a first set and a second set. Then, a training phase and a testing phase are performed. The training phase is used to learn the embeddings and train the synergistic effect predictor based on the drug structure data of the molecular graph obtained from the first set and the gene expression data of the cell line. The testing phase is used to learn the embeddings to obtain the node embeddings based on the drug structure data of the molecular graph obtained from the second set and the gene expression data of the cell line. The prediction is then performed based on the obtained node embeddings and the trained synergistic effect predictor.
[0060] See Figure 2 As shown in the embodiment of the present invention, the training phase specifically includes:
[0061] S101: Using the first set as the training set, preprocess each drug-drug-cell line triplet in the training set to obtain drug structure data based on molecular graphs and gene expression data of cell lines.
[0062] S102: GCN (Graph Convolution Network) is used to perform deep structural feature learning on the obtained drug structure data based on molecular graphs to obtain drug structure features. FCN (Fully Connected Network) is used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features.
[0063] S103: Use all obtained drug structure features and cell line gene features as the initial features of nodes in the hypergraph, organize all drug-drug-cell line triples in the training set to construct hyperedges, and use HGCN (Hypergraph Convolution Network) to learn the embeddings to obtain node embeddings.
[0064] S104: Based on the obtained node embedding, all drug-drug-cell line triples in the training set are fed into the synergistic effect predictor for training to generate a prediction model.
[0065] See Figure 3 As shown in the embodiment of the present invention, the training phase specifically includes:
[0066] S201: Using the second set as the test set, preprocess each drug-drug-cell line triplet in the test set to obtain drug structure data based on molecular diagrams and gene expression data of cell lines.
[0067] S202: GCN is used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, and FCN is used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features.
[0068] S203: All obtained drug structural features and cell line gene features are used as the initial features of nodes in the hypergraph. HGCN is used for embedding learning to obtain node embeddings.
[0069] S204: The obtained node embeddings are fed into the constructed synergy effect predictor for prediction.
[0070] In this embodiment of the invention, each sample in the first set and the second set contains biochemical information; the sample is a drug-drug-cell line triplet, denoted as D. a D b C; Biochemical information includes SMILES (The simplified molecular-inputline-entry system) data for drugs and gene expression data for cell lines.
[0071] In this embodiment of the invention, preprocessing is performed, and the specific steps include:
[0072] A: Extraction of drug-drug-cell line ternary group D a and D b Drug structure data based on molecular graphs; the original drug biochemical information SMILES is a type of sequence data. Each drug SMILES can be transformed into a molecular graph G with atoms as nodes and chemical bonds as edges. d =(X d A d ), where X d A represents the characteristic matrix of all atoms. d Let X represent the adjacency matrix. d The dimensions are fixed at 75.
[0073] B: Extract gene expression data from cell line C in the drug-drug-cell line triplet. Logarithmic operations and Z-score normalization were performed on the original cell line gene expression data, resulting in 651-dimensional gene expression data.
[0074] In this embodiment of the invention, a GCN is used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features. Specifically, the obtained molecular graph-based drug structure data is input into a two-layer GCN for deep structural feature learning to obtain drug structure features.
[0075] The GCN is used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features. Specifically, the obtained molecular graph-based drug structure data is input into a 2-layer GCN for deep structural feature learning to obtain drug structure features.
[0076] In this GCN, the first layer has a hidden layer dimension of 128, and the second layer has a hidden layer dimension of 100. Each layer of the GCN uses ReLU (Rectified Linear Function) as the activation function and is normalized by a Batch Normalization layer. The first layer of the GCN also includes a dropout layer (ratio = 0.3), and the second layer includes a global max pooling layer. The dropout layer is a structure that can be used to reduce overfitting in neural networks.
[0077] In this embodiment of the invention, FCN is used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features. Specifically, it includes: inputting the obtained cell line gene expression data into a 2-layer FCN for deep gene feature learning to obtain cell line gene features.
[0078] The FCN was used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features. Specifically, the obtained cell line gene expression data was input into a 2-layer FCN for deep gene feature learning to obtain cell line gene features.
[0079] In FCN, the first layer has a hidden layer dimension of 128, the second layer has a hidden layer dimension of 100, and after the first layer, FCN also includes a Tanh activation function, a BatchNormalization layer, and a dropout layer. The second layer of FCN uses ReLU as the activation function.
[0080] In this embodiment of the invention, all obtained drug structural features and cell line gene features are used as initial features of nodes in the hypergraph. All drug-drug-cell line triples in the training set are organized to construct hyperedges. HGCN is used for embedding learning to obtain node embeddings. The specific steps include:
[0081] S1031: All obtained drug structural features and cell line gene features are used as the initial features of nodes in the hypergraph, and the feature dimension of each node is 100.
[0082] S1032: For all drug-drug-cell line triples in the training set, use a hyperedge to connect drug-drug-cell line triples with a co-operation score higher than 30.
[0083] S1033: Input the constructed hypergraph into the 3-layer HGCN for embedding learning to obtain node embeddings.
[0084] In this embodiment of the invention, the network propagation rules of HGCN are specifically as follows:
[0085] H (l) =σ(O -1 YWB -1 Y T H (l-1) θ (l-1) )
[0086]
[0087] Where H represents the node embedding matrix, σ represents the nonlinear activation function, O represents the node diagonal matrix, B represents the hyperedge degree diagonal matrix, Y represents the hyperedge correlation matrix, W represents the hyperedge weight matrix, θ represents the learnable weight matrix, the superscript l represents the layer number in the network iteration process, the superscript T represents the matrix transpose operation, the subscript i represents any node in the node set, v represents the number of nodes in the node set, the subscript e represents any hyperedge in the hyperedge set, and ε represents the number of hyperedges in the hyperedge set.
[0088] The hidden layer dimensions of HGCN's first, second, and third layers are all 256, and each layer uses ReLU as the activation function. After the first and second layers, there is also a BatchNormalization layer.
[0089] In this embodiment of the invention, the obtained node embeddings are fed into the constructed synergy effect predictor for prediction. Specific steps include:
[0090] S2041: The drug-cooperative hypergraph encoder is used to learn the node embedding of the drug-drug-cell line triplet to be predicted, and the vectors are concatenated to obtain the input vector of the cooperative effect predictor. The input vector has a dimension of 768.
[0091] S2042: A synergistic effect predictor is constructed using a 3-layer MLP. The first layer of the MLP (Multilayer Perceptron) has a hidden layer dimension of 384, the second layer has a hidden layer dimension of 192, and the third layer has a hidden layer dimension of 1. The first and second layers of the MLP are followed by a Tanh activation function, a BatchNormalization layer, and a dropout layer. The third layer of the MLP is followed by a Sigmoid activation function.
[0092] S2043: The node embeddings learned by the drug co-synergistic hypergraph encoder are categorized as drugs and cell lines, and then fed into the constructed synergistic effect predictor for prediction.
[0093] This involves a similarity network reconstruction task, categorizing the learned node embeddings into drugs and cell lines. The prior drug similarity network is calculated based on the cosine similarity of the drug's SMILES sequence features, while the prior cell line similarity network is calculated using Pearson correlation based on the cell line's gene expression features (P-value <= 0.01). The drug reconstruction similarity network uses the drug's hypergraph node embedding Z... D Its transpose Z D T Perform an inner product and approximate the result with a prior drug similarity network. The cell line reconstructed similarity network uses the cell line's hypergraph node embedding Z. C Its transpose Z C T Perform an inner product and make the result of the inner product approximate a prior cell line similarity network.
[0094] It should be noted that the drug synergy hypergraph encoder is composed of HGCN, whose function is to transform complex high-dimensional graph structure data into simple low-dimensional vectors, thereby learning and representing the multivariate, complex, and high-order relationships existing in the synergy prediction of anticancer drugs. The synergy effect predictor is composed of MLP (Multilayer Perceptron), whose function is to map high-dimensional embeddings to low dimensions, thereby realizing the transformation of embeddings into synergy effect scores.
[0095] The intrinsic structure of drug similarity and cell line similarity should be preserved in the embeddings learned from the hypergraph. To enhance the model's generalization ability, this invention additionally considers an auxiliary task based on reconstructing the drug and cell line similarity network from the learned embeddings.
[0096] See Figure 4 As shown, in the ablation experiments analyzing the various components of this invention (English name: Hypergraph Synergy), two performance evaluation metrics, AUC (Area Under Receiver Operating Characteristic Curve) and AUPR (Area Under Precision-Recall curve), were considered. Attempts were made to remove the biochemical feature encoder (without biochemical feature), the similarity network reconstruction auxiliary task (without similarity), the drug structure feature-based graph neural network module (without molecular structure feature), and the drug-cooperative hypergraph encoder (without multi-way relation). Experimental results show that removing any one component of this invention reduces system performance to varying degrees, especially removing the drug-cooperative hypergraph encoder (without multi-way relation), thus demonstrating that using the drug-cooperative hypergraph encoder can significantly improve prediction performance.
[0097] This invention provides a unified deep learning-based framework for predicting the synergistic effects of anticancer drugs. For the first time, it models drug synergy data from a hypergraph perspective, combining the biochemical information of drugs and cell lines with the multi-directional associations of drug synergy data. It mainly comprises three modules: a biochemical feature encoder, a drug synergy hypergraph encoder, and a synergistic effect predictor. In the biochemical feature encoder, structural features of drugs and gene expression features of cell lines are extracted from the raw data. In the drug synergy hypergraph encoder, the extracted biochemical features are used as the initial features of the hypergraph nodes, and a drug synergy hypergraph network is constructed to extract high-order, deep association embeddings. In the synergistic effect predictor, an auxiliary task based on similarity network reconstruction of drugs and cell lines is designed and jointly optimized with a supervised task to obtain the corresponding synergistic effect prediction results. This invention can not only help clinical medicine discover novel synergistic drug combinations for cancer treatment but also contribute to the discovery of potential mechanisms of synergistic effects of anticancer drugs.
[0098] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.
[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
Claims
1. A method for synergistic prediction of anticancer drugs based on multivariate data-augmented hypergraph neural networks, characterized in that, include: The dataset containing all drug-drug-cell line triplets is divided into a first set and a second set, and then the training and testing phases are performed. The training phase is used to learn the drug structure data of the molecular map based on the first set, as well as the gene expression data of the cell line, to perform embedding learning and complete the training of the synergistic effect predictor. The testing phase is used to learn the embeddings of drug structure data from molecular graphs obtained from the second set, as well as gene expression data from cell lines, to obtain node embeddings, and to make predictions based on the obtained node embeddings and the trained synergistic effect predictor. The training phase specifically includes: The first set was used as the training set. Each drug-drug-cell line triplet in the training set was preprocessed to obtain drug structure data based on molecular graphs and gene expression data of cell lines. GCN was used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, and FCN was used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features. All obtained drug structure features and cell line gene features are used as the initial features of nodes in the hypergraph. All drug-drug-cell line triples in the training set are organized to construct hyperedges. HGCN is used for embedding learning to obtain node embeddings. Based on the obtained node embeddings, all drug-drug-cell line triples in the training set are fed into the synergistic effect predictor for training to generate a prediction model.
2. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 1, characterized in that, The testing phase specifically includes: The second set was used as the test set. Each drug-drug-cell line triplet in the test set was preprocessed to obtain drug structure data based on molecular diagrams and gene expression data of cell lines. GCN was used to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, and FCN was used to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features. All obtained drug structure features and cell line gene features were used as initial features for nodes in the hypergraph. HGCN was used for embedding learning to obtain node embeddings. The obtained node embeddings are fed into the constructed synergy effect predictor for prediction.
3. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 2, characterized in that: Each sample in the first and second sets contains biochemical information; The sample is a drug-drug-cell line triplet, denoted as [missing information]. , , ; The biochemical information includes SMILES data of the drug and gene expression data of the cell line.
4. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 3, characterized in that, The preprocessing process includes the following steps: Extraction of drug-drug-cell line ternary group and Molecular graph-based drug structure data; Extraction of drug-drug-cell line ternary group Gene expression data of cell lines.
5. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 3, characterized in that, The method employs GCN to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structure features, specifically including: The obtained drug structure data based on molecular graphs is input into a 2-layer GCN for deep structural feature learning to obtain drug structure features. The method employs GCN to perform deep structural feature learning on the obtained molecular graph-based drug structure data to obtain drug structural features, specifically including: The obtained drug structure data based on molecular graphs is input into a 2-layer GCN for deep structural feature learning to obtain drug structure features. In GCN, the first layer has a hidden layer dimension of 128, the second layer has a hidden layer dimension of 100, each layer of GCN uses ReLU as the activation function and is normalized by a Batch Normalization layer, and the first layer of GCN also includes a dropout layer, and the second layer of GCN also includes a global max pooling layer.
6. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 3, characterized in that, The method employs FCN to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features, specifically including: The obtained gene expression data of the cell line is input into a 2-layer FCN for deep gene feature learning to obtain the gene features of the cell line. The method of using FCN to perform deep gene feature learning on the obtained cell line gene expression data to obtain cell line gene features specifically includes: The obtained gene expression data of the cell line is input into a 2-layer FCN for deep gene feature learning to obtain the gene features of the cell line. In FCN, the first layer of the hidden layer has a dimension of 128, and the second layer of the hidden layer has a dimension of 100. After the first layer of FCN, there are also Tanh activation function, Batch Normalization layer and drop-out layer. The second layer of FCN uses ReLU as the activation function.
7. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 3, characterized in that, The process involves using all obtained drug structural features and cell line gene features as initial features for nodes in the hypergraph, organizing all drug-drug-cell line triples in the training set to construct hyperedges, and using HGCN for embedding learning to obtain node embeddings. Specific steps include: All obtained drug structural features and cell line gene features are used as the initial features of nodes in the hypergraph, and the feature dimension of each node is 100. For all drug-drug-cell line triples in the training set, use a superedge to connect drug-drug-cell line triples with a cooperability score higher than 30. The constructed hypergraph is input into a 3-layer HGCN for embedding learning to obtain node embeddings.
8. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 7, characterized in that, The specific network propagation rules of HGCN are as follows: in, The embedding matrix represents the nodes. Represents a non-linear activation function. A diagonal matrix representing the nodes. Represents a diagonal matrix of hypermarginality. The incidence matrix of the hyperedge is represented. The weight matrix represents the weight of the hyperedge. Represents the learnable weight matrix, with superscript... Indicates the number of layers in the network iteration process, superscript This indicates a transpose operation on a matrix, with the subscript... Represents any node in the set of nodes. The index represents the number of nodes in the node set. Denotes any superedge in the set of superedges. This represents the number of superedges in the set of superedges.
9. The method for synergistic prediction of anticancer drugs based on multivariate data augmentation hypergraph neural networks as described in claim 3, characterized in that, The specific steps of embedding the obtained nodes into the constructed synergistic effect predictor for prediction include: A drug-cooperative hypergraph encoder is used to learn the node embeddings of the drug-drug-cell line triplet to be predicted, and the vectors are concatenated to obtain the input vector of the cooperative effect predictor. A synergistic effect predictor is constructed using a 3-layer MLP. The first layer of the MLP has a hidden layer dimension of 384, the second layer has a hidden layer dimension of 192, and the third layer has a hidden layer dimension of 1. The first and second layers of the MLP are followed by a Tanh activation function, a Batch Normalization layer, and a dropout layer. The third layer of the MLP is followed by a Sigmoid activation function. The node embeddings learned by the drug co-synergistic hypergraph encoder are categorized into drugs and cell lines, and then fed into the constructed co-synergistic effect predictor for prediction.