Drug response prediction method and device, computer equipment and readable storage medium

Abstract

The embodiment of the invention provides a drug response prediction method and device, computer equipment and a readable storage medium. Comprising the following steps: performing gene sequencing to obtain multilevel omics characteristics of target cells; determining the plurality of target genes as a plurality of graph nodes, taking a plurality of attribute association relationships between any two target genes as a plurality of interaction edges, and generating a multi-edge graph; each target gene is associated with the corresponding multi-level omics feature as a node feature; for each graph node, carrying out message passing through a plurality of interaction edges, updating node features in each graph node according to obtained adjacent node features, and carrying out feature mapping on the updated multi-edge graph to obtain a first cell feature vector; and predicting a first fusion vector obtained by fusing the first cell feature vector and the drug feature vector of the drug compound through a full connection layer of the target model to obtain a predicted drug response result of the drug compound to the target cell. Therefore, the accuracy of drug response prediction can be improved.

Description

Technical Field

[0001] The present application relates to the field of bioinformatics analysis technology, and in particular to a drug response prediction method, apparatus, computer equipment, and readable storage medium. Background Art

[0002] Drug response prediction refers to the use of computational models and biological data to predict the response of individuals or cells to specific drugs, in the hope of understanding the factors that may affect the efficacy and safety of drugs, thereby predicting the biological response of individuals or cell lines after using a certain drug, such as drug efficacy, toxicity, etc., thereby improving the accuracy and effectiveness of treatment.

[0003] Related technologies typically use single-modality omics data to predict drug responses. For example, using one of the genomic, transcriptomic, and proteomic data sets to predict drug responses in specific cell lines or organisms reduces the difficulty and improves the efficiency of data integration. However, using a single omics data set to predict drug responses only captures biological information at a specific level within the biological system and may not reflect the actual expression levels of genes or the functional status of proteins, leading to inaccurate predictions. Summary of the Invention

[0004] The main purpose of the embodiments of the present application is to provide a drug response prediction method, apparatus, computer device and readable storage medium, which can improve the accuracy of drug response prediction.

[0005] To achieve the above objectives, a first aspect of an embodiment of the present application provides a drug response prediction method, the method comprising:

[0006] Performing gene sequencing on multiple target genes in target cells to obtain multi-level omics characteristics of the target cells;

[0007] Obtaining a gene expression vector of the target cell, determining multiple target genes in the gene expression vector as multiple graph nodes, and using multiple attribute associations between any two target genes as multiple interaction edges to generate a multi-edge graph; wherein the multi-level omics features corresponding to each target gene association serve as node features of the corresponding graph node;

[0008] Through multiple graph neural network layers in a preset target model, in each graph neural network layer, for each graph node in the multi-edge graph, multiple message transmissions are performed through the multiple interaction edges connected to adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes, and node features in each graph node are updated according to the adjacent node features, and feature mapping is performed on the updated multi-edge graph to obtain a first cell feature vector;

[0009] Obtaining a drug feature vector of the drug compound corresponding to the target cell through the graph isomorphism network layer of the target model, and fusing the first cell feature vector and the drug feature vector to obtain a first fusion vector;

[0010] The first fusion vector is predicted through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

[0011] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a drug response prediction device, comprising:

[0012] A sequencing module is used to sequence multiple target genes in target cells to obtain multi-level genomic features of the target cells;

[0013] A generation module is configured to obtain a gene expression vector of the target cell, determine multiple target genes in the gene expression vector as multiple graph nodes, and use multiple attribute associations between any two target genes as multiple interaction edges to generate a multi-edge graph; wherein the multi-level omics features corresponding to each target gene association serve as node features of the corresponding graph node;

[0014] An updating module is configured to, through a plurality of graph neural network layers in a preset target model, perform multiple message passing in each graph neural network layer for each graph node in the multi-edge graph through the plurality of interaction edges connected to adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes, update node features in each graph node according to the adjacent node features, and perform feature mapping on the updated multi-edge graph to obtain a first cell feature vector;

[0015] a fusion module, configured to obtain a drug feature vector of the drug compound corresponding to the target cell through the graph isomorphism network layer of the target model, and fuse the first cell feature vector and the drug feature vector to obtain a first fusion vector;

[0016] A prediction module is used to predict the first fusion vector through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

[0017] In some embodiments, the drug response prediction device further comprises an extraction module for:

[0018] Extracting high-dimensional features from the gene expression vector using a target encoder of the target model to determine a second cell feature vector, wherein the second cell feature vector is used to characterize the interaction relationship between each target gene and the corresponding gene expression pattern;

[0019] Fusing the second cell feature vector with the first fusion vector to obtain a second fusion vector;

[0020] Then, predicting the first fusion vector through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell includes:

[0021] The second fusion vector is predicted through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

[0022] In some embodiments, the multi-level omics features include gene expression features, copy number variation features, gene mutation status features, and methylation status features, and the sequencing module is further used to:

[0023] Obtaining a gene expression vector corresponding to the target cell using a genome sequencer, and using the gene expression amount of each target gene in the gene expression vector as a gene expression feature of the corresponding target gene;

[0024] Sequencing the genome of the target cells using multiple genome sequencing methods to obtain the copy number variation status, gene mutation status, and methylation status of each target gene;

[0025] Converting the copy number variation state into a binary copy matrix, and determining the copy number variation feature of each target gene based on the binary copy matrix;

[0026] Converting the gene mutation state into a binary mutation matrix, and determining the gene mutation state characteristics of each target gene based on the binary mutation matrix;

[0027] The methylation status is converted into a binary methylation matrix, and the methylation status characteristics of each target gene are determined based on the binary methylation matrix.

[0028] In some embodiments, the multiple attribute association relationships include an association relationship between proteins encoded by any two target genes, a gene pathway relationship between any two genes, and a similarity relationship between any two genes. The drug response prediction device further includes an acquisition module for:

[0029] Obtain the predicted protein sequence of each target gene, and query the predicted protein sequences of any two target genes through the gene expression database to obtain the correlation relationship between the proteins produced by the any two target genes;

[0030] Querying the gene pathway relationship between the two target genes through a bioinformatics database to obtain the gene pathway relationship between the two target genes;

[0031] The Pearson correlation coefficient between the two target genes is determined by the gene expression vector corresponding to the target cell, and the similarity relationship between the two target genes is determined based on the Pearson correlation coefficient.

[0032] In some implementations, the update module is further configured to:

[0033] By using the current graph neural network layer in the preset target model, for each graph node in the multi-edge graph updated by the previous graph neural network layer, multiple message transmissions are performed through the multiple interaction edges connected to the adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes;

[0034] Updating the node features in each graph node according to the adjacent node features through the node-level gated recurrent unit of the current graph neural network layer to obtain the updated node features of each graph node;

[0035] After the node features of the multiple graph nodes of the multi-edge graph are updated, the global node features of the multi-edge graph are updated through the graph-level gated recurrent unit of the current graph neural network layer, and the updated multi-edge graph is passed to the next graph neural network layer;

[0036] Repeatedly passing through the current graph neural network layer in the preset target model, for each graph node in the updated multi-edge graph updated by the previous graph neural network layer, multiple message transmissions are performed through the multiple interactive edges connected to the adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes; through the node-level gated loop unit of the current graph neural network layer, the node features in each graph node are updated according to the adjacent node features to obtain the updated node features of each graph node; after the node features of the multiple graph nodes in the updated multi-edge graph are updated, the global node features of the multi-edge graph are updated through the graph-level gated loop unit of the current graph neural network layer until the next graph neural network layer is the last layer of the multiple graph neural network layers to obtain an updated multi-edge graph;

[0037] Feature mapping is performed on the updated multi-edge graph to obtain a first cell vector.

[0038] In some embodiments, the fusion module is further configured to:

[0039] Constructing a drug molecular graph of the drug compound using drug atoms of the drug compound corresponding to the target cell as drug nodes and bonds between drug atoms as drug edges between drug nodes; wherein the characteristics of each drug node include at least the type of drug atom of the drug node, and the characteristics of each drug edge include at least the type of bond between the drug atoms;

[0040] Aggregating and updating the features of each drug node and the corresponding adjacent drug nodes through the graph isomorphism network layer of the target model to obtain an updated drug node;

[0041] A plurality of features corresponding to a plurality of drug nodes are globally pooled to obtain a drug feature vector of the drug molecule graph.

[0042] In some embodiments, the drug response prediction device further comprises a training module for:

[0043] Obtaining a sample training set, and performing gene sequencing on multiple sample target genes in each sample cell in the sample training set to obtain a sample multi-level omics feature of the sample cell;

[0044] Obtaining a sample gene expression vector of the sample cell, determining multiple sample target genes in the sample gene expression vector as multiple sample graph nodes, and using multiple attribute associations between any two sample target genes as multiple sample interaction edges to generate a sample multi-edge graph; wherein the sample multi-level omics feature corresponding to each sample target gene association is used as the sample node feature of the corresponding sample graph node;

[0045] Through multiple graph neural network layers in a preset model, in each graph neural network layer, for each sample graph node in the sample multi-edge graph, multiple message transmissions are performed through the multiple sample interaction edges connected to the adjacent sample graph nodes to obtain adjacent sample node features from the adjacent sample graph nodes, and the sample node features in each sample graph node are updated according to the adjacent sample node features, and feature mapping is performed on the updated sample multi-edge graph to obtain the sample first cell feature vector;

[0046] Obtaining a sample drug feature vector of the sample drug compound corresponding to the sample cell through the graph isomorphism network layer of the preset model, and fusing the sample first cell feature vector and the sample drug feature vector to obtain a sample first fusion vector;

[0047] Predicting the first fusion vector of the sample through the fully connected layer of the preset model to obtain a sample predicted drug response result of the sample drug to each sample cell;

[0048] A drug response label of the sample drug to each sample cell is obtained, and the preset model is trained by minimizing the difference based on the sample predicted drug response result and the drug response label to obtain a target model.

[0049] To achieve the above-mentioned purpose, the third aspect of the embodiments of the present application proposes a computer device, which includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the drug response prediction method described in any one of the embodiments of the first aspect of the present application.

[0050] To achieve the above-mentioned purpose, the fourth aspect of the embodiments of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the drug response prediction method described in any one of the embodiments of the first aspect of the present application.

[0051] In an embodiment of the present application, a plurality of target genes in a target cell are sequenced to obtain a multi-level omics feature of the target cell; a gene expression vector of the target cell is obtained, and the plurality of target genes in the gene expression vector are determined as a plurality of graph nodes, and the plurality of attribute association relationships between any two target genes are used as a plurality of interaction edges to generate a multi-edge graph; wherein the multi-level omics feature corresponding to each target gene association is used as the node feature of the corresponding graph node; through a plurality of graph neural network layers in a preset target model, in each graph neural network layer, for each graph node in the multi-edge graph, multiple message transmissions are performed through a plurality of interaction edges connected to adjacent graph nodes to obtain adjacent node features from adjacent graph nodes, and the node features in each graph node are updated according to the adjacent node features, and feature mapping is performed on the updated multi-edge graph to obtain a first cell feature vector; a drug feature vector of a drug compound corresponding to the target cell is obtained through a graph isomorphism network layer of the target model, and the first cell feature vector and the drug feature vector are fused to obtain a first fusion vector; the first fusion vector is predicted through a fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell. In this way, it is possible to construct a multi-edge graph by integrating multi-level omics data, so that the multi-edge graph can accurately reflect the biological state of the cell, and the multiple attribute associations between the target genes as multiple interaction edges can make the multi-edge graph more comprehensively reflect the multiple interactions between the target genes in the target cells, thereby improving the accuracy of the prediction. After the multi-edge graph is constructed, the complex relationships between the graph nodes and the global structure of the network can be captured through multiple message passing, thereby providing more accurate input features for subsequent prediction tasks, and thus obtaining more accurate prediction results. In summary, the present application can improve the accuracy of drug response prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the architecture of the drug response prediction system provided in the embodiment of the present application;

[0053] Figure 2 is a flow chart of the drug response prediction method provided in an embodiment of the present application;

[0054] Figure 3 This is an overall flow chart of the drug response prediction method provided in the embodiments of the present application;

[0055] Figure 4 Schematic diagram of the functional modules of the drug response prediction device provided in an embodiment of the present application;

[0056] Figure 5 This is a schematic diagram of the hardware structure of the computer device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0058] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0060] Drug response prediction refers to the use of computational models and biological data to predict the response of individuals or cells to specific drugs, in the hope of understanding the factors that may affect the efficacy and safety of drugs, thereby predicting the biological response of individuals or cell lines after using a certain drug, such as drug efficacy, toxicity, etc., thereby improving the accuracy and effectiveness of treatment.

[0061] Related technologies typically use single-modality omics data or simply splice features to predict drug response. For example, using a single omics dataset—genomic, transcriptomic, or proteomic—to predict drug responses in specific cell lines or organisms reduces the complexity and improves data integration efficiency. However, while single-omics data prediction methods simplify the data analysis process to a certain extent, they also have significant limitations. First, the use of single-modality omics data limits the generalization ability of the model, as the model often struggles to apply pre-trained information to unseen data modalities, resulting in limited predictive accuracy. Furthermore, when integrating multimodal omics data, related technologies often employ simple feature-level splicing strategies. This strategy fails to fully explore and utilize the deep interactions and intrinsic connections between different data modalities, overlooks the complex interactions between genes, and effectively integrates multi-omics data, ultimately limiting the accuracy of drug response prediction. Therefore, to overcome these issues, new methods and technologies are needed to more effectively integrate multi-omics data and improve the accuracy and reliability of drug response prediction.

[0062] Based on this, the embodiments of the present application provide a drug response prediction method, apparatus, computer device, and readable storage medium, which can improve the accuracy of drug response prediction.

[0063] The drug response prediction method, apparatus, computer device, and readable storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the drug response prediction system in the embodiments of the present application is described.

[0064] Please refer to Figure 1 In some embodiments, an embodiment of the present application provides a drug response prediction system, including a terminal 11 and a server 12 .

[0065] For example, terminal 11 can be any device with computing power and a network connection, such as a personal computer, laptop, tablet, smartphone, or dedicated bioinformatics workstation. Terminal 11 can be responsible for collecting and preprocessing multi-omics data, including information such as gene expression, gene mutation status, copy number variation, and methylation status. The preprocessed data is then uploaded to server 12 for further analysis and processing. Terminal 11 can also provide a user interface, allowing users to input data, set parameters, initiate prediction tasks, and view response prediction results.

[0066] Server 12 can be a high-performance computing server or cloud computing platform, which possesses powerful computing resources and storage capabilities and can handle a large number of complex data analysis tasks. Server 12 receives data from terminal 11 and uses advanced algorithms to perform in-depth data analysis, enabling effective integration of multi-omics data and drug response prediction. The prediction results are then returned to terminal 11 for user review and analysis. Server 12 can also build and train pre-defined models for drug response prediction, iterating and optimizing model parameters to improve prediction accuracy.

[0067] The drug response prediction method in the embodiments of the present application can be illustrated by the following examples.

[0068] It should be noted that in each specific embodiment of the present application, when it comes to the need to perform relevant processing based on data related to user identity or characteristics such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first. Moreover, the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or consent, the necessary user-related data for the normal operation of the embodiment of the present application will be obtained.

[0069] In the embodiment of the present application, the invention will be described from the perspective of a drug response prediction device, which can be integrated into a computer device. Figure 2 , Figure 2 This is a flowchart of the steps of the drug response prediction method provided in an embodiment of the present application. In this embodiment of the present application, the drug response prediction device is specifically integrated into a terminal or server as an example. When the processor on the terminal or server executes the program instructions corresponding to the drug response prediction method, the specific process is as follows:

[0070] Step 101: perform gene sequencing on multiple target genes in target cells to obtain multi-level genomic features of the target cells.

[0071] In some embodiments, in order to comprehensively and accurately reflect the biological state of cells, gene sequencing can be performed on multiple target genes in target cells to obtain multi-level omics characteristics of target cells, so as to capture the complex interactions between genes and achieve more reliable drug response prediction.

[0072] Target cells are specific cells being studied for drug response prediction. They can come from different tissues or organs, such as tumor cells, normal cells, and immune cells. The choice of target cells depends on the purpose and context of the study. For example, if studying the effects of an anticancer drug, the target cells might be a corresponding tumor cell line.

[0073] Among them, the target gene can be all genes in the target cell. The target gene is usually closely related to the function of the cell, the occurrence and development of the disease, or the response of the drug. By sequencing and analyzing the target gene, important information about the cell state can be obtained. In some embodiments, the target gene can also be a specific gene selected from the target cell, such as a gene associated with a certain disease, a gene related to drug metabolism, or a gene that plays a key role in a specific biological pathway, thereby achieving targeted research on drug response prediction.

[0074] Multi-level omics features can be obtained through different omics technologies (such as genomics, transcriptomics, epigenetics, etc.) to reflect information on target cells at multiple biological levels. Multi-level omics features include but are not limited to gene expression features, copy number variation features, gene mutation status features, and methylation status features.

[0075] For example, a specific cell line or sample can be selected according to the research purpose. For example, if the research is about lung cancer drug response, a lung cancer cell line (such as A549 cells) can be selected. Furthermore, high-throughput sequencing technology (such as Illumina sequencing platform) can be used to sequence the target gene in the target cell to obtain the sequence information of the gene, including gene expression level, copy number variation, gene mutation status and methylation status, etc.

[0076] By sequencing multiple target genes in target cells and obtaining multi-level omics features, we can comprehensively and accurately reflect the biological state of cells, capture the complex interactions between genes, and improve the generalization ability, robustness and accuracy of the prediction model, thereby achieving more reliable drug response prediction.

[0077] In some embodiments, multi-level omics features may include gene expression features, copy number variation features, gene mutation status features, and methylation status features. In order to comprehensively and systematically obtain the multi-level omics features of target cells, genome sequencing of target cells may be performed to capture rich information of each gene at multiple biological levels, thereby achieving reliable drug response prediction. For example, step 101 may include:

[0078] (101.1) Obtaining a gene expression vector corresponding to the target cell using a genome sequencer, and using the gene expression level of each target gene in the gene expression vector as a gene expression signature of the corresponding target gene;

[0079] (101.2) Sequencing the genome of the target cells using multiple genome sequencing methods to obtain the copy number variation status, gene mutation status, and methylation status of each target gene;

[0080] (101.3) converting the copy number variation state into a binary copy matrix, and determining the copy number variation characteristics of each target gene based on the binary copy matrix;

[0081] (101.4) converting the gene mutation state into a binary mutation matrix, and determining the gene mutation state characteristics of each target gene based on the binary mutation matrix;

[0082] (101.5) The methylation status is converted into a binary methylation matrix, and the methylation status characteristics of each target gene are determined based on the binary methylation matrix.

[0083] Among them, a genome sequencer can be a high-throughput instrument for determining deoxyribonucleic acid (DNA) sequences, which can quickly and accurately obtain genome sequence information through various sequencing technologies (such as Illumina sequencing, PacBio sequencing, Nanopore sequencing, etc.).

[0084] The gene expression vector can be a numerical vector representing the expression level of each target gene in the target cell. Each element in the gene expression vector corresponds to a target gene, and each element represents the gene expression level of the corresponding target gene in the cell. The gene expression vector can be obtained using ribonucleic acid (RNA) sequencing technology.

[0085] Among them, gene expression can refer to the expression level of the target gene in the target cell under specific conditions, which can be absolute expression (such as counts per million mapped reads) or relative expression (such as normalized number of fragments per million or number of transcripts per million transcripts). Gene expression reflects the activity and functional status of the target gene in the target cell.

[0086] The gene expression feature may be the gene expression level, or a feature extracted from the gene expression level that can describe the gene expression level.

[0087] Copy number variation (CNV) refers to changes in the number of gene copies in a specific region of the genome. CNV can manifest as an increase or decrease in the number of copies of a target gene. CNV may affect the expression and function of the target gene, thereby affecting the biological behavior of the target cell.

[0088] The gene mutation status can refer to base changes in the target gene's gene sequence. Gene mutations can include single nucleotide polymorphisms (SNPs), insertions, or deletions. Gene mutations can affect the function of the target gene, leading to changes in protein structure and function.

[0089] The methylation status can be the methylation modification on DNA. DNA methylation is an epigenetic modification that typically occurs at the 5' carbon position of cytosine, forming 5-methylcytosine. Methylation status can affect the expression of target genes, and different methylation patterns may turn gene expression on or off.

[0090] The binary copy matrix can be a binary matrix used to represent the copy number variation status of the target gene. Each element in the binary copy matrix has a value of 0 or 1, where 1 indicates that the target gene has a copy number variation, and 0 indicates that the target gene has no copy number variation.

[0091] The copy number variation feature may be a feature extracted through a binary copy matrix and capable of describing the gene copy number variation state.

[0092] The binary mutation matrix can be a binary matrix used to represent the mutation status of the target gene. Each element in the binary mutation matrix takes a value of 0 or 1, where 1 indicates the presence of a mutation in the target gene and 0 indicates the absence of a mutation. The binary mutation matrix format simplifies data processing and facilitates subsequent analysis.

[0093] The gene mutation status feature can be a feature extracted from a binary mutation matrix that can describe the gene mutation status. The gene mutation status feature can be used to help understand the impact of the gene mutation on the gene function of the target gene and the behavior of the target cell.

[0094] The binary methylation matrix can be a binary matrix used to represent the methylation status of the target gene. Each element in the binary methylation matrix has a value of 0 or 1, where 1 indicates that the target gene is methylated and 0 indicates that the target gene is not methylated.

[0095] The methylation status feature can be a feature extracted from a binary methylation matrix that can describe the methylation status of a gene. The methylation status feature can be used to help understand the impact of methylation status changes on the gene function of the target gene and the behavior of the target cell.

[0096] For example, the expression level of each target gene in the cell can be obtained by RNA sequencing (RNA-seq), and then the corresponding gene expression signature can be obtained.

[0097] Furthermore, gene copy number changes in corresponding regions of the genome can be obtained through gene chips or high-throughput sequencing technologies (such as whole-genome sequencing (WGS) and high-throughput sequencing), thereby obtaining the copy number variation status of each target gene. This is then converted into a binary copy matrix based on the gene copy number changes, where "1" indicates an increase or decrease in gene copy number and "0" indicates no change. This provides the copy number variation signature for each target gene.

[0098] Furthermore, whole exome sequencing (WES) or whole genome sequencing (WGS) can be used to identify base changes in the gene sequence and determine the gene mutation status. The resulting mutation status data is then converted into a binary mutation matrix, where "1" indicates a mutation and "0" indicates no mutation. This allows the gene mutation status characteristics of each target gene to be determined.

[0099] Furthermore, bisulfite sequencing or methylation microarrays can be used to obtain DNA methylation modification information and determine the methylation status. The methylation status data obtained from sequencing is then converted into a binary methylation matrix, where "1" indicates gene methylation and "0" indicates unmethylation. This allows the methylation status characteristics of each target gene to be determined.

[0100] Through the above methods, the multi-level omics characteristics of target cells can be comprehensively and systematically obtained to facilitate the subsequent construction of rich multi-edge graphs, thereby providing rich and accurate data support for drug response prediction.

[0101] Step 102: Obtain the gene expression vector of the target cell, and determine multiple target genes in the gene expression vector as multiple graph nodes, and use the multiple attribute association relationships between any two target genes as multiple interaction edges to generate a multi-edge graph; wherein, the multi-level omics features corresponding to each target gene association are used as node features of the corresponding graph node.

[0102] In some embodiments, in order to comprehensively and accurately reflect the complex interactions between target genes in target cells, multiple omics data can be integrated to construct a multi-edge graph corresponding to the target cells to better capture the complex characteristics of the gene network and provide important data support and scientific basis for drug response prediction.

[0103] The graph nodes may represent target genes in target cells, and each target gene serves as a graph node.

[0104] Among them, the interaction edge can be an edge between the nodes of the connection graph, which represents the multiple attribute association relationships between any two target genes, that is, the multiple interaction relationships between the target genes.

[0105] Among them, a multi-edge graph can be a special graph structure, in which each graph node can be connected by multiple interaction edges, and each interaction edge represents a different type of relationship.

[0106] The node feature may be the attribute information carried by each graph node, and the node feature of each graph node may include multi-level omics features.

[0107] Exemplarily, a multi-edge graph can be represented by G = (V, E), where each graph node V represents a target gene in a target cell, having characteristics such as gene expression characteristics, copy number variation characteristics, gene mutation status characteristics, methylation status characteristics, spatial omics, etc.; the interaction edge E represents the relationship between the target genes, including multiple interaction edges between any two graph nodes, including but not limited to the association relationship (Protein-Protein Interaction, PPI) between the proteins encoded by any two target genes, the gene pathway relationship between any two genes, and the similarity relationship (Pearson Correlation Coefficient, PCC) between any two genes, multiple attribute association relationships between any two target genes, etc. Each relationship represents an interaction edge, and there can be multiple interaction edges between any two graph nodes.

[0108] Furthermore, when acquiring the spatial omics characteristics of each target gene, a high-resolution microscope can be used to image sample tissue sections, obtaining images of individual cells or cell colonies. Image distortion caused by the optical system, such as nonlinear deformation and chromatic aberration, can be corrected, and noise interference in the image can be removed to obtain a high-quality preprocessed image. Subsequently, computer vision technology is used to extract the spatial location information and other relevant spatial features of each target gene from the preprocessed image, such as the gene's distribution within the cell and its positional relationship with other molecules. The extracted spatial features are then used as the spatial omics features of the corresponding graph node. By incorporating spatial omics information into the multi-edge graph, the graph structure can be made to more completely and realistically reflect the spatial distribution and interactions of target genes in the network, thereby improving the accuracy and reliability of drug response prediction.

[0109] Through these methods, we can comprehensively and systematically integrate the multi-level omics features of each target gene in the target cell. Furthermore, by using the multiple attribute associations between any two target genes as multiple interaction edges, the multi-edge graph can more accurately reflect the complex interactions between genes. This not only captures the biological behavior of genes at different levels, but also improves the accuracy of model predictions, providing richer and more accurate data support for drug response prediction.

[0110] In some embodiments, in order for the multi-edge graph to comprehensively and accurately reflect the complex interactions between genes, the association relationship between the proteins encoded by any two target genes, the gene pathway relationship between any two genes, and the similarity relationship between any two genes can be used as interaction edges between corresponding graph nodes to provide richer and more accurate data support for drug response prediction. For example, the multiple attribute association relationships between any two target genes can be obtained by the following methods:

[0111] (A.1) Obtain the predicted protein sequence of each target gene, and query the predicted protein sequences of any two target genes through the gene expression database to obtain the correlation relationship between the proteins produced by any two target genes;

[0112] (A.2) Querying the gene pathway relationship between any two target genes through a bioinformatics database to obtain the gene pathway relationship between any two target genes;

[0113] (A.3) Determine the Pearson correlation coefficient between any two target genes using the gene expression vectors corresponding to the target cells, and determine the similarity relationship between any two target genes based on the Pearson correlation coefficient.

[0114] The protein sequence can be predicted by bioinformatics methods to predict the amino acid sequence of the protein encoded by the target gene based on the DNA or mRNA sequence.

[0115] The gene expression database may be a database for storing and managing a large amount of gene expression data, and the gene expression database provides information on the expression levels of genes under different conditions.

[0116] Protein-protein associations can be interactions between two or more proteins within a cell through physical contact, chemical reactions, or other means. Protein-protein associations can be direct (e.g., protein-protein complexes) or indirect (e.g., interactions mediated by other molecules).

[0117] Among them, the bioinformatics database can be a database used to store and manage a large amount of bioinformatics data, including gene sequences, protein sequences, gene annotations, gene pathways, gene mutations, etc.

[0118] Among them, the gene pathway relationship can be a metabolic or signal transduction pathway in which the target gene participates in the biological process in the target cell.

[0119] Among them, the Pearson correlation coefficient can be used to measure the degree of linear correlation between two target genes. The value of the Pearson correlation coefficient ranges from -1 to 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation.

[0120] The similarity relationship can be the degree of similarity between any two target genes, and the similarity relationship can be quantified by the Pearson correlation coefficient or other similarity metrics (such as Euclidean distance, cosine similarity, etc.).

[0121] For example, when obtaining predicted protein sequences, the gene identifier can be used to query the Ensembl database (or other databases) for the corresponding gene information, download the mRNA sequences of the two target genes, and translate the mRNA sequences into predicted protein sequences using online tools (such as the Expasy Translate Tool) or local software (such as EMBOSS Transeq). Alternatively, a detection tool can be used to directly detect the target gene to obtain the predicted protein sequence. This application does not limit the specific acquisition method.

[0122] For example, the gene expression database can be a STRING database. By inputting the gene identification codes of any two target genes or the identification codes of the proteins encoded by the target genes into the STRING database, the association relationship between the proteins produced by any two target genes can be obtained. Alternatively, by inputting the gene identification code of one target gene or the identification code of the protein encoded by the target gene, other genes with which the target gene has a protein association relationship can be obtained. The target genes present in the multi-edge graph can be screened based on these other genes, and the association relationship between the proteins of the two target genes can be calibrated.

[0123] For example, the bioinformatics database may be the KEGG database or the Reactome database. Taking the KEGG database as an example, the gene names of any one or two target genes may be input to determine whether the two target genes are in the same pathway. For example, TP53 and BRCA1 are both involved in the cell cycle regulation pathway.

[0124] For example, the Pearson correlation coefficient between any two target genes can be directly calculated using the gene expression vectors corresponding to the target cells. The method for obtaining the gene expression vectors has been discussed above and will not be repeated here. The similarity relationship between any two target genes can be determined based on the Pearson correlation coefficient. For example, if the Pearson correlation coefficient is close to 1, it indicates that the two target genes are highly positively correlated in expression level. If the Pearson correlation coefficient is close to -1, it indicates that the two target genes are highly negatively correlated in expression level. If the Pearson correlation coefficient is close to 0, it indicates that there is no significant correlation in expression level between the two target genes. The specific correlation can be further determined based on the set threshold.

[0125] Through the above methods, multiple attribute associations between any two target genes can be obtained and integrated from multiple perspectives, so as to more comprehensively understand the complex interactions between target genes, improve the accuracy of target model prediction, and thus provide richer data support for drug response prediction.

[0126] Step 103, through multiple graph neural network layers in the preset target model, in each graph neural network layer, for each graph node in the multi-edge graph, multiple message transmissions are performed through multiple interactive edges connected to adjacent graph nodes to obtain adjacent node features from adjacent graph nodes, and the node features in each graph node are updated according to the adjacent node features, and feature mapping is performed on the updated multi-edge graph to obtain a first cell feature vector.

[0127] In some embodiments, in order to enable the target model to extract useful information from complex multi-omics data, multiple message aggregations can be performed on the multi-edge graph to comprehensively and accurately capture the complex interactions between genes, thereby improving the accuracy of drug response prediction.

[0128] The target model can be an overall machine learning model for drug response prediction. The target model can include multiple components, such as a graph neural network (GNN) or a graph isomorphism network.

[0129] The graph neural network layer can be the basic building block in the target model for feature propagation and aggregation. Each graph neural network layer uses a message passing mechanism to enable each graph node to update its own features based on the features of its neighboring nodes.

[0130] The adjacent node feature may be a feature of a neighboring node that is directly connected to the current node through at least one interaction edge.

[0131] The first cell feature vector may be a feature representation of the target cell extracted from the multi-edge graph after being processed by the graph neural network.

[0132] Exemplarily, each graph node in a multi-edge graph corresponds to a node feature. For each graph node in the multi-edge graph, the graph neural network can perform multiple message transmissions through the interaction edges between each graph node and the adjacent graph nodes, and each message transmission involves collecting information from the adjacent graph nodes.

[0133] Specifically, in each graph neural network layer, features are aggregated based on the multiple interaction edges between graph nodes and adjacent graph nodes. For example, the first interaction edge formed by the association between the proteins encoded by any two target genes is used for feature aggregation, the second interaction edge formed by the gene pathway relationship between any two genes is used for feature aggregation, and the third interaction edge formed by the similarity relationship between any two genes is used for feature aggregation, and so on.

[0134] Furthermore, after aggregating the adjacent node features of all adjacent graph nodes, each graph node will update its own features through the node-level gated recurrent unit (nGRU), so that the target model can more flexibly process the contextual information of different graph nodes.

[0135] In some embodiments, multiple graph node updates may be performed on the multi-edge graph in each graph neural network layer, for example, 3 times, 8 times, etc., to better capture the complex relationships between graph nodes.

[0136] Furthermore, the target model can include multiple graph neural network layers, and the above process can be performed in multiple graph neural network layers. After the node features of all graph nodes in each graph neural network layer are updated, the multi-edge graph in the current graph neural network layer can be updated through the graph-level gated recurrent unit (gGRU) to obtain an updated multi-edge graph, and the updated multi-edge graph is passed to the next graph neural network layer for updating to ensure that the low-level node features (close to the input) are retained to the high level, and to ensure that each layer gradually learns more complex graph patterns, from local gene interactions to global cell line behaviors. The update process has been introduced above and will not be repeated here.

[0137] Furthermore, after multiple iterations, each graph node corresponds to a final updated node feature. At this time, a readout function (such as global average pooling or maximum pooling) can be used to aggregate the node features of all graph nodes into a feature vector of fixed dimension to obtain the first cell feature vector.

[0138] Through the message-passing mechanism of a multi-layer graph neural network, each graph node can acquire information from its neighboring nodes and gradually update its own node features. This allows the target model to better capture the complex relationships between nodes in the graph, improving the richness and accuracy of feature representation. The resulting first cell feature vector integrates the feature information of all graph nodes, better reflecting the multi-omics characteristics of the cell line, thereby improving the accuracy of drug response prediction.

[0139] In some embodiments, to effectively integrate multi-omics data and capture the complex relationships between graph nodes in a multi-edge graph, iterative message passing and feature updates can be performed through a multi-layer graph neural network. Node-level and graph-level gated recurrent units are used to finely integrate local and global information, ultimately mapping the features of the multi-edge graph to a first cell vector representing the entire cell line to improve the accuracy of drug response prediction. For example, step 103 may include:

[0140] (103.1) For each graph node in the multi-edge graph updated by the previous graph neural network layer, multiple message transmissions are performed through multiple interaction edges connected to adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes through the current graph neural network layer in the preset target model;

[0141] (103.2) Using the node-level gated recurrent unit of the current graph neural network layer, the node features in each graph node are updated according to the features of the adjacent nodes to obtain the updated node features of each graph node;

[0142] (103.3) After the node features of multiple graph nodes in the multi-edge graph are updated, the global node features of the multi-edge graph are updated through the graph-level gated recurrent unit of the current graph neural network layer, and the updated multi-edge graph is passed to the next graph neural network layer;

[0143] (103.4) Repeatedly passing through the current graph neural network layer in the preset target model, for each graph node in the updated multi-edge graph updated by the previous graph neural network layer, multiple message transmissions are performed through multiple interaction edges connected to adjacent graph nodes to obtain adjacent node features from adjacent graph nodes; updating the node features in each graph node according to the adjacent node features through the node-level gated recurrent unit of the current graph neural network layer to obtain updated node features of each graph node; after the node features of multiple graph nodes in the updated multi-edge graph are updated, updating the global node features of the multi-edge graph through the graph-level gated recurrent unit of the current graph neural network layer until the next graph neural network layer is the last layer of the multiple graph neural network layers to obtain an updated multi-edge graph;

[0144] (103.5) Perform feature mapping on the updated multi-edge map to obtain a first cell vector.

[0145] The node-level gated recurrent unit (GRU) is a recurrent neural network structure used in graph neural networks, designed to process node feature updates for each graph node in a multi-edge graph. The GRU combines a gating mechanism with a recurrent structure, enabling each graph node to selectively update its node features upon receiving information from neighboring graph nodes.

[0146] Among them, the graph-level gated recurrent unit can be a global feature update mechanism for graph neural networks, aiming to integrate the node features of all graph nodes in the entire multi-edge graph.

[0147] For example, if there is a multi-edge graph, where graph nodes A, B, and C represent target gene 1, target gene 2, and target gene 3 respectively. The node feature of graph node A is The node characteristics of graph node B are The initial features of graph node C are The adjacent graph nodes of graph node A are graph nodes B and C. Then, graph node A can obtain its current node features from graph nodes B and C. and Graph node A transmits messages through multiple interaction edges connected to graph nodes B and C (including interaction edge 1 consisting of the association relationship between proteins encoded by any two target genes, interaction edge 2 consisting of the gene pathway relationship between any two genes, and interaction edge 3 consisting of the similarity relationship between any two genes, etc.), and obtains the adjacent node features m of graph nodes B and C.AB and m AC .

[0148] Furthermore, the message m received by the graph node A through the interaction edge 1 can be processed through the node-level gated recurrent unit of the current graph neural network layer. AB and m AC Perform aggregation to generate the feature representation of the new graph node A Aggregation can be summed or averaged, such as or The specific aggregation method is determined by the setting of the node-level gated recurrent unit. Similarly, the messages received through interaction edge 2 and interaction edge 3 are also aggregated in the above manner, which will not be detailed here.

[0149] In some embodiments, the node features of the current graph node can be updated through the node-level gated loop unit after the current graph node receives messages transmitted by all interaction edges. Alternatively, the node features of the current graph node can be updated through the node-level gated loop unit after each corresponding message is received through an interaction edge. For example, there are 3 (or other numbers, such as 2 or 5) interaction edges, which is updated 3 times. The number of updates is the same as the number of interaction edges.

[0150] Furthermore, after the node features of all graph nodes (A, B, C) of the multi-edge graph are updated, the global node features of the entire multi-edge graph can be updated using a graph-level gated recurrent unit to ensure that low-level features (close to the input) can be retained to high levels to avoid information loss.

[0151] Furthermore, the updated multi-edge graph of the current graph neural network layer can be passed to the next graph neural network layer to continue feature aggregation and updating. For example, in the second-layer graph neural network, the above message passing and feature updating process is repeated based on the updated multi-edge graph of the first-layer graph neural network to further update the feature representations of nodes A, B, and C. For example, in the third-layer graph neural network, the above message passing and feature updating process is repeated based on the updated multi-edge graph of the second-layer graph neural network to further update the feature representations of nodes A, B, and C, and so on. This process is not further described here.

[0152] For example, after all graph neural network layers are processed, a readout function (such as global average pooling or maximum pooling) can be used to aggregate the features of all nodes into a fixed-dimensional feature vector. For example, the first cell feature vector can be Indicates that the fixed dimension is 128, and so on. The dimension can be set according to the actual situation.

[0153] In some embodiments, the above introduction to the scheme is only used as an example. In actual situations, the number of layers of the neural network, the number of graph nodes in the multi-edge graph, and the relationship between the graph nodes may be different, but the processing method is the same as above. Please refer to the above for details.

[0154] By integrating node-level and graph-level gated recurrent units (GRUs) into a graph neural network, the target model's ability to process complex graph-structured data is effectively enhanced. The GRUs and the introduction of a gating mechanism enable each graph node to selectively update its own features based on information from adjacent graph nodes, enhancing the accuracy and relevance of node features. The GRUs, on the other hand, are responsible for integrating and updating global features across the entire graph structure, ensuring the effective propagation of information within multi-edge graphs and a smooth transition from low-level features to higher-level features, thus avoiding information loss. This mechanism not only improves the model's expressive power but also enhances the target model's ability to understand and learn complex relationships within multi-edge graphs. Ultimately, through iterative processing and feature aggregation within a multi-layer graph neural network, the resulting fixed-dimensional feature vectors more comprehensively reflect the intrinsic properties of the graph data, providing high-quality feature representations for downstream tasks such as drug response prediction.

[0155] Step 104 : Obtain the drug feature vector of the drug compound corresponding to the target cell through the graph isomorphism network layer of the target model, and fuse the first cell feature vector and the drug feature vector to obtain a first fusion vector.

[0156] In some embodiments, in order to comprehensively evaluate the interaction between the drug and the target cell, thereby predicting the effect of the drug or the response of the target cell to the drug, the drug feature vector of the drug compound corresponding to the target cell can be obtained through the graph isomorphism network layer of the target model, and the drug feature vector can be fused with the cell feature vector to achieve efficient and accurate drug-cell interaction prediction.

[0157] Among them, the graph isomorphism network layer can be a neural network layer in the target model for capturing structural information in the multi-edge graph.

[0158] Drug compounds can be chemical substances or biomolecules used to treat diseases. Drug compounds can be small molecules, proteins, nucleic acids, etc., and exert therapeutic effects by interacting with targets in the body (such as receptors, enzymes, ion channels, etc.).

[0159] The drug feature vector may be a fixed-dimensional feature representation generated by processing the drug molecule graph through a graph isomorphism network layer.

[0160] The first fusion vector may be a feature vector generated by fusing the first cell feature vector and the drug feature vector. The first fusion vector integrates the multi-omics features of the target cell and the structural information of the drug molecule, and is used for subsequent drug response prediction tasks.

[0161] In some embodiments, a drug feature vector can be first extracted from the drug compound using a graph isomorphism network (GIN), and then fused with the cell feature vector to obtain a comprehensive fused vector. For example, the molecular structure of the drug compound can be represented as a drug molecule graph, where drug nodes represent drug atoms and drug edges represent chemical bonds between drug atoms.

[0162] Furthermore, the drug molecule graph can be input into a multi-layer graph isomorphism network layer for processing. In each graph isomorphism network layer, each drug node can update its own features based on the features of its adjacent drug nodes. After that, each graph isomorphism network layer can further refine and integrate the features of drug molecules. After the updated drug molecule graph is output by the last layer, a global pooling operation (such as average pooling or maximum pooling) can be performed on the drug molecule graph to obtain a global drug feature vector from the high-dimensional features of the drug molecule graph, thereby capturing the key information of the entire drug molecule graph.

[0163] Furthermore, the first cell feature vector and the drug feature vector obtained above can be spliced, that is, the first cell feature vector and the drug feature vector are connected end-to-end to form a comprehensive first fusion vector.

[0164] Furthermore, the fused first fusion vector can be input into one or more fully connected layers of the target model to further adjust the weights of the features to optimize the performance of the prediction task.

[0165] Through the above methods, the information of drug compounds and target cells can be effectively integrated to provide a more comprehensive and accurate feature representation for drug response prediction.

[0166] In some embodiments, to enable the target model to extract useful information from complex multi-omics data and drug molecular structures, the drug molecular graph of the drug compound may be aggregated and updated multiple times, thereby improving the accuracy of drug response prediction. For example, step 104 of "obtaining the drug feature vector of the drug compound corresponding to the target cell through the graph isomorphism network layer of the target model" may include:

[0167] (104.1) Constructing a drug molecular graph for a drug compound using drug atoms of a drug compound corresponding to a target cell as drug nodes and bonds between drug atoms as drug edges between drug nodes; wherein the characteristics of each drug node include at least the type of drug atom in the drug node, and the characteristics of each drug edge include at least the type of bond between drug atoms;

[0168] (104.2) Aggregate and update the features of each drug node and its corresponding adjacent drug nodes through the graph isomorphism network layer of the target model to obtain an updated drug node;

[0169] (104.3) Globally pool multiple features corresponding to multiple drug nodes to obtain the drug feature vector of the drug molecule graph.

[0170] The drug atom may be an atom in a drug molecule, such as carbon (C), nitrogen (N), oxygen (O), etc.

[0171] The bond between drug atoms can be a chemical bond connecting two drug atoms, such as a single bond, a double bond, an aromatic bond, etc.

[0172] Among them, the drug atom type can be the atoms in the drug molecule, which constitute the basic units of the drug molecule. The drug atoms include but are not limited to carbon (C), hydrogen (H), nitrogen (N), oxygen (O), sulfur (S), phosphorus (P), etc.

[0173] The type of bond may be a chemical bond connecting two drug atoms in a drug molecule.

[0174] In some embodiments, the graph isomorphism network has multiple layers, and at each layer, each drug node updates its own characteristics based on the characteristics of adjacent drug nodes. For example, a carbon atom node will consider the information of its connected oxygen and hydrogen atoms.

[0175] In some implementations, at each graph isomorphic network layer, updates can be performed in the following ways:

[0176]

[0177] in, is the representation of drug node v at the kth layer, N(v) is the set of drug nodes adjacent to drug node v, and ε is a learnable hyperparameter. After each graph isomorphism network layer processes the features of all drug nodes in the drug molecule graph, an updated drug molecule graph is obtained. The updated drug molecule graph is then repeatedly updated and passed on to the last layer of the graph isomorphism network.

[0178] Furthermore, after processing through the multi-layer graph isomorphism network layer, each drug node has an updated feature vector. To obtain a global feature vector for the entire drug molecule graph, the features of all drug nodes can be globally pooled. For example, using average pooling, the average of the feature vectors of all drug nodes is calculated to obtain a global feature vector representing the entire drug molecule graph.

[0179] By constructing a drug molecule graph and applying a graph isomorphism network for multi-level feature aggregation and updating, the complex characteristics of individual atoms in drug molecules and the chemical bonds between them can be accurately captured and represented. This not only enriches the representation of the drug molecule graph, but also effectively integrates the structural information of the entire molecule through a global pooling operation, generating a global feature vector that can fully represent the characteristics of the drug molecule. This, in turn, accurately represents the structural information of the drug molecule, enhancing the discriminative ability of the target model and the accuracy of prediction.

[0180] Step 105 , predicting the first fusion vector through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

[0181] Among them, the predicted drug response results can be the biological effects of drug compounds on target cells, expressed in quantitative or qualitative form. The predicted drug response results can be IC50 (half maximal inhibitory concentration), EC50 (half effective concentration), cell survival rate, proliferation inhibition rate, apoptosis rate, etc.

[0182] In some embodiments, in order to achieve accurate prediction of drug response, the first fusion vector can be predicted through the fully connected layer of the target model to effectively integrate these multimodal data, ensure that the target model can fully understand the drug-cell interaction, and improve the accuracy of the prediction.

[0183] Exemplarily, the fully connected layer can be composed of one or more linear transformations and activation functions. After the first fusion vector is input to the fully connected layer of the target model, the fully connected layer can output the predicted drug response result, such as the IC50 predicted value, through the linear transformation and activation function.

[0184] In an embodiment of the present application, a plurality of target genes in a target cell are sequenced to obtain a multi-level omics feature of the target cell; a gene expression vector of the target cell is obtained, and the plurality of target genes in the gene expression vector are determined as a plurality of graph nodes, and the plurality of attribute association relationships between any two target genes are used as a plurality of interaction edges to generate a multi-edge graph; wherein the multi-level omics feature corresponding to each target gene association is used as the node feature of the corresponding graph node; through a plurality of graph neural network layers in a preset target model, in each graph neural network layer, for each graph node in the multi-edge graph, multiple message transmissions are performed through a plurality of interaction edges connected to adjacent graph nodes to obtain adjacent node features from adjacent graph nodes, and the node features in each graph node are updated according to the adjacent node features, and feature mapping is performed on the updated multi-edge graph to obtain a first cell feature vector; a drug feature vector of a drug compound corresponding to the target cell is obtained through a graph isomorphism network layer of the target model, and the first cell feature vector and the drug feature vector are fused to obtain a first fusion vector; the first fusion vector is predicted through a fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell. In this way, it is possible to construct a multi-edge graph by integrating multi-level omics data, so that the multi-edge graph can accurately reflect the biological state of the cell, and the multiple attribute associations between the target genes as multiple interaction edges can make the multi-edge graph more comprehensively reflect the multiple interactions between the target genes in the target cells, thereby improving the accuracy of the prediction. After the multi-edge graph is constructed, the complex relationships between the graph nodes and the global structure of the network can be captured through multiple message passing, thereby providing more accurate input features for subsequent prediction tasks, and thus obtaining more accurate prediction results. In summary, the present application can improve the accuracy of drug response prediction.

[0185] In some embodiments, to improve the representational capability, robustness, and prediction accuracy of the target model, the gene expression matrix corresponding to the target cells can be input into the target encoder of the target model for feature extraction to obtain a second cell feature vector, thereby fully utilizing multimodal data to supplement the first fusion vector and optimize the effect of the drug response prediction task. For example, before step 105, the following steps may also be included:

[0186] (B.1) extracting high-dimensional features from the gene expression vector using a target encoder of the target model to determine a second cell feature vector, wherein the second cell feature vector is used to characterize the interaction relationship between each target gene and the corresponding gene expression pattern;

[0187] (B.2) fusing the second cell feature vector with the first fusion vector to obtain a second fusion vector;

[0188] The first fusion vector is predicted through the fully connected layer of the target model to obtain the predicted drug response results of the drug compound to the target cell, including:

[0189] The second fusion vector is predicted through the fully connected layer of the target model to obtain the predicted drug response results of the drug compound to the target cells.

[0190] Among them, the target encoder can be a scBERT encoder, which can extract high-dimensional features from the gene expression vectors corresponding to the target cells to better characterize the gene expression patterns and their interaction relationships.

[0191] The second cell feature vector may be a feature vector obtained by performing high-dimensional feature extraction on the gene expression vector through the target encoder, and is used to characterize the interaction relationship between each target gene and the corresponding gene expression pattern.

[0192] Gene expression patterns can be the changes in the expression levels of a group of genes in target cells under specific conditions (e.g., different time points, different treatment conditions, different cell types, etc.). Gene expression patterns can reflect the state, function, and biological processes of target cells.

[0193] The second fusion vector may be a feature vector obtained by concatenating the second cell feature vector and the first fusion vector. In some embodiments, the second fusion vector may be used in a final drug response prediction task.

[0194] In some embodiments, the gene expression vector can be mapped to an embedding vector using the scBERT encoder, and high-dimensional features in the gene expression vector can be extracted through multiple encoder layers to capture long-range dependencies and complex interaction patterns in the gene expression data. Subsequently, the high-dimensional features can be reduced to a dimension consistent with the first fusion vector (e.g., 128 dimensions) using one or more linear transformations to obtain a second cell feature vector.

[0195] Furthermore, the second cell feature vector is concatenated with the first fusion vector to obtain a second fusion vector. The second cell feature vector captures the long-range dependencies and complex interaction patterns in the gene expression data and has higher abstractness and representativeness.

[0196] The second cell feature vector extracted by the target encoder can effectively deal with the noise and variation in gene expression data and improve the prediction performance of the target model.

[0197] In some embodiments, in order to enable the model to learn the inherent laws and complex relationships of predicting drug responses from the sample training set, thereby accurately predicting the drug response of unknown samples, a preset model can be trained to obtain a target model to ensure that the target model can capture subtle patterns and trends in the data, ultimately achieving effective prediction of drug-cell interactions. Exemplarily, the target model can be trained by the following methods:

[0198] (C.1) Obtain a sample training set and perform gene sequencing on multiple target genes in each sample cell in the sample training set to obtain a multi-level omics profile of the sample cells;

[0199] (C.2) Obtain the sample gene expression vector of the sample cell, and identify multiple sample target genes in the sample gene expression vector as multiple sample graph nodes. Use the multiple attribute associations between any two sample target genes as multiple sample interaction edges to generate a sample multi-edge graph; wherein, the sample multi-level omics feature corresponding to each sample target gene association is used as the sample node feature of the corresponding sample graph node;

[0200] (C.3) Through multiple graph neural network layers in the preset model, in each graph neural network layer, for each sample graph node in the sample multi-edge graph, multiple message transmissions are performed through multiple sample interaction edges connected to adjacent sample graph nodes to obtain adjacent sample node features from the adjacent sample graph nodes, and the sample node features in each sample graph node are updated according to the adjacent sample node features, and feature mapping is performed on the updated sample multi-edge graph to obtain the sample first cell feature vector;

[0201] (C.4) Obtaining the sample drug feature vector of the sample drug compound corresponding to the sample cell through the graph isomorphism network layer of the preset model, and fusing the sample first cell feature vector and the sample drug feature vector to obtain the sample first fusion vector;

[0202] (C.5) Predicting the first fusion vector of the sample through the fully connected layer of the preset model to obtain the sample predicted drug response result of the sample drug for each sample cell;

[0203] (C.6) Obtain the drug response label of the sample drug for each sample cell, and perform difference minimization training on the preset model based on the sample predicted drug response results and the drug response label to obtain the target model.

[0204] The sample training set may be a set of data sets used to train a preset model, which includes multiple sample cells, sample drugs, and corresponding drug response labels.

[0205] The preset model can be a pre-designed deep learning model used to learn feature representation and prediction rules from a sample training set. The preset model includes at least a graph neural network layer, a graph isomorphism network layer, a fully connected layer, and a target encoder.

[0206] The sample predicted drug response result may be a result obtained by predicting the characteristic vector of the sample cell using a preset model, indicating the response value (such as IC50 value) of the drug to each sample cell predicted by the preset model.

[0207] The drug response label can be the actual drug response value of the sample cell, which can be obtained through experiments and used to evaluate the predictive performance of the preset model.

[0208] In some embodiments, multiple cells from the same cell line can be used as sample cells to provide diverse training data, thereby improving the generalization ability of the model. Furthermore, different cells from the same cell line under different conditions (such as different disease states) can be used as sample cells to help the model capture this heterogeneity and improve the model's ability to recognize different cell states.

[0209] In some embodiments, the process of obtaining the sample predicted drug response result of the sample drug to each sample cell for each sample cell is the same as the processing process of obtaining the predicted drug response result of the drug compound to the target cell for each target cell as described above. There is a difference in the objects of processing. The embodiment of the present application does not elaborate on the processing process of the sample cells. For details, please refer to the aforementioned processing process of the target cells.

[0210] In some embodiments, when the model is applied, there is a need to extract high-dimensional features of the gene expression vector through the target encoder of the target model, determine the second cell feature vector, fuse the second cell feature vector with the first fusion vector to obtain a second fusion vector, and then predict the second fusion vector through the fully connected layer of the target model to obtain the predicted drug response results of the drug compound to the target cell. Then, when training the preset model, the target encoder of the preset model can be used to extract high-dimensional features of the sample gene expression vector, determine the sample second cell feature vector, fuse the sample second cell feature vector with the sample first fusion vector to obtain the sample second fusion vector, and then predict the sample second fusion vector through the fully connected layer of the preset model to obtain the sample predicted drug response results of the sample drug to each sample cell.

[0211] In some embodiments, the target loss can be calculated by predicting drug response results and drug response labels using samples, and then the preset model is trained to minimize the difference based on the target loss. That is, the parameters of the preset model are adjusted until a preset convergence condition is reached, thereby obtaining the target model. Furthermore, the preset convergence condition can be that the number of training times reaches a preset number, such as 100 or 500 times, or the target loss is less than a preset threshold, etc.

[0212] Exemplarily, the target loss can be mean square error loss, cross entropy loss, mean square logarithmic error loss, etc. The corresponding loss function can be selected according to the actual situation. The embodiment of the present application does not impose any specific restrictions on this.

[0213] Training the preset model with diverse training data can enable the model to recognize and adapt to the heterogeneity in different cell states, enhance the generalization ability of the model, and thus improve the recognition and prediction accuracy of different cell states.

[0214] Please refer to Figure 3 , for example, the following is combined Figure 3 This section introduces the overall process of this application.

[0215] For example, in obtaining a multi-edge map h of a target cell x (t-1) Afterwards, multiple graph neural network layers can be used to pass messages through multiple interactive edges for each graph node to update the node features corresponding to each graph node in the multi-edge graph. After the multi-edge graph is updated by multiple graph neural network layers, the updated multi-edge graph h can be obtained. x (t-1) ,Afterwards, feature mapping is performed on the updated multi-edge ,graph to obtain the first cell feature vector.

[0216] Furthermore, the drug feature vector of the drug compound can be obtained through the graph isomorphism network, and the first cell feature vector and the drug feature vector can be concatenated to obtain a first fusion vector.

[0217] Furthermore, through the target encoder (i.e., scBERT Encoder), high-dimensional features can be extracted from multiple gene expression matrices corresponding to multiple cells to obtain the second cell feature vector corresponding to each cell.

[0218] Furthermore, the first fusion vector can be concatenated with the second cell feature vector corresponding to the target cell to obtain a second fusion vector. By inputting the second fusion vector into the fully connected layer of the target model for prediction, the predicted drug response result of the drug compound on the target cell can be obtained.

[0219] See also Figure 4The present application also provides a drug response prediction device that can implement the above-mentioned drug response prediction method. The drug response prediction device includes:

[0220] A sequencing module 41 is used to sequence multiple target genes in target cells to obtain multi-level genomic features of the target cells;

[0221] A generation module 42 is configured to obtain a gene expression vector of a target cell, identify multiple target genes in the gene expression vector as multiple graph nodes, and use multiple attribute associations between any two target genes as multiple interaction edges to generate a multi-edge graph; wherein the multi-level omics features corresponding to each target gene association serve as node features of the corresponding graph node;

[0222] An updating module 43 is configured to, through multiple graph neural network layers in a preset target model, perform multiple message transmissions on each graph node in the multi-edge graph in each graph neural network layer through multiple interaction edges connected to adjacent graph nodes to obtain adjacent node features from adjacent graph nodes, update node features in each graph node according to the adjacent node features, and perform feature mapping on the updated multi-edge graph to obtain a first cell feature vector;

[0223] A fusion module 44 is configured to obtain a drug feature vector of the drug compound corresponding to the target cell through the graph isomorphism network layer of the target model, and fuse the first cell feature vector and the drug feature vector to obtain a first fusion vector;

[0224] The prediction module 45 is configured to predict the first fusion vector through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

[0225] By using a drug response prediction device to predict drug response and obtain predicted drug response results, medical professionals can choose the most effective drugs at an early stage, thereby greatly shortening the treatment time, reducing the consumption of medical resources, and improving the success rate of new drug research and development, which greatly promotes the development of precision medicine.

[0226] The specific implementation of the drug response prediction device is basically the same as the specific embodiment of the drug response prediction method described above, and will not be repeated here. Under the premise of meeting the requirements of the embodiment of this application, the drug response prediction device can also be equipped with other functional modules to implement the drug response prediction method in the above embodiment.

[0227] The present application also provides a computer device comprising a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described drug response prediction method. The computer device can be any intelligent terminal, including a tablet computer and an in-vehicle computer.

[0228] See also Figure 5 , Figure 5 The hardware structure of a computer device according to another embodiment is shown. The computer device includes:

[0229] The processor 51 may be implemented as a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0230] The memory 52 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 52 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program codes are stored in the memory 52 and are called by the processor 51 to execute the drug response prediction method of the embodiments of this application.

[0231] Input / output interface 53, used to implement information input and output;

[0232] Communication interface 54, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0233] bus 55 , which transmits information between the various components of the device (e.g., processor 51 , memory 52 , input / output interface 53 , and communication interface 54 );

[0234] The processor 51 , the memory 52 , the input / output interface 53 and the communication interface 54 are connected to each other in communication within the device via a bus 55 .

[0235] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned drug response prediction method is implemented.

[0236] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0237] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0238] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0239] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0240] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0241] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0242] It should be understood that in this application, "at least one (item)" and "several" refer to one or more, and "plurality" refers to two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0243] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of the above units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0244] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0245] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0246] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes multiple instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0247] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A drug response prediction method, characterized in that: The method comprises: Performing gene sequencing on multiple target genes in target cells to obtain multi-level omics characteristics of the target cells; Obtaining a gene expression vector of the target cell, and determining a plurality of target genes in the gene expression vector as a plurality of graph nodes, and using a plurality of attribute associations between any two target genes as a plurality of interaction edges to generate a multi-edge graph; wherein the multi-level omics features corresponding to each target gene association are used as node features of the corresponding graph nodes; Through a plurality of graph neural network layers in a preset target model, in each graph neural network layer, for each graph node in the multi-edge graph, multiple message transmissions are performed through the plurality of interaction edges connected to adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes, and node features in each graph node are updated according to the adjacent node features, and feature mapping is performed on the updated multi-edge graph to obtain a first cell feature vector; Obtaining a drug feature vector of the drug compound corresponding to the target cell through the graph isomorphic network layer of the target model, and fusing the first cell feature vector and the drug feature vector to obtain a first fusion vector; The first fusion vector is predicted through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

2. The drug response prediction method according to claim 1, characterized in that: Before predicting the first fusion vector through the fully connected layer of the target model to obtain the predicted drug response result of the drug compound to the target cell, the method further includes: Extracting high-dimensional features from the gene expression vector by using a target encoder of the target model to determine a second cell feature vector, wherein the second cell feature vector is used to characterize the interaction relationship between each target gene and the corresponding gene expression pattern; Fusing the second cell feature vector with the first fusion vector to obtain a second fusion vector; Then, predicting the first fusion vector through the fully connected layer of the target model to obtain the predicted drug response result of the drug compound to the target cell includes: The second fusion vector is predicted through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

3. The drug response prediction method according to claim 1, characterized in that: The multi-level omics features include gene expression features, copy number variation features, gene mutation status features and methylation status features. The gene sequencing of multiple target genes in the target cells to obtain the multi-level omics features of the target cells includes: Obtaining a gene expression vector corresponding to the target cell by a genome sequencer, and taking the gene expression amount of each target gene in the gene expression vector as the gene expression feature of the corresponding target gene; Sequencing the genome of the target cell by multiple genome sequencing methods to obtain the copy number variation status, gene mutation status and methylation status of each target gene; Converting the copy number variation state into a binary copy matrix, and determining the copy number variation characteristics of each target gene based on the binary copy matrix; Converting the gene mutation state into a binary mutation matrix, and determining the gene mutation state characteristics of each target gene based on the binary mutation matrix; The methylation status is converted into a binary methylation matrix, and the methylation status characteristics of each target gene are determined based on the binary methylation matrix.

4. The drug response prediction method according to claim 1, characterized in that: The multiple attribute association relationships include the association relationship between proteins encoded by any two target genes, the gene pathway relationship between any two genes, and the similarity relationship between any two genes. The multiple attribute association relationships between any two target genes are obtained by: Obtaining the predicted protein sequence of each target gene, and querying the predicted protein sequences of any two target genes through a gene expression database to obtain the correlation relationship between the proteins produced by the any two target genes; Querying the gene pathway relationship between the two arbitrary target genes through a biological information database to obtain the gene pathway relationship between the two arbitrary target genes; The Pearson correlation coefficient between the two arbitrary target genes is determined by the gene expression vector corresponding to the target cell, and the similarity relationship between the two arbitrary target genes is determined based on the Pearson correlation coefficient.

5. The drug response prediction method according to claim 1, characterized in that: The method includes: performing multiple message passing for each graph node in the multi-edge graph in each graph neural network layer through the multiple interactive edges connected to the adjacent graph nodes, obtaining adjacent node features from the adjacent graph nodes, updating the node features in each graph node according to the adjacent node features, and performing feature mapping on the updated multi-edge graph to obtain a first cell feature vector, including: Through the current graph neural network layer in the preset target model, for each graph node in the multi-edge graph updated by the previous graph neural network layer, multiple message transmissions are performed through the multiple interaction edges connected to the adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes; By using the node-level gated recurrent unit of the current graph neural network layer, the node features in each graph node are updated according to the adjacent node features to obtain the updated node features of each graph node; After the node features of the multiple graph nodes of the multi-edge graph are updated, the global node features of the multi-edge graph are updated through the graph-level gated recurrent unit of the current graph neural network layer, and the updated multi-edge graph is passed to the next graph neural network layer; Repeat the current graph neural network layer in the preset target model, and for each graph node in the updated multi-edge graph updated by the previous graph neural network layer, perform multiple message transmissions through the multiple interactive edges connected to the adjacent graph nodes to obtain adjacent node features from the adjacent graph nodes; update the node features in each graph node according to the adjacent node features through the node-level gated loop unit of the current graph neural network layer to obtain the updated node features of each graph node; after the node features of the multiple graph nodes in the updated multi-edge graph are updated, update the global node features of the multi-edge graph through the graph-level gated loop unit of the current graph neural network layer until the next graph neural network layer is the last layer of the multiple graph neural network layers to obtain an updated multi-edge graph; Perform feature mapping on the updated multi-edge graph to obtain a first cell vector.

6. The drug response prediction method according to claim 1, characterized in that: The step of obtaining the drug feature vector of the drug compound corresponding to the target cell through the graph isomorphic network layer of the target model includes: The drug atoms of the drug compound corresponding to the target cell are used as drug nodes, and the bonds between the drug atoms are used as drug edges between the drug nodes to construct a drug molecular graph of the drug compound; wherein the characteristics of each drug node at least include the type of drug atoms of the drug node, and the characteristics of each drug edge at least include the type of bonds between the drug atoms; Aggregating and updating the features of each drug node and the corresponding adjacent drug nodes through the graph isomorphic network layer of the target model to obtain an updated drug node; A plurality of features corresponding to a plurality of drug nodes are globally pooled to obtain a drug feature vector of the drug molecule graph.

7. The drug response prediction method according to claim 1, characterized in that: The target model is trained in the following way: Acquire a sample training set, and perform gene sequencing on multiple sample target genes in each sample cell in the sample training set to obtain a sample multi-level omics feature of the sample cell; Obtaining a sample gene expression vector of the sample cell, and determining a plurality of sample target genes in the sample gene expression vector as a plurality of sample graph nodes, and using a plurality of attribute association relationships between any two sample target genes as a plurality of sample interaction edges to generate a sample multi-edge graph; wherein the sample multi-level omics features corresponding to each sample target gene association are used as sample node features of the corresponding sample graph node; Through a plurality of graph neural network layers in a preset model, in each graph neural network layer, for each sample graph node in the sample multi-edge graph, multiple message transmissions are performed through the plurality of sample interaction edges connected to adjacent sample graph nodes to obtain adjacent sample node features from the adjacent sample graph nodes, and sample node features in each sample graph node are updated according to the adjacent sample node features, and feature mapping is performed on the updated sample multi-edge graph to obtain a sample first cell feature vector; Obtaining a sample drug feature vector of the sample drug compound corresponding to the sample cell through the graph isomorphism network layer of the preset model, and fusing the sample first cell feature vector and the sample drug feature vector to obtain a sample first fusion vector; Predicting the first fusion vector of the sample through the fully connected layer of the preset model to obtain a sample predicted drug response result of the sample drug to each sample cell; The drug response label of the sample drug to each sample cell is obtained, and the preset model is trained to minimize the difference based on the sample predicted drug response result and the drug response label to obtain a target model.

8. A drug response prediction device, characterized in that: The device comprises: A sequencing module is used to perform gene sequencing on multiple target genes in target cells to obtain multi-level omics features of the target cells; A generation module is used to obtain the gene expression vector of the target cell, and determine multiple target genes in the gene expression vector as multiple graph nodes, and use multiple attribute associations between any two target genes as multiple interaction edges to generate a multi-edge graph; wherein the multi-level omics features corresponding to each target gene association are used as node features of the corresponding graph node; An updating module, configured to perform multiple message transmissions for each graph node in the multi-edge graph in each graph neural network layer through the multiple interactive edges connected to the adjacent graph nodes, obtain adjacent node features from the adjacent graph nodes, update node features in each graph node according to the adjacent node features, and perform feature mapping on the updated multi-edge graph to obtain a first cell feature vector; A fusion module, used to obtain a drug feature vector of the drug compound corresponding to the target cell through the graph isomorphic network layer of the target model, and fuse the first cell feature vector with the drug feature vector to obtain a first fusion vector; A prediction module is used to predict the first fusion vector through the fully connected layer of the target model to obtain a predicted drug response result of the drug compound to the target cell.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the drug response prediction method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the drug response prediction method according to any one of claims 1 to 7 is implemented.

Images