Cancer decision-making aid method based on knowledge-enhanced hypergraph contrastive learning

Abstract

The present invention provides a cancer decision-making assistance method based on knowledge-enhanced hypergraph comparative learning, which relates to the field of intelligent medical technology, and includes: extracting features from original drug and cell line data, constructing a training hypergraph with drugs and cell lines as vertices and synergistic relationships as hyperedges; using a hypergraph neural network for learning to obtain an embedded representation of the combined drug and cell line; and enhancing the hypergraph to perform comparative learning on the model; extracting features from the test data set based on the learned model and updating the hypergraph to obtain an embedded representation of the test vertex; and obtaining the synergistic relationship of each drug in the test data set based on the embedded representation of the test vertex. The present invention takes into account the high-order dependency between drugs and cell lines, obtains a good representation that is better used to predict synergistic relationships, and provides a good foundation for clinical decision-making assistance.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical technology, and in particular to a cancer decision-making assistance method based on knowledge-enhanced hypergraph contrast learning. Background Art

[0002] Drug synergy is crucial to improving the effectiveness of cancer treatment. The complex mechanisms of cancer pose a major challenge to effective treatment. Even if drug treatment is effective in the early stages, drug resistance may eventually develop. Studies have shown that traditional single-compound treatments that target a single molecule are not always effective. Therefore, predicting drug synergy is of great significance for assisting clinical decision-making.

[0003] Predicting drug synergy is a complex and arduous task. This approach is impractical, considering that selecting effective drug combinations through clinical trials and high-throughput screening is time-consuming and labor-intensive. Computational methods have emerged as a viable alternative, mainly including mathematical methods and machine learning methods. Although mathematical methods are hampered by their limited ability to handle large data sets and their reliance on prior knowledge, traditional machine learning methods, such as random forests, support vector machines, and logistic regression, can achieve relatively good results. However, these methods still lag behind cutting-edge phenotype-based deep learning methods. These methods predict drug synergy by training models in cell lines, usually representing samples as a quadruple of two drugs, a cell line, and a synergy label. However, these methods usually simply connect these four pieces of information together, which may ignore the complex interrelationships between them. In recent years, methods based on graph models have shown their effectiveness in capturing high-order dependencies between drug combinations and cell lines. Hypergraphs provide a naturally suitable framework for modeling drug synergy data.

[0004] However, existing hypergraph-based drug synergy relationship prediction methods face the problem of sparse labels, which leads to sparse hyperedges of the hypergraph used in training, limiting the model effect. Summary of the invention

[0005] The purpose of the present invention is to provide a cancer decision-making support method based on knowledge-enhanced hypergraph contrast learning, which takes into account the high-order dependencies between drugs and cell lines, obtains good representations for better prediction of synergistic relationships, and provides a good basis for clinical decision-making support.

[0006] A cancer decision-making assistance method based on knowledge-enhanced hypergraph contrast learning, comprising:

[0007] S1, obtaining an original drug synergy relationship data set including a plurality of sample data, wherein the sample data is a four-tuple data of drug-drug-cell line-actual synergy relationship;

[0008] S2, constructing a training hypergraph based on each of the sample data, and performing an enhancement operation on the training hypergraph to obtain a first enhanced view and a second enhanced view, and performing comparative training on a graph convolutional network, a fully connected network, a hypergraph neural network, and a prediction network based on the training hypergraph, the first enhanced view, and the second enhanced view to obtain the trained graph convolutional network, the fully connected network, the hypergraph neural network, and the prediction network;

[0009] S3, based on the trained graph convolutional network, feature extraction is performed on the drugs in the test data set to obtain several molecular representations of the drugs to be tested; based on the trained fully connected network, feature extraction is performed on several cell lines in the test data set to obtain several molecular representations of the cell lines to be tested; based on the trained graph convolutional network, feature extraction is performed on the drugs in each of the sample data to obtain several molecular representations of sample drugs; based on the trained fully connected network, feature extraction is performed on the cell lines in each of the sample data to obtain several molecular representations of sample cell lines;

[0010] S4, constructing a sample hypergraph with each of the sample drug molecular representations and each of the sample cell line molecular representations as sample vertices, and updating the sample hypergraph with each of the test drug molecular representations and each of the test cell line molecular representations as test vertices to obtain a test hypergraph;

[0011] S5, inputting the hypergraph to be tested into the trained hypergraph neural network to obtain an embedded representation of each vertex to be tested;

[0012] S6, inputting the embedded representation of each of the vertices to be tested into the trained prediction network to obtain the synergistic relationship of each drug in the test data set.

[0013] Optionally, the S2 includes:

[0014] S21, using a graph convolutional network to extract features of drugs in the sample data to obtain a number of training drug molecular representations; using a fully connected network to extract features of cell lines in each of the sample data to obtain a number of training cell line molecular representations;

[0015] S22, taking each of the training drug molecular representations and each of the training cell line molecular representations as training vertices, adding edges between the training drug molecular representations corresponding to the drugs with antagonistic effects to obtain a drug relationship matrix, adding edges between the training cell line molecular representations corresponding to the cell lines whose similarities are greater than a similarity set value to obtain a cell line similarity matrix, adding hyperedges between the training drug molecular representations and the training cell line molecular representations corresponding to the synergistic relationships greater than a relationship threshold to obtain a hyperedge matrix and the training hypergraph, and performing an enhancement operation on the training hypergraph to obtain the first enhanced view and the second enhanced view;

[0016] S23, obtaining, based on the hypergraph neural network, an embedded representation of each training vertex in the training hypergraph, an embedded representation of each training vertex in the first enhanced view, and an embedded representation of each training vertex in the second enhanced view;

[0017] S24, taking the embedding representation of each training vertex in the training hypergraph corresponding to each sample data as a set to obtain a plurality of training embedding sets, inputting each training embedding set into the prediction network to obtain a plurality of training prediction synergy relationships; converting the embedding representation of each training vertex in the first enhanced view and the embedding representation of each training vertex in the second enhanced view based on the projection head to obtain first comparative learning reference data and second comparative learning reference data;

[0018] S25, obtaining a target loss function value based on the training prediction synergy relationship, the first comparative learning reference data, and the second comparative learning reference data, and optimizing and updating the graph convolution network, the fully connected network, the hypergraph neural network, the prediction network, and the projection head based on the target loss function value;

[0019] S26, repeat S21-S25 to obtain the trained graph convolution network, the trained fully connected network, the trained hypergraph neural network and the trained prediction network. Optionally, the molecular graph of the drug is represented as ,in Representative drug molecules Atomic dimensional feature vector set, Represents the connection matrix between atoms, indicating the atomic bonds in the drug molecule, represents a set of real numbers; the cell line uses 651 gene expressions; optionally, the graph convolutional network includes K update layers and a maximum pooling layer; the update layer expression is: ;

[0020] Where: Indicates Drugs, Indicates The first atoms, is the activation function of the update layer, Indicates The first Atom in Layer update layer update features, Indicates The first Atom in Layer update layer update features, For the The drug in The parameter matrix of the layer update layer, , For the The degree matrix of drugs, the maximum pooling layer expression is: ;

[0021] Where: Indicates Drug molecular characterization of each drug, Indicates Drug No. Atom in Layer Update: Update features of the layer; ;

[0022] Where:

[0023] The molecular characterization expression of the cell line is: Indicates Cell line molecular characterization of 10 cell lines, is the activation function of the fully connected network, Indicates the expression of 651 genes, represents the parameter matrix of the 0th layer of the fully connected network, Represents the fully connected network The parameter matrix of the layer, is the number of layers in the fully connected network.

[0024] Optionally, the hypergraph neural network includes a plurality of hyperedge convolutional layers;

[0025] When the vertex in the hypergraph to be tested is a sample vertex, the hyperedge convolution layer expression is: ;

[0026] When the vertex in the hypergraph to be tested is the vertex to be tested, the expression of the hyperedge convolution layer is: ;

[0027] Where: Indicates Layer hyperconvolutional layer, Indicates that the vertex of the hypergraph to be tested is in Molecular representation of hyperconvolutional layers, The vertex of the hypergraph is Molecular representation of hyperconvolutional layers, is the hyperedge matrix, represents the hyperedge degree matrix, represents the vertex degree matrix, represents the hyperedge weight matrix, Indicates The feature extraction matrix of the hyperconvolutional layer, is transposed.

[0028] Optionally, the training embedding set includes two training drug embedding representations and one training cell line embedding representation, the two training drug embedding representations correspond to two drugs in the sample data, and the two training drug embeddings are respectively defined as a first training drug embedding representation and a second training drug embedding representation;

[0029] The prediction network expression is: ;

[0030] Where: represents the training prediction synergy relationship, represents the first trained drug embedding representation, represents the second training drug embedding representation, represents the trained cell line embedding representation, represents the weight matrix of the prediction network, represents the activation function of the prediction network, represents the first bias of the prediction network, represents the second bias of the prediction network. Optionally, the objective loss function value expression is: ;

[0031] Where: is the target loss function value, is the weight parameter, is the drug contrast learning loss value, is the learning loss value for cell line comparison, To predict the network loss value.

[0032] Optionally, the predicted network loss value expression is: ;

[0033] Where: is the number of training embedding sets, Indicates The training prediction synergy relationship corresponding to the training embedding set is Indicates The actual synergy relationship corresponding to the training embedding set.

[0034] Optionally, the first comparative learning reference data includes first drug comparative learning reference data and first cell line comparative learning reference data, and the second comparative learning reference data includes second drug comparative learning reference data and second cell line comparative learning reference data;

[0035] The drug contrast learning loss value expression is: ; ; In the formula, represents the cell line contrast loss function, is a cell line set, is the total number of cell lines contained in each of the sample data, Indicates the i-th value in the first cell line comparison learning reference data, which is the value of the first comparison corresponding to the i-th cell line. Represents the i-th value in the second cell line comparison learning reference data, which is the value of the second comparison corresponding to the i-th cell line. The cell line similarity matrix Line The value of the column represents Cell lines and Whether the similarity of the cell lines is greater than the similarity set value. The effects of the present invention are as follows:

[0036] The present invention is a cancer decision-making support method based on knowledge-enhanced hypergraph contrast learning. Taking into account the high-order dependencies between drugs and cell lines, hypergraphs are used for modeling to obtain good representations that are better used to predict synergistic relationships.

[0037] The present invention is a cancer decision-making assistance method based on knowledge-enhanced hypergraph contrastive learning. Contrastive learning is introduced on the basis of a hypergraph. The hypergraph contrastive learning can be guided by knowledge to alleviate the data sparsity problem, which is conducive to obtaining a good embedded representation of drug combinations and cell lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the cancer decision-making assistance method based on knowledge-enhanced hypergraph comparative learning of the present invention;

[0039] Figure 2 This is a learning rate super-parameter experiment effect diagram of the present invention;

[0040] Figure 3 This is a diagram showing the effect of a proportional hyperparameter experiment of the enhanced operation of the present invention;

[0041] Figure 4 This is a diagram showing the effect of the super-parameter experiment of the weight parameter a of the present invention. DETAILED DESCRIPTION

[0042] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0043] Figure 1 Flowchart of the cancer decision-making assistance method based on knowledge-enhanced hypergraph comparative learning of the present invention. Figure 1 As shown, the present invention provides a cancer decision-making assistance method based on knowledge-enhanced hypergraph contrast learning, which includes:

[0044] S1, obtain the original drug synergy relationship dataset including several sample data, the sample data is the four-tuple data of drug-drug-cell line-actual synergy relationship. The original drug synergy relationship dataset is obtained from the Oncology-Screen dataset and the NCI-ALMANAC dataset. Specifically, the molecular graph of the drug is represented as ,in Representative drug molecules Atomic dimensional feature vector set, Represents the connection matrix between atoms, indicating the atomic bonds in the drug molecule, represents a set of real numbers; the cell lines used 651 gene expressions.

[0045] S2, construct a training hypergraph based on each sample data, and enhance the training hypergraph to obtain a first enhanced view and a second enhanced view, and perform comparative training on the graph convolutional network, the fully connected network, the hypergraph neural network and the prediction network based on the training hypergraph, the first enhanced view and the second enhanced view to obtain a trained graph convolutional network, a fully connected network, a hypergraph neural network and a prediction network.

[0046] Specifically, S2 includes:

[0047] S21, use a graph convolutional network to extract features of drugs in the sample data to obtain several training drug molecular representations; use a fully connected network to extract features of cell lines in each sample data to obtain several training cell line molecular representations.

[0048] Preferably, the graph convolutional network includes K update layers and maximum pooling layers.

[0049] The update layer expression is: Where: Indicates Drugs, Indicates The first atoms, is the activation function of the update layer, Indicates The first Atom in Layer update layer update features, Indicates The first Atom in Layer update layer update features, For the The drug in The parameter matrix of the layer update layer, , For the The degree matrix of drugs, For the The atomic adjacency matrix of each drug, The molecular diagram of the drug , For the The identity matrix of each drug.

[0050] The maximum pooling layer expression is: .

[0051] Where: Indicates Drug molecular characterization of each drug, Indicates Drug No. The updated features of the atoms in the Kth update layer.

[0052] The molecular characterization expression of the cell line is: .

[0053] Where: represents the cell line molecular characterization of the a-th cell line, is the activation function of the fully connected network, Indicates the expression of 651 genes, represents the parameter matrix of the 0th layer of the fully connected network, Represents the parameter matrix of the uth layer of the fully connected network, where u is the number of layers of the fully connected network. S22, taking each training drug molecular representation and each training cell line molecular representation as training vertices, adding edges between the training drug molecular representations corresponding to drugs with antagonistic effects to obtain a drug relationship matrix, adding edges between the training cell line molecular representations corresponding to cell lines whose similarity is greater than a similarity set value to obtain a cell line similarity matrix, adding hyperedges between the training drug molecular representations and the training cell line molecular representations corresponding to synergistic relationships greater than a relationship threshold to obtain a hyperedge matrix and a training hypergraph, and performing an enhancement operation on the training hypergraph to obtain a first enhanced view and a second enhanced view.

[0054] S23, based on the hypergraph neural network, obtain the embedded representation of each training vertex in the training hypergraph, the embedded representation of each training vertex in the first enhanced view, and the embedded representation of each training vertex in the second enhanced view.

[0055] S24, taking the embedding representation of each training vertex in the training hypergraph corresponding to each sample data as a set to obtain several training embedding sets, inputting each training embedding set into the prediction network to obtain several training prediction synergy relationships; based on the projection head, converting the embedding representation of each training vertex in the first enhanced view and the embedding representation of each training vertex in the second enhanced view to obtain first comparative learning reference data and second comparative learning reference data.

[0056] During the training phase, each vertex in the training hypergraph is input into the hypergraph neural network to obtain the embedded representation of each training vertex.

[0057] Specifically, the embedding representations corresponding to the two drugs and one cell line in each sample data are taken as a set. The training embedding set includes two training drug embedding representations and one training cell line embedding representation, the two training drug embedding representations correspond to the two drugs in the sample data, and the two training drug embeddings are defined as the first training drug embedding representation and the second training drug embedding representation, respectively.

[0058] The prediction network expression is: Where: represents the training prediction synergy relationship, represents the first trained drug embedding representation, represents the second training drug embedding representation, represents the trained cell line embedding representation, represents the weight matrix of the prediction network, represents the activation function of the prediction network, represents the first bias of the prediction network, Represents the second bias of the prediction network.

[0059] S25, obtaining a target loss function value based on the training prediction synergy relationship, the first comparative learning reference data and the second comparative learning reference data, and optimizing and updating the graph convolution network, the fully connected network, the hypergraph neural network, the prediction network and the projection head based on the target loss function value.

[0060] Preferably, the objective loss function value expression is: ;

[0061] Where: is the target loss function value, is the weight parameter, is the drug contrast learning loss value, is the learning loss value for cell line comparison, To predict the network loss value.

[0062] The expression for predicting network loss value is: ;

[0063] Where: is the number of training embedding sets, Indicates The training prediction synergy relationship corresponding to the training embedding set is Indicates The actual synergy relationship corresponding to the training embedding set.

[0064] The first comparative learning reference data includes first drug comparative learning reference data and first cell line comparative learning reference data, and the second comparative learning reference data includes second drug comparative learning reference data and second cell line comparative learning reference data;

[0065] The drug contrast learning loss value expression is: ; ; Where: represents the drug contrast loss function, D is the drug set, and is the total number of drugs contained in each sample data. is the cosine similarity function, is the ith value in the first drug comparison learning reference data, is the value of the first comparison corresponding to the ith drug, is the i-th value in the second drug comparison learning reference data, is the value of the second comparison corresponding to the i-th drug, is the drug relationship matrix Line The value of the column represents Drugs and Does the drug have an antagonistic effect? Represents the temperature coefficient, which is a hyperparameter.

[0066] The expression of cell line contrast learning loss value is: ; ; In the formula, represents the cell line comparison loss function, C is the cell line set, and is the total number of cell lines contained in each sample data. Indicates the i-th value in the first cell line comparison learning reference data, which is the value of the first comparison corresponding to the i-th cell line. Represents the i-th value in the second cell line comparison learning reference data, which is the value of the second comparison corresponding to the i-th cell line. The cell line similarity matrix Line The value of the column represents cell lines and Whether the similarity of the cell lines is greater than the similarity set value.

[0067] S26, repeat S21-S25 to obtain a trained graph convolutional network, a trained fully connected network, a trained hypergraph neural network and a trained prediction network.

[0068] S3, based on the trained graph convolutional network, feature extraction is performed on the drugs in the test data set to obtain several molecular representations of the drugs to be tested; based on the trained fully connected network, feature extraction is performed on several cell lines in the test data set to obtain several molecular representations of the cell lines to be tested; based on the trained graph convolutional network, feature extraction is performed on the drugs in each sample data to obtain several sample drug molecular representations; based on the trained fully connected network, feature extraction is performed on the cell lines in each sample data to obtain several sample cell line molecular representations.

[0069] S4, constructing a sample hypergraph with each sample drug molecular representation and each sample cell line molecular representation as sample vertices, and updating the sample hypergraph with each test drug molecular representation and each test cell line molecular representation as test vertices to obtain a test hypergraph.

[0070] Specifically, taking each sample drug molecular representation and each sample cell line molecular representation as sample vertices, edges are added between the sample drug molecular representations corresponding to drugs with antagonistic effects to obtain a drug relationship matrix, edges are added between the sample cell line molecular representations corresponding to cell lines whose similarity is greater than a similarity set value to obtain a cell line similarity matrix, hyperedges are added between the sample drug molecular representations and the sample cell line molecular representations corresponding to synergistic relationships greater than a relationship threshold to obtain a hyperedge matrix and a sample hypergraph, and taking each drug molecular representation to be tested and each cell line molecular representation to be tested as vertices to be tested to update the sample hypergraph to obtain a hypergraph to be tested.

[0071] S5, input the hypergraph to be tested into the trained hypergraph neural network to obtain the embedded representation of each vertex to be tested.

[0072] Specifically, the hypergraph neural network includes several hyperedge convolutional layers.

[0073] When the vertex in the hypergraph to be tested is a sample vertex, the expression of the hyperedge convolution layer is: .

[0074] When the vertex in the hypergraph to be tested is the vertex to be tested, the expression of the hyperedge convolution layer is: .

[0075] Where: Indicates Layer hyperconvolutional layer, Indicates that the vertex of the hypergraph to be tested is in Molecular representation of hyperconvolutional layers, The vertex of the hypergraph is Molecular representation of hyperconvolutional layers, is the hyperedge matrix, represents the hyperedge degree matrix, represents the vertex degree matrix, represents the hyperedge weight matrix, Indicates The feature extraction matrix of the hyperconvolutional layer, is transposed.

[0076] S6, inputting the embedded representation of each vertex to be tested into the trained prediction network to obtain the synergistic relationship of each drug in the test data set.

[0077] Specifically, the method of the present invention (KnowSynergy) is compared with existing methods. Existing methods include: DeepSynergy, which uses ECFP_6, physicochemical properties and toxicity characteristics to characterize cell lines, and uses gene expression profiles for description, and predicts synergy scores through a three-layer feedforward neural network; Li's method, which integrates biological knowledge through a network propagation method to generate useful features, and then uses a random forest classifier for prediction; Celebi's method, which trains an XGBoost model using molecular information and related pharmacological data; DTF, which uses a tensor decomposition method and combines a deep neural network as a binary classifier; ComboFM, which models the high-order interactions of drug pairs with cell lines and the dose-response matrix as a fifth-order tensor, and uses a factorization machine for prediction; HypergraphSynergy, which applies a hypergraph neural network (HGNN) on the constructed hypergraph for node embedding learning, and then uses a multi-layer perceptron (MLP) for prediction; HyperSynergy, which proposes a prior-guided hypernetwork architecture suitable for new small sample predictions.

[0078] In the experimental example, a five-fold cross-validation was used. First, the data set was divided into a cross-validation set (90%) and an independent test set (10%). The cross-validation set also needed to be divided into five folds. Three scenarios were used to verify the effectiveness of the proposed method: Scenario 1, randomly dividing the data set; Scenario 2, dividing according to cell lines to ensure that cell lines that appeared in the training set would not appear in the validation set; Scenario 3, dividing according to drug pairs to ensure that drug pairs that appeared in the training set would not appear in the validation set. The evaluation indicators used Root Mean Square Error (RMSE), which represents the standard deviation between the predicted value and the true value, the smaller the better; Determination Coefficient R 2, which is used to evaluate the degree of explanation of the regression model on the variation of the dependent variable (target variable), reflects the goodness of fit of the model, and the larger the better; Pearson correlation coefficient PCC is a statistical indicator to measure the degree of linear correlation between two variables, which is used to describe the strength and direction of the linear relationship between the two variables. The larger the better. The results are shown in Table 1.

[0079] Table 1 Cross-validation regression experiment results

[0080]

[0081] Table 1 shows the performance of each method in the two datasets. It can be seen from Table 1 that the proposed method outperforms the existing methods on all datasets and ranks first in the average ranking. In particular, on the NCI-ALMANAC dataset, RMSE, R² and PCC are improved by an average of 27.61%, 67.67% and 29.12% respectively, and on the Oncology-Screen dataset, these indicators are improved by 40.33%, 44.02% and 20.02% respectively, which shows that the proposed method has good learning ability.

[0082] Table 2 Independent test results

[0083]

[0084] Table 2 shows the effect of the model in the independent test set. It can be seen from Table 2 that the method of the present invention surpasses the comparison method, indicating that the present invention can also achieve good results for unknown data. In particular, on the NCI-ALMANAC dataset, the present invention surpasses the leading method by 2.10% and 1.47% in R² and PCC, respectively. On the O'Neil dataset, the present invention shows an improvement of 3.35%, 4.49% and 2.65% in these indicators. The significant improvement of the model highlights its advantage in accurately predicting the synergy score of unseen data.

[0085] Reference Figure 2 , Figure 3 and Figure 4 , which is the result of the experiment on different learning rates, enhancement ratios and loss function weights α of the present invention. It can be seen that the sensitivity of the present invention to the hyper-parameters is not large. According to the results, this example selects a learning rate of 0.0001, an enhancement ratio of 0.2 and a loss function weight α of 0.2.

[0086] The above embodiments are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A cancer decision-making assistance method based on knowledge-enhanced hypergraph contrastive learning, characterized in that: It includes: S1, obtaining an original drug synergy relationship data set including a plurality of sample data, wherein the sample data is a four-tuple data of drug-drug-cell line-actual synergy relationship; S2, constructing a training hypergraph based on each of the sample data, and performing an enhancement operation on the training hypergraph to obtain a first enhanced view and a second enhanced view, and performing comparative training on a graph convolutional network, a fully connected network, a hypergraph neural network, and a prediction network based on the training hypergraph, the first enhanced view, and the second enhanced view to obtain the trained graph convolutional network, the fully connected network, the hypergraph neural network, and the prediction network; S3, based on the trained graph convolutional network, feature extraction is performed on the drugs in the test data set to obtain several molecular representations of the drugs to be tested; based on the trained fully connected network, feature extraction is performed on several cell lines in the test data set to obtain several molecular representations of the cell lines to be tested; based on the trained graph convolutional network, feature extraction is performed on the drugs in each of the sample data to obtain several molecular representations of sample drugs; based on the trained fully connected network, feature extraction is performed on the cell lines in each of the sample data to obtain several molecular representations of sample cell lines; S4, constructing a sample hypergraph with each of the sample drug molecular representations and each of the sample cell line molecular representations as sample vertices, and updating the sample hypergraph with each of the test drug molecular representations and each of the test cell line molecular representations as test vertices to obtain a test hypergraph; S5, inputting the hypergraph to be tested into the trained hypergraph neural network to obtain an embedded representation of each vertex to be tested; S6, inputting the embedded representation of each of the vertices to be tested into the trained prediction network to obtain the synergistic relationship of each drug in the test data set; The S2 includes: S21, using a graph convolutional network to extract features of drugs in the sample data to obtain a number of training drug molecular representations; using a fully connected network to extract features of cell lines in each of the sample data to obtain a number of training cell line molecular representations; S22, taking each of the training drug molecular representations and each of the training cell line molecular representations as training vertices, adding edges between the training drug molecular representations corresponding to the drugs with antagonistic effects to obtain a drug relationship matrix, adding edges between the training cell line molecular representations corresponding to the cell lines whose similarities are greater than a similarity set value to obtain a cell line similarity matrix, adding hyperedges between the training drug molecular representations and the training cell line molecular representations corresponding to the synergistic relationships greater than a relationship threshold to obtain a hyperedge matrix and the training hypergraph, and performing an enhancement operation on the training hypergraph to obtain the first enhanced view and the second enhanced view; S23, obtaining, based on the hypergraph neural network, an embedded representation of each training vertex in the training hypergraph, an embedded representation of each training vertex in the first enhanced view, and an embedded representation of each training vertex in the second enhanced view; S24, taking the embedding representation of each training vertex in the training hypergraph corresponding to each sample data as a set to obtain a plurality of training embedding sets, inputting each training embedding set into the prediction network to obtain a plurality of training prediction synergy relationships; converting the embedding representation of each training vertex in the first enhanced view and the embedding representation of each training vertex in the second enhanced view based on the projection head to obtain first comparative learning reference data and second comparative learning reference data; S25, obtaining a target loss function value based on the training prediction synergy relationship, the first comparative learning reference data, and the second comparative learning reference data, and optimizing and updating the graph convolution network, the fully connected network, the hypergraph neural network, the prediction network, and the projection head based on the target loss function value; S26, repeatedly executing S21-S25 to obtain the trained graph convolution network, the trained fully connected network, the trained hypergraph neural network and the trained prediction network; The objective loss function value expression is: Where: L is the target loss function value, α is the weight parameter, L d is the drug contrast learning loss value, L c is the cell line comparison learning loss value, L s To predict the network loss value; The expression for predicting network loss value is: Where: m is the number of training embedding sets, S q represents the training prediction synergy relationship corresponding to the qth training embedding set, Y q represents the actual synergy relationship corresponding to the qth training embedding set; The first comparative learning reference data includes first drug comparative learning reference data and first cell line comparative learning reference data, and the second comparative learning reference data includes second drug comparative learning reference data and second cell line comparative learning reference data; The drug contrast learning loss value expression is: Where: l d (p' d,i ,p” d,i ) represents the drug comparison loss function, D is the drug set, is the total number of drugs contained in each sample data, sim is the cosine similarity function, p' d,i is the i-th value in the first drug comparison learning reference data, is the value of the first comparison corresponding to the i-th drug, and p” d,i is the i-th value in the second drug comparison learning reference data, is the value of the second comparison corresponding to the i-th drug, is the value of the i-th row and x-th column in the drug relationship matrix, representing whether the i-th drug and the x-th drug have antagonistic effects. τ represents the temperature coefficient, which is a hyperparameter. The expression of cell line contrast learning loss value is: In the formula, l c (p' c,i ,p” c,i ) represents the cell line comparison loss function, C is the cell line set, is the total number of cell lines contained in each sample data, and p' c,i Indicates the i-th value in the first cell line comparison learning reference data, which is the value of the first comparison corresponding to the i-th cell line, p” c,i Represents the i-th value in the second cell line comparison learning reference data, which is the value of the second comparison corresponding to the i-th cell line. is the value of the ith row and xth column in the cell line similarity matrix, representing whether the similarity between the ith cell line and the xth cell line is greater than the similarity set value.

2. The cancer decision-making assistance method based on knowledge-enhanced hypergraph contrastive learning according to claim 1, characterized in that: The molecular graph of the drug is represented by G d =(M d ,A d ), where M d ∈R n×f represents the set of f-dimensional feature vectors of n atoms in the drug molecule, A d ∈R n×f represents the connection matrix between atoms, represents the atomic bonds in the drug molecule, and R represents a set of real numbers; the cell line uses 651 gene expressions.

3. The cancer decision-making assistance method based on knowledge-enhanced hypergraph contrastive learning according to claim 2 is characterized in that: The graph convolutional network includes K update layers and a maximum pooling layer; The update layer expression is: Where: i represents the i-th drug, j represents the j-th atom of the i-th drug, σ d is the activation function of the update layer, represents the updated features of the jth atom of the ith drug at the kth update layer, represents the updated features of the jth atom of the ith drug at the k-1th update layer, is the parameter matrix of the update layer of the i-th drug at the k-1th layer, Deg i is the degree matrix of the ith drug, A i is the atomic adjacency matrix of the ith drug, A i A in the molecular diagram of the drug d ∈R n×f A d , I i is the identity matrix of the ith drug; The maximum pooling layer expression is: Where: X d,i represents the drug molecule characterization of the i-th drug, represents the updated features of the nth atom of the ith drug at the Kth update layer; The molecular characterization expression of the cell line is: Where: X c,a represents the cell line molecular representation of the a-th cell line, σ c is the activation function of the fully connected network, g represents the expression of 651 genes, represents the parameter matrix of the 0th layer of the fully connected network, Represents the parameter matrix of the u-th layer of the fully connected network, where u is the number of layers in the fully connected network.

4. The cancer decision-making assistance method based on knowledge-enhanced hypergraph contrastive learning according to claim 1, characterized in that: The hypergraph neural network includes a plurality of hyperedge convolutional layers; When the vertex in the hypergraph to be tested is a sample vertex, the hyperedge convolution layer expression is: When the vertex in the hypergraph to be tested is the vertex to be tested, the expression of the hyperedge convolution layer is: X l =s t (X l-1 I l-1 ); Where: l represents the lth hyper-edge convolution layer, X l represents the molecular representation of the vertices of the hypergraph to be tested in the lth hyperedge convolution layer, X l-1 represents the molecular representation of the vertices of the hypergraph at the l-1th hyperedge convolution layer, H is the hyperedge matrix, represents the hyperedge degree matrix, represents the vertex degree matrix, W t represents the hyperedge weight matrix, Θ l-1 represents the feature extraction matrix of the l-1th hyper-edge convolution layer, and T is the transpose.

5. The cancer decision-making assistance method based on knowledge-enhanced hypergraph contrastive learning according to claim 1, characterized in that: The training embedding set includes two training drug embedding representations and one training cell line embedding representation, the two training drug embedding representations correspond to two drugs in the sample data, and the two training drug embeddings are respectively defined as a first training drug embedding representation and a second training drug embedding representation; The prediction network expression is: S=W1σ1(W1(Z d1 ||Z d2 ||Z c )+b0)+b1; Where: S represents the training prediction synergy, Z d1 represents the first training drug embedding representation, Z d2 represents the second training drug embedding representation, Z c represents the embedded representation of the trained cell line, W1 represents the weight matrix of the prediction network, σ1 represents the activation function of the prediction network, b0 represents the first bias of the prediction network, and b1 represents the second bias of the prediction network.

Images