Knowledge map construction method, map encoder training method, and drug research and development method

Abstract

The present disclosure relates to a knowledge map construction method, a map encoder training method, and a drug research and development method. The knowledge map construction method comprises: extracting a plurality of entities and relationships among the plurality of entities from data related to drugs and diseases; and on the basis of the extracted entities and relationships, constructing an initial knowledge map, wherein the entities are used for constituting nodes in the initial knowledge map, and the relationships are used for constituting edges between the nodes in the initial knowledge map. The plurality of entities comprise prescription entities and drug entities that belong to a first group and disease entities belonging to a second group, and further comprise at least one of the following: one or more first entities that belong to the first group and are located at lower levels than the drug entities; or one or more second entities that belong to the second group and are located at lower levels than the disease entities.

Description

Knowledge graph construction methods, graph encoder training methods, and drug development methods Technical Field

[0001] This disclosure relates to the field of drug development, and more specifically, to methods for constructing knowledge graphs, methods for training graph encoders based on knowledge graphs, methods for constructing knowledge graphs using graph encoders, and drug development methods using knowledge graphs and graph encoders, as well as electronic devices, non-transient storage media, and computer program products associated with these methods. Background Technology

[0002] Drugs in naturally occurring forms (referred to as natural drugs in this text) can be obtained and used directly from nature, for example, through physical methods (crushing, drying) or extraction methods (solvent extraction, distillation). Some natural drugs with known components and structures can also be obtained through artificial synthesis, for example, through chemical synthesis or semi-synthesis, which is usually based on the chemical structure of the active ingredient in the natural drug.

[0003] Traditional Chinese medicine (TCM) refers to medicines used in China based on traditional medical theories for the prevention, treatment, and diagnosis of diseases, or for health maintenance and conditioning of the body. It extensively utilizes natural medicines, including plant-based, animal-based, and mineral-based drugs. Besides TCM, other ethnic medicines such as Tibetan and Miao medicines (medicines developed by various ethnic groups based on their own traditional medical theories and natural resources), other national medicines such as Burmese and Indian medicines (medicines developed by different countries based on their own traditional medical theories and natural resources), and folk medicines (medicines not part of the official medical system) also widely utilize natural medicines. Summary of the Invention

[0004] A brief overview of this disclosure is given below to provide a basic understanding of some aspects of it. However, it should be understood that this overview is not an exhaustive summary of this disclosure. It is not intended to identify key or essential parts of this disclosure, nor is it intended to limit the scope of this disclosure. Its purpose is merely to present certain concepts of this disclosure in a simplified form as a prelude to the more detailed description that follows.

[0005] According to a first aspect of this disclosure, a method for training a graph encoder is provided, comprising: acquiring a knowledge graph constructed based on a plurality of entities extracted from data related to drugs and diseases and relationships between the plurality of entities, wherein the entities constitute nodes in the knowledge graph and the relationships constitute edges between nodes in the knowledge graph; and training a graph encoder constructed based on graph deep learning using the acquired knowledge graph through unsupervised learning, the graph encoder being configured to extract feature vectors of nodes in the knowledge graph. The plurality of entities includes prescription entities and drug entities belonging to a first group and disease entities belonging to a second group, and the plurality of entities further includes at least one of the following: a first one or more entities belonging to the first group at a lower level than the drug entities; or a second one or more entities belonging to the second group at a lower level than the disease entities.

[0006] In some embodiments, the first one or more entities include at least one of a biochemical component entity and a molecular structure entity, and the second one or more entities include at least one of a target entity, a protein entity, and a site entity.

[0007] In some embodiments, the graph encoder is configured to extract feature vectors of nodes in the knowledge graph by: initializing the feature vectors of the nodes; and aggregating feature vectors from other nodes and / or edges in the knowledge graph to update the feature vectors of the nodes.

[0008] In some embodiments, the feature vector of a node is determined based on the description of the entity corresponding to the node in the drug and disease related data, and the feature vector of an edge is determined based on the description of the relationship corresponding to the edge in the drug and disease related data, the relationship being directional.

[0009] In some embodiments, the graph encoder is configured to extract feature vectors of nodes in the knowledge graph by:

[0010] The feature vectors of nodes are initialized using the following formula.

[0011] in, Let x represent the initial feature vector of node v in the knowledge graph G. v s represents a vector determined based on all attributes of node v. v Represents a vector determined based on selected attributes of node v; and

[0012] The feature vector of the node is updated using the following formula:

[0013] in, Let f represent the updated feature vector of node v after passing through k graph deep learning network layers, F represent the graph deep learning method used to update the feature vector of the node, w represent the aggregation weight, AGGREGATE represent the aggregation method, V represent the aggregation source and is a subset or the entire set of G, and e represent the feature vector of the edge between node v and node u.

[0014] In some embodiments, the loss function for unsupervised learning is determined based on at least one of the following: the similarity between a node and a positive sample and the similarity between a node and a negative sample; or pseudo-labels obtained by data augmentation of the node.

[0015] In some embodiments, the unsupervised learning is configured to update the parameters of the graph encoder based on the following loss function:

[0016] Where L represents the loss function, sim represents the similarity calculation function, node u is considered a positive sample of node v, node i is considered a negative sample of node v, m represents the number of negative samples, and h represents the number of negative samples. v h u h i These represent the feature vectors extracted by the graph encoder for nodes v, u, and i, respectively.

[0017] In some embodiments, the unsupervised learning is configured to update the parameters of the graph encoder based on the following loss function:

[0018] Where L represents the loss function, The pseudo-label of node v, h v L represents the feature vector extracted by the graph encoder for node v. pred This represents the loss between the pseudo-label and the feature vector of the graph encoder for node v.

[0019] In some embodiments, the method further includes: inputting feature vectors extracted by the graph encoder for nodes in the knowledge graph into a neural network model to obtain a predicted representation of the category of the node; obtaining a label of the category of the node based on the description of the entity corresponding to the node in the drug and disease related data; and determining that the graph encoder has completed training in response to the error between the predicted representation of the category of the node and the label satisfying a first preset condition.

[0020] In some embodiments, the method further includes: in response to the error between the predicted representation of the node's category and the label not satisfying the first preset condition, adjusting one or more of the model structure parameters and model training parameters of the graph encoder, and training the adjusted graph encoder using the knowledge graph through the unsupervised learning.

[0021] In some embodiments, the graph encoder is constructed based on a graph neural network (GNN) or a graph transformer.

[0022] According to a second aspect of this disclosure, a method for constructing a knowledge graph is provided, comprising: extracting a plurality of entities and relationships between the plurality of entities from data related to drugs and diseases; and constructing an initial knowledge graph based on the extracted entities and relationships, wherein the entities constitute nodes in the initial knowledge graph, and the relationships constitute edges between nodes in the initial knowledge graph. The plurality of entities includes prescription entities and drug entities belonging to a first group and disease entities belonging to a second group, and the plurality of entities further includes at least one of the following: a first one or more entities belonging to the first group and located at a lower level than the drug entities; or a second one or more entities belonging to the second group and located at a lower level than the disease entities.

[0023] In some embodiments, the first one or more entities include at least one of a biochemical component entity and a molecular structure entity, and the second one or more entities include at least one of a target entity, a protein entity, and a site entity.

[0024] In some embodiments, the relationship between the prescription entity and the drug entity includes an inclusion relationship or a subordinate relationship, and the relationship between the prescription entity or the drug entity and the disease entity includes a promoting relationship or an inhibiting relationship.

[0025] In some embodiments, the drug comprises a drug in a naturally occurring form. In some embodiments, the extraction is performed using a Natural Language Processing (NLP) model.

[0026] In some embodiments, the knowledge graph includes prescription nodes, drug nodes, biochemical component nodes, molecular structure nodes, disease nodes, target nodes, and protein nodes. In some embodiments, the attributes of the prescription node include one or more of name, efficacy, indications, and constituent components; the attributes of the drug node include one or more of name, properties, meridian tropism, functions and indications, and dosage; the attributes of the biochemical component node include one or more of name, compound category, general molecular formula, molecular composition description, and extraction method; the attributes of the molecular structure node include one or more of name, molecular formula, molecular weight, and specific efficacy; the attributes of the disease node include one or more of name, basic symptoms, and diagnostic results; the attributes of the target node include one or more of name and mechanism of action; and the attributes of the protein node include one or more of name and amino acid sequence.

[0027] In some embodiments, the method further includes: extracting a first feature vector corresponding to a first node of a first entity in the initial knowledge graph using a graph encoder trained according to any embodiment of the first aspect of the present disclosure; and supplementing the initial knowledge graph based on the first feature vector.

[0028] In some embodiments, supplementing the initial knowledge graph based on the first feature vector includes: adding a second node corresponding to the second entity to the initial knowledge graph, and extracting a second feature vector of the second node using the graph encoder; and adding an edge between the first node and the second node to the initial knowledge graph in response to the similarity between the first feature vector and the second feature vector satisfying a second preset condition. In some embodiments, the first entity and the second entity are at the same level and belong to the same group. In some embodiments, the first entity and the second entity are at different levels and belong to the same group.

[0029] In some embodiments, an edge exists between the third node corresponding to the third entity in the initial knowledge graph and the first node, and supplementing the initial knowledge graph based on the first feature vector includes: extracting a fourth feature vector of the fourth node corresponding to the fourth entity in the initial knowledge graph using the graph encoder; and adding an edge between the third node and the fourth node in the initial knowledge graph in response to the similarity between the first feature vector and the fourth feature vector satisfying a third preset condition. In some embodiments, the first entity and the third entity belong to different groups, and the first entity and the fourth entity belong to the same group. In some embodiments, the first entity and the fourth entity belong to the first group, and the third entity belongs to the second group. In some embodiments, the fourth entity is located at a lower level than the first entity. In some embodiments, the fourth entity is located at a higher level than the first entity. In some embodiments, the fourth entity is located at the same level as the first entity.

[0030] In some embodiments, an edge exists between the third node corresponding to the third entity in the initial knowledge graph and the first node, and supplementing the initial knowledge graph based on the first feature vector includes: extracting the third feature vector of the third node, the fourth feature vector of the fourth node corresponding to the fourth entity in the initial knowledge graph, and the fifth feature vector of the fifth node corresponding to the fifth entity in the initial knowledge graph using the graph encoder; in response to the similarity between the first feature vector and the fourth feature vector satisfying a fourth preset condition and the similarity between the third feature vector and the fifth feature vector satisfying a fifth preset condition, adding an edge between the fourth node and the fifth node in the initial knowledge graph. In some embodiments, the first entity and the third entity belong to different groups, the first entity and the fourth entity belong to the same group, and the third entity and the fifth entity belong to the same group. In some embodiments, the first entity and the fourth entity belong to the first group, and the third entity and the fifth entity belong to the second group. In some embodiments, the fourth entity is located at a lower level than the first entity, and the fifth entity is located at a lower level than the third entity. In some embodiments, the fourth entity is located at a higher level than the first entity, and the fifth entity is located at a higher level than the third entity. In some embodiments, the fourth entity is located at the same level as the first entity, and the fifth entity is located at the same level as the third entity.

[0031] In some embodiments, the added edge attributes include weights determined based on the corresponding similarity, wherein the determined weights increase as the corresponding similarity increases.

[0032] In some embodiments, the initial knowledge graph is the knowledge graph used to train the graph encoder.

[0033] According to a third aspect of this disclosure, a drug development method is provided, comprising: acquiring a knowledge graph constructed by the method according to any embodiment of a second aspect of this disclosure; and inputting the knowledge graph into an artificial intelligence prediction model including a graph encoder trained by the method according to any embodiment of a first aspect of this disclosure to obtain a prediction result.

[0034] According to a fourth aspect of this disclosure, an electronic device is provided, comprising: one or more processors; and a memory storing computer-executable instructions, which, when executed by the one or more processors, cause the one or more processors to perform the method according to any embodiment of any of the first to third aspects of this disclosure.

[0035] According to a fifth aspect of this disclosure, a non-transient storage medium having computer-executable instructions stored thereon is provided, which, when executed by a computer, cause the computer to perform the method described in any embodiment of any of the first to third aspects of this disclosure.

[0036] According to a sixth aspect of this disclosure, a computer program product is provided, the computer program product including instructions that, when executed by a processor, implement the method according to any embodiment of any of the first to third aspects of this disclosure. Attached Figure Description

[0037] The foregoing and other features and advantages of this disclosure will become clear from the following description of embodiments illustrated in conjunction with the accompanying drawings. The drawings, incorporated herein and forming a part of the specification, are further used to explain the principles of this disclosure and to enable those skilled in the art to make and use it. Wherein:

[0038] Figure 1 illustrates a flowchart of a method for constructing a knowledge graph according to some embodiments of the present disclosure;

[0039] Figure 2 schematically illustrates a non-limiting example of a hierarchical grouping organization entity according to some embodiments of the present disclosure;

[0040] Figures 3 and 4 schematically illustrate non-limiting examples of knowledge graphs according to some embodiments of the present disclosure;

[0041] Figure 5 shows a flowchart of a method for training a graph encoder according to some embodiments of the present disclosure;

[0042] Figure 6 shows a flowchart of a method for training a graph encoder according to some embodiments of the present disclosure;

[0043] Figure 7 illustrates a flowchart of a method for constructing a knowledge graph according to some embodiments of the present disclosure;

[0044] Figure 8 schematically illustrates a non-limiting example of a knowledge graph according to some embodiments of the present disclosure;

[0045] Figure 9 illustrates a flowchart of a drug development method according to some embodiments of the present disclosure;

[0046] Figure 10 shows a schematic block diagram of an electronic device according to some embodiments of the present disclosure;

[0047] Figure 11 shows a schematic block diagram of a computer system on which embodiments of the present disclosure may be implemented.

[0048] Note that in the embodiments described below, the same reference numerals are sometimes used across different figures to denote the same parts or parts with the same function, and repeated descriptions are omitted. In some cases, similar reference numerals and letters are used to denote similar items, so once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0049] For ease of understanding, the positions, dimensions, and extents of the structures shown in the accompanying drawings and other materials may not represent actual positions, dimensions, and extents. Therefore, this disclosure is not limited to the positions, dimensions, and extents disclosed in the accompanying drawings and other materials. Detailed Implementation

[0050] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present disclosure.

[0051] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this disclosure or its application or use. That is, the structures and methods herein are shown in an exemplary manner to illustrate different embodiments of the structures and methods in this disclosure. However, those skilled in the art will understand that they merely illustrate exemplary ways that can be used to implement this disclosure, and not exhaustive ways. Furthermore, the drawings are not necessarily drawn to scale, and some features may be enlarged to show details of specific components.

[0052] In addition, techniques, methods and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods and equipment should be considered part of the specification.

[0053] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0054] While there is a wealth of experience in the application of natural medicines in medical practice across various countries and regions, this experience often lacks explanation, analysis, verification, and support based on modern biochemical scientific theories. For example, a natural medicine may contain hundreds of chemical components, many of which have unknown structures and functions, and these components may exhibit synergistic or antagonistic effects, making it difficult to elucidate the efficacy of natural medicines and identify their active ingredients. Despite their long history of use, some components of natural medicines may possess potential toxicity, and without scientific toxicological evaluation, their safety remains questionable. Therefore, despite the long history and rich potential of natural medicines, their development is fraught with challenges due to the lack of clarity regarding their components, efficacy, and toxicity.

[0055] How to utilize information technology for drug development, especially natural drug development, is a question that the field of drug development has been exploring. Related technologies include the following methods: (1) using traditional database collision detection methods to perform blind selection with large amounts of data; (2) using algebraic or clustering methods to study and continue using traditional prescriptions; (3) predicting results by training convolutional neural network models; and (4) conducting extensive wet laboratory screening. These methods all suffer from one or more of the following problems: time-consuming, labor-intensive, costly, lacking scientific validation processes (such as molecular and cellular level analysis and biochemical validation), and unreliable prediction accuracy.

[0056] This disclosure proposes drug development technologies based on knowledge graphs and graph deep learning. Utilizing artificial intelligence to analyze massive amounts of data related to drugs and diseases, it can not only predict the effective biochemical components and mechanisms of action of drugs known to be effective against diseases, but also predict potential drugs effective against diseases. It is understood that this disclosure can be applied to both natural and non-natural drugs (i.e., drugs that do not exist in a natural form, such as those produced according to modern medical theories), and its application to natural drugs is particularly advantageous, as it can promote the analysis of the components and study the mechanisms of action of natural drugs, and uncover potential natural drug resources and mechanisms of action.

[0057] Various embodiments of the methods according to this disclosure are described in detail below with reference to the accompanying drawings. It should be understood that actual methods may include other steps, which will not be discussed herein and are not shown in the drawings in order to avoid obscuring the essential points of this disclosure.

[0058] Figure 1 illustrates a method 100 for constructing a knowledge graph according to some embodiments of the present disclosure. As shown in Figure 1, method 100 includes: in step S102, extracting a plurality of entities and relationships between the plurality of entities from data related to drugs and diseases; and in step S104, constructing an initial knowledge graph based on the extracted entities and relationships, wherein entities are used to constitute nodes in the initial knowledge graph, and relationships are used to constitute edges between nodes in the initial knowledge graph.

[0059] Here, "medicine" can include medicines in their naturally occurring form. As a non-limiting example of using traditional Chinese medicine (TCM) as a natural medicine, data related to medicines and diseases can include structured and / or unstructured data from various TCM clinical data (e.g., patient electronic health records, medical records, and medical history), TCM and prescription literature (e.g., ancient texts such as *Compendium of Materia Medica* and *Treatise on Febrile Diseases*, modern research papers, patents, etc.), and TCM databases (such as TCMSP (Traditional Chinese Medicine Systems Pharmacology Database and Analysis Platform), TCMID (Traditional Chinese Medicine Integrated Database), ETCM (The Encyclopedia of Traditional Chinese Medicine), SymMap, and HIT (Herb Ingredients' Targets)). For example, data related to medicines and diseases can come from large databases of TCM and other substances and their chemical components. These other substances can include: (1) substances not defined in nature but potentially generated in TCM model simulations; (2) substances potentially generated during the metabolism of TCM taken or applied topically; and (3) substances potentially generated due to the interaction between TCM components and molecules in the body.

[0060] In some implementations, the term "medicine" may also additionally include drugs that do not exist in a naturally occurring form. A non-limiting example using Western medicine as a non-natural drug is possible; in this case, the constructed knowledge graph could be used for the joint research and development of traditional Chinese and Western medicines.

[0061] For example, for data that may be heterogeneous and / or dissimilar to each other, entities and relationships between entities can be extracted using Natural Language Processing (NLP) models. Any known or future NLP model can be used, including but not limited to Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, Bidirectional LSTMs (BiLSTMs) combined with Conditional Random Fields (CRFs), and Bidirectional Encoder Representations from Transformers (BERT).

[0062] This disclosure employs a hierarchical, grouped organization when designing the entities to be extracted. Specifically, the extracted entities may include prescription entities and drug entities belonging to a first group, and disease entities belonging to a second group. In some embodiments, the extracted entities may also include at least one of the following: a first one or more entities belonging to the first group at a lower level than drug entities; or a second one or more entities belonging to the second group at a lower level than disease entities. For example, the first one or more entities may include at least one of biochemical component entities and molecular structure entities, and the second one or more entities may include at least one of target entities, protein entities, and site entities.

[0063] The grouping here is used to distinguish between disease-related entities and drug-related entities. The hierarchical structure here is used to represent a progressive approach from the macroscopic to the specific. The hierarchical structure of the first group can represent the progressive process from macroscopic prescription design to microscopic molecular structure, facilitating the analysis of the biochemical components and molecular structure of drugs and revealing the mechanism of action of prescriptions. The hierarchical structure of the second group can represent the progressive process from macroscopic disease characterization to microscopic molecular mechanisms, facilitating the analysis of the mechanisms of disease occurrence and development. Through the mining of entity relationships within and / or between groups and hierarchical levels, the interaction between drugs and diseases at different levels can be studied. The depth of the hierarchy in each of the first and second groups can be designed according to specific needs to study the interaction between drugs and diseases at the desired level. In this disclosure, the extracted entities may include one or more entities belonging to the first group at a lower level than drug entities (which may be distributed across one or more corresponding levels) and / or one or more entities belonging to the second group at a lower level than disease entities (which may also be distributed across one or more corresponding levels), thereby enabling the analysis of drug-disease interaction mechanisms at a more microscopic level than the drug-disease hierarchy.

[0064] Figure 2 schematically illustrates a non-limiting example of hierarchical grouping of entities according to some embodiments of the present disclosure. As shown in Figure 2, in the first group, from top to bottom, the entities include prescription entities, drug entities, biochemical component entities, and molecular structure entities; in the second group, from top to bottom, the entities include disease entities, target entities, protein entities, and site entities. Although Figure 2 depicts the first to fourth levels as consecutive levels, this is merely exemplary and not limiting. Additional levels may be provided between or outside these levels as needed, or one or more of these levels may be omitted. Furthermore, although Figure 2 depicts prescription entities and disease entities, drug entities and target entities, biochemical component entities and protein entities, and molecular structure entities and site entities as being at the same level, this is merely exemplary and not limiting. The hierarchical correspondence between different groups may be adjusted as needed. For example, in some embodiments, drug entities and disease entities, biochemical component entities and target entities, and molecular structure entities and protein entities may be at the same level, and comorbid entities located at a higher level than disease entities may be set to be at the same level as prescription entities.

[0065] Various influences and associations between entities can be extracted as relationships between entities. For example, the relationships between prescription entities and drug entities, drug entities and biochemical component entities, biochemical component entities and molecular structure entities, target entities and protein entities, and protein entities and site entities can each include inclusion or subordination relationships. Relationships between prescription entities or drug entities and disease entities can include promoting (e.g., but not limited to worsening) or inhibiting (e.g., but not limited to treatment, conditioning, prevention, etc.) relationships. Relationships between drug entities, biochemical component entities, or molecular structure entities and target entities can include activation or inhibition relationships. Relationships between drug entities can include synergistic or antagonistic relationships. Further details are omitted here.

[0066] Figure 3 illustrates a non-limiting example knowledge graph constructed based on extracted entities and the relationships between them. For visual distinction rather than limitation, nodes corresponding to entities in the first group are depicted as gray circles, and nodes corresponding to entities in the second group are depicted as white circles. For example, Figure 3 can be used for drug development down to the biochemical component-target level. Two nodes corresponding to two entities whose relationships have been extracted can be connected by an edge. Nodes corresponding to entities whose relationships with other entities have not been extracted can be temporarily depicted as isolated nodes. In the example of Figure 3, edges are implemented as undirected edges, which can reflect the association between the entities corresponding to the nodes connected by the edge. Additionally or alternatively, in some embodiments, such as Figure 4 described later, edges can also be implemented as directed edges, which can further reflect the directionality of the association between the entities corresponding to the nodes connected by the edge.

[0067] For example, Figure 4 shows another non-limiting example of a knowledge graph, which, compared to Figure 3, can be used to further delve into drug development at the molecular structure-protein level. In the example in Figure 4, the knowledge graph includes prescription nodes, drug nodes, biochemical component nodes, molecular structure nodes, disease nodes, target nodes, and protein nodes. One or more nodes (e.g., comorbidity nodes, site nodes, etc.) can also be added or removed depending on the specific circumstances.

[0068] The attributes of a node can be determined based on the description of the entity corresponding to that node in drug- and disease-related data. In some embodiments, the attributes of a prescription node include one or more of name, efficacy, indications, and components; the attributes of a drug node include one or more of name, properties, meridian tropism, functions and indications, and dosage; the attributes of a biochemical component node include one or more of name, compound category, molecular formula, molecular composition description, and extraction method; the attributes of a molecular structure node include one or more of name, molecular formula, molecular weight, and specific efficacy; the attributes of a disease node include one or more of name, basic symptoms, and diagnostic results; the attributes of a target node include one or more of name and mechanism of action; and the attributes of a protein node include one or more of name and amino acid sequence. For example, Table 1 below shows some non-limiting examples of nodes and their attributes.

[0069] Table 1

[0070] The attributes of edges can be determined based on the description of the entity relationships corresponding to the edges in drug and disease-related data. As a non-limiting example, Figure 4 illustrates some edge attributes. It can be understood that although Figure 4 depicts both edges indicating inclusion and edges indicating dependency relationships between nodes of entities with inclusion / dependency relationships, only one of them may be retained.

[0071] Next, the (initial) knowledge graph thus constructed can be used to train a graph encoder to infuse it with knowledge from the drug development field, ensuring the reliability of the training content. Referring to Figure 5, a method 200 for training a graph encoder according to some embodiments of this disclosure is illustrated.

[0072] As shown in Figure 5, method 200 includes: in step S202, acquiring a knowledge graph; and in step S204, using the acquired knowledge graph to train a graph encoder constructed based on graph deep learning through unsupervised learning, the graph encoder being configured to extract feature vectors of nodes in the knowledge graph. In some embodiments, the graph encoder may be constructed based on a graph neural network (GNN) or a graph transformer, etc.

[0073] Specifically, the acquired knowledge graph can be constructed based on multiple entities extracted from drug and disease-related data and the relationships between these entities, where entities constitute nodes in the knowledge graph, and relationships constitute edges between nodes. The multiple entities may include prescription entities and drug entities belonging to a first group and disease entities belonging to a second group, and may also include at least one of the following: a first or more entities belonging to the first group at a lower level than drug entities; or a second or more entities belonging to the second group at a lower level than disease entities. In some embodiments, the first or more entities include at least one of biochemical component entities and molecular structure entities, and the second or more entities include at least one of target entities, protein entities, and site entities. For example, the initial knowledge graph constructed by the method 100 according to any of the foregoing embodiments can be obtained as a knowledge graph for training a graph encoder. Therefore, various embodiments regarding the knowledge graph for training a graph encoder can be similarly described above and will not be repeated here. Furthermore, before training the graph encoder with the knowledge graph, the graph encoder may have been pre-trained, for example, but not limited to, using graph structure data from a general domain.

[0074] In some embodiments, a graph encoder can be configured to extract feature vectors of nodes in a knowledge graph by: initializing the feature vectors of the nodes; and aggregating feature vectors from other nodes and / or edges in the knowledge graph to update the node's feature vector. For example, aggregation can be performed using message passing mechanisms, random walk-based methods, global attention-based methods, etc., which can be selected according to specific needs.

[0075] The feature vectors of each node and edge can be initialized in any suitable manner. Theoretically, any vector can be used as a starting point. However, initializing the feature vectors of each node and edge based on their corresponding properties can be advantageous, as these properties provide important information about the features of the corresponding node or edge, helping the graph encoder to learn patterns and relationships in the graph structure more effectively. In some embodiments, the feature vectors of nodes can be determined based on the description of the entity corresponding to the node in drug- and disease-related data. In some embodiments, the feature vectors of edges can be determined based on the description of the relationship corresponding to the edge in drug- and disease-related data. Such relationships can be directional, for example. These descriptions are typically in text form, but in some cases may include other forms such as images, audio, and video. The vectorized representation of the description can be obtained based on any algorithm or model currently known or developed in the future. For example, for text descriptions, Word2Vec, GloVe, Doc2Vec, BERT, GPT, etc., can be used.

[0076] In some embodiments, the feature vector of a node is initialized using the following formula.

[0077] in, Let x represent the initial feature vector of node v in the knowledge graph G. v This represents a vector determined based on all the attributes of node v.

[0078] In other embodiments, the feature vector of a node can be initialized using the following formula:

[0079] in, Let x represent the initial feature vector of node v in the knowledge graph G. v s represents a vector determined based on all attributes of node v. v This represents a vector determined based on selected attributes of node v.

[0080] Compared to initializing the node's feature vector using formula (1), initializing the node's feature vector using formula (2) can further emphasize the attributes that the graph encoder wants to learn, minimizing the loss of such attribute information. As a non-restrictive example, the x of the prescription node... v It can be determined based on name, efficacy, indications, and composition, and s v The x-axis of the drug node can be determined based on efficacy, indications, and composition. v It can be determined based on name, properties, meridian tropism, functions and indications, usage and dosage, and s v The x-axis of biochemical components can be determined based on their properties, functions, and indications. vIt can be determined based on name, compound category, general molecular formula, molecular composition description, extraction method, and s v The x-axis of molecular structure nodes can be determined based on the general molecular formula and molecular composition description. v It can be determined based on name, molecular formula, molecular weight, specific utility, and s v The x-axis of a disease node can be determined based on its molecular formula and specific utility. v It can be determined based on the name, basic symptoms, and diagnosis results, and s v The x-axis of the target node can be determined based on basic symptoms and diagnostic results. v It can be determined based on name and mechanism of action, and s v The x-axis of a protein node can be determined based on its mechanism of action. v It can be determined based on name, amino acid sequence and s v It can be determined based on the amino acid sequence.

[0081] Furthermore, in some embodiments, the feature vector of the node is updated using the following formula:

[0082] in, Let F represent the updated feature vector of node v after passing through k graph deep learning network layers, F represent the graph deep learning method used to update the feature vector of the node, w represent the aggregation weight, AGGREGATE represent the aggregation method, and V represent the aggregation source and is a subset or the entire set of G.

[0083] In some embodiments, the feature vector of the node is updated using the following formula:

[0084] in, Let f represent the updated feature vector of node v after passing through k graph deep learning network layers, F represent the graph deep learning method used to update the feature vector of the node, w represent the aggregation weight, AGGREGATE represent the aggregation method, V represent the aggregation source and is a subset or the entire set of G, and e represent the feature vector of the edge between node v and node u.

[0085] Compared to updating the feature vector of a node using formula (3), updating the feature vector of a node using formula (4) can simultaneously aggregate node features and edge features, which can capture graph structure information more comprehensively and improve the expressive and generalization capabilities of the graph encoder.

[0086] For example, F can include, but is not limited to, GNN, Graph Transformer, etc.; AGGREGATE can include, but is not limited to, summation, averaging, etc.; V can include, but is not limited to, node neighborhood, etc.; and w can include, but is not limited to, similarity function, attention coefficient, etc.

[0087] Overfitting can be avoided during aggregation, thus improving model accuracy. Weights during updates are used to assign different update proportions to different types of nodes, which helps avoid oversmoothing. For example, if a node has similar attributes to one neighbor but opposite attributes to another, during aggregation, the weight of the neighbor with the opposite attribute is reduced, while the weight of the neighbor with the similar attribute is increased. This ensures high-quality feature vectors are obtained, improving the accuracy of the graph encoder output.

[0088] In addition, to prevent gradient explosion and gradient minimization, activation functions, including but not limited to the ReLU (Rectified Linear Unit) function, can be used during the aggregation process. Specifically, the aggregated features can be linearly transformed to learn the relationships between them, and then an activation function can be applied to perform a non-linear mapping on the linearly transformed features to enhance the expressive power of the graph encoder.

[0089] The loss function used in unsupervised learning can be specifically selected as needed. In some embodiments, the loss function for unsupervised learning can be determined based on at least one of the following: the similarity between a node and a positive sample and the similarity between a node and a negative sample; or pseudo-labels obtained by data augmentation of the node.

[0090] For example, a contrastive learning-based loss function can be used, which learns node representations by constructing positive and negative sample pairs, maximizing the similarity of positive sample pairs, and minimizing the similarity of negative sample pairs. In some embodiments, unsupervised learning can be configured to update the parameters of the graph encoder based on the following loss function:

[0091] Where L represents the loss function, sim represents the similarity calculation function (such as, but not limited to, cosine similarity, matrix inner product, etc.), node u is considered a positive sample of node v, node i is considered a negative sample of node v, m represents the number of negative samples, and h v h u h i Let v, u, and i represent the feature vectors extracted by the graph encoder for nodes v, u, and i, respectively. During training, the learnable parameters in the graph encoder can be updated by minimizing this loss function. By reducing this loss function, the distance between a node and positive samples in the embedding space can be shortened, while the distance between the node and negative samples can be lengthened.

[0092] In some embodiments, m is greater than 1. By comparing positive samples with multiple negative samples, the graph encoder can more effectively learn the features that distinguish between positive and negative samples, thereby improving the quality of representation learning. Introducing multiple negative samples helps balance the training process, avoids overfitting to a single positive sample pair, and thus improves the generalization ability of the graph encoder. The choice of the number of negative samples m can be specifically determined by factors such as specific task requirements, dataset characteristics, and computational resource constraints.

[0093] It is understandable that the loss function based on contrastive learning is not limited to the form shown in equation (5), as long as it can maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs.

[0094] Alternatively, a pseudo-label-based loss function can be used, which can be employed to improve model performance using unlabeled data. By augmenting the unlabeled data and combining it with model predictions, pseudo-labels can be generated and incorporated into the training process. Specifically, data augmentation can be performed on the graph structure data of the knowledge graph through various transformations. Then, the initial graph encoder is used to predict the augmented unlabeled data, and a pseudo-label is assigned to each unlabeled data point based on the prediction results. In some examples, a confidence threshold can be set, and predictions with a confidence level exceeding this threshold can be used as pseudo-labels to ensure their reliability. In some embodiments, the prediction with the highest probability can be directly selected as the pseudo-label.

[0095] In some embodiments, unsupervised learning can be configured to update the parameters of the graph encoder based on the following loss function:

[0096] Where L represents the loss function, The pseudo-label of node v, h v L represents the feature vector extracted by the graph encoder for node v. pred This represents the loss between the pseudo-label and the feature vector of the graph encoder for node v.

[0097] For example, L pred Any suitable form can be used, including but not limited to cross-entropy loss, mean squared error (MSE) loss, entropy minimization loss, etc.

[0098] In some embodiments, L predA pseudo-label-based contrastive learning loss can be used, for example, by combining equations (5) and (6) above. For instance, different views can be generated by data augmentation of unlabeled nodes, and pseudo-labels can be assigned to each view. Then, a contrastive learning loss function is designed to maximize the similarity between different views of the same node (as positive samples) and minimize the similarity between different nodes (as negative samples). For example, it can be...

[0099] Figure 6 illustrates a method 200' for training a graph encoder according to some embodiments of the present disclosure. Compared to method 200', method 200 may further include: in step S206, inputting feature vectors extracted by the graph encoder for nodes in a knowledge graph into a neural network model to obtain a predicted representation of the node's category; in step S208, obtaining a label for the node's category based on the description of the entity corresponding to the node in drug and disease-related data. In some embodiments, method 200' includes: in step S210, determining that the graph encoder has completed training in response to the error between the predicted representation of the node's category and the label satisfying a first preset condition. In some embodiments, method 200' includes: in step S212, in response to the error between the predicted representation of the node's category and the label not satisfying the first preset condition, adjusting one or more of the model structure parameters and model training parameters of the graph encoder, and training the adjusted graph encoder using the knowledge graph through unsupervised learning.

[0100] Here, a neural network model is deployed downstream of the graph encoder and configured to perform the task of predicting node classes, so as to evaluate the performance of the graph encoder by prediction accuracy. In some embodiments, the neural network model includes multiple linear transformation layers. For example, prediction accuracy can be evaluated using the following MSE loss function:

[0101] Among them, z j This indicates that the neural network model is based on the predicted representation output by the graph encoder for the feature vector extracted for node j, y j This represents the vector representation of the label determined based on the description of the entity corresponding to node j in the drug and disease-related data. The vector representation of the label can be generated using any suitable vectorization method, including but not limited to one-hot encoding; alternatively, label vectorization can be achieved using artificial intelligence language models such as NLP models.

[0102] In some embodiments, each entity (e.g., the entity's name) can be directly treated as a separate category. In other embodiments, the entities can be clustered, and each cluster (e.g., the cluster's name) can be treated as a separate category. This can help reduce the number of labels, thereby balancing assessment speed and accuracy and improving assessment efficiency. In some embodiments, a subset rather than all entities (e.g., drug entities and disease entities) can be selected for assessment, thereby balancing assessment speed and accuracy and improving assessment efficiency.

[0103] The first preset condition can be set according to the specific circumstances, and no special restrictions are imposed here. For example, when the error (e.g., the MSE loss function mentioned above) does not exceed a preset threshold, the first preset condition can be considered to be satisfied.

[0104] The constructed knowledge graph can be further supplemented using a trained graph encoder. Figure 7 illustrates a method 100' for constructing a knowledge graph according to some embodiments of the present disclosure. Compared with method 100', method 100 may further include: in step S106, extracting a first feature vector corresponding to a first node of a first entity in the initial knowledge graph using a graph encoder trained (e.g., the method 200, 200' according to any of the foregoing embodiments); and in step S108, supplementing the initial knowledge graph based on the first feature vector.

[0105] For illustrative purposes only, Figure 8 schematically illustrates a non-limiting example of a knowledge graph that supplements the (initial) knowledge graph shown in Figure 4 using method 100'. For visual clarity rather than limitation, the newly added nodes and edges are depicted with bold dashed lines, and their attributes are shown in a different font. Furthermore, the example in Figure 8 can similarly employ the hierarchical and grouping approach shown in Figure 2.

[0106] In some embodiments, step S108 includes: adding a second node corresponding to the second entity in the initial knowledge graph, and extracting a second feature vector of the second node using a graph encoder; and adding an edge between the first node and the second node in the initial knowledge graph in response to the similarity between the first feature vector and the second feature vector satisfying a second preset condition.

[0107] In some examples, the first entity and the second entity are at the same level and belong to the same group. For example, referring to Figure 8, suppose a new "drug 5" is introduced. After obtaining its feature vector through a graph encoder, it is matched with existing nodes in the knowledge graph for similarity. If the similarity between the feature vector of "drug 5" and the feature vector of "drug 4" exceeds a preset similarity threshold, it is predicted that there is an edge between "drug 5" and "drug 4", thus adding the node "drug 5" and the edge between it and "drug 4" to the knowledge graph. Alternatively, an antagonistic relationship between "drug 5" and "drug 4" can be extracted from drug and disease-related data using, for example, an NLP model, so the attribute "antagonistic" can be added to the edge between them. This allows for the expansion of the drug knowledge graph.

[0108] In other examples, the first entity and the second entity may be at different levels but belong to the same group. For example, referring to Figure 8, suppose a new "biochemical component 4" is introduced. After obtaining its feature vector through a graph encoder, it is matched with existing nodes in the knowledge graph for similarity. If the similarity between the feature vector of "biochemical component 4" and the feature vector of "drug 4" exceeds a preset similarity threshold, it is predicted that there is an edge between "biochemical component 4" and "drug 4". This adds the node "biochemical component 4" and its edge with "drug 4" to the knowledge graph. For example, it can also predict that the edge's attribute is "drug 4" contains "biochemical component 4". This can help identify the biochemical components of a drug.

[0109] In some embodiments, there is an edge between the third node corresponding to the third entity in the initial knowledge graph and the first node, and step S108 may include: extracting the fourth feature vector of the fourth node corresponding to the fourth entity in the initial knowledge graph using a graph encoder; and adding an edge between the third node and the fourth node in the initial knowledge graph in response to the similarity between the first feature vector and the fourth feature vector satisfying a third preset condition.

[0110] In some examples, the first entity and the third entity belong to different groups, while the first entity and the fourth entity belong to the same group. In some examples, the first entity and the fourth entity belong to the first group, and the third entity belongs to the second group. In some examples, the fourth entity is at a lower level than the first entity. In some examples, the fourth entity is at a higher level than the first entity. In some examples, the fourth entity is at the same level as the first entity.

[0111] For example, referring to Figure 8, the knowledge graph already contains an edge "inhibition (prevention)" pointing from "drug 4" to "disease 3". If the similarity between the feature vector of "drug 4" and the feature vector of "prescription 3" exceeds a preset similarity threshold, then the existence of an edge "inhibition (prevention)" pointing from "prescription 3" to "disease 3" is predicted, thus supplementing the knowledge graph with the edge between "prescription 3" and "disease 3". This can facilitate finding usable prescriptions starting from drugs.

[0112] Referring again to Figure 8, the knowledge graph already contains an edge "Inhibition (Treatment)" pointing from "Prescription 1" to "Disease 1". If the similarity between the feature vector of "Prescription 1" and the feature vector of "Prescription 2" exceeds a preset similarity threshold, then an edge "Inhibition (Treatment)" pointing from "Prescription 2" to "Disease 1" is predicted, thus supplementing the knowledge graph with an edge between "Prescription 2" and "Disease 1". This can facilitate the development of new uses for prescriptions.

[0113] Referring again to Figure 8, the knowledge graph already contains an edge "inhibition (conditioning)" pointing from "drug 3" to "disease 2". If the similarity between the feature vector of "drug 3" and the feature vector of "biochemical component 3" exceeds a preset similarity threshold, then the existence of an edge "inhibition (conditioning)" pointing from "biochemical component 3" to "disease 2" is predicted, thus supplementing the knowledge graph with the edge between "biochemical component 3" and "disease 2". This can help identify the effective biochemical components of the drug.

[0114] Referring again to Figure 8, the knowledge graph already contains an edge "inhibition (treatment)" pointing from "molecular structure 2" to "disease 4". If the similarity between the feature vector of "drug 1" and the feature vector of "molecular structure 2" exceeds a preset similarity threshold, then the existence of an edge "inhibition (treatment)" pointing from "drug 1" to "disease 4" is predicted, thus supplementing the knowledge graph with the edge between "drug 1" and "disease 4". Therefore, new drugs or new uses for drugs can be discovered at the molecular structure level.

[0115] In some embodiments, there is an edge between the third node corresponding to the third entity in the initial knowledge graph and the first node, and step S108 may include: extracting the third feature vector of the third node, the fourth feature vector of the fourth node corresponding to the fourth entity in the initial knowledge graph, and the fifth feature vector of the fifth node corresponding to the fifth entity in the initial knowledge graph using a graph encoder; in response to the similarity between the first feature vector and the fourth feature vector satisfying a fourth preset condition and the similarity between the third feature vector and the fifth feature vector satisfying a fifth preset condition, adding an edge between the fourth node and the fifth node in the initial knowledge graph.

[0116] In some examples, the first entity and the third entity belong to different groups; the first entity and the fourth entity belong to the same group; and the third entity and the fifth entity belong to the same group. In some examples, the first entity and the fourth entity belong to the first group, and the third entity and the fifth entity belong to the second group. In some examples, the fourth entity is at a lower, higher, or the same level as the first entity. In some examples, the fifth entity is at a lower, higher, or the same level as the third entity.

[0117] For example, referring to Figure 8, the knowledge graph already contains an edge "inhibition (modulation)" pointing from "drug 3" to "disease 2". If the similarity between the feature vector of "drug 3" and the feature vector of "biochemical component 3" exceeds a preset similarity threshold, and the similarity between the feature vector of "disease 2" and the feature vector of "target 2" exceeds a preset similarity threshold, then an edge pointing from "biochemical component 3" to "target 2" is predicted, thus supplementing the knowledge graph with the edge between "biochemical component 3" and "target 2". This can help to elucidate the interaction mechanism between drugs and diseases at the microscopic level (e.g., between biochemical components / molecular structures and targets / proteins / sites).

[0118] Referring again to Figure 8, the knowledge graph already contains an edge "inhibition" pointing from "Molecular Structure 1" to "Target 5". If the similarity between the feature vector of "Drug 1" and the feature vector of "Molecular Structure 1" exceeds a preset similarity threshold, and the similarity between the feature vector of "Disease 5" and the feature vector of "Target 5" also exceeds a preset similarity threshold, then an edge "inhibition" pointing from "Drug 1" to "Disease 5" is predicted, thus supplementing the edge between "Drug 1" and "Disease 5" in the knowledge graph. This can facilitate the development of new drugs or new uses for drugs from the interaction mechanisms at the microscopic level (e.g., between biochemical components / molecular structures and targets / proteins / sites).

[0119] Each of the second to fifth preset conditions can be set according to specific circumstances, without any special restrictions. For example, when the similarity (such as, but not limited to, cosine similarity, matrix inner product, etc.) exceeds the corresponding preset similarity threshold, it can be considered that the corresponding preset condition is met.

[0120] In some embodiments, the attributes of the added edges may include weights determined based on the corresponding similarity, wherein the determined weights increase as the corresponding similarity increases. This can further improve the accuracy of the graph encoder output, which in turn can better facilitate the supplementation of the knowledge graph and improve its construction.

[0121] The knowledge graph obtained by the aforementioned method can aid in drug development. Figure 9 shows a flowchart of a drug development method 300 (hereinafter referred to as method 300) according to some embodiments of the present disclosure. As shown in Figure 9, method 300 includes: in step S302, acquiring (e.g., by constructing a knowledge graph according to methods 100, 100' of any of the foregoing embodiments); and in step S304, inputting the knowledge graph into an artificial intelligence prediction model including (e.g., by training a graph encoder according to methods 200, 200' of any of the foregoing embodiments) to obtain prediction results.

[0122] Artificial intelligence (AI) prediction models can be configured to perform various prediction tasks related to drug development. In some embodiments, AI prediction models can be configured to predict diseases, targets, proteins, sites, or other levels of disease-related entities that may be effectively treated, based on drug formulations, drugs, biochemical components, molecular structures, or other levels of drug-related entities. In some embodiments, AI prediction models can be configured to predict drug formulations, drugs, biochemical components, molecular structures, or other levels of drug-related entities that may be effective against them, based on diseases, targets, proteins, sites, or other levels of drug-related entities. In some embodiments, AI prediction models can be configured to predict the potential properties of drug formulations, drugs, biochemical components, molecular structures, or other levels of drug-related entities. In some embodiments, AI prediction models can be configured to predict the potential properties of disease-related entities based on diseases, targets, proteins, sites, or other levels of drug-related entities.

[0123] The artificial intelligence prediction model can employ any suitable artificial intelligence algorithm model, whether currently known or developed in the future, without any particular limitation. Furthermore, although the graph encoder in the artificial intelligence prediction model has been well trained using the methods 200, 200' described according to any of the foregoing embodiments, the artificial intelligence prediction model can be further trained according to the specific downstream task, including but not limited to supervised learning, due to differences in specific downstream tasks.

[0124] This disclosure also provides an electronic device that may include: one or more processors; and a memory storing computer-executable instructions, which, when executed by the one or more processors, cause the one or more processors to perform the method according to any of the foregoing embodiments.

[0125] Referring to FIG10, a schematic block diagram of an electronic device 400 according to some embodiments of the present disclosure is shown. As shown in FIG10, the electronic device 400 includes one or more processors 402 and a memory 404 storing computer-executable instructions that, when executed by the processor 402, cause the processor 402 to perform the methods described according to any of the foregoing embodiments (e.g., methods 100, 100', 200, 200', 300, etc.). The processor 402 may, for example, be a central processing unit (CPU) of the electronic device 400. The processor 402 may be any type of general-purpose processor or may be a processor specifically designed for building knowledge graphs and / or training graph encoders and / or drug discovery, such as an application-specific integrated circuit (“ASIC”). The memory 404 may be coupled to the processor 402 and may include various computer-readable media accessible by the processor 402. In various embodiments, the memory 404 described herein may include volatile and non-volatile media, removable and non-removable media. For example, memory 404 may include any combination of the following: random access memory (“RAM”), dynamic RAM (“DRAM”), static RAM (“SRAM”), read-only memory (“ROM”), flash memory, cache memory, and / or any other type of non-transitory computer-readable medium. Memory 404 may store instructions that, when executed by processor 402, cause processor 402 to perform the methods described according to any of the foregoing embodiments of this disclosure.

[0126] The electronic device 400 is configured to perform the method according to any of the foregoing embodiments, and therefore reference can be made to the foregoing description of various embodiments of various methods according to this disclosure, which will not be repeated here.

[0127] This disclosure also provides a non-transient storage medium having computer-executable instructions stored thereon, which, when executed by a computer, cause the computer to perform the methods described according to any of the foregoing embodiments of this disclosure.

[0128] This disclosure also provides a computer program product that may include instructions that, when executed by a processor, can implement the methods described according to any of the foregoing embodiments of this disclosure. The instructions may be any set of instructions that will be executed directly by one or more processors, such as machine code, or any set of instructions that will be executed indirectly, such as a script. The instructions may be stored in an object code format for direct processing by one or more processors, or stored in any other computer language, including scripts or sets of independent source code modules that are interpreted on demand or compiled in advance.

[0129] Figure 11 shows a schematic block diagram of a computer system 500 on which embodiments of the present disclosure may be implemented. The computer system 500 includes a bus 502 or other communication mechanism for transmitting information, and a processing means 504 coupled to the bus 502 for processing information. The computer system 500 also includes a memory 506 coupled to the bus 502 for storing instructions to be executed by the processing means 504; the memory 506 may be random access memory (RAM) or other dynamic storage device. The memory 506 may also be used to store temporary variables or other intermediate information during the execution of instructions to be executed by the processing means 504. The computer system 500 also includes a read-only memory (ROM) 508 or other static storage device coupled to the bus 502 for storing static information and instructions for the processing means 504. A storage device 510, such as a magnetic disk or optical disk, is provided and coupled to the bus 502 for storing information and instructions. Computer system 500 may be coupled via bus 502 to output device 512 for providing output to a user, such as, but not limited to, a display (e.g., a cathode ray tube (CRT) or liquid crystal display (LCD)), speakers, etc. Input device 514, such as a keyboard, mouse, microphone, etc., is coupled to bus 502 for transmitting information and command selections to processing device 504. Computer system 500 may execute embodiments of the present disclosure. Consistent with certain implementations of the present disclosure, results are provided by computer system 500 in response to processing device 504 executing one or more sequences of one or more instructions contained in memory 506. Such instructions may be read into memory 506 from another computer-readable medium, such as storage device 510. Execution of the sequence of instructions contained in memory 506 causes processing device 504 to perform the methods described herein. Alternatively, the teachings may be implemented using hard-wired circuitry in place of or in combination with software instructions. Therefore, implementations of the present disclosure are not limited to any particular combination of hardware circuitry and software. In various embodiments, computer system 500 can be connected across a network to one or more other computer systems, such as computer system 500, to form a networked system via network interface 516. This network may include a private network or a public network such as the Internet. In a networked system, one or more computer systems can store data and supply data to other computer systems. As used herein, the term "computer-readable medium" refers to any medium that participates in providing instructions to processing device 504 for execution. Such media can take many forms, including but not limited to non-volatile media, volatile media, and transmission media. Non-volatile media include, for example, optical discs or magnetic disks such as storage device 510. Volatile media include dynamic memory such as memory 506. Transmission media include coaxial cables, copper wires, and optical fibers, including wiring containing bus 502.Common forms of computer-readable media or computer program products include, for example, floppy disks, flexible disks, hard disks, magnetic tapes, or any other magnetic media, CD-ROMs, digital video discs (DVDs), Blu-ray discs, any other optical media, thumb drives, memory cards, RAM, PROMs and EPROMs, fast EPROMs, any other memory chips or cartridges, or any other tangible media from which a computer can read. Various forms of computer-readable media may be involved when carrying one or more sequences of one or more instructions to processing device 504 for execution. For example, instructions may initially be carried on a disk of a remote computer. The remote computer may load the instructions into its dynamic memory and transmit the instructions over a telephone line using a modem. A modem local to computer system 500 may receive data over a telephone line and convert the data into an infrared signal using an infrared transmitter. An infrared detector coupled to bus 502 may receive the data carried in the infrared signal and place the data on bus 502. Bus 502 carries the data to memory 506, from which processing device 504 retrieves and executes the instructions. Optionally, the instructions received by the memory 506 may be stored on the storage device 510 before or after execution by the processing device 504.

[0130] According to various embodiments, instructions configured to be executed by a processing device to perform a method are stored on a computer-readable medium. The computer-readable medium may be a device for storing digital information. For example, a computer-readable medium includes a compact disc read-only memory (CD-ROM) as known in the art for storing software. The computer-readable medium is accessed by a processor adapted to execute the instructions configured to be executed.

[0131] The foregoing has described one or more exemplary embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0132] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a server system. Of course, this disclosure does not exclude the possibility that, with the future development of computer technology, the computer implementing the functions of the above embodiments can be, for example, a personal computer, a laptop computer, an in-vehicle human-machine interaction device, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0133] While one or more embodiments of this disclosure provide the method operation steps as described in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or terminal product execution, the methods shown in the embodiments or drawings can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment).

[0134] The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitation, the presence of other identical or equivalent elements in the process, method, product, or apparatus that includes said elements is not excluded. For example, the use of terms such as "first" or "second" to denote names does not indicate any particular order.

[0135] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, when implementing one or more embodiments of this disclosure, the functions of each module can be implemented in one or more software and / or hardware, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0136] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0137] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0138] Those skilled in the art will understand that one or more embodiments of this disclosure may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of this disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] One or more embodiments of this disclosure can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a particular task or implement a particular abstract data type. One or more embodiments of this disclosure can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules can reside in local and remote computer storage media, including storage devices.

[0140] The same or similar parts between the various embodiments of this disclosure can be referred to mutually, and each embodiment focuses on describing the differences from other embodiments. In particular, for the apparatus embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to the description of the method embodiments. In the description of this disclosure, the descriptions of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., mean that the specific feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of this disclosure. In this disclosure, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this disclosure and the features of the different embodiments or examples.

[0141] Additionally, when used in this disclosure, the terms “here,” “above,” “below,” “below,” “in the following,” “overall,” and similar terms should refer to the entirety of this disclosure and not any particular part thereof. Furthermore, unless expressly stated otherwise or otherwise understood in the context in which they are used, conditional language used herein, such as “may,” “possibly,” “for example,” “like,” etc., is generally intended to express that certain embodiments include, while other embodiments do not, certain features, elements, and / or states. Therefore, such conditional language is not generally intended to imply that one or more embodiments require features, elements, and / or states in any way, or whether such features, elements, and / or states are included or performed in any particular embodiment.

[0142] The above description is merely an embodiment of one or more embodiments of this disclosure and is not intended to limit the scope of the one or more embodiments of this disclosure. Various modifications and variations can be made to the one or more embodiments of this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of the claims.

Claims

1. A method for training a graph encoder, comprising: A knowledge graph is obtained, which is constructed based on multiple entities extracted from data related to drugs and diseases and the relationships between the multiple entities, wherein the entities are used to form nodes in the knowledge graph and the relationships are used to form edges between nodes in the knowledge graph. as well as A graph encoder based on graph deep learning is trained using the acquired knowledge graph through unsupervised learning. The graph encoder is configured to extract feature vectors from nodes in the knowledge graph. The plurality of entities includes prescription entities and drug entities belonging to the first group and disease entities belonging to the second group, and the plurality of entities also includes at least one of the following: One or more entities belonging to the first group that are located at a lower level than the drug entity; or The second group includes one or more entities that are located at a lower level than the disease entity.

2. The method according to claim 1, wherein, The first one or more entities include at least one of biochemical component entities and molecular structure entities, and the second one or more entities include at least one of target entities, protein entities and site entities.

3. The method according to claim 1, wherein, The graph encoder is configured to extract feature vectors of nodes in the knowledge graph by: Initialize the feature vector of the node; and The feature vectors of the node are aggregated from other nodes and / or edges in the knowledge graph to update the feature vector of the node.

4. The method according to claim 3, wherein, The feature vector of a node is determined based on the description of the entity corresponding to the node in the drug- and disease-related data, and The feature vector of an edge is determined based on the description of the relationship corresponding to the edge in the drug and disease related data, and the relationship is directional.

5. The method according to claim 3 or 4, wherein, The graph encoder is configured to extract feature vectors of nodes in the knowledge graph by: The feature vectors of nodes are initialized using the following formula. in, Let x represent the initial feature vector of node v in the knowledge graph G. v s represents a vector determined based on all attributes of node v. v Represents a vector determined based on selected attributes of node v; and The feature vector of the node is updated using the following formula: in, Let f represent the updated feature vector of node v after passing through k graph deep learning network layers, F represent the graph deep learning method used to update the feature vector of the node, w represent the aggregation weight, AGGREGATE represent the aggregation method, V represent the aggregation source and is a subset or the entire set of G, and e represent the feature vector of the edge between node v and node u.

6. The method according to claim 1, wherein, The loss function for unsupervised learning is determined based on at least one of the following: Similarity between a node and positive samples, and similarity between a node and negative samples; or Pseudo-labels are obtained by augmenting the data of nodes.

7. The method according to claim 1, further comprising: The feature vectors extracted by the graph encoder for nodes in the knowledge graph are input into the neural network model to obtain a predicted representation of the node's category; The category label of the node is obtained based on the description of the entity corresponding to the node in the drug and disease related data; When the error between the predicted representation of the node's category and the label satisfies a first preset condition, it is determined that the graph encoder has completed training.

8. The method according to claim 7, further comprising: If the error between the predicted category representation of the node and the label does not meet the first preset condition, one or more of the model structure parameters and model training parameters of the graph encoder are adjusted, and the adjusted graph encoder is trained by the unsupervised learning using the knowledge graph.

9. The method according to claim 1, wherein, The graph encoder is built based on a graph neural network (GNN) or a graph transformer.

10. A method for constructing a knowledge graph, comprising: Extract multiple entities and the relationships between said entities from drug and disease-related data; as well as An initial knowledge graph is constructed based on the extracted entities and relations, wherein the entities constitute the nodes in the initial knowledge graph, and the relations constitute the edges between the nodes in the initial knowledge graph. The plurality of entities includes prescription entities and drug entities belonging to the first group and disease entities belonging to the second group, and the plurality of entities also includes at least one of the following: One or more entities belonging to the first group that are located at a lower level than the drug entity; or The second group includes one or more entities that are located at a lower level than the disease entity.

11. The method of claim 10, wherein: The first one or more entities include at least one of biochemical component entities and molecular structure entities, and the second one or more entities include at least one of target entities, protein entities and site entities; and / or The relationship between the prescription entity and the drug entity includes an inclusion relationship or a subordinate relationship, and the relationship between the prescription entity or the drug entity and the disease entity includes a promoting relationship or an inhibiting relationship.

12. The method according to claim 10 or 11, wherein, The drug includes drugs in naturally occurring forms, and wherein the extraction is performed using a natural language processing (NLP) model.

13. The method according to claim 12, in, The knowledge graph includes prescription nodes, drug nodes, biochemical component nodes, molecular structure nodes, disease nodes, target nodes, protein nodes, and... The attributes of the prescription node include one or more of the following: name, efficacy, indications, and components; the attributes of the drug node include one or more of the following: name, properties, meridian tropism, functions and indications, and dosage; the attributes of the biochemical component node include one or more of the following: name, compound category, molecular formula, molecular composition description, and extraction method; the attributes of the molecular structure node include one or more of the following: name, molecular formula, molecular weight, and specific efficacy; the attributes of the disease node include one or more of the following: name, basic symptoms, and diagnostic results; the attributes of the target node include one or more of the following: name and mechanism of action; and the attributes of the protein node include one or more of the following: name and amino acid sequence.

14. The method of claim 10, further comprising: The graph encoder trained according to any one of claims 1 to 9 extracts the first feature vector of the first node corresponding to the first entity in the initial knowledge graph; and The initial knowledge graph is supplemented based on the first feature vector.

15. The method according to claim 14, wherein, The initial knowledge graph is supplemented based on the first feature vector, including: A second node corresponding to the second entity is added to the initial knowledge graph, and the second feature vector of the second node is extracted using the graph encoder; and In response to the similarity between the first feature vector and the second feature vector satisfying a second preset condition, an edge between the first node and the second node is added to the initial knowledge graph.

16. The method of claim 15, wherein: The first entity and the second entity are at the same level and belong to the same group; or The first entity and the second entity are located at different levels but belong to the same group.

17. The method of claim 14, wherein, The initial knowledge graph contains an edge between the third node corresponding to the third entity and the first node, and the initial knowledge graph is supplemented based on the first feature vector, including: The graph encoder is used to extract the fourth feature vector of the fourth node corresponding to the fourth entity in the initial knowledge graph; In response to the similarity between the first feature vector and the fourth feature vector satisfying a third preset condition, an edge between the third node and the fourth node is added to the initial knowledge graph.

18. The method according to claim 17, wherein, The first entity and the third entity belong to different groups, while the first entity and the fourth entity belong to the same group.

19. The method according to claim 18, wherein, The first entity and the fourth entity belong to the first group, and the third entity belongs to the second group.

20. The method of claim 18, wherein: The fourth entity is located at a lower level than the first entity; or The fourth entity is located at a higher level than the first entity; or The fourth entity is located at the same level as the first entity.

21. The method according to claim 14, wherein, The initial knowledge graph contains an edge between the third node corresponding to the third entity and the first node, and the initial knowledge graph is supplemented based on the first feature vector, including: The graph encoder is used to extract the third feature vector of the third node, the fourth feature vector of the fourth node corresponding to the fourth entity in the initial knowledge graph, and the fifth feature vector of the fifth node corresponding to the fifth entity in the initial knowledge graph. In response to the similarity between the first feature vector and the fourth feature vector satisfying a fourth preset condition and the similarity between the third feature vector and the fifth feature vector satisfying a fifth preset condition, an edge between the fourth node and the fifth node is added to the initial knowledge graph.

22. The method according to claim 21, wherein, The first entity and the third entity belong to different groups, the first entity and the fourth entity belong to the same group, and the third entity and the fifth entity belong to the same group.

23. The method according to claim 22, wherein, The first entity and the fourth entity belong to the first group, and the third entity and the fifth entity belong to the second group.

24. The method of claim 22, wherein: The fourth entity is located at a lower level than the first entity, and the fifth entity is located at a lower level than the third entity; or The fourth entity is located at a higher level than the first entity, and the fifth entity is located at a higher level than the third entity; or The fourth entity is located at the same level as the first entity, and the fifth entity is located at the same level as the third entity.

25. The method according to any one of claims 15 to 24, wherein, The added edge attributes include weights determined based on the corresponding similarity, wherein the determined weights increase as the corresponding similarity increases.

26. The method according to any one of claims 14 to 24, wherein, The initial knowledge graph is the knowledge graph used to train the graph encoder.

27. A drug development method, comprising: Obtain a knowledge graph constructed by the method according to any one of claims 10 to 26; as well as The knowledge graph is input into an artificial intelligence prediction model comprising a graph encoder trained by any one of claims 1 to 9 to obtain prediction results.

28. An electronic device comprising: One or more processors; as well as A memory storing computer-executable instructions, which, when executed by the one or more processors, cause the one or more processors to perform the method according to any one of claims 1 to 27.

29. A non-transient storage medium having stored thereon computer-executable instructions, which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 27.

30. A computer program product comprising instructions that, when executed by a processor, implement the method according to any one of claims 1 to 27.

Images