Drug treatment effect prediction method, device, computing device and storage medium

Abstract

A method, apparatus, computing device, and storage medium for predicting drug treatment effects are provided. The method may include: obtaining input data; processing the input data using a drug treatment effect prediction model to obtain a predicted treatment effect. The model includes a first neural network, a heterogeneous graph neural network, and a second neural network. The heterogeneous graph neural network includes at least one drug node, at least one gene node, at least one protein node, and edges representing the relationships between the drug node, the gene node, and the protein node. The gene node and the protein node are configured with corresponding feature vectors. Processing the input data includes: extracting a drug feature vector; feeding the drug feature vector to the drug node to update the corresponding feature vector; combining the updated corresponding feature vectors of the drug node, the gene node, and the protein node; and generating output data from the combined feature vector.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and in particular to a method, apparatus, computing device, and storage medium for predicting drug treatment effects. Background Art

[0002] Screening drugs for specific diseases using purely experimental methods is difficult and can require significant labor and experimental costs, as well as a long time to test as many drugs or drug combinations as possible. Therefore, it is desirable to develop a method for predicting the therapeutic effects of drugs that can pre-select potentially effective single drugs or synergistically acting drugs.

[0003] The approaches described in this section are not necessarily approaches that have been previously conceived or employed. Unless otherwise indicated, it should not be assumed that any approach described in this section is prior art simply by virtue of its inclusion in this section. Similarly, unless otherwise indicated, the issues raised in this section should not be considered as having been recognized as prior art. Summary of the Invention

[0004] According to one aspect of the present disclosure, a method for predicting drug therapeutic effects is provided, comprising: obtaining input data, wherein the input data comprises a corresponding chemical structure formula or amino acid sequence of at least one drug; and processing the input data using a drug therapeutic effect prediction model to obtain a therapeutic effect of the at least one drug predicted by the drug therapeutic effect prediction model, wherein the drug therapeutic effect prediction model comprises a first neural network, a heterogeneous graph neural network, and a second neural network, wherein the heterogeneous graph neural network comprises at least one drug node respectively representing the at least one drug, at least one gene node respectively representing at least one gene, at least one protein node respectively representing at least one protein, and edges representing a relationship between nodes in at least one subset of the at least one drug node, the at least one gene node, and the at least one protein node, wherein the at least one gene node is configured with a corresponding feature vector, and the at least one protein node is configured with a corresponding feature vector. There is a corresponding feature vector, wherein processing the input data includes: using the first neural network to extract the feature vector corresponding to the at least one drug from the corresponding chemical structure formula or amino acid sequence-amino acid interaction relationship of the at least one drug; feeding the feature vector corresponding to the at least one drug to the at least one drug node, so that the heterogeneous graph neural network updates the feature vector corresponding to the at least one drug node, the feature vector corresponding to the at least one gene node, and the feature vector corresponding to the at least one protein node through message passing; combining the updated corresponding feature vector of the at least one drug node, the updated corresponding feature vector of the at least one gene node, and the updated corresponding feature vector of the at least one protein node into a combined feature vector; and using the second neural network to generate output data from the combined feature vector, wherein the output data indicates the therapeutic effect of the at least one drug.

[0005] According to another aspect of the present disclosure, a drug treatment effect prediction device is provided, comprising: an input data acquisition unit for obtaining input data, wherein the input data comprises a corresponding chemical structural formula of at least one drug; and an input data processing unit for processing the input data using a drug treatment effect prediction model to obtain the treatment effect of the at least one drug predicted by the drug treatment effect prediction model, wherein the drug treatment effect prediction model comprises a first neural network, a heterogeneous graph neural network, and a second neural network, wherein the heterogeneous graph neural network comprises at least one drug node respectively representing the at least one drug, at least one gene node respectively representing at least one gene, at least one protein node respectively representing at least one protein, and edges representing the relationship between the nodes in at least one subset of the at least one drug node, the at least one gene node, and the at least one protein node, wherein the at least one gene node is configured with a corresponding feature vector, and the at least one The protein node is configured with a corresponding feature vector, wherein the input data processing unit includes a unit for performing the following operations: using the first neural network to extract the feature vector corresponding to the at least one drug from the corresponding chemical structure formula or amino acid sequence of the at least one drug; feeding the feature vector corresponding to the at least one drug to the at least one drug node, so that the heterogeneous graph neural network updates the feature vector corresponding to the at least one drug node, the feature vector corresponding to the at least one gene node, and the feature vector corresponding to the at least one protein node through message passing; combining the updated corresponding feature vector of the at least one drug node, the updated corresponding feature vector of the at least one gene node, and the updated corresponding feature vector of the at least one protein node into a combined feature vector; and using the second neural network to generate output data from the combined feature vector, wherein the output data indicates the therapeutic effect of the at least one drug.

[0006] According to yet another aspect of the present disclosure, a computing device is provided, comprising: a memory, a processor, and a computer program stored on the memory, wherein the processor is configured to execute the computer program to implement the steps of the method according to an embodiment of the present disclosure.

[0007] According to yet another aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the method according to the embodiment of the present disclosure are implemented.

[0008] According to yet another aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the steps of the method according to the embodiment of the present disclosure are implemented.

[0009] These and other aspects of the disclosure will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Further details, features and advantages of the present disclosure are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0011] Figure 1 is a schematic diagram illustrating an example system in which the various methods described herein may be implemented, according to an exemplary embodiment;

[0012] Figure 2 is a flow chart illustrating a method for predicting drug treatment effects according to an exemplary embodiment;

[0013] Figure 3 is a schematic diagram of a model according to an embodiment of the present disclosure;

[0014] Figure 4 is a data flow diagram according to an embodiment of the present disclosure;

[0015] Figure 5 is a schematic block diagram illustrating a drug treatment effect prediction apparatus according to an exemplary embodiment;

[0016] Figure 6 is a block diagram illustrating an exemplary computer device that can be used with the exemplary embodiments. DETAILED DESCRIPTION

[0017] In this disclosure, unless otherwise specified, the use of terms such as "first" and "second" to describe various elements is not intended to limit the positional relationship, temporal relationship, or importance relationship of these elements. Such terms are only used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of the element, while in some cases, based on the context of the description, they may also refer to different instances.

[0018] The terms used in the description of the various examples described in this disclosure are for the purpose of describing specific examples only and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element can be one or more. As used herein, the term "plurality" means two or more, and the term "based on" should be interpreted as "based at least in part on". In addition, the terms "and / or" and "at least one of..." cover any one of the listed items and all possible combinations.

[0019] As used herein, the term "heterogeneous graph" refers to a graph containing different types of nodes and edges. Different types of nodes and edges often have different types of attributes. These attributes are intended to characterize the characteristics of each type of node and edge. The term "heterogeneous graph neural network" refers to a graph neural network (GNN) with a network structure of a heterogeneous graph. Examples of graph neural networks include, but are not limited to, graph attention networks (GAN), relational graph convolutional neural networks (R-GCN, also known as RGCN or RGCNs or R-GCNs), graph convolutional neural networks (GCN), etc. "Message passing" is a general framework and programming paradigm for implementing GNN. It summarizes the implementation of various GNN models from the perspective of aggregation and updating. The message passing paradigm is as follows:

[0020] Assume that the feature of node v is The feature on the edge (u, v) is The message passing paradigm defines the following node-wise and edge-wise computations at step t+1:

[0021] Edge calculation:

[0022] Node calculation:

[0023] In the equation above, φ is the message function defined on each edge, which generates a message by combining the edge's features with the features of the nodes at its ends. The aggregation function ρ aggregates the messages received by the node. The update function ψ updates the node's features by combining the aggregated messages with the node's own features.

[0024] Message passing on heterogeneous graphs can be divided into two parts: (1) computing and aggregating messages for each relation. (2) aggregating messages from different relations for each node. More information on message passing on heterogeneous graphs can be found, for example, at https: / / docs.dgl.ai / en / latest / guide / message-heterograph.html#guide-message-passing-heterograph.

[0025] Exemplary embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0026] Figure 1is a schematic diagram illustrating an example system 100 in which the various methods described herein may be implemented, according to an exemplary embodiment.

[0027] refer to Figure 1 , the system 100 includes a client device 110 , a server 120 , and a network 130 communicatively coupling the client device 110 and the server 120 .

[0028] The client device 110 includes a display 114 and a client application (APP) 112 that can be displayed via the display 114. The client application 112 can be an application that needs to be downloaded and installed before running or a small program (liteapp) that is a lightweight application. In the case where the client application 112 is an application that needs to be downloaded and installed before running, the client application 112 can be pre-installed on the client device 110 and activated. In the case where the client application 112 is a small program, the user 102 can directly run the client application 112 on the client device 110 by searching for the client application 112 in the host application (for example, by the name of the client application 112, etc.) or scanning a graphic code (for example, a barcode, a QR code, etc.) of the client application 112, without installing the client application 112. In some embodiments, the client device 110 can be any type of mobile computer device, including a mobile computer, a mobile phone, a wearable computer device (for example, a smart watch, a head-mounted device, including smart glasses, etc.) or other types of mobile devices. In some embodiments, client device 110 may alternatively be a stationary computer device, such as a desktop computer, a server computer, or other type of stationary computer device.

[0029] The server 120 is typically a server deployed by an Internet Service Provider (ISP) or an Internet Content Provider (ICP). The server 120 may represent a single server, a cluster of multiple servers, a distributed system, or a cloud server that provides basic cloud services (such as cloud databases, cloud computing, cloud storage, and cloud communications). It will be understood that although Figure 1 1. The server 120 is shown communicating with only one client device 110, but the server 120 may provide background services to multiple client devices simultaneously.

[0030] Examples of network 130 include a local area network (LAN), a wide area network (WAN), a personal area network (PAN), and / or a combination of communication networks such as the Internet. Network 130 can be a wired or wireless network. In some embodiments, data exchanged through network 130 is processed using technologies and / or formats including Hypertext Markup Language (HTML), Extensible Markup Language (XML), etc. In addition, encryption technologies such as Secure Sockets Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), Internet Protocol Security (IPsec) can also be used to encrypt all or some links. In some embodiments, customized and / or dedicated data communication technologies can also be used to replace or supplement the above-mentioned data communication technologies.

[0031] For the purpose of the embodiments of this disclosure, Figure 1 In the example, client application 112 may be an application program that provides various functions for drug effect prediction. Accordingly, server 120 may be a server used in conjunction with the application program. Server 120 may provide services to client application 112 running on client device 110. Alternatively, server 120 may provide computing functions, etc., to client device 110, which in turn may provide services to client application 112 running on client device 110.

[0032] Figure 2 2 is a flow chart illustrating a method 200 for predicting drug treatment effects according to an exemplary embodiment. The method 200 may be executed on a client device (e.g., Figure 1 , that is, the execution subject of each step of the method 200 may be a client device 110 shown in FIG. Figure 1 In some embodiments, the method 200 may be performed on a server (e.g., Figure 1 In some embodiments, the method 200 may be performed by a client device (eg, the client device 110) and a server (eg, the server 120) in combination.

[0033] In the following, each step of the method 200 is described in detail.

[0034] refer to Figure 2 In step 210, input data is obtained, where the input data includes a corresponding chemical structure formula or amino acid sequence of at least one drug.

[0035] In step 220, the input data is processed using a drug treatment effect prediction model to obtain the treatment effect of the at least one drug predicted by the drug treatment effect prediction model. The drug treatment effect prediction model includes a first neural network, a heterogeneous graph neural network, and a second neural network. The heterogeneous graph neural network includes at least one drug node representing the at least one drug, at least one gene node representing at least one gene, at least one protein node representing at least one protein, and edges representing the relationship between the nodes in at least one subset of the at least one drug node, the at least one gene node, and the at least one protein node. The at least one gene node is configured with a corresponding feature vector, and the at least one protein node is configured with a corresponding feature vector.

[0036] The step 220 of processing the input data may include: extracting a feature vector corresponding to the at least one drug from the corresponding chemical formula or amino acid sequence of the at least one drug using the first neural network; feeding the feature vector corresponding to the at least one drug to the at least one drug node so that the heterogeneous graph neural network updates the feature vector corresponding to the at least one drug node, the feature vector corresponding to the at least one gene node, and the feature vector corresponding to the at least one protein node through message passing; combining the updated corresponding feature vector of the at least one drug node, the updated corresponding feature vector of the at least one gene node, and the updated corresponding feature vector of the at least one protein node into a combined feature vector; and generating output data from the combined feature vector using the second neural network, the output data indicating the therapeutic effect of the at least one drug.

[0037] According to such an embodiment, the effect of a drug can be effectively predicted based on the relationship between proteins, drugs and genes.

[0038] It is understood that, for example, for drugs in the form of small molecule compounds, a chemical structural formula can be used to represent them. For macromolecular drugs, an amino acid sequence can be used to represent them. As will be understood by those skilled in the art, for drugs, vertices are used to represent atoms in the chemical structural formula or amino acids in the amino acid sequence and edges are used to represent chemical bonds between atoms or interactions between amino acids, and the corresponding chemical structural formula or amino acid sequence of the at least one drug is converted into a corresponding isomorphic graph, and the present disclosure is not limited thereto. In addition, it is understood that, depending on the specific usage scenario of the model, at least one gene can be a gene associated with a disease, disease tissue cells, tumor cell lines, etc., and at least one protein can be a protein associated with a disease, disease tissue cells, tumor cell lines, etc., and the present disclosure is not limited thereto.

[0039] As cancer research continues to deepen, more and more anti-cancer drugs are being developed and designed. However, given the vast number of existing anti-cancer drugs, screening them solely through experimental methods is difficult, associated with significant experimental costs, and requires a long time to test as many drugs or combinations as possible. Computational methods can quickly pre-select one or more potentially effective drugs, saving experimental time and costs.

[0040] The models used in early machine learning are generally simpler, with less computational effort and a certain degree of interpretability. However, the prediction accuracy of the corresponding model is lower than that of the deep learning model, and it is more dependent on the input data, so improving the model performance is a relatively difficult task. Although the gradient boosting tree algorithm and the random forest algorithm are still used in the biological field, the application of deep learning algorithms has far exceeded that of traditional machine learning algorithms, especially the methods based on attention mechanisms and graph neural networks can significantly improve the prediction accuracy of the model. One obvious drawback is that although most deep learning models show good prediction accuracy during training, they often perform poorly on independent test sets. When identifying new drug combinations, models such as the DeepDDS model show obvious deficiencies. According to the solution of the embodiment of the present disclosure, it is possible to better explore the relationship between drugs and specific diseases based on the relationship between drugs, proteins and genes, and thus to make more accurate predictions.

[0041] In some related technologies, the prediction results obtained by simply integrating the omics data of cell lines with the chemical characterization of drugs lack biological knowledge and interpretability. Whether a drug can have an effect on a cell line depends largely on whether the drug can inhibit specific oncogenes in the cell line. The embodiments of the present disclosure transmit messages through a heterogeneous graph containing four types of edges: drug-target, protein-protein, gene-protein, and protein-gene, to obtain global features of proteins and genes, which are combined with drug features to enter the fully connected layer to predict drug treatment effect labels or drug combination treatment effect labels. The relationships between drugs, targets, proteins, and genes can be various relationships that can be understood by those skilled in the art, such as the relationship between drugs acting on targets, the relationship between genes being transcribed into proteins, etc.; and will be described in detail below in conjunction with specific exemplary embodiments.

[0042] While various operations are depicted in the drawings as following a particular order, this should not be understood as requiring that these operations be performed in the particular order shown or in sequential order, nor should it be understood that all illustrated operations must be performed to achieve desirable results.

[0043] Reference below Figure 3 and Figure 4 Some exemplary embodiments of the present disclosure are described. Figure 3, which shows a schematic diagram of a model according to an embodiment of the present disclosure, Figure 4 A data flow diagram according to an embodiment of the present disclosure is shown.

[0044] like Figure 3 As shown, the model 300 may include a first neural network 310, a heterogeneous graph neural network 320, and a second neural network 330. The first neural network 310 may be used to extract a feature vector corresponding to the at least one drug from the corresponding chemical structure formula or amino acid sequence of the at least one drug.

[0045] The extracted feature vector corresponding to at least one drug can be fed to at least one drug node of the heterogeneous graph neural network. Figure 4 , wherein the model 400 is shown in the form of data nodes, and for the sake of simplicity, the first neural network portion is omitted. The extracted feature vector corresponding to the at least one drug can be fed to one or more corresponding drug nodes 421 in the heterogeneous graph 420, so that the heterogeneous graph neural network updates the feature vectors of each node therein through message passing, including the corresponding feature vector of the at least one drug node 421, the corresponding feature vector of the at least one protein node 422, and the corresponding feature vector of the at least one gene node 423. As will be further described below, the feature vectors of the at least one protein node 422 and the at least one gene node 423 can be randomly initialized or configured in other ways, and the present disclosure is not limited thereto.

[0046] It can be understood that although two drug nodes, six protein nodes, and six gene nodes are shown in the figure, the above numbers are only examples and the present disclosure is not limited thereto.

[0047] Continue to refer Figure 4 , where the model 400 may also include a neural network 430, which corresponds to, for example Figure 3 The neural network 430 may combine the updated corresponding feature vector of the at least one drug node 421, the updated corresponding feature vector of the at least one gene node 423, and the updated corresponding feature vector of the at least one protein node 422 into a combined feature vector, and generate output data from the combined feature vector, the output data indicating the therapeutic effect of the at least one drug. As an example, Figure 4As shown, the neural network 430 may include a combination subnetwork 431 and a prediction subnetwork 432, wherein the combination subnetwork 431 is used to combine feature vectors, and the prediction subnetwork 432 (for example, a multi-layer perceptron or one or more fully connected layers) can generate a prediction result based on the combined feature vectors. It is understood that the present disclosure is not limited to this, and the neural network 430 does not necessarily include the above-mentioned subnetwork parts, and may include more or fewer subnetworks or submodules; in other examples, the step of combining feature vectors may be implemented by other model parts besides the neural network 430, and the present disclosure is not limited to this. For example, combining feature vectors may include summing, superposition, splicing or other common combination methods, and the present disclosure is not limited to this.

[0048] According to some embodiments, the relationship between each node in at least a subset of the at least one drug node, the at least one gene node and the at least one protein node includes at least one of the following: interaction between proteins, relationship between drugs and proteins as targets, effect of proteins on gene expression, and expression of genes to proteins.

[0049] According to such an embodiment, information can be transferred by using a biological knowledge network constructed by using information such as the relationship between proteins, the relationship between drugs and targets, and the information between transcription factors (generally proteins) and gene expression, thereby making full use of biological knowledge and obtaining biological explainability.

[0050] Continue to refer Figure 4 , where four types of relationships are shown with different lines. In this example, relationship type 1 can represent the relationship between a drug and a target, relationship type 2 can represent a protein interaction, relationship type 3 can represent a gene-to-protein relationship, and relationship type 4 can represent a protein-to-gene relationship, such as a regulatory role as a transcription factor or other influence.

[0051] Protein-Protein Interaction Networks (PPI) are composed of proteins that participate in various aspects of life processes such as biological signal transmission, gene expression regulation, energy and material metabolism, and cell cycle regulation through interactions with each other. Systematic analysis of the interaction relationship between a large number of proteins in biological systems is of great significance for understanding the working principles of proteins in biological systems, understanding the reaction mechanisms of biological signals and energy material metabolism under special physiological conditions such as diseases, and understanding the functional connections between proteins. According to one or more embodiments of the present disclosure, the PPI network can be used to construct or configure the relationship between proteins.

[0052] According to one or more embodiments of the present disclosure, compared to simply connecting drug features with cell line omics information, or only using gene-gene interaction networks to directly predict drug combination synergy, a more complex biological network is constructed by using PPI information, drug-target information, and gene transcription information to transfer information, so that the network contains more biological prior knowledge. PPI information can represent known and predicted protein interaction networks, and can come from genome predictions, high-throughput experiments, (conservative) co-expression experiments, automated text mining, database-related knowledge, etc. As an example, the STRING database currently covers more than 240,000 proteins from 5,090 species.

[0053] As a specific non-limiting example, the model of the present disclosure can use 19,264 genes and 89,176 high-confidence protein interaction edges, and can have a combined score combined_score>850. It is understandable that the above is only an example, and the present disclosure is not limited thereto. The model of the present disclosure can be applied to a larger or smaller number of genes, proteins, etc. Drug target information can come from the DGIdb database, which is a drug-gene interaction database that contains association information between genes and known or potential drugs as drug-target information in a heterogeneous graph. Gene transcriptional regulation information can come from the GTRD database, which contains more comprehensive transcriptional regulation information. Figure 4 As shown, in the heterogeneous graph, protein-gene information can represent the relationship between transcription factors regulating target gene information, or the relationship between other proteins that can affect gene expression.

[0054] According to some embodiments, the first neural network may be a graph neural network, and extracting the feature vector corresponding to the at least one drug from the corresponding chemical structure formula of the at least one drug using the first neural network may include: converting the corresponding chemical structure formula of the at least one drug or the amino acid sequence into a corresponding isomorphic graph by using vertices to represent atoms in the chemical structure formula or amino acids in the amino acid sequence and using edges to represent chemical bonds between atoms or interaction relationships between amino acids in the amino acid sequence; and extracting the feature vector corresponding to the at least one drug from the corresponding isomorphic graph using the first neural network.

[0055] According to one or more embodiments, drug characterization based on SMILES information can be adopted. SMILES (Simplified molecular input line entry system), a simplified molecular linear input specification, is a specification that uses ASCII strings to clearly describe molecular structures. For example, the RDKIT software package or other methods known to those skilled in the art can be used to convert drug SMILES information into a molecular graph, with atoms representing nodes and chemical bonds representing edges, and the molecular graph is converted into a homogeneous graph. The initial features of atoms can be calculated using, for example, the Chem software package as node features. According to one or more embodiments, since each drug or drug combination corresponds to different drug-target information, the heterogeneous graph network corresponding to each drug or drug combination has specificity, which improves the accuracy of the prediction, and the model exhibits strong generalization performance, and still maintains good prediction accuracy on an independent test set. In some other embodiments, for the structure of macromolecular drugs, ligand-receptor relationship prediction can be used to predict the structure of macromolecular drugs.

[0056] According to some embodiments, the first neural network may include a graph attention network.

[0057] According to such an embodiment, a graph attention network (GAT) can be used when initializing drug feature vectors to more accurately extract features worthy of attention. Through the self-attention mechanism, higher-level features of nodes can be learned, and node features in the entire graph can be aggregated as the final drug representation for input into the corresponding drug node.

[0058] According to other embodiments, the first neural network may include a graph convolutional neural network.

[0059] According to some embodiments, the first neural network may be a deep learning network. In such embodiments, extracting a feature vector corresponding to the at least one drug from a corresponding chemical structure or amino acid sequence of the at least one drug using the first neural network may include extracting a feature vector corresponding to the at least one drug from an amino acid sequence of the at least one drug and an interaction relationship between amino acids in the amino acid sequence using the deep learning network.

[0060] According to some embodiments, the first neural network may be an encoder-decoder network, such as a transformer network. In such embodiments, extracting a feature vector corresponding to the at least one drug from a corresponding chemical structure or amino acid sequence of the at least one drug using the first neural network may include extracting a feature vector corresponding to the at least one drug from the amino acid sequence of the at least one drug using the encoder-decoder network.

[0061] According to some other embodiments, the first neural network may be another type of neural network. For example, a feature vector corresponding to an amino acid sequence of a compound may be obtained by using a protein three-dimensional structure prediction model similar to omegaFold, but the present disclosure is not limited thereto.

[0062] It is understood that the amino acid sequence of a drug (e.g., a macromolecular drug) can be an amino acid sequence string, an amino acid sequence number string, or other forms that are understandable to those skilled in the art. The interaction relationship between amino acids can be the interaction relationship between each amino acid in the amino acid sequence (such as, but not limited to, hydrogen bonds, van der Waals forces, or other interactions between amino acids that are understandable to those skilled in the art), and can be expressed in numerical values, labels, or other ways that are understandable to those skilled in the art, and the present disclosure is not limited thereto.

[0063] According to some embodiments, the corresponding feature vector configured for the at least one protein node is at least partially based on the values of N gene pre-trained representation vectors in the pre-trained model, where N is an integer greater than 1. The pre-trained model may include an encoder network, a decoder network, and a gene representation network. The pre-trained model can be used to generate complete gene expression data from masked gene expression data, wherein the complete gene expression data includes data on N gene expression quantities associated with a specific disease or a specific cell line, and the masked gene expression data may be data obtained by masking at least one gene expression quantity in the complete gene expression data. The encoder network can be used to generate N gene expression embedding vectors from the masked gene expression data; the gene representation network can be used to generate N gene pre-trained representation vectors. Each gene pre-trained representation vector can also be referred to as a pre-trained gene representation encoding vector, and each gene pre-trained representation vector has the same dimension as the corresponding gene expression embedding vector; the decoder network can be used to generate the complete gene expression data from N added embedding vectors, wherein each added embedding vector is obtained by adding a corresponding one of the N gene expression embedding vectors to a corresponding one of the N gene pre-trained representation vectors.

[0064] In such an embodiment, the corresponding feature vectors of the protein nodes can be initialized by a pre-trained model. As an example, such a model architecture can be regarded as a self-supervised language model for learning gene expression, for generating complete genes based on masked genes. Such a model architecture can be similar to the BERT model, which includes an encoder and a decoder. As a more specific example, an expression matrix (B×N, where B is the sample size and N is the number of genes) with a random 15% position mask is created as input. For each position of each sample, the expression value is converted to an int and projected onto a randomly initialized D-dimensional embedding vector, which can be referred to as the first embedding vector here, or the gene expression embedding vector as described above, that is, N gene expression embedding vectors are generated. Each gene is tokenized by retrieving the corresponding D-dimensional random embedding vector. Here, the D-dimensional random embedding vector can be referred to as the second embedding vector, or the gene pre-trained representation vector as described above, which acts similarly to the position relationship encoding vector and can represent the relationship between genes after training. Corresponding to the N gene expression embedding vectors, there can also be N gene pre-trained representation vectors. The corresponding two embedding vectors are added element-by-element and input to the encoder. Any encoder and decoder design known to those skilled in the art can be used. For example, the encoder portion can include multiple attention layers to capture gene-gene dependencies and output a latent feature matrix. The decoder can use a sequential MLP layer to project the hidden features to the original shape (size B×N). Layer normalization can be merged between the encoder and decoder. The training process can be supervised by calculating the mean squared error (MSE) loss between the true matrix and the output matrix, or by other supervision strategies known to those skilled in the art, and the present disclosure is not limited thereto.

[0065] As described above, the second embedding vector or gene pre-training representation vector after such training can reflect the implicit relationship between gene expression or protein. In such an embodiment, the protein features input during the training of the drug treatment effect prediction model can include the protein representation in TCGA, and can also include the protein representation obtained based on pre-training (that is, the gene pre-training representation vector as described above). TCGA (The Cancer Genome Atlas, Cancer Genome Atlas) includes multi-omics data of different patients with 33 different cancer types, mainly including gene expression, mutation, DNA methylation, copy number change data, etc. The inventors found that by configuring the feature vector of the corresponding protein node using the gene pre-training representation vector obtained by such training, a model with faster convergence and more accurate prediction effect can be obtained.

[0066] According to some embodiments, the complete gene expression data includes at least one single-cell complete gene expression data, and each single-cell complete gene expression data represents a gene expression profile in a single cell associated with a specific disease or a specific cell line. The gene expression profile represents the type and abundance information of gene expression of a specific disease or a specific cell line in a specific state. For example, a plurality of sample data can be used to train the above-mentioned pre-training model, and each sample data is sample data based on the measurement results of the gene expression profile or protein in a single cell. In such an embodiment, the pre-training model can be referred to as a single-cell pre-training model. Experiments have shown that using a single-cell pre-training model to configure the node can obtain better prediction results.

[0067] Gene representations can be calculated using the single-cell pre-trained model described above. The single-cell pre-trained model takes gene expression as input, performs feature encoding via the Transformer, compares the trained feature representations with the true gene expression, calculates the mean square error (MSE) loss function, and performs backpropagation to ultimately obtain the trained gene representations. After incorporating the pre-trained model, known protein multi-omics data is combined with the pre-trained data to incorporate more information into the input protein feature representations. Experiments have shown that models incorporating single-cell pre-training perform significantly better than models without pre-training.

[0068] According to some embodiments, the N gene expression levels, the at least one gene, and the at least one protein are in one-to-one correspondence with each other, and the corresponding feature vectors assigned to the at least one protein node are the pre-trained representation vectors of the N genes.

[0069] In such an embodiment, if the data used in the pre-trained model completely corresponds to the data of the main model, the vectors of the pre-trained model are used to initialize the protein features in the main model. In other embodiments, for example, if there is a certain degree of mismatch between the data of the pre-trained model and the main model, the feature vectors of the corresponding protein nodes can be used to initialize the corresponding embedding vectors of the trained model. For protein nodes without corresponding embedding vectors, random initialization or other initialization methods can be used.

[0070] According to some embodiments, the at least one drug includes at least two drugs. The therapeutic effect is selected from the group including the following: synergy and antagonism. According to such an embodiment, the synergistic effect of two or more drugs can be predicted. It is understood that the selection of the therapeutic effect from the group including synergy and antagonism does not necessarily mean that the therapeutic effect is a binary label. For example, the therapeutic effect can include a numerical value, such as a numerical value between 0 and 1, where the closer to 0, the higher the antagonism, the closer to 1, the higher the synergy, and so on.

[0071] Since tumors with the same gene mutations rarely occur clinically, in order to further improve the utilization rate of drugs and reduce the possibility of drug resistance, drug combination therapy is an effective way to solve this problem. Compared with single drug therapy, drug combination therapy can not only alleviate drug resistance, but also improve the efficacy of drugs and reduce the toxicity and side effects of drugs. When single drug therapy is ineffective, drug combination therapy is often a common solution. For example, triple-negative breast cancer is a malignant tumor with strong invasiveness, high metastasis rate and poor efficacy. The use of a single drug treatment has little effect, but drug combination therapy can significantly increase the apoptosis rate of triple-negative breast cancer. Therefore, finding an effective drug combination is an important treatment strategy for some diseases. However, it is more difficult to experiment with each drug combination one by one. Therefore, according to one or more embodiments of the present disclosure, the workload required to find drug combinations can be greatly reduced.

[0072] In other embodiments, the at least one drug may include only one drug, and the therapeutic effect may be a label or a numerical value, for example, a binary label of "effective" or "ineffective", or a score representing the therapeutic effect, and the present disclosure is not limited thereto.

[0073] According to some embodiments, the at least one gene includes a plurality of genes associated with a specific disease or a specific cell line, and the at least one protein includes a plurality of proteins corresponding to the plurality of genes.

[0074] According to such an embodiment, the relationship between a drug and a specific cell line can be learned based on the gene representation and protein representation in a specific cell line, for example, by setting the initial gene representation input to the corresponding gene expression profile, thereby enabling more accurate prediction of drug effects. For example, each gene node can represent a corresponding gene, and the value of each gene node can be the corresponding gene expression level (e.g., the amount of the mRNA in the specific cell line).

[0075] According to one or more embodiments of the present disclosure, a deep learning model based on single-cell pre-training is proposed for predicting drug combination synergy, in which the chemical structure of the drug is represented by a homogeneous graph, with vertices being atoms and edges being chemical bonds. A graph attention network (GAT) is used to calculate the feature vector of a single drug. A heterogeneous graph containing four types of edges—protein-protein, drug-target, protein-gene, and gene-protein—is then constructed. Drug features derived from graph convolution are input into the heterogeneous graph for message passing. Finally, the protein, drug, and gene features are integrated to predict the drug combination label.

[0076] It is understood that throughout the text, unless otherwise specified, the term "gene" may refer to RNA, such as mRNA. The term "gene expression" refers to the process of synthesizing genetic information from a gene into a functional gene product. The term "gene expression amount" may refer to the amount of mRNA produced during gene expression, and the present disclosure is not limited thereto.

[0077] According to one or more embodiments, the protein representation obtained by pre-training and the drug representation are input into the heterogeneous graph network for training. For example, the training data can include the synergistic effects of 353,372 drug combinations on cell lines, with data labels of 0 or 1. The training samples can include 18,614 different drugs and 169 cell lines, involving various common cancer types such as lung cancer, gastric cancer, prostate cancer, and breast cancer. The model is trained using 5-fold cross-validation. The test data can use an independent test data set containing 1,151 samples, including 103 drugs and 82 cell lines.

[0078] As shown in Table 1, the comparative example (baseline) shows the accuracy of the model prediction on the independent test set when the heterogeneous graph is not used. Example 1 shows the prediction accuracy after using the heterogeneous graph model, and it can be seen that the model prediction accuracy is improved. Example 2 is the model accuracy after adding protein multi-omics data from TCGA as protein characterization, and it can be seen that the prediction accuracy is further significantly improved. Example 3 is the protein characterization after adding single-cell pre-training, and the model performance is further improved.

[0079] Table 1 Comparison of AUC results of different models

[0080] Model Comparative Example Example 1 Example 2 Example 3 AUC 0.75 0.79 0.82 0.85

[0081] In addition, Table 2 shows the prediction effect of the model on the test dataset under different pre-trained representations and different heterogeneous graph algorithms. Table 2 shows the comparison of AUC and F1 scores under different combinations of pre-trained representations and TCGA data representations, where Example 4 is based on the combination of TCGA and the pre-trained representation of pre-trained model 1, Example 5 is based on the combination of TCGA and the pre-trained representations of pre-trained model 1 and pre-trained model 2, Example 6 is based on the combination of TCGA and the pre-trained representations of pre-trained model 1 and pre-trained model 3, and Example 7 is based on the combination of TCGA and the pre-trained representations of pre-trained model 1 and pre-trained model 4.

[0082] Table 2 Comparison of prediction results using different pre-training representations

[0083] Model Example 4 Example 5 Example 6 Example 7 AUC 0.847 0.839 0.848 0.85 F1 score 0.839 0.832 0.844 0.847

[0084] As can be seen from Table 2, the addition of different pre-trained representations affects the model's predictive ability. After adding two different pre-trained representations, the model's performance is slightly better than that of the model using only one pre-trained representation. Overall, the model with the addition of pre-trained representations performs better on the test dataset than the model using only TCGA multi-omics features. However, regardless of whether pre-trained representations or TCGA multi-omics features are used, the solution based on the features of one or more embodiments of the present disclosure will achieve better prediction results than traditional solutions in the related art.

[0085] According to one or more embodiments of the present disclosure, it is possible to collect a large amount of data from datasets covering various drug-related aspects, then leverage the advantages of large-scale pre-trained models to generate information representations of drugs, proteins, and diseases, and build a message-passing graph based on this. Leveraging the flexibility of graph-structured learning, it is possible to accurately search for drugs or drug combinations for specific diseases. Furthermore, because it utilizes the biological relationships between drugs, targets, proteins, and genes, it has a certain degree of universality and is conducive to large-scale promotion.

[0086] According to one or more embodiments of the present disclosure, a drug treatment effect prediction model can employ an end-to-end deep learning framework. By leveraging multimodal data, graph neural networks, and large-scale unsupervised training, it integrates and learns multimodal drug action prediction information, thereby accurately predicting drug effects or drug synergy. The model can use a drug's chemical structure graph and cell line protein expression as input and apply a pre-trained molecular graph transformer to convert the drug graph information into an embedding. Simultaneously, a protein language model is used to generate an embedding for each protein in the expression. To obtain more and richer features, disease information is included. The model can also obtain disease embeddings from the precision medicine knowledge graph PrimeKG. Next, a graph neural network is used to represent the generated embeddings as nodes. To perform reasoning on unknown (unseen) drugs, the model can also include drug-drug similarity relationships, drug-target / drug-drug interaction relationships, etc. to generate pseudo edges, thereby forming a graph with richer information. In addition, the model can also include a drug effect predictor (e.g., single drug effect or synergistic effect) using a multilayer perceptron (MLP) to predict drug effects.

[0087] Figure 5 2 is a schematic block diagram illustrating a drug treatment effect prediction apparatus 500 according to an exemplary embodiment. The drug treatment effect prediction apparatus 500 may include an input data obtaining unit 510 and an input data processing unit 520.

[0088] The input data obtaining unit 510 is configured to obtain input data, where the input data includes a corresponding chemical structural formula of at least one drug.

[0089] The input data processing unit 520 is used to process the input data using a drug treatment effect prediction model to obtain the treatment effect of the at least one drug predicted by the drug treatment effect prediction model, wherein the drug treatment effect prediction model includes a first neural network, a heterogeneous graph neural network, and a second neural network, and the heterogeneous graph neural network includes at least one drug node representing the at least one drug, at least one gene node representing at least one gene, at least one protein node representing at least one protein, and edges representing the relationship between the nodes in at least one subset of the at least one drug node, the at least one gene node, and the at least one protein node, the at least one gene node is configured with a corresponding feature vector, and the at least one protein node is configured with a corresponding feature vector.

[0090] The input data processing unit 520 includes a unit for performing the following operations: using the first neural network to extract the corresponding feature vector of the at least one drug from the corresponding chemical structure formula of the at least one drug; feeding the corresponding feature vector of the at least one drug to the at least one drug node, so that the heterogeneous graph neural network updates the corresponding feature vector of the at least one drug node, the corresponding feature vector of the at least one gene node, and the corresponding feature vector of the at least one protein node through message passing; combining the updated corresponding feature vector of the at least one drug node, the updated corresponding feature vector of the at least one gene node, and the updated corresponding feature vector of the at least one protein node into a combined feature vector; using the second neural network to generate output data from the combined feature vector, wherein the output data indicates the therapeutic effect of the at least one drug.

[0091] It should be understood that Figure 5 The modules of the apparatus 500 shown in FIG. 5 can be used in conjunction with the reference Figure 2 The steps in the method 200 described above correspond to each other. Therefore, the operations, features and advantages described above for the method 200 are also applicable to the apparatus 500 and the modules included therein. For the sake of brevity, some operations, features and advantages are not repeated here.

[0092] According to an embodiment of the present disclosure, a computing device is also disclosed, including a memory, a processor, and a computer program stored on the memory, wherein the processor is configured to execute the computer program to implement the steps of the drug treatment effect prediction method and its variant examples according to the embodiment of the present disclosure.

[0093] According to an embodiment of the present disclosure, a non-transitory computer-readable storage medium is also disclosed, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the drug treatment effect prediction method according to the embodiment of the present disclosure and its variant examples are implemented.

[0094] According to an embodiment of the present disclosure, a computer program product is also disclosed, including a computer program, wherein when the computer program is executed by a processor, it implements the steps of the drug treatment effect prediction method according to the embodiment of the present disclosure and its variant examples.

[0095] Although specific functions are discussed above with reference to specific modules / units, it should be noted that the functions of the various modules / units discussed herein may be divided into multiple modules / units, and / or at least some of the functions of multiple modules / units may be combined into a single module / unit. The specific module / unit discussed herein performing an action includes the specific module / unit itself performing the action, or alternatively the specific module / unit calling or otherwise accessing another component or module / unit that performs the action (or performs the action in conjunction with the specific module / unit). Therefore, the specific module / unit that performs an action may include the specific module / unit itself that performs the action and / or another module / unit that the specific module / unit calls or otherwise accesses to perform the action. As used herein, the phrase "entity A initiates action B" may mean that entity A issues an instruction to perform action B, but entity A itself does not necessarily perform the action B.

[0096] It should also be understood that various techniques may be described herein in the general context of software hardware elements or program modules. Figure 5 The various modules / units described can be implemented in hardware or in hardware in combination with software and / or firmware. For example, these modules / units can be implemented as computer program codes / instructions, which are configured to be executed in one or more processors and stored in a computer-readable storage medium. Alternatively, these modules / units can be implemented as hardware logic / circuits. For example, in some embodiments, one or more of the described modules or units can be implemented together in a system on chip (SoC). SoC can include an integrated circuit chip (which includes a processor (e.g., a central processing unit (CPU), a microcontroller, a microprocessor, a digital signal processor (DSP), etc.), a memory, one or more communication interfaces, and / or one or more components in other circuits), and can optionally execute the received program code and / or include embedded firmware to perform functions.

[0097] According to one aspect of the present disclosure, a computing device is provided, comprising a memory, a processor, and a computer program stored in the memory. The processor is configured to execute the computer program to implement the steps of any one of the method embodiments described above.

[0098] According to one aspect of the present disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any method embodiment described above are implemented.

[0099] According to one aspect of the present disclosure, a computer program product is provided, which includes a computer program. When the computer program is executed by a processor, the steps of any one of the method embodiments described above are implemented.

[0100] In the following, combined Figure 6 Illustrative examples of such a computer device, non-transitory computer-readable storage medium, and computer program product are described.

[0101] Figure 6 6 shows an example configuration of a computer device 600 that can be used to implement the methods described herein. Figure 1 The server 120 and / or client device 110 shown in FIG may include an architecture similar to the computer device 600. The above-mentioned drug treatment effect prediction device / apparatus may also be fully or at least partially implemented by the computer device 600 or similar devices or systems.

[0102] The computer device 600 can be a variety of different types of devices, such as a server of a service provider, a device associated with a client (e.g., a client device), a system on a chip, and / or any other suitable computer device or computing system. Examples of the computer device 600 include, but are not limited to, a desktop computer, a server computer, a laptop or netbook computer, a mobile device (e.g., a tablet computer, a cellular or other wireless phone (e.g., a smartphone), a notepad computer, a mobile station), a wearable device (e.g., glasses, a watch), an entertainment device (e.g., an entertainment appliance, a set-top box communicatively coupled to a display device, a game console), a television or other display device, a car computer, and the like. Thus, the computer device 600 can range from a full-resource device with a large amount of memory and processor resources (e.g., a personal computer, a game console) to a low-resource device with limited memory and / or processing resources (e.g., a traditional set-top box, a handheld game console).

[0103] The computer device 600 may include at least one processor 602, memory 604, communication interface(s) 606, a display device 608, other input / output (I / O) devices 610, and one or more mass storage devices 612, all capable of communicating with one another, such as via a system bus 614 or other appropriate connections.

[0104] The processor 602 may be a single processing unit or multiple processing units, all of which may include a single or multiple computing units or multiple cores. The processor 602 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operational instructions. Among other capabilities, the processor 602 may be configured to retrieve and execute computer-readable instructions stored in the memory 604, mass storage device 612, or other computer-readable media, such as program code for an operating system 616, program code for application programs 618, program code for other programs 620, and the like.

[0105] The memory 604 and the mass storage device 612 are examples of computer-readable storage media for storing instructions that are executed by the processor 602 to implement the various functions described above. For example, the memory 604 may generally include both volatile memory and non-volatile memory (e.g., RAM, ROM, etc.). In addition, the mass storage device 612 may generally include a hard drive, a solid-state drive, removable media, including external and removable drives, memory cards, flash memory, floppy disks, optical disks (e.g., CDs, DVDs), storage arrays, network attached storage, storage area networks, etc. The memory 604 and the mass storage device 612 may all be collectively referred to herein as memory or computer-readable storage media, and may be non-transitory media capable of storing computer-readable, processor-executable program instructions as computer program code that can be executed by the processor 602 as a specific machine configured to implement the operations and functions described in the examples herein.

[0106] A number of program modules can be stored on the mass storage device 612. These programs include an operating system 616, one or more application programs 618, other programs 620, and program data 622, and they can be loaded into the memory 604 for execution. Examples of such applications or program modules can include, for example, computer program logic (e.g., computer program code or instructions) for implementing the following components / functionality: the client application 112, the method 200 (including any suitable steps of the method 200), and / or other embodiments described herein.

[0107] Although Figure 66 as being stored in the memory 604 of the computer device 600, but the modules 616, 618, 620, and 622, or portions thereof, may be implemented using any form of computer-readable media that can be accessed by the computer device 600. As used herein, "computer-readable media" includes at least two types of computer-readable media, namely, computer storage media and communication media.

[0108] Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information, such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVDs), or other optical storage devices, magnetic cassettes, magnetic tape, magnetic disk storage devices or other magnetic storage devices, or any other non-transmission media that can be used to store information for access by a computer device.

[0109] In contrast, communication media may embody computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism. Computer storage media as defined herein does not include communication media.

[0110] The computer device 600 may also include one or more communication interfaces 606 for exchanging data with other devices, such as through a network, a direct connection, etc., as previously discussed. Such communication interfaces may be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), a wired or wireless (such as an IEEE 802.11 wireless LAN (WLAN)) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth TM The communication interface 606 may include a wireless network interface, a near field communication (NFC) interface, and the like. The communication interface 606 may facilitate communication within a variety of network and protocol types, including wired networks (e.g., LAN, cable, etc.) and wireless networks (e.g., WLAN, cellular, satellite, etc.), the Internet, and the like. The communication interface 606 may also provide for communication with external storage devices (not shown) such as storage arrays, network attached storage, storage area networks, and the like.

[0111] In some examples, a display device 608 such as a monitor may be included for displaying information and images to the user. Other I / O devices 610 may be devices that receive various inputs from the user and provide various outputs to the user, and may include a touch input device, a gesture input device, a camera, a keyboard, a remote control, a mouse, a printer, an audio input / output device, and the like.

[0112] Although the present disclosure has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative and exemplary and not restrictive; the disclosure is not limited to the disclosed embodiments. Variations to the disclosed embodiments will be understood and effected by those skilled in the art in practicing the claimed subject matter by studying the drawings, the disclosure and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps that are not listed, and the word "a" or "an" does not exclude a plurality. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

1. A method for predicting drug treatment effects, comprising: Obtaining input data, wherein the input data includes a corresponding chemical formula or amino acid sequence of at least one drug; as well as The input data is processed using a drug treatment effect prediction model to obtain a treatment effect of the at least one drug predicted by the drug treatment effect prediction model, wherein the drug treatment effect prediction model includes a first neural network, a heterogeneous graph neural network, and a second neural network, the heterogeneous graph neural network includes at least one drug node representing the at least one drug, at least one gene node representing at least one gene, at least one protein node representing at least one protein, and edges representing a relationship between nodes in at least one subset of the at least one drug node, the at least one gene node, and the at least one protein node, the at least one gene node being configured with a corresponding feature vector, and the at least one protein node being configured with a corresponding feature vector, wherein processing the input data includes: extracting a feature vector corresponding to the at least one drug from a corresponding chemical structure formula or amino acid sequence of the at least one drug using the first neural network; Feeding the feature vector corresponding to the at least one drug to the at least one drug node, so that the heterogeneous graph neural network updates the feature vector corresponding to the at least one drug node, the feature vector corresponding to the at least one gene node, and the feature vector corresponding to the at least one protein node through message passing; combining the updated corresponding feature vector of the at least one drug node, the updated corresponding feature vector of the at least one gene node, and the updated corresponding feature vector of the at least one protein node into a combined feature vector; and Output data is generated from the combined feature vector using the second neural network, the output data being indicative of the therapeutic effect of the at least one drug.

2. The method according to claim 1, wherein The relationship between each node in at least a subset of the at least one drug node, the at least one gene node and the at least one protein node includes at least one item selected from the group consisting of: an interaction relationship between proteins, a relationship between a drug and a target protein, an effect of a protein on gene expression, and gene-to-protein expression.

3. The method according to claim 1, wherein The first neural network is a graph neural network, and wherein extracting a feature vector corresponding to the at least one drug from a corresponding chemical structure formula or amino acid sequence of the at least one drug using the first neural network comprises: Converting the corresponding chemical formula of the at least one drug or the amino acid sequence into a corresponding isomorphic graph by using vertices to represent atoms in the chemical formula or amino acids in the amino acid sequence and using edges to represent chemical bonds between atoms or interaction relationships between amino acids; and The first neural network is used to extract a feature vector corresponding to the at least one drug from the corresponding isomorphic graph.

4. The method according to claim 3, wherein: The first neural network includes a graph attention network; or the first neural network includes a graph convolutional neural network.

5. The method according to claim 1, wherein The first neural network is a deep learning network, and wherein, using the first neural network to extract the feature vector corresponding to the at least one drug from the corresponding chemical structure or amino acid sequence of the at least one drug includes: using the deep learning network to extract the feature vector corresponding to the at least one drug from the amino acid sequence of the at least one drug and the interaction relationship between amino acids in the amino acid sequence.

6. The method according to claim 1, wherein The first neural network is an encoding-decoding network, and wherein, using the first neural network to extract the feature vector corresponding to the at least one drug from the corresponding chemical structure or amino acid sequence of the at least one drug includes: using the encoding-decoding network to extract the feature vector corresponding to the at least one drug from the amino acid sequence of the at least one drug.

7. The method according to claim 1, wherein The corresponding feature vector assigned to the at least one protein node is based at least in part on the values of N pre-trained gene representation vectors in a pre-trained model, wherein the pre-trained model includes an encoder network, a decoder network, and a gene representation network, and the pre-trained model is used to generate complete gene expression data from masked gene expression data, wherein the complete gene expression data includes N gene expression data associated with a specific disease or a specific cell line, where N is an integer greater than 1, and the masked gene expression data is data obtained by masking at least one gene expression value in the complete gene expression data, wherein: The encoder network is used to generate N gene expression embedding vectors from the masked gene expression data; The gene representation network is used to generate N gene pre-trained representation vectors, each gene pre-trained representation vector has the same dimension as the corresponding gene expression embedding vector; The decoder network is used to generate the complete gene expression data from N added embedding vectors, wherein each added embedding vector is obtained by adding a corresponding one of the N gene expression embedding vectors to a corresponding one of the N gene pre-trained representation vectors.

8. The method according to claim 7, wherein: The complete gene expression data includes at least one single-cell complete gene expression data, and each single-cell complete gene expression data represents a gene expression profile in a single cell associated with the specific disease or the specific cell line.

9. The method according to claim 7, wherein: The N gene expression levels, the at least one gene, and the at least one protein correspond to each other one by one, and the corresponding feature vectors assigned to the at least one protein node are the N gene pre-trained representation vectors.

10. The method according to any one of claims 1 to 9, wherein The at least one drug comprises at least two drugs, and wherein the therapeutic effect is selected from the group consisting of: synergy and antagonism.

11. The method according to any one of claims 1 to 9, wherein The at least one gene includes a plurality of genes associated with a specific disease or a specific cell line, and wherein the at least one protein includes a plurality of proteins corresponding to the plurality of genes.

12. A device for predicting drug treatment effects, comprising: an input data obtaining unit, configured to obtain input data, wherein the input data includes a corresponding chemical structural formula or amino acid sequence of at least one drug; as well as An input data processing unit is configured to process the input data using a drug treatment effect prediction model to obtain the treatment effect of the at least one drug predicted by the drug treatment effect prediction model, wherein the drug treatment effect prediction model includes a first neural network, a heterogeneous graph neural network, and a second neural network, the heterogeneous graph neural network includes at least one drug node representing the at least one drug, at least one gene node representing at least one gene, at least one protein node representing at least one protein, and edges representing the relationship between the at least one drug node, the at least one gene node, and at least one subset of the at least one protein node, the at least one gene node being configured with a corresponding feature vector, and the at least one protein node being configured with a corresponding feature vector, wherein the input data processing unit includes a unit for performing the following operations: extracting a feature vector corresponding to the at least one drug from a corresponding chemical structure formula or amino acid sequence of the at least one drug using the first neural network; Feeding the feature vector corresponding to the at least one drug to the at least one drug node, so that the heterogeneous graph neural network updates the feature vector corresponding to the at least one drug node, the feature vector corresponding to the at least one gene node, and the feature vector corresponding to the at least one protein node through message passing; combining the updated corresponding feature vector of the at least one drug node, the updated corresponding feature vector of the at least one gene node, and the updated corresponding feature vector of the at least one protein node into a combined feature vector; and Output data is generated from the combined feature vector using the second neural network, the output data being indicative of the therapeutic effect of the at least one drug.

13. A computing device comprising: a memory, a processor, and a computer program stored on said memory, The processor is configured to execute the computer program to implement the steps of the method according to any one of claims 1 to 11.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

15. A computer program product comprising a computer program, wherein When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images