Anticancer drug sensitivity prediction method and system based on bigraph and hypergraph learning

Abstract

The invention discloses an anti-cancer drug sensitivity prediction method and system based on bipartite graph and hypergraph learning, and belongs to the technical field of computer software. The method comprises the following steps: acquiring cell line and drug data and drug and cell line reaction original IC50 values by using an open source database, and standardizing the IC50 values; constructing a bipartite graph of the cell line and the drug, and obtaining embedded representation of the drug and the cell line by applying a graph convolutional network; super nodes are introduced to construct a cell line-drug isomerism hypergraph, and node features are learned through GCN; and finally, fusing the output of the GCN layer by applying an attention mechanism, and predicting the final association probability of the cell line-drug pair in combination with a full-connection layer. According to the method, combined exploration and collaborative optimization of the bipartite graph and the hypergraph are achieved, meanwhile, potential and complex relations between the medicine and the cell line are explored through the hypergraph, the quality of node feature representation is jointly improved, and therefore the association score between the medicine and the cell line is predicted more accurately.

Description

Technical Field

[0001] The present invention belongs to the field of computer software technology, and in particular relates to a method and system for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning. Background Art

[0002] Cancer is a major challenge in the field of global health. Its morbidity and mortality continue to grow, especially among young people, making cancer treatment one of the key concerns in the medical and biological fields. In traditional treatment, clinicians often rely on a "one-size-fits-all" approach to screen, collect medical histories, perform physical examinations, treat, monitor side effects, and prognose cancer patients. At the same time, once cancer occurs, it will mutate within and at metastatic sites. In addition, traditional treatment methods are costly and time-consuming, and medication needs to be determined based on the patient's symptoms, making the treatment method very passive. With the rise of the concept of precision medicine, cancer treatment is gradually moving towards a new stage of individualization and precision. Precision medicine aims to combine the patient's omics characteristics and disease status to tailor personalized treatment plans for them in order to improve the treatment effect. The realization of this goal is inseparable from the deep integration and analysis of clinical data and genomics data, so as to dig out biomarkers to guide treatment and provide more accurate predictions and guidance for treatment.

[0003] Deep learning, with its high flexibility and excellent generalization ability, has shown great application potential in the field of anticancer drug response prediction. Among them, graph neural networks (GNNs) have significantly improved the accuracy of drug response prediction by efficiently processing and fusing information in complex graph data. However, when representing the complex relationship between drugs and cell lines, traditional graph structures have limitations and it is difficult to fully capture this many-to-many relationship and potential high-order interactions. In order to meet this challenge, researchers are constantly exploring new methods and techniques. Liu et al. proposed a method called GraphCDR, which combines graph neural networks and contrastive learning techniques to model the interaction between cancer cell lines and drugs as a heterogeneous graph structure, thereby enhancing the prediction performance of drug sensitivity response. However, traditional graph structures, such as simple bipartite graphs, can still only represent pairwise associations and are difficult to fully reflect the complex many-to-many relationship between cell lines and drugs. In actual biological systems, the relationship between cell lines and drugs is intricate. One drug may act on multiple cell lines, and one cell line may be affected by multiple drugs. This complex relationship network and the potential associations therein are far from being fully represented by simple graph structures. In addition, current research methods, such as using bipartite graphs or hypergraphs alone, are often limited to the perspective of a certain graph structure and cannot achieve joint optimization of representation or features, thus affecting the quality of feature expression. Therefore, how to more effectively represent and capture the complex relationship between cell lines and drugs, and how to mine more potential cell line-drug associations, have become key challenges facing current research. Summary of the invention

[0004] In the prediction of anticancer drug sensitivity, traditional graph structures are difficult to fully capture the complex many-to-many relationship between drugs and cell lines and the potential association between drugs and cell lines. The present invention aims to provide a prediction method and system based on bipartite graphs and hypergraph learning, which combines bipartite graphs and hypergraph learning to more accurately represent the complex relationship between cell lines and drugs and discover more potential associations, thereby providing strong support for the precision treatment of cancer.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: The present invention provides a method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning, comprising: S1. Obtain drug and cell line data from open source databases and preprocess to obtain IC for standardized drug and cell line processing 50 value; S2, IC based on drug and cell line normalization 50Values, extract the genomic features of the cell line and the chemical features of the drug, construct the cell line-drug bipartite graph and the drug-cell line bipartite graph, aggregate the neighbor node features of the cell line and the drug through the graph convolutional network, and obtain the initial cell line and drug node feature representation; S3, based on the cell line-drug bipartite graph and the drug-cell line bipartite graph, super nodes are introduced to construct a cell line-drug heterogeneous hypergraph, and a graph convolutional network is applied to update the node representation to obtain the updated node feature representation of cell lines and drugs; S4, the updated node feature representations of cell lines and drugs are further extracted using the attention mechanism and one-dimensional convolutional network and input into the deep learning model to predict the association probability of drug-cell line pairs.

[0006] The S1 includes: S11, using the open-source database Cancer Drug Sensitivity Genomics as a benchmark dataset and the Cancer Cell Line Encyclopedia as a test dataset; S12, extract basic information of cell lines and drugs from the benchmark dataset and the test dataset, and extract the IC of each cell line for each drug 50 value; S13, IC 50 The values ​​were normalized to obtain the IC values ​​normalized for drugs and cell lines. 50 value.

[0007] The standardization process is Z-score standardization or Min-Max standardization.

[0008] Furthermore, the Z-score normalization process is as follows:

[0009] in, represents the gene expression value of the th gene, and represents the mean and standard deviation of gene expression.

[0010] S2 includes: S21, IC based on drug and cell line standardization 50 Values, extract the genomic features of the cell line and the chemical features of the drug, apply linear transformation, obtain the cell line projection feature matrix and the drug projection feature matrix

[0011] S22, taking drugs and cell lines as nodes and drug and cell line reaction information as edges, constructing cell line-drug bipartite graph and drug-cell line bipartite graph, applying graph convolutional network aggregation to aggregate the neighbor node features of cell lines and drugs on the cell line-drug bipartite graph and drug-cell line bipartite graph, and obtaining aggregated cell line neighbor node feature representation and aggregated drug neighbor node feature representation; S23, based on the aggregated cell line neighbor node feature representation and the aggregated drug neighbor node feature representation, define the aggregation function of the cell line and the drug, and generate the cell line and drug node feature representation.

[0012] Further, in S21, the cell line projection feature matrix is:

[0013] is the cell line projection feature matrix, represents the original gene expression matrix, m is the number of cancer cells, is the dimension of gene expression features, each row corresponds to the feature vector of a cell line, is the parameter set of the linear transformation of the cell line; Further, in S21, the drug projection feature matrix is:

[0014] represents the drug molecule fingerprint matrix, n represents the number of drugs, is the characteristic dimension of drug molecule fingerprint, is the drug projection feature matrix, is the parameter set of the linear transformation of the drug; Further, in S22, the aggregated drug neighbor node feature is expressed as:

[0015] in, Indicates the first ) Representation of cell line nodes at iteration is the initial property of the cell line, represents the adjacency matrix plus the self-connectivity matrix, express The transpose of , , represents the aggregation weight parameter of the cell line, represents the aggregation weight parameter of the drug; Further, in S22, the neighbor node feature of the aggregated cell line is represented as: ; in, Indicates the first ) iterations, is the initial property of the drug, represents the adjacency matrix plus the self-connectivity matrix, express The transpose of , , represents the aggregation weight parameter of the cell line, represents the aggregation weight parameter of the drug; Further, in S22, the cell line and drug aggregation function is:

[0016]

[0017] in is the ReLU activation function, and They are the cell lines and drug embedding obtained after multi-layer propagation, SC and SD are the characteristics of the cell lines and drugs themselves, , .

[0018] The S3 includes: S31, based on the cell line-drug bipartite graph and the drug-cell line bipartite graph, super nodes are introduced to construct a cell line-drug heterogeneous hypergraph, and the cosine similarity is calculated to obtain the adjacency matrix between the hypergraph nodes and the drugs and cell lines. The node feature representation of the cell line-drug heterogeneous hypergraph is updated using the graph convolutional network to obtain the node feature representation of the cell line-drug heterogeneous hypergraph; S32, extracting node feature representations of cell lines and drugs according to the node feature representations of the cell line-drug heterogeneous hypergraph, and obtaining updated node feature representations of cell lines and drugs.

[0019] Furthermore, the cell line-drug heterogeneous hypergraph can represent the relationships between cell lines, drugs, and abstract concepts such as specific biological functions and pathways; Further, in S31, the adjacency matrix between the hypergraph nodes and the cell line is:

[0020] in, Representative cell lines, Represents the introduced super nodes, represents the adjacency matrix between the hypergraph nodes and the cell lines, represent The 2-norm of represent The 2-norm of .

[0021] Further, in S31, the adjacency matrix between the hypergraph nodes and the drugs is:

[0022] in, Represents drugs, Represents the introduced super nodes, represents the adjacency matrix between the hypergraph nodes and the drugs, represent The 2-norm of represent The 2-norm of .

[0023] Furthermore, in S31, the node features of the cell line-drug heterogeneous hypergraph are expressed as:

[0024] in, represents the hypergraph adjacency matrix with the identity matrix I added, where is the number of super nodes, It is the LeakyReLU activation function, which mainly ensures the validity of node information. yes The degree matrix of represents the embedding of the nodes in the lth GCN layer, represents the learnable weight matrix, is the ReLU activation function, It is expressed as: ; Further, in S32, the node features of the updated cell lines and drugs are expressed as:

[0025] in, is the node feature representation of the updated cell line, express The first line of .

[0026] Further, in S32, the node feature of the updated drug is expressed as:

[0027] in, is the node feature representation of the updated drug, express The second line of .

[0028] S4 includes: S41, node feature representations of cell lines and drugs updated according to the outputs of each layer of the graph convolutional network, using the attention mechanism to obtain hierarchical node feature representations of different importance; S42, using a one-dimensional convolutional network to perform convolution operations on hierarchical cell lines and drug nodes of different importance to further extract features in the node feature representation, and obtain the node feature representation after the convolution operation; S43, perform element-wise product operation on the node feature representations of cell lines and drugs after convolution operation to obtain the interaction features between them, input the interaction features into the deep learning model to predict the association probability of drug-cell line pairs, and use the cross entropy loss function to calculate the error between the model prediction result and the true label.

[0029] Further, in S41, the hierarchical node feature is represented as follows:

[0030]

[0031] in, Indicates passing The node feature representation obtained by the layer output is represents the attention mechanism, The specific implementation is as follows:

[0032] Among them, Q, K, and V represent Query, Key, and Value respectively. For cell lines, , , , , , is the trainable parameter matrix, represents the dimension of K; Further, in S42, the convolution operation is specifically expressed as follows:

[0033]

[0034] Among them, w represents the weight of the convolution kernel, k represents the length of the convolution kernel, b represents the bias term, and f represents the Softmax function.

[0035] Further, in S43, the association probability of the drug-cell line pair is expressed as:

[0036] in, is the element-wise product of the cell line node vector and the drug node vector, It is a one-layer feed-forward neural network whose output is activated by the Sigmoid activation function.

[0037] The cross entropy loss function is:

[0038] in, is the cross entropy loss function, Y and represent the positive and negative connection sets of all drug-cell line pairs, respectively. Indicates the true label of the connection, marked as 1 or 0, Represents the association probability of a drug-cell line pair.

[0039] Furthermore, the deep learning model is a feedforward neural network model, which consists of one or more hidden layers, and the final output layer uses a Sigmoid activation function to map the output value between 0 and 1 to represent the association probability.

[0040] The method also includes the steps of verifying the predicted association probability and ranking or classifying the drug sensitivity according to the magnitude or threshold of the association probability.

[0041] The present invention provides a system used in the above-mentioned anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning, comprising: Data acquisition and preprocessing module, used to acquire drug and cell line data, preprocess for cleaning and standardizing data; The bipartite graph construction and feature representation acquisition module extracts the genomic features of the cell line and the chemical features of the drug based on the preprocessed data, constructs the cell line-drug bipartite graph and the drug-cell line bipartite graph, and applies the graph convolutional network (GCN) to aggregate the neighbor node features of the cell line and the drug, thereby obtaining the feature representation of the cell line and the drug; Construct heterogeneous hypergraphs and update node feature representation modules. Based on the cell line-drug bipartite graph and drug-cell line bipartite graph, introduce super nodes, construct cell line-drug heterogeneous hypergraphs, and apply graph convolutional networks to obtain updated node feature representations of cell lines and drugs. The module for predicting the association probability of drug-cell line pairs is used to further extract and update the node feature representations of the updated cell lines and drugs and input them into the deep learning model. The model is used to predict the association probability of drug-cell line pairs, that is, the sensitivity of the drug to a specific cell line.

[0042] The present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning when executing the computer program.

[0043] The present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects: The anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning provided by the present invention has shown significant technical effects in terms of feature fusion, capture of associated information, model optimization and prediction, coping with sample imbalance and practical application value. This method not only improves the accuracy and reliability of prediction, but also provides strong support for the research and development and application of anticancer drugs, and has important scientific significance and practical application value. This method integrates multiple characteristics of cell lines and drugs, which helps to more comprehensively understand the characteristics of cell lines and drugs, so as to more accurately predict the interaction between them; through graph convolutional network learning, the node representation node feature representation not only considers the direct association relationship, but also incorporates potential and complex relationship information, thereby improving the quality; the joint learning of the bipartite graph and the hypergraph can more comprehensively capture the association information between cell lines and drugs, reduce the possibility of information loss, help improve the accuracy and reliability of prediction, and provide strong support for the research and development and application of anticancer drugs; through the joint exploration and collaborative optimization of the bipartite graph and the hypergraph, the dynamic update of the node feature representation is realized, and the node representation Based on the joint learning algorithm of bipartite graph and hypergraph, the connection relationship and node feature representation are continuously updated and optimized. The dynamic update method can more comprehensively capture the complex relationship between cell lines and drugs, avoiding the information loss or misleading that may be caused by static representation; at the same time, the quality of node feature representation is improved. The node representation not only considers the direct association relationship, but also incorporates potential and complex relationship information, which helps to more accurately predict the association score between cell lines and drugs, and provides a more reliable basis for the prediction of anticancer drug sensitivity; this method improves the accuracy and reliability of prediction by deeply exploring the complex relationship between cell lines and drugs, providing strong support for the research and development and application of anticancer drugs.

[0045] Furthermore, IC 50The standardized processing method of values ​​helps to eliminate the differences between data and improve the generalization ability of the model; super nodes are introduced to construct a cell line-drug heterogeneous hypergraph, and the node representation is updated through a graph convolutional network, which further explores the potential relationship between cell lines and drugs and improves the quality of node feature representation; the hierarchical node feature representations of different importance are further extracted and fused through the attention mechanism and one-dimensional convolutional network to obtain richer feature information; the interactive features are input into the deep learning model to predict the association probability of drug-cell line pairs, thereby achieving accurate prediction of anticancer drug sensitivity.

[0046] The system provided by the present invention comprehensively improves the accuracy and reliability of anticancer drug sensitivity prediction by integrating bipartite graph and hypergraph learning. The system uses modules such as data preprocessing, bipartite graph construction, feature aggregation, and heterogeneous hypergraph learning to deeply explore the complex relationship between cell lines and drugs. By dynamically updating node feature representation and fusion features, the system effectively captures potential and complex correlation information and avoids information loss. The introduction of deep learning models further improves the prediction accuracy and provides strong support for the research and development of anticancer drugs. The result verification and output module ensures accurate evaluation of model performance and provides a reliable decision-making basis for clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart of the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning of the present invention; Figure 2 This is an algorithm framework diagram of the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning of the present invention; Figure 3 The ROC curve diagram drawn for cross-validation using different K values ​​in the GDSC data set of the present invention; Figure 4 ROC curves drawn for cross-validation of different modules in the GDSC data set of the present invention; Figure 5 The influence of different parameters on the prediction performance of the present invention, A is the evaluation index score of different feature dimensions, B is the evaluation index score of different numbers of hidden layers, and C is the evaluation index score of different numbers of super nodes; Figure 6 This is a diagram showing the influence of different positive and negative sample ratios on the prediction performance in the GDSC data set of the present invention. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0049] The anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning proposed in the present invention predicts anticancer drug reactions in an end-to-end manner, and introduces a virtual node, namely a supernode, to construct a heterogeneous hypergraph of cell lines and drugs to explore potential cell line-drug pairs. First, the cell line gene expression and drug molecular fingerprints are linearly transformed so that they have the same dimension. Next, a cell line-drug bipartite graph is constructed, in which the node features represent drugs and cell lines, and the edges represent known reaction information. By applying a graph convolutional network, useful information is extracted from the bipartite graph to obtain an embedded representation of drugs and cell lines. Next, a cell line-drug heterogeneous hypergraph network is constructed using supernodes and GCN is used to aggregate hypergraph information. Finally, the attention mechanism is applied to fuse the output of the GCN layer, and the final score is obtained in combination with the fully connected layer.

[0050] Example 1 See attached Figure 1~2 As shown, it is a framework diagram of the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning of the present invention, and the specific steps are as follows: S1, using the open source database Genomics of Cancer Drug Sensitivity (GDSC) as a benchmark dataset and the Cancer Cell Line Encyclopedia (CCLE) as an extended test dataset to obtain cell line and drug data as well as raw IC of drug and cell line responses 50 value, and for IC 50 The values ​​are standardized.

[0051] S2, extract cell line and drug features, build a cell line-drug bipartite graph based on known drug reactions in cell lines, aggregate the neighbor node features of cell lines through graph convolutional networks, and use the same method to build a drug-cell line bipartite graph based on known cell line reactions in drugs to aggregate the neighbor node features of drugs.

[0052] S3, based on the cell line-drug bipartite graph and the drug-cell line bipartite graph, introduces super nodes to construct a cell line-drug heterogeneous hypergraph, applies a graph convolutional network to update the node representation, and obtains the updated node feature representation of cell lines and drugs.

[0053] S4, based on the cell line-drug bipartite graph and the drug-cell line bipartite graph, introduces super nodes to construct a cell line-drug heterogeneous hypergraph, applies a graph convolutional network to update the representation of the nodes, and obtains the updated node feature representation of cell lines and drugs.

[0054] The step S1 includes the following steps: S11, the GDSC dataset includes 256 drugs, 1001 cancer cell lines, and IC50 values ​​of drugs and cell lines. Since there are large differences in the mean and variance distribution of the original IC50 values ​​of drugs and cell lines in the GDSC dataset, it is necessary to eliminate the statistical differences between different drugs before using the data, that is, to perform standardization. The standardization process is as follows:

[0055] represents the gene expression value of the th gene, and represents the mean and standard deviation of gene expression.

[0056] S12, after preprocessing, the preprocessing method was similar to (Sharifi-Noghabi, et al, 2019), 962 cell lines and 228 drugs were obtained in the GDSC database, and the IC50 values ​​between drugs and cell lines were divided into sensitive and resistant categories according to the threshold, finally including 20851 sensitive reactions, 156512 resistant reactions and 41973 unknown responses.

[0057] The step S2 comprises the following steps: S21, the gene expression characteristics of the cell line are encoded into an f-dimensional representation matrix as follows:

[0058] is the cell line projection feature matrix, represents the original gene expression matrix, m is the number of cancer cells, is the dimension of gene expression features, each row corresponds to the feature vector of a cell line, is the parameter set of the linear transformation of the cell line. Similarly, the drug molecular fingerprint is linearly transformed to obtain the drug feature representation as shown below:

[0059] represents the drug molecule fingerprint matrix, n represents the number of drugs, is the characteristic dimension of drug molecule fingerprint, is the drug projection feature matrix, is the set of parameters for the linear transformation of drugs.

[0060] S22, based on the known drug reactions in the cell lines, a cell line-drug bipartite graph is constructed, and the neighbor node features of the cell lines are aggregated through the graph convolutional network and represented as:

[0061] Similarly, the characteristics of drug neighbor nodes can be expressed as:

[0062] in and Respectively represent the first ) iterations, and are the initial properties of the cell line and drug in step S1, represents the adjacency matrix plus the self-connectivity matrix, express The transpose of , , represents the aggregation weight parameter of the cell line, Represents the aggregation weight parameter of the drug.

[0063] S23, and finally the cell line and drug aggregation functions were obtained as follows:

[0064]

[0065] The step S3 comprises the following steps: S31, after the hypergraph is introduced, the adjacency matrix between the hypergraph nodes and the drugs and cell lines is obtained by calculating the cosine similarity, which is specifically expressed as and , and then use GCN to update the node feature representation:

[0066] in represents the hypergraph adjacency matrix with the identity matrix I added, where is the number of super nodes, It is the LeakyReLU activation function, which mainly ensures the validity of node information. yes The degree matrix of . represents the embedding of the nodes in the lth GCN layer. represents the learnable weight matrix, is the ReLU activation function. , which can be expressed as:

[0067] S32, according to the output of each layer of GCN, the cell lines can be obtained respectively and medications The node feature representation is:

[0068] in express The first line of express The second line of .

[0069] The step S4 comprises the following steps: S41, use the attention mechanism to obtain hierarchical node feature representations of different importance:

[0070]

[0071] in Indicates passing The node feature representation obtained by the layer output is Represents the attention mechanism. Unlike the fully connected layer whose weights are fixed after model training, the weights of the attention mechanism will still change with the input after training. The specific implementation is as follows:

[0072] Among them, Q, K, and V represent Query, Key, and Value respectively. For cell lines, , , , , , is a trainable parameter matrix. Represents the dimension of K. The representation of drugs is similar to that of cell lines.

[0073] S42, using a one-dimensional convolutional network to perform convolution operations on cell line and drug nodes, specifically expressed as:

[0074]

[0075] Among them, w represents the convolution kernel, k represents the length of the convolution kernel, b represents the bias term, and f represents the Softmax function.

[0076] S43, calculate the element-wise product of the cell line node embedding and the drug node embedding, and predict the association probability of the drug-cell line pair through FNN:

[0077] In this formula is the element-wise product of the cell line node vector and the drug node vector, It is a feedforward neural network whose output is activated by the Sigmoid activation function. At the same time, in order to optimize the model, the model uses the cross entropy loss function to calculate the loss during training:

[0078] where Y and represent the positive and negative connection sets of all drug-cell line pairs, respectively. Indicates the true label of the connection, marked as 1 or 0.

[0079] Example 2 In order to evaluate the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning of the present invention, five models with better performance, DeepCDR, DeepDSC, MOFGCN, and GraphCDR, were used for comparative study. DeepCDR is an anticancer drug response (CDR) predictor based on a hybrid graph convolutional network, which integrates multi-omics features of cell lines and drugs. DeepDSC uses stacked deep autoencoders to extract cell features from gene expression data, and then combines them with drug chemistry features for drug response prediction. MOFGCN introduces a method based on multi-omics fusion and graph convolution to predict drug response. The model uses multi-omics data, including information such as gene expression and copy number variation, combined with a graph convolutional network to capture the complex relationship between drugs and cell lines. GraphCDR uses contrastive learning to predict the response of cancer drugs. The method of the present invention combines the characteristics of graph neural networks, learns the graph structure representation of drugs and cell lines, and uses a contrastive learning framework to optimize feature representation to improve the accuracy of drug response prediction.

[0080] Drug sensitivity is predicted by estimating the association probability between cell lines and drugs. In this experimental design, the sensitive reaction of cell lines to drugs is used as positive samples, and the drug resistance reaction is used as negative samples. These samples are randomly divided into five equal parts, and a five-fold cross-validation method is used. In each round of validation, 1 / 5 of the positive samples and the same number of negative samples are used as test data, and the remaining 4 / 5 positive samples and negative samples are used as training data. The experimental results are shown in Table 1: Table 1: Randomized cross validation

[0081] As can be seen from the data in Table 1, when classifying whether a drug is sensitive or resistant to a cell line, the method of the present invention always outperforms all baseline methods of the two data sets. In the GDSC data set, the AUC value of the anticancer drug sensitivity prediction method (HGCNDCR) based on bipartite graph and hypergraph learning of the present invention reached 95.67%, which was significantly better than other baseline methods; in the CCLE data set, the AUC value of the anticancer drug sensitivity prediction method (HGCNDCR) based on bipartite graph and hypergraph learning of the present invention reached 97.65%, which was also better than all baseline methods; it shows that it is very important for the method of the present invention to consider the element interaction between cell lines and drugs and explore the potential association between cell lines and drugs. The method of the present invention has excellent performance in predicting the sensitive or resistant response of drugs to cell lines. Since the graph convolutional network method can explicitly encode drugs and cell lines by aggregating information from neighbors with similar response signals. Therefore, the methods based on graph convolutional networks (such as MOFGCN and GraphCDR) perform better in the data set than the baseline methods based on deep learning (such as DeepDSC).

[0082] In order to evaluate the impact of different K-CV cross-validation on model performance, this experiment used different K values ​​for cross-validation in the GDSC dataset. The K values ​​were 2, 5, and 10, respectively. The corresponding ROC curves were drawn. The corresponding AUC specific results are shown in the attached Figure 3 shown.

[0083] By the attached Figure 3 The data shows that as the number of training data sets increases (i.e., the K value increases), the prediction accuracy improves; the validation effect of 10-fold cross validation is better than that of 2-fold and 5-fold cross validation, and the average AUC value is also higher. However, it is worth noting that 10-fold cross validation requires more computing resources.

[0084] The present invention's anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning (HGCNDCR) includes learning drug and cell line features based on graph convolutional network, introducing super nodes to build hypergraphs, and introducing attention mechanisms to predict final drug sensitivity responses. In order to verify the importance of each module in the present invention's anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning (HGCNDCR), three variant models HGCNCDR (w / o GCN), HGCNCDR (w / oHyper) and HGCNCDR (w / o LayerAtt) were developed for comparison. Specifically, HGCNCDR (w / o GCN) removes the graph convolutional network layer and directly represents the linearly transformed drugs and cell lines as embedded association hypergraphs. HGCNCDR (w / o Hyper) aims to study the contribution of super nodes to model performance, and uses the original cell line-drug heterogeneous graph to replace the cell line-drug hypergraph in the experiment. HGCNCDR (w / o LayerAtt) removes hierarchical attention before classification. Figure 4 The ROC curves of HGCNCDR (w / o GCN), HGCNCDR (w / o Hyper) and HGCNCDR (w / o LayerAtt) are compared with the original model. It can be found that the graph convolutional neural network module and the layer attention module have a certain impact on the prediction results of the model, indicating that the embedding representation learned by GCN and the output of different layers have different importance on the results. At the same time, the AUC of HGCNCDR-Hyper is lower than that of the original model, which shows the effectiveness of introducing super nodes to construct a hypergraph.

[0085] Hyperparameters can affect the prediction performance of the model. Therefore, a 5-fold cross validation was used in the experiment to analyze the impact of some parameters on the prediction results and compare the evaluation index values. The hyperparameters involved in the present invention mainly include feature dimensions, the number of hidden layers, and the number of super nodes, which are specifically analyzed as follows.

[0086] First, the feature embedding dimension is analyzed. The embedding size of the feature affects the dimension of its corresponding learnable parameter matrix. Figure 5In Figure 1, A shows the evaluation index scores corresponding to different feature dimensions [32, 64, 128, 256]. When the feature dimension is 128, the AUC score is optimal. In order to ensure the prediction performance of the model and the quality of feature information, the feature dimension of all nodes in this model is set to 128. Secondly, the impact of the number of hidden layers is explored. Generally speaking, increasing the number of GCN layers can aggregate more node information. However, increasing the number of GCN layers too much can easily lead to over-smoothing and over-fitting problems; B shows the evaluation index scores of different numbers of hidden layers [1, 2, 3, 4]; experiments show that when the number of GCN layers gradually increases to 3 layers, the performance indicators of the model begin to decline, so the number of GCN layers is finally set to 2; finally, the impact of different numbers of super nodes on model performance is studied. HGCNCDR uses supernodes to construct a cell line-drug heterogeneous hypergraph to connect cell line-drug pairs. Therefore, changes in the number of supernodes will change the structure of the hypergraph. C shows the evaluation index scores of different numbers of supernodes [32, 64, 128, 256]. Since the representation of supernodes is randomly initialized, too many supernodes are not conducive to extracting useful information and even complicate the graph structure. Therefore, the number of supernodes is finally selected as 64.

[0087] The number of positive samples in the GDSC dataset is 20851, and the number of negative samples is 157175. The sample distribution is extremely unbalanced. In order to evaluate the performance of the model under different numbers of negative samples, the experiment uses the precision-recall curve (PR curve) and the area under the curve AUPR value as evaluation indicators. Considering the extremely unbalanced situation of positive and negative samples in the dataset, the PR curve is more discriminatory than the ROC curve. In the experiment, the negative samples in the training set are fixed to p times the positive samples during sampling, where the values ​​of p are 1, 3, and 5, respectively, and then a 5-fold cross-validation is performed. In actual situations, the positive and negative ratios of samples are difficult to calculate accurately, so the p value in the experiment is an approximate value. Attachment Figure 6 The effect of different positive-negative sample ratios on the prediction performance is demonstrated. When the number of positive and negative samples is equal, the model has the best performance, with a corresponding AUC value of 0.9686 and an AUPR value of 0.9664. When all negative samples are taken, the AUC value is 0.9567 and the AUPR value is 0.7883. As the sample distribution becomes more and more balanced, the AUC value fluctuates less, showing good classification performance. The PR curve involves the calculation of the precision, and the sample distribution has a greater impact on its value. The AUPR value continues to rise, increasing by 17.8%, indicating that the sample distribution has a certain influence on the performance prediction of the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning of the present invention.

[0088] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning, characterized in that: include: S1. Obtain drug and cell line data from open source databases and preprocess to obtain IC for drug and cell line standardization 50 value; S2, IC based on drug and cell line normalization 50 Values, extract the genomic features of the cell line and the chemical features of the drug, construct the cell line-drug bipartite graph and the drug-cell line bipartite graph, aggregate the neighbor node features of the cell line and the drug through the graph convolutional network, and obtain the cell line and drug node feature representation; S3, based on the cell line-drug bipartite graph and the drug-cell line bipartite graph, super nodes are introduced to construct a cell line-drug heterogeneous hypergraph, and a graph convolutional network is applied to update the node representation to obtain the updated node feature representation of cell lines and drugs; S4, the updated node feature representations of cell lines and drugs are further extracted using the attention mechanism and one-dimensional convolutional network and input into the deep learning model to predict the association probability of drug-cell line pairs.

2. The method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to claim 1, characterized in that: The S1 includes: S11, using the open-source database Cancer Drug Sensitivity Genomics as a benchmark dataset and the Cancer Cell Line Encyclopedia as a test dataset; S12, extract basic information of cell lines and drugs from the benchmark dataset and the test dataset, and extract the IC of each cell line for each drug 50 value; S13, IC 50 The values ​​were normalized to obtain the IC values ​​normalized for drugs and cell lines. 50 value.

3. The method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to claim 2, characterized in that: The standardization process is Z-score standardization or Min-Max standardization.

4. The method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to claim 1, characterized in that: S2 includes: S21, IC based on drug and cell line standardization 50 Values, extract the genomic features of the cell line and the chemical features of the drug, apply linear transformation, and obtain the cell line projection feature matrix and the drug projection feature matrix; S22, taking drugs and cell lines as nodes and drug and cell line reaction information as edges, constructing cell line-drug bipartite graph and drug-cell line bipartite graph, applying graph convolutional network aggregation to aggregate the neighbor node features of cell lines and drugs on the cell line-drug bipartite graph and drug-cell line bipartite graph, and obtaining aggregated cell line neighbor node feature representation and aggregated drug neighbor node feature representation; S23, based on the aggregated cell line neighbor node feature representation and the aggregated drug neighbor node feature representation, define the aggregation function of the cell line and the drug, and generate the cell line and drug node feature representation.

5. The method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to claim 1, characterized in that: The S3 includes: S31, based on the cell line-drug bipartite graph and the drug-cell line bipartite graph, super nodes are introduced to construct a cell line-drug heterogeneous hypergraph, and the cosine similarity is calculated to obtain the adjacency matrix between the hypergraph nodes and the drugs and cell lines. The node feature representation of the cell line-drug heterogeneous hypergraph is updated using the graph convolutional network to obtain the node feature representation of the cell line-drug heterogeneous hypergraph; S32, extracting node feature representations of cell lines and drugs according to the node feature representations of the cell line-drug heterogeneous hypergraph, and obtaining updated node feature representations of cell lines and drugs.

6. The method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to claim 1, characterized in that: S4 includes: S41, inputting the updated node feature representations of the cell lines and drugs into the graph convolutional network deep learning model, and using the attention mechanism to obtain hierarchical node feature representations of different importance according to the node feature representations of the cell lines and drugs updated by the outputs of each layer of the graph convolutional network; S42, using a one-dimensional convolutional network to perform convolution operations on hierarchical cell lines and drug nodes of different importance to further extract features in the node feature representation, and obtain the node feature representation after the convolution operation; S43, perform element-wise product operation on the node feature representations of cell lines and drugs after convolution operation to obtain the interaction features between them, input the interaction features into the deep learning model to predict the association probability of drug-cell line pairs, and use the cross entropy loss function to calculate the error between the model prediction result and the true label.

7. The method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to claim 1, characterized in that: The method also includes the steps of verifying the predicted association probability and ranking or classifying the drug sensitivity according to the magnitude or threshold of the association probability.

8. The system used in the method for predicting anticancer drug sensitivity based on bipartite graph and hypergraph learning according to any one of claims 1 to 7, characterized in that: include: Data acquisition and preprocessing module, used to acquire, clean, and normalize drug and cell line benchmark and test datasets; The bipartite graph construction and feature representation acquisition module extracts the genomic features of the cell line and the chemical features of the drug based on the preprocessed data, constructs the cell line-drug bipartite graph and the drug-cell line bipartite graph, and applies the graph convolutional network (GCN) to aggregate the neighbor node features of the cell line and the drug, thereby obtaining the feature representation of the cell line and the drug; Construct heterogeneous hypergraphs and update node feature representation modules. Based on the cell line-drug bipartite graph and drug-cell line bipartite graph, introduce super nodes, construct cell line-drug heterogeneous hypergraphs, and apply graph convolutional networks to obtain updated node feature representations of cell lines and drugs. The module for predicting the association probability of drug-cell line pairs is used to input the updated node feature representations of cell lines and drugs into the deep learning model, and use the model to predict the association probability of drug-cell line pairs, that is, the sensitivity of drugs to specific cell lines.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the anticancer drug sensitivity prediction method based on bipartite graph and hypergraph learning described in any one of claims 1 to 7 are implemented.

Images