Graph neural network-based candidate drug efficacy prediction and selection method, medium and device

Abstract

The present application relates to a candidate drug efficacy prediction and selection method based on a graph neural network, a medium and equipment, first, a biological medicine causal chain is obtained, then an efficacy prediction model is constructed and trained according to the causal chain. Next, after the model is constructed, the target differential protein is determined, and the relationship is determined according to the type of the differential protein. Finally, the differential protein is taken as the tail entity, the relationship is taken as the edge, the candidate drug is taken as the head entity, the triple is constructed, the prediction model is input, and the efficacy prediction result of each candidate drug is output. Overall, the present application solves the problem that the prior art cannot efficiently integrate and mine the multi-dimensional association relationship among drugs, target points and diseases, constructs an interpretable, scalable and updatable knowledge network, reduces the difficulty of drug discovery, shortens the research and development cycle and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph neural networks, in particular to a candidate drug efficacy prediction and selection method based on a graph neural network, a medium and equipment. BACKGROUND

[0002] Drug development is a systematic project with long cycle, high cost and extremely high failure rate. On the one hand, the pathogenesis of known diseases is complex and diverse, and the causes can reach thousands of kinds, and different causes often act on completely different molecular targets in cells; on the other hand, the number of candidate drug entities (including small molecule compounds, polypeptides, proteins, nucleic acids, etc.) available for screening or optimization also reaches tens of thousands. If traditional in vivo and in vitro pharmacological experiments are carried out for each pair of “drug-target” or even “target-disease” combinations, not only a large amount of manpower, reagents, animals and clinical resources need to be invested, but also there are bottlenecks such as long experimental cycle, data update lag and poor scalability, resulting in low development efficiency and high cost. More seriously, with the increasing refinement of disease typing and the exponential expansion of candidate molecular space, completely relying on “trial and error” type experimental verification has been unable to meet the urgent need for rapid and accurate development of new drugs.

[0003] Chinese application with publication number CN 114420221 A discloses a “multi-task drug screening method and system based on knowledge graph assistance”. However, the knowledge graph of this application only contains drug, compound, target, protein entity, completely missing the “disease” entity, and cannot establish the correlation between drug and disease efficacy. At the same time, this application mainly trains “drug-target interaction” and “compound-protein interaction” cross-entropy classifiers, and takes the average of the probabilities of the two paths as the final score. It extracts and trains the respective features, without effectively modeling the correlation in the knowledge graph, lacking modeling and utilization of the direction of the relationship and the polarity of biological regulation (such as up / down regulation, activation / inhibition, etc.) as well as individualized weighting aggregation among multiple targets. This application is disconnected from the core needs of “finding effective drugs for diseases” in actual drug development, and cannot directly support clinical drug decision-making.

[0004] Chinese application CN 112562791 A discloses a "drug target action deep learning prediction system based on knowledge graph, computer device and storage medium", the knowledge graph only contains "drug and target" two entities, and the relationship is only the fixed "Drug X has Target Y" (single interaction relationship), completely missing "disease" and "differential protein" entities and regulation / correlation heterogeneous logic relationship, and cannot carry directional biological semantics. In addition, it uses the symmetric bilinear DistMult method, which will lead to the lack of asymmetry and directionality in node and relationship embedding in biomedical networks, making it difficult to use edge properties as the core factor to drive multi-node and multi-hop drug efficacy reasoning. This single structure makes the application only predict the interaction between drugs and targets at the molecular level, cannot associate the complete omics background of the disease, and cannot answer the core question of "whether the drug is effective for a certain disease". It is seriously out of touch with the "drug development based on disease pathogenesis" scenario.

[0005] More importantly, the existing technology method is to extract individual features of drugs, target points, etc., and fuse and train the respective features, rather than training and updating the knowledge graph itself, without paying attention to its complex associations and influences. Therefore, how to efficiently integrate and mine the multi-dimensional association relationship between drugs, target points and diseases, and construct an interpretable, scalable and updatable knowledge network to reduce the difficulty of drug discovery and shorten the research and development cycle has become a core technical problem to be solved in the field. SUMMARY

[0006] To solve at least one of the above technical problems, the present application provides a candidate drug efficacy prediction method based on a graph neural network, comprising:

[0007] Obtaining a causal chain of biological medicine;

[0008] According to the causal chain, a drug efficacy prediction model is constructed and trained, including: an input layer for constructing a biomedical directed graph according to the causal chain and generating drug, relationship and target triplets; the nodes of the biomedical directed graph at least include three medical entities of drug, target and disease; the edges represent the relationship between entities; a graph neural network for feature diffusion and updating of entities and relationships between entities of the biomedical directed graph to obtain updated entities and relationships between entities; an output layer for scoring the triplets according to the updated entities and relationships between entities to output the drug efficacy prediction result;

[0009] Determining a target differential protein and determining the relationship according to the type of the target differential protein;

[0010] The difference protein is taken as a tail entity, the relationship is taken as an edge, and the candidate drug is taken as a head entity, a triple is constructed, a prediction model is input, and a drug efficacy prediction result of each candidate drug is output.

[0011] Further, in the input layer, a node embedding unit and an edge embedding unit are included to construct the biomedical directed graph.

[0012] The node embedding unit is configured to set a unique identifier for each medical entity according to a category, so as to define nodes of the biomedical directed graph.

[0013] The edge embedding unit is configured to extract different relationship categories, relationship properties and relationship directions between entities according to the causal chain, so as to define types, polarities and directions of different edges, and construct the biomedical directed graph.

[0014] Further, in the input layer, a triple generator is included to generate drug, relationship and target triplets; and the input layer includes:

[0015] The positive sample input unit is configured to input the drug, relationship and target triplets as positive samples;

[0016] The negative sample generation unit is configured to replace the head entity or / and the tail entity of the positive sample to generate negative samples.

[0017] Further, the positive sample input unit is further configured to:

[0018] According to the data source of each positive sample, the basic strength of each positive sample is determined;

[0019] According to the data quantity of each positive sample, the additional strength of each positive sample is determined;

[0020] According to the basic strength and the additional strength of each positive sample, the evidence strength of each positive sample is determined comprehensively;

[0021] According to the evidence strength of each positive sample, a probability bias mechanism is introduced, so that the probability of the positive sample with high evidence strength is higher than the probability of the positive sample with low evidence strength.

[0022] Further, the negative sample generation unit is further configured to:

[0023] According to the evidence strength of the positive sample, a quantity bias mechanism is introduced, so that the number of negative samples generated by the positive sample with high evidence strength is more than the number of negative samples generated by the positive sample with low evidence strength.

[0024] Further, the graph neural network is configured to:

[0025] The nodes and edges in the biomedical directed graph are converted into initial embedding vectors through unique identifiers to obtain initial node embedding vectors and initial edge embedding vectors, and sparse feature modeling is realized;

[0026] The information of the adjacent nodes and edges is aggregated, and the initial node embedding vector and the initial edge embedding vector are updated layer by layer to obtain the updated node embedding vector and the edge embedding vector.

[0027] Further, an output layer is configured to:

[0028] According to the updated entities and the relationships between the entities, a boundary loss function is constructed to maximize the score gap between the positive samples and the negative samples.

[0029] The boundary loss function is solved to obtain the efficacy prediction result.

[0030] Further, the relationship is determined according to the type of the target differential protein, including:

[0031] According to the type of the target differential protein, for the high expression protein, a negative relationship is used as an edge, and for the low expression protein, a positive relationship is used as an edge.

[0032] On the other hand, the application also provides a candidate drug selection method based on a graph neural network, including:

[0033] A target differential protein set of the current disease is determined.

[0034] If the target differential protein set includes only one element, the above-mentioned any method is used to determine the efficacy prediction result of each drug on the target, and each drug is sorted according to the efficacy prediction result, and a plurality of candidate drugs are sequentially selected.

[0035] If the target differential protein set includes a plurality of elements, the above-mentioned any method is used to determine the efficacy prediction result of each drug on each target, and the drugs are sorted by weighted summation, and a plurality of candidate drugs are sequentially selected.

[0036] Further, it also includes:

[0037] According to the efficacy prediction result, the drugs corresponding to each target are sorted, and a plurality of candidate drugs are sequentially selected.

[0038] The target differential proteins are divided into up-regulated proteins and down-regulated proteins to obtain an up-regulated protein drug set and a down-regulated protein drug set.

[0039] The expression change fold of each target differential protein in the disease and normal state is obtained, and the efficacy prediction result and the absolute value of the change fold are weighted and averaged to obtain a comprehensive score of each drug.

[0040] According to the comprehensive score of each drug, the candidate drugs in the up-regulated protein drug set and the down-regulated protein drug set are sorted and sequentially selected.

[0041] In another aspect, the present application also provides a computer storage medium storing executable program codes; the executable program codes are used to execute any of the candidate drug efficacy prediction methods or any of the candidate drug selection methods.

[0042] In another aspect, the present application also provides a terminal device comprising a memory and a processor; the memory stores program codes executable by the processor; the program codes are used to execute any of the candidate drug efficacy prediction methods or any of the candidate drug selection methods.

[0043] The present application provides a candidate drug efficacy prediction and selection method based on a graph neural network, a medium and a device. First, a biomedical causal chain is obtained. Then, according to the causal chain, a drug efficacy prediction model is constructed and trained: an input layer is used to construct a biomedical directed graph according to the causal chain and generate drug, relationship and target triplets; a graph neural network is used to update the features of the entities and the relationships between the entities of the biomedical directed graph, to obtain updated entities and relationships between the entities; an output layer is used to score the triplets according to the updated entities and relationships between the entities, to output the drug efficacy prediction results. Next, after the model is constructed, a target differential protein is determined, and the relationship is determined according to the type of the differential protein. Finally, the differential protein is taken as a tail entity, the relationship is taken as an edge, and the candidate drugs are taken as head entities to construct triplets, input the prediction model, and output the drug efficacy prediction results of each candidate drug.

[0044] The key of the present application lies in that: 1. According to the causal chain, the drug, the target and the disease are unified to be a semantic-rich, complete-expressed directed graph, which is consistent with the nature of the directed multi-level correlation of the biomedical system, helps to reveal the complex disease mechanism, and clearly defines the relationship between entities (such as "drug-target", "target-disease", "drug-disease", etc.), and provides interpretable structured prior knowledge for subsequent modeling; 2. Based on the graph neural network, the correlation characteristics of the drug, the target and the disease are iteratively propagated in the graph through message passing, multi-hop reasoning is realized, and the relationship type, polarity and direction are explicitly used as edge attributes to enhance the understanding ability of the model to the biological semantics. The key core lies in that: the feature diffusion is the diffusion of the whole correlation relationship and the causal chain, rather than the update of a single feature, which is the biggest difference from the prior art. A prediction model is obtained, which takes the relationship chain of the head entity, the relationship and the tail entity as input, and the efficacy result as output; 3. The unique encoding of the entity and the relationship can be used to learn the correlation relationship between different entities, without extracting complex individual features of each entity, which can effectively solve the feature sparsity problem faced by large-scale data; 4. On this basis, the differential protein set of the disease is determined; the proteins significantly related to the disease are screened out, the irrelevant signals are filtered out, the redundancy of the graph is reduced, and the focusing degree of the model is improved; the differential proteins are taken as tail entities, the relationship is taken as an edge, and the candidate drugs are taken as head entities, and the model input is constructed, which is input into the prediction model, and the efficacy prediction results of each candidate drug are output. All candidate drugs can be batched and traversed to realize high-throughput virtual screening, greatly improving the efficiency. At the same time, each prediction result can be traced back to a specific (drug→relationship→protein) path, supporting mechanism explanation and experimental verification design, and personalized efficacy prediction can be generated for different patient / subtype differential protein profiles, moving towards precision medicine. Overall, the present application solves the problem that the prior art cannot efficiently integrate and mine the multi-dimensional correlation relationship among the drug, the target and the disease, and constructs an interpretable, extensible and updatable knowledge network to reduce the difficulty of drug discovery, shorten the research and development cycle and other problems. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 A flowchart of one embodiment of the candidate drug efficacy prediction method based on the graph neural network;

[0046] Figure 2 A structural schematic diagram of one embodiment of the efficacy prediction model;

[0047] Figure 3 An inhibitor category schematic diagram of one embodiment;

[0048] Figure 4 An activator category schematic diagram of one embodiment;

[0049] Figure 5 An intersection diagram of the up-regulation protein inhibitor and the down-regulation protein activator for an embodiment. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0051] It should be noted that if the embodiments of the present application involve directional indications, such as up, down, left, right, front, back, etc., the directional indications are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture, and if the certain posture changes, the directional indications also change accordingly. In addition, if the embodiments of the present application involve descriptions such as “first, second”, “S1, S2”, “step one, step two” and the like, such descriptions are only for description purposes, and cannot be understood as indicating or implying the relative importance of the indicated technical features or the number of indicated technical features or indicating the execution order of the method, etc. A person skilled in the art can understand that any modification within the technical concept of the present application without deviating from the main points of the present application should be included in the protection scope of the present application.

[0052] As shown in Figure 1 The present application provides a candidate drug efficacy prediction method based on a graph neural network, which comprises:

[0053] S1: Obtain a causal chain of biological medicine;

[0054] Specifically, but not limited to, integrate professional databases such as DrugBank, GNBR, STRING, etc., track and obtain the causal chain between multiple types of entities such as drugs, target points, and diseases, to subsequently build a multi-source heterogeneous biomedical directed graph.

[0055] S2: According to the causal chain, build and train an efficacy prediction model, as shown in Figure 2 which comprises:

[0056] (I) Input layer

[0057] The input layer is used to build a biomedical directed graph according to the causal chain, and generate drug, relationship, and target point triplets. The nodes of the biomedical directed graph at least include three medical entities: drugs, target points, and diseases; and the edges represent the relationship between the entities.

[0058] Specifically, according to the acquired causal chain, biomedical entities such as drugs (DrugBank ID), targets (UniProt ID), and diseases (MeSH ID) are defined as nodes of the biomedical directed graph; and according to the relationship between entities, edges of the biomedical directed graph are defined to construct the biomedical directed graph and generate drug, relationship, and target triplets.

[0059] Preferably, in the input layer, a node embedding unit and an edge embedding unit are included to construct the biomedical directed graph.

[0060] (1) The node embedding unit is configured to set a unique identifier for each medical entity according to the category to define the nodes of the biomedical directed graph.

[0061] Specifically, taking drugs, targets, and diseases as examples, each type can have thousands of entities. The optional drugs are named a1 to aN, where N represents the number of drugs, and each drug entity is coded with a unique identifier and embedded in the biomedical directed graph as a single node.

[0062] For example, the drug, target, and disease information and the number of optional drugs can be obtained from any public professional database as shown in Table 1, and a unified naming specification (such as UniProt ID for proteins, DrugBank ID for drugs, and MeSH ID for diseases) is used to solve the naming conflict of heterogeneous data.

[0063] Table 1: Number of drug-targets based on each database

[0064]

[0065] (2) The edge embedding unit is configured to extract different relationship categories, relationship properties, and relationship directions between entities according to the causal chain to define the types, polarities, and directions of different edges and construct the biomedical directed graph.

[0066] Specifically: taking the two-way relationship between drugs, targets and diseases as an example, including inhibition, activation, up-regulation, down-regulation, pathogenesis, correlation, etc. as shown in Table 2, different types of edges can be defined with different texts, colors and line types, and different labels can be attached to different edges to clearly define which type of relationship this type of edge represents, and to define different types of edges. When the relationship has opposite properties such as "positive / negative", such as inhibition / activation, up-regulation / down-regulation, etc., the edges can be supplemented with "polarity", such as "positive / negative" or "binary label". The core is to distinguish the "tendency" of the relationship. Positive polarity: represents a "positive, supportive, beneficial" relationship. Negative polarity: represents a "negative, opposing, harmful" relationship. When the relationship has a flow direction, for example: a drug acts on a target, but the target does not act on the drug, there is an initiator (starting point) of the relationship between the two and a receiver (end point) of the relationship. The edges can be further supplemented with "direction". The arrow points from the starting node to the end node, representing "the starting point has a certain relationship with the end point, such as the drug-target relationship", and this relationship is unidirectional (unless an additional reverse edge is added). Reversing the direction can completely change the relationship semantics. If you need to express "bidirectional relationship", you need to add two edges with opposite directions between the two nodes (not a single undirected edge) to construct a biomedical directed graph.

[0067] Table 2: Relationship table between entities

[0068]

[0069] In this embodiment, the preferred embodiment of constructing a biomedical directed graph is given, 1, set a unique identifier for each medical entity by category to define the nodes of the biomedical directed graph; it does not introduce the characteristics of drugs, targets and diseases into the graph to diffuse and update their complex characteristics, but sets a unique identifier to simply encode the nodes, and only uses the position of the node identifier in the graph, the connection relationship, etc. to diffuse and update the characteristics, realizing sparse feature modeling, which can greatly reduce the model complexity and improve the precision; 2, according to the causal chain, different relationship categories, relationship properties and relationship flow directions between entities are extracted to define different edge types, polarities and directions, and a biomedical directed graph is constructed; it is a hierarchical progressive definition from "basic" to "detailed". First define a single node, then define the relationship between nodes; through the "type", "polarity" and "direction" of the edge, which are not independent but progressive "semantic supplement" relationships, the complete definition of the edge is formed, which can fully reflect the relationship between entities and construct a semantic-rich and complete biomedical directed graph.

[0070] Although the node, the type, the polarity, the direction of the edge of the directed graph are basic concepts, in the existing medical knowledge graph, the characteristics of the drug, the target and the disease are fused and trained, not the directed graph of the node + edge, and these detailed problems are not considered, which is the key of the present application. The present application extracts the causal chain of biological medicine and constructs a biomedical directed graph. The diversity, the polarity and the integrity of the direction of this type make the overall semantics of the directed graph more rich and complete. On this basis, the present application is different from the prior art: extracting the single characteristics of a node, comparing the similarity, and fusing the characteristics; the present application updates the characteristics of the entity and the relationship between the entities of the directed graph, can extract and mine the multi-dimensional relationship among the drug, the target and the disease, and improve the prediction accuracy and accuracy.

[0071] More preferably, in the input layer, a triple generator is included to generate a drug, relationship, target triple; including:

[0072] (3) a positive sample input unit for inputting a drug, relationship, target triple positive sample;

[0073] Preferably, the positive sample input unit is also used for:

[0074] According to the data source of each positive sample, the basic strength of each positive sample is determined; for example, the data source can be selected as the type of data source: clinical trial data sample, such as: patient sample in random control test (RCT) “medication group + target marker positive + efficacy positive”, sample of target related index associated with efficacy after medication in real world study (RWS), which can directly reflect the human body effect, has the highest basic strength, and can be given the highest probability; animal model data sample, such as: gene knockout (KO) mouse verification sample (such as drug failure after knocking out target, refuting the necessity of target), disease model drug efficacy positive sample, etc., which can observe the effect of tissue / organ level by simulating the physiological environment in vivo, but there are species differences between animals and human body, the basic strength does not directly act on the human body, and can be given a secondary probability; in vitro experiment data, such as: drug-target binding experiment (SPR, ITC) positive sample, cell pathway activation / inhibition experiment positive sample, which can quickly verify the molecular level causality, but lacks the complexity of in vivo environment, and can be given a secondary probability; in addition, the data source can also be selected as the supplier of the data source, such as the basic weight of the positive sample given by some authoritative agencies, authoritative periodical papers, which is higher than the basic weight of the positive sample given by some newspapers, small agencies;

[0075] According to the data quantity of each positive sample, the additional strength of each positive sample is determined; for example, when extracting the positive sample data chain, a certain positive sample appears repeatedly, and multiple evidence chains point to the sample, so the evidence strength is greater, and the additional strength of the positive sample can be determined according to the data quantity pointing to the positive sample to supplement the strength. From the perspective of random samples, it can also be proved that the same sample has more quantity and is more likely to be selected;

[0076] According to the basic strength and the additional strength of each positive sample, the evidence strength of each positive sample is comprehensively determined;

[0077] According to the evidence strength of each positive sample, a probability bias mechanism is introduced, so that the probability of a positive sample with high evidence strength is higher than that of a positive sample with low evidence strength.

[0078] (4) A negative sample generation unit is configured to replace the head entity or / and the tail entity of the positive sample to generate a negative sample.

[0079] Preferably, the negative sample generation unit further introduces a quantity bias mechanism according to the evidence strength of the positive sample, so that the number of negative samples generated by a positive sample with high evidence strength is greater than the number of negative samples generated by a positive sample with low evidence strength.

[0080] In this embodiment, an optimal embodiment of generating a drug-relation-target triple is given, 1, receiving a positive sample, and adopting a strategy of randomly sampling the head entity or the tail entity to replace the head entity or the tail entity of the positive sample to generate a negative sample; 2, multi-angle determination of the evidence strength of the positive sample, introduction of a probability bias mechanism, so that the probability of a positive sample with high evidence strength being selected is higher than that of a positive sample with low evidence strength; 3, further according to the evidence strength of the positive sample, a quantity bias mechanism is introduced, so that the number of negative samples generated by a positive sample with high evidence strength is greater than the number of negative samples generated by a positive sample with low evidence strength; not only can the model training data be enriched, but more importantly, the model training can be closer to the positive sample with high evidence strength, and the subsequent efficacy prediction accuracy is improved.

[0081] (II) Graph neural network

[0082] The graph neural network is configured to perform feature diffusion update on entities of a biomedical directed graph and relationships between the entities to obtain updated entities and relationships between the entities.

[0083] Preferably, for the complex graph structure of the above multiple types of edges, a relation-aware graph neural network can be selected, including L-layer message passing and updating, which captures different types of relationship information by defining different message propagation mechanisms. Specifically, the state update of each node depends not only on the state of its neighbor nodes, but also on the type of edge connecting these nodes. Preferably, when aggregating neighbor node information, a weight matrix is introduced for each relationship, allowing the model to learn the different effects of different types of relationships on the center node representation.

[0084] Preferably, the step comprises:

[0085] (1) converting a plurality of nodes and edges in the biomedical directed graph into initial embedding vectors through unique identification, obtaining initial node embedding vectors and initial edge embedding vectors, and realizing sparse feature modeling;

[0086] Specifically, since GNN (such as R-GCN, CompGCN, GAT) can learn the low-dimensional embedding of entities and relationships, capture high-order neighbor information, and break through the linear limitations of traditional KG embedding (such as TransE), in the process of biomedical directed graph embedding modeling, graph neural network (Graph Neural Network, GNN) can be selected to diffuse the features of entities and relationships in the knowledge graph, i.e., nodes and edges, for the prediction task of drugs. In the form of one-hot encoding, initial embedding vectors are generated for each node and edge through unique identification, serving as input features of the model, and realizing sparse feature modeling.

[0087] More specifically, based on one-hot encoding, original embedding vectors are constructed for nodes and edges in the knowledge graph. For head entity, relationship, and tail entity triplets , the original embedding vectors of the head entity and the tail entity can be represented as , and the original embedding vector of the relationship can be represented as Through learning and optimization of these embedding vectors, the model can capture the deep semantic information between entities and relationships, thereby improving the accuracy of drug prediction.

[0088] (2) aggregating the information of adjacent nodes and edges, and updating the initial node embedding vectors and the initial edge embedding vectors layer by layer to obtain updated node embedding vectors and edge embedding vectors;

[0089] Specifically, in graph neural networks, a feature propagation method based on message passing mechanism can be adopted to iteratively propagate the features of drugs, targets, and diseases in the graph, realizing multi-hop reasoning (e.g., drug → target A → pathway → target B → disease). At the same time, the "relationship type" is explicitly used as the edge attribute (e.g., "inhibition", "activation", etc., "upregulation" and "downregulation" relationships) to enhance the model's understanding of biological semantics. The information of adjacent nodes is aggregated, and the node embedding vector is updated layer by layer to obtain the updated node embedding vector and edge embedding vector. By aggregating the adjacent information, the model can learn the local structural features of nodes in the graph, enhancing the representational ability of the embedding vector. As the number of layers increases, the node embedding vector can integrate information from more distant neighbors, thereby improving the model's expressive ability. The mathematical expression for feature propagation can be as shown in Equation (1). At the same time, since the edge embedding vector is only related to the type and is not related to the initial representation content, the edge embedding update method can be as shown in Equation (2).

[0090] (1)

[0091] (2)

[0092] in, Indicates the first Layer node characteristics, For trainable parameters, The sigmoid is a non-linear activation function. Represents a node Neighbors Indicates the first Layer edge features. Through this formula, the features of a node not only contain its own information, but also effectively aggregate the features of its neighboring nodes, thereby capturing the local structure and global semantics in the knowledge graph. Graph neural networks can propagate and aggregate the features of nodes and edges layer by layer, so that the structural and semantic information in the knowledge graph can be fully expressed.

[0093] This embodiment presents a preferred implementation of a graph neural network. The model assigns trainable initial vectors to each type of entity and relation, and then aggregates neighbor information layer by layer on the graph using a message-passing-based graph neural network. Each layer combines the entity's own features with "messages" transmitted through different relation types, and then undergoes learnable transformations and non-linear activations to obtain progressively context-rich representations. To express directionality and semantic differences, inhibition, activation, and association relations participate in propagation as independent relation vectors, enabling the model to explicitly distinguish different edge types and directions when updating node representations. As the number of layers increases, nodes not only absorb local information from first-order neighbors but also integrate a wider range of structural and semantic cues.

[0094] III. Output layer

[0095] The output layer is configured to score the triplets according to the updated entities and the relationships between the entities to output the efficacy prediction result.

[0096] Specifically, the step includes:

[0097] (1) constructing a margin loss function according to the updated entities and the relationships between the entities, such as the node embedding vectors and the edge embedding vectors, so that the score difference between the positive samples and the negative samples is maximized;

[0098] (2) solving the margin loss function to obtain the efficacy prediction result.

[0099] Specifically, in order to optimize the embedding learning of the knowledge graph, a classical TransE (Translating Embeddings) framework can be optionally adopted as a representation target to map the entities (such as drugs, targets, and diseases) and relationships in the knowledge graph to a low-dimensional continuous vector space, and guide the triplets (head entity, relationship, and tail entity) to satisfy to ensure that the knowledge graph can still maintain its semantic information after being embedded into the low-dimensional vector space, and ensure that the embedding of the knowledge graph in the low-dimensional vector space can preserve its original semantic information, and then construct a margin loss function according to the updated node embedding vectors and edge embedding vectors, so that the score difference between the positive samples and the negative samples is maximized. By minimizing the margin loss (the opposite number of the score difference), the model can better distinguish between positive and negative samples, improve the accuracy of the prediction, and finally solve the margin loss function to obtain the efficacy prediction result.

[0100] In an example, a margin loss function can be defined during model training to maximize the score difference between the positive samples and the negative samples. The specific form of the loss function can be as shown in equations (3)-(5):

[0101] (3)

[0102] (4)

[0103] (5)

[0104] wherein, and represent the positive sample set and the negative sample set, respectively, is a positive sample scoring function, is a negative sample scoring function, is a hyperparameter; , are the embeddings of the head entity, the relationship, and the tail entity of the positive sample, respectively. , These are the embeddings of the negative sample head entity, relation, and tail entity, respectively.

[0105] In this embodiment, a preferred embodiment of the output layer is given. During the training phase, a distance-based objective is used to constrain entities and relationships. For each real sample, the head or tail entity is randomly replaced to generate a negative sample. The score difference between the real and negative samples is calculated and the corresponding ranking loss is minimized, which enables the model to continuously bring the real association closer and distance the erroneous association.

[0106] S3: Identify the target differentially expressed protein and determine the relationship based on the type of differentially expressed protein;

[0107] Specifically, since differentially expressed proteins are often pathogenic drivers or therapeutic response biomarkers, providing a biological basis for drug action, differential expression analysis methods can be used, but are not limited to, in combination with transcriptomic and proteomic data to reflect disease-specific molecular characteristics, screen for protein molecules that change significantly in disease states, filter out irrelevant signals, reduce map redundancy, improve model focus, and verify their reliability through statistical models (such as t-tests or analysis of variance) to construct target differentially expressed proteins for the disease. If the disease includes multiple target differentially expressed proteins, a set of target differentially expressed proteins can be constructed to provide key inputs for subsequent target identification and mechanism analysis.

[0108] Taking the CryAB (R120G) mutation as an example, when screening and feature extraction of differentially expressed proteins to determine the differentially expressed protein set for a disease, proteomics analysis of CryAB (R120G) mutant and wild-type mouse heart tissues can be performed to identify differentially expressed proteins. The limma package was used for differential analysis, and the screening conditions were as follows: And FDR < 0.05, the target differential protein set is obtained.

[0109] More preferably, based on the target protein type, negative relationships are used as edges for highly expressed proteins and positive relationships are used as edges for low-expressed proteins; specifically, negative relationships can be, but are not limited to, the aforementioned downregulation, inhibition, etc.; positive relationships can be, but are not limited to, the aforementioned upregulation, activation, action, etc.

[0110] More preferably, when there are multiple relations, the same head and tail entities will be assembled into different triples, such as (head entity, relation 1, tail entity) and (head entity, relation 2, tail entity), and the two triples will be fed into the model as independent samples for calculation.

[0111] S4: Construct triples by treating differentially expressed proteins as tail entities, relationships as edges, and iterating through candidate drugs as head entities. Input these triples into the prediction model and output the efficacy prediction results for each candidate drug.

[0112] Specifically, but not limited to, the differential proteins in the differential protein set obtained in step S3 are selected as tail entities, and the differential protein expression direction is aligned with the drug action relationship to realize the prediction of biological logic consistency. For high expression proteins, a negative relationship is used as an edge, that is, the drug should "inhibit" the protein or its upstream. For low expression proteins, a positive relationship is used as an edge, that is, the drug should "activate" the protein or its pathway. Then, the candidate drugs are used as head entities, and the model input is constructed and input into the prediction model to calculate the relevance of the triplets, thereby realizing high-throughput virtual screening of all candidate drugs, greatly improving the efficiency, and finally outputting the efficacy prediction results of each candidate drug. At this time, each efficacy prediction result obtained can be traced back to a specific (drug→ relationship→ protein) path, supporting mechanism interpretation and experimental verification design, and personalized efficacy prediction can be generated for different patient / subtype differential protein profiles, moving towards precision medicine.

[0113] More preferably, according to the number of elements in the target differential protein set, a single target and multi-target strategy can be used for joint analysis during drug screening.

[0114] In this embodiment, a candidate drug efficacy prediction method based on a graph neural network is given. First, a biomedical causal chain is obtained. Then, according to the causal chain, a drug efficacy prediction model is constructed and trained: an input layer is used to construct a biomedical directed graph according to the causal chain and generate drug, relationship, and target triplets; a graph neural network is used to update the features of the entities and the relationships between the entities in the biomedical directed graph, and obtain updated entities and relationships between the entities; an output layer is used to score the triplets according to the updated entities and relationships between the entities, and output the drug efficacy prediction results. Next, after the model is constructed, the target differential proteins are determined, and the relationships are determined according to the types of the differential proteins. Finally, the differential proteins are used as tail entities, the relationships are used as edges, and the candidate drugs are used as head entities to construct triplets, input the prediction model, and output the drug efficacy prediction results of each candidate drug.

[0115] The key of the present application lies in that: 1. According to the causal chain, the drug, target and disease are unified to be a semantic-rich and complete-expressed directed graph, which is consistent with the nature of the directed multi-level association of the biomedical system, helps to reveal the complex disease mechanism, and clearly defines the relationship between entities (such as "drug-target", "target-disease", "drug-disease", etc.), and provides interpretable structured prior knowledge for subsequent modeling; 2. Based on the graph neural network, the correlation characteristics of the drug, target and disease are iteratively propagated in the graph through message passing, multi-hop reasoning is realized, and the relationship type, polarity and direction are explicitly used as edge attributes to enhance the model's understanding of biological semantics. The key core lies in that: the feature diffusion is the diffusion of the graph structure, the whole correlation relationship and the causal chain, rather than the update of a single feature, which is the biggest difference from the prior art. A prediction model is obtained, which takes the relationship chain of the head entity, relationship and tail entity as input, and the efficacy result as output; 3. The unique encoding of the entity and the relationship can be used to learn the correlation between different entities, without extracting complex individual features of each entity, which can effectively solve the feature sparsity problem faced by large-scale data; 4. On this basis, the differential protein set of the disease is determined; the proteins significantly related to the disease are screened out, the irrelevant signals are filtered out, the redundancy of the graph is reduced, and the focusing degree of the model is improved; the differential proteins are taken as tail entities, the relationship is taken as an edge, and the candidate drugs are taken as head entities, and the model input is constructed, which is input into the prediction model, and the efficacy prediction results of each candidate drug are output. All candidate drugs can be batched and traversed to realize high-throughput virtual screening, greatly improving the efficiency. At the same time, each prediction result can be traced back to a specific (drug→relationship→protein) path, supporting mechanism explanation and experimental verification design, and personalized efficacy prediction can be generated for different patient / subtype differential protein profiles, moving towards precision medicine. Overall, the present application solves the problem that the prior art cannot efficiently integrate and mine the multi-dimensional correlation between drugs, targets and diseases, and constructs an interpretable, extensible and updatable knowledge network to reduce the difficulty of drug discovery, shorten the research and development cycle, etc.

[0116] Preferably, after obtaining the efficacy prediction results of each candidate drug, the method further comprises:

[0117] S5: converting the efficacy prediction results into a probability representation to obtain optimized efficacy prediction results.

[0118] Specifically, since the efficacy prediction result obtained in step S4 can be any real number (such as -4.3 ~ 6.7), different batches, different disease model scales are inconsistent, therefore, it can be converted into a probability representation, compressed to the interval [0, 1], eliminate the dimension and distribution difference, obtain the optimized efficacy prediction result, so that the efficacy prediction results of different types of drugs can be directly compared, and the output of the original model is more intuitive.

[0119] For example, for a certain target and relationship , the efficacy score of N candidate drugs is Optionally, the correlation of all drugs is normalized to a probability by softmax as shown in equations 3-5:

[0120] 3-5

[0121] On the other hand, the application also provides a candidate drug selection method based on a graph neural network, comprising:

[0122] P1: determining a target differential protein set of the current disease;

[0123] Specifically, the same method as step S3 can be used to determine the target differential protein set of the current disease. Optionally but not limited to, it is judged whether the set includes several elements, that is, the corresponding disease includes several target differential proteins, which provides branching logic basis for subsequent steps, ensures that single-target diseases use traditional single-target optimization strategies, and multiple-target diseases use multiple-target synergistic strategies, thereby improving the research and development efficiency.

[0124] P2: if the target differential protein set includes only one element, then using any of the above methods, determining the efficacy prediction result of each drug on the target, and sorting the drugs according to the efficacy prediction result, and sequentially selecting several candidate drugs;

[0125] Specifically, when the corresponding disease only contains one differential protein, the efficacy prediction result of each drug on the differential protein can be obtained according to the candidate drug efficacy prediction method based on the graph neural network, and the drugs can be sorted according to the efficacy prediction result, and several candidate drugs can be sequentially selected as the final determined drug set for further screening or use by the staff.

[0126] P3: if the target differential protein set includes several elements, then using any of the above methods, determining the efficacy prediction result of each drug on each target, and sequentially selecting several candidate drugs.

[0127] Specifically, when the corresponding disease contains multiple differential proteins, the weight of each target point can be assigned according to the specific site of the target point in the human body or the density and the like, and the final score result of each drug is obtained by weighted summation of the drug efficacy prediction results of each target point. Finally, the drugs are sorted according to the final score result, and a number of candidate drugs are selected in sequence as the final determined drug set for further screening or use by the staff.

[0128] For example, in multi-target screening, the candidate drugs can be comprehensively evaluated in combination with the differential protein information related to each target point. First, for each target point, the top 50 candidate drugs are screened, and after deduplication, a drug set is obtained. Then, the drug prediction score is calculated by weighting according to the importance weight of the differential protein. For a certain candidate drug d, the comprehensive score calculation formula can be shown in formula (6):

[0129] (6)

[0130] Among them, represents the weight of the target point, which is weighted according to the fold change of the differential protein to reflect its importance in the disease process, reflects the prediction score of the drug, reflects the number of up-regulated or down-regulated proteins.

[0131] Preferably, since the performance of each differential protein may be different or even opposite under the same disease state, such as some proteins may be abnormally highly expressed (up-regulated proteins), while others may be functionally low or insufficiently expressed (down-regulated proteins), therefore, when selecting drugs, it is necessary to distinguish them for targeted treatment. Preferably, step P3 further comprises:

[0132] P31: According to the drug efficacy prediction result, the drugs corresponding to each target point are sorted, and a number of candidate drugs are selected in sequence;

[0133] P32: The differential proteins are divided into up-regulated proteins and down-regulated proteins to obtain an up-regulated protein drug set and a down-regulated protein drug set;

[0134] Specifically, the differential proteins can be divided into up-regulated proteins and down-regulated proteins to obtain an up-regulated protein drug set and a down-regulated protein drug set according to the candidate drugs obtained in step P31, so as to facilitate targeted treatment of up-regulated proteins and down-regulated proteins.

[0135] In the example, optional selection of the inhibitor (such as antagonist, degrading agent) reduces the activity or expression level of the up-regulated protein, and the agonist or expression enhancer restores the function or expression level of the down-regulated protein, so as to ensure that the direction of the drug action is consistent with the molecular mechanism of the disease, avoid "reverse regulation" to cause poor curative effect or side effects, and at the same time, complex diseases (such as cancer, neurodegenerative diseases) usually involve multiple abnormal targets, by classifying the drugs according to the regulation direction of the target, a synergistic combination scheme can be systematically designed, for example, the cancer-promoting protein (up-regulation) is inhibited and the cancer-inhibiting protein (down-regulation) is activated at the same time, so as to realize "two-way regulation" to enhance the curative effect.

[0136] P33: Obtain the expression change fold of each protein in the disease and normal state, and weighted average the absolute value of the change fold and the drug efficacy prediction result to obtain the comprehensive score of each drug;

[0137] Specifically, since the change degree of the protein can reflect the severity of the disease to a certain extent, the expression change fold of each protein in the disease and normal state is selected for weighting of each drug, and the absolute value of the change fold and the drug efficacy prediction result are weighted and averaged to obtain the comprehensive score Sd of each drug.

[0138] P34: According to the comprehensive score of each drug, the up-regulated protein drug set and the down-regulated protein drug set are sorted in sequence, and a plurality of drugs are selected as the final determined drug set.

[0139] Specifically, a plurality of candidate drugs with high scores can be selected from the up-regulated protein drug set and the down-regulated protein drug set, respectively, to obtain a drug set for acting on the up-regulated protein and a drug set for acting on the down-regulated protein, which are the final determined drug set, and provide a reliable theoretical basis for subsequent experimental verification.

[0140] In this embodiment, a candidate drug selection method based on a graph neural network is given, which provides a bifurcation logic basis for subsequent steps by judging whether the corresponding disease contains multiple differential proteins, and by distinguishing the complexity of the disease (single protein vs. multiple proteins) early, the subsequent drug screening process can be simplified, and the research and development efficiency can be improved; if not, the drugs are sorted according to the drug efficacy prediction result, and a plurality of candidate drugs are selected in sequence as the final determined drug set, since only a single target needs to be considered, the interference factors of drug screening are less, and therefore the top-ranked drugs can be directly selected as candidates after sorting, thereby shortening the drug research and development cycle; if yes, the drug efficacy prediction results of each target are weighted and summed to obtain the final score result of each drug, and a plurality of candidate drugs are selected in sequence as the final determined drug set, and more robust drug scores are obtained by weighting multi-dimensional drug efficacy data (such as activity, selectivity, toxicity prediction, etc.). The research and development efficiency of the prior art is improved, and the research and development cost is reduced.

[0141] For example, based on the 27 differentially expressed proteins obtained by screening, a target-drug interaction network was constructed for drug repositioning for CryAB mutant cardiomyopathy. The network construction rules are as follows: for the 16 up-regulated proteins, the "down-regulation regulation" relationship edges between them and potential inhibitors are established; for the 11 down-regulated proteins, the "up-regulation regulation" relationship edges between them and potential activators are established. Using the above-mentioned candidate drug efficacy prediction method based on graph neural network, the control matching degree of drug molecules to the target network was calculated through the pre-trained vector space model (loss function using marginal loss). Specifically, for each up-regulated protein, compounds that can significantly inhibit its function are predicted; for each down-regulated protein, compounds that can effectively activate its activity are predicted. Finally, the top 20 inhibitors with the highest matching degree were screened out, as shown in Figures 3-4 Figure 3: Deep learning-based multi-target drug repositioning screening. Up-regulated protein inhibitor weighted ranking: based on the transfer learning of the known drug-target relationship by the TransE deep learning model, the top 20 candidate drugs that can inhibit the differentially up-regulated proteins are screened, and the drug efficacy score reflects the targeting regulation potential; Down-regulated protein activator weighted ranking: for differentially down-regulated proteins, the top 20 drug scores that can promote their functions are calculated, and the height of the column chart represents the strength of the multi-target synergistic effect.

[0142] As shown in Figure 5 Figure 4: There are 11 intersection drugs between up-regulated protein inhibitors and down-regulated protein activators, and the intersection drugs are shown in Table 3. Further analysis shows that the intersection drugs cover chemotherapeutic drugs and antiparasitic drugs, and their effects may simultaneously correct protein expression imbalance through multi-target effect. It is worth noting that raloxifene (selective estrogen receptor modulator) and digoxin were simultaneously predicted as key co-regulation drugs, the former may affect lipid metabolism through estrogen receptor beta-mediated PPAR gamma synergistic effect, and the latter may enhance myocardial contractility by changing intracellular calcium homeostasis, which is consistent with the functional compensation phenotype formed by the inhibition of the contraction pathway found in the previous study.

[0143] Table 3: Intersection drug table

[0144]

[0145] In this specific example, the present application employs the above-mentioned drug efficacy prediction method, constructs a biomedical directed graph by integrating multiple biomedical databases, and successfully develops a drug efficacy prediction model for CryAB mutation-related dilated cardiomyopathy (DRC) by combining graph neural network technology. The model is based on the PyTorch framework and is trained efficiently using NVIDIA A100 GPU. It has the ability to quickly discover potential therapeutic drugs. During the training process, the batch size is set to 32, the total number of training rounds is set to 20 rounds, and after each round of training, the validation set is evaluated to monitor the performance of the model and prevent overfitting. Compared with traditional drug screening methods, the present application significantly improves the efficiency and accuracy of drug discovery, and can accurately identify candidate drugs that can intervene in the abnormal aggregation of CryAB protein, relieve stress damage to myocardial cells, and energy metabolism disorders. Its beneficial effects lie in providing a new targeted treatment strategy for CryAB mutation DRC, which currently has no radical cure, and is expected to promote the development of individualized treatment, improve patient prognosis, and reduce the risk of heart failure progression, which has important clinical translation value and social significance.

[0146] In another aspect, the present application also provides a computer storage medium storing executable program codes; the executable program codes are used to execute any of the above-mentioned drug efficacy prediction methods or selection methods of candidate drugs based on graph neural networks.

[0147] In another aspect, the present application also provides a terminal device comprising a memory and a processor; the memory stores program codes executable by the processor; the program codes are used to execute any of the above-mentioned drug efficacy prediction methods or selection methods of candidate drugs based on graph neural networks.

[0148] For example, the program codes can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. The one or more modules / units can be a series of computer program instructions that can complete a specific function, which are used to describe the execution process of the program codes in the terminal device.

[0149] The terminal device can be a desktop computer, a notebook, a palm computer, and a cloud server, etc. The terminal device can include, but is not limited to, a processor, a memory. Those skilled in the art can understand that the terminal device can also include input / output devices, network access devices, buses, etc.

[0150] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0151] The memory can be an internal storage unit of the terminal device, such as a hard disk or a memory. The memory can also be an external storage device of the terminal device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory can also include both the internal storage unit and the external storage device of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output.

[0152] The above candidate drug selection method, computer storage medium and terminal device are based on the above-mentioned candidate drug efficacy prediction method, and the technical effects and advantages will not be repeated here. The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above-mentioned embodiments are not described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0153] The above-mentioned embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, a number of variations and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for predicting drug efficacy of a candidate drug based on a graph neural network, characterized in that, The application relates to a method for predicting drug efficacy based on a causal chain of biological medicine. The method comprises the following steps: acquiring a causal chain of biological medicine; constructing and training a drug efficacy prediction model according to the causal chain, which comprises an input layer, a graph neural network and an output layer; the input layer is used for constructing a biomedical directed graph and generating drug, relationship and target triplets according to the causal chain; nodes of the biomedical directed graph at least include three medical entities of drugs, targets and diseases; edges represent the relationships between the entities; the graph neural network is used for updating the features of the entities of the biomedical directed graph and the relationships between the entities to obtain updated entities and relationships between the entities; and the output layer is used for scoring the triplets according to the updated entities and relationships between the entities to output a drug efficacy prediction result; determining a target differential protein and determining a relationship according to the type of the target differential protein; constructing triplets by taking the differential protein as a tail entity, taking the relationship as an edge and taking a candidate drug as a head entity, inputting the triplets into the prediction model and outputting drug efficacy prediction results of the candidate drugs. In the input layer, a node embedding unit and an edge embedding unit are included to construct the biomedical directed graph. The node embedding unit is used for setting a unique identifier for each medical entity according to the category to define the nodes of the biomedical directed graph. The edge embedding unit is used for extracting different relationship categories, relationship properties and relationship directions between the entities according to the causal chain to define the types, polarities and directions of different edges and construct the biomedical directed graph.

2. The candidate drug drug efficacy prediction method according to claim 1, characterized by, In the input layer, a triplet generator is included to generate drug, relationship and target triplets; the triplet generator comprises a positive sample input unit and a negative sample generation unit. The positive sample input unit is used for inputting positive sample triplets of drugs, relationships and targets. The negative sample generation unit is used for replacing the head entity or / and the tail entity of the positive sample to generate negative samples.

3. The candidate drug drug efficacy prediction method according to claim 2, characterized by, The positive sample input unit is also used for: determining the basic strength of each positive sample according to the data source of the positive sample; determining the additional strength of each positive sample according to the data quantity of the positive sample; 4. The candidate drug drug efficacy prediction method according to claim 3, characterized by, comprehensively determining the evidence strength of each positive sample according to the basic strength and the additional strength of the positive sample; introducing a probability bias mechanism according to the evidence strength of each positive sample, so that the probability of the positive sample with high evidence strength is higher than that of the positive sample with low evidence strength. The negative sample generation unit is also used for: introducing a quantity bias mechanism according to the evidence strength of the positive sample, so that the number of negative samples generated by the positive sample with high evidence strength is more than that of negative samples generated by the positive sample with low evidence strength. The graph neural network is used for:

5. The candidate drug drug efficacy prediction method according to claim 4, characterized by, converting a plurality of nodes and edges in the biomedical directed graph into initial embedding vectors through unique identifiers to obtain initial node embedding vectors and initial edge embedding vectors and realize sparse feature modeling; aggregating the information of adjacent nodes and edges, updating the initial node embedding vectors and the initial edge embedding vectors layer by layer to obtain updated node embedding vectors and edge embedding vectors.

6. The candidate drug drug efficacy prediction method according to claim 1, wherein, The output layer is used for: constructing a boundary loss function according to the updated entities and relationships between the entities to maximize the score gap between the positive samples and the negative samples; solving the boundary loss function to obtain a drug efficacy prediction result.

7. The candidate drug drug efficacy prediction method according to claim 1, wherein, Determining the relationship according to the type of the target differential protein comprises: ​ ​ 8. The candidate drug efficacy prediction method according to any one of claims 1 to 7, characterized in that, ​ According to the type of the target differential protein, for the high expression protein, a negative relationship is used as an edge, and for the low expression protein, a positive relationship is used as an edge. 9.A method for candidate drug selection based on a graph neural network, characterized in that, The method comprises the following steps: determining a target differential protein set of the current disease; if the target differential protein set only includes one element, the method of any one of claims 1-8 is used to determine the drug efficacy prediction result of each drug on the target point, and each drug is sorted according to the drug efficacy prediction result, and a plurality of candidate drugs are sequentially selected; if the target differential protein set includes a plurality of elements, the method of any one of claims 1-8 is used to determine the drug efficacy prediction result of each drug on each target point, and the drug efficacy prediction results of each drug are weighted and summed to sort each drug, and a plurality of candidate drugs are sequentially selected.

10. The candidate drug selection method according to claim 9, characterized in that, Further comprising: According to the drug efficacy prediction result, the drugs corresponding to each target point are sorted, and a plurality of candidate drugs are sequentially selected; dividing the target differential protein into up-regulated proteins and down-regulated proteins to obtain an up-regulated protein drug set and a down-regulated protein drug set; obtaining the expression change fold of each target differential protein in the disease and normal state, and weighting and averaging the drug efficacy prediction result and the absolute value of the change fold to obtain a comprehensive score of each drug; According to the comprehensive score of each drug, the candidate drugs in the up-regulated protein drug set and the down-regulated protein drug set are sorted and sequentially selected.

Images