Knowledge graph-based self-supervised drug-disease association prediction method and system

Abstract

The invention provides a self-supervised drug-disease association prediction method and system based on a knowledge graph, and belongs to the technical field of drug discovery in bioinformatics. Constructing and training a drug-disease association prediction model, wherein the drug-disease association prediction comprises main task drug-disease association prediction and auxiliary task self-supervised learning; the main task learns vector representation of each node from the biomedical knowledge graph through a heterogeneous graph neural network; predicting the association probability of the disease vector representation and the drug vector representation by using the disease vector representation and the drug vector representation; a loss function combining the main task and the auxiliary task in model training is used as a loss function of the model; providing an additional training label for the main task through the auxiliary task; and the auxiliary task optimization gradient is corrected in each round of training. The problems that only a single comparative learning auxiliary task is designed in an existing method, self-supervision based on structures, semantics or attributes is lacked, and the model prediction accuracy is reduced due to the fact that the gradient optimization direction and magnitude between a self-supervision auxiliary task and a main task are seriously inconsistent are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of drug discovery in bioinformatics, and in particular to a self-supervised drug-disease association prediction method and system based on a knowledge graph. Background Art

[0002] The traditional drug development process includes four main stages: target identification and verification, compound screening and lead optimization, preclinical evaluation and clinical verification. According to statistics, the development cycle of a new drug approved by the FDA is at least 10 to 15 years, but due to adverse drug reactions, about 90% of experimental drugs fail to pass Phase I clinical trials. How to achieve cost reduction and efficiency improvement in drug development has become an urgent problem to be solved. The technology of finding new indications or new uses for existing drugs is called drug repositioning. In recent years, thanks to the high-throughput data accumulated in the biomedical field, subsequent analysis has become possible; and the new generation of information technologies including artificial intelligence, cloud computing, and big data have provided algorithm support; computational drug repositioning methods have been further developed. In order to effectively integrate and organize multi-source biomedical data sources, knowledge graphs are introduced into drug repositioning. Knowledge graphs can be regarded as a knowledge base with a directed graph structure. It maps concepts and entities in the real world to nodes in the graph structure, and represents the relationships between concepts, concepts and entities, and entities as edges in the graph structure to achieve the purpose of describing objective facts. It is widely used due to its advantages in knowledge expression, fusion, and reasoning, but the existing drug repositioning models based on biomedical knowledge graphs still face the problem of data sparsity.

[0003] To alleviate this problem, existing technologies use self-supervised learning technology to inject additional training pseudo-labels into prediction tasks. However, in existing technologies, such methods usually only design a single contrastive learning auxiliary task, ignoring other types of self-supervised learning based on structure, semantics or attributes. Therefore, it is impossible to fully obtain the intrinsic implicit attributes of various types of biological entities and the complex semantic associations between drugs and diseases. In addition, the serious inconsistency in the direction and magnitude of gradient optimization between the self-supervised task and the main task leads to reduced model prediction accuracy.

[0004] Therefore, there is a need for a knowledge graph-based self-supervised drug-disease association prediction method that can obtain implicit attributes of multiple types of entities and correct the gradient optimization direction. Summary of the invention

[0005] In view of this, the present invention provides a self-supervised drug-disease association prediction method and system based on knowledge graph, which solves the problem that the single type of self-supervised learning task cannot fully obtain the intrinsic implicit attributes of various types of biological entities, and the serious inconsistency in the direction and magnitude of gradient optimization between the self-supervised task and the main task leads to reduced model prediction accuracy.

[0006] To this end, the present invention provides the following technical solutions:

[0007] Self-supervised drug-disease association prediction method based on knowledge graph, including:

[0008] Construct and train a drug-disease association prediction model; the model includes:

[0009] The main task is drug-disease association prediction and the auxiliary task is self-supervised learning;

[0010] The main tasks include:

[0011] Learn the vector representation of each entity from the biomedical knowledge graph through a heterogeneous graph neural network;

[0012] Using the disease vector representation and the drug vector representation, the association probability between the two is predicted through an association predictor;

[0013] The model training includes:

[0014] The loss function of the combined main task and auxiliary task is used as the loss function of the model;

[0015] Provide additional training labels for the main task through auxiliary tasks;

[0016] Correct the direction and magnitude of gradient optimization for the self-supervised auxiliary task at each training epoch.

[0017] Furthermore, the auxiliary tasks include:

[0018] The disease node binary classification task is used to determine whether the disease is a common disease or a rare disease based on the disease label information;

[0019] Drug type multi-classification task, used to identify the class to which a drug belongs based on anatomical, therapeutic, and chemical classification systems;

[0020] The node similarity comparison task makes the representation vectors corresponding to semantically similar entities closer in the representation space, and makes the representation vectors corresponding to semantically unrelated entities farther in the representation space.

[0021] Furthermore, learning the vector representation corresponding to each entity from the biomedical knowledge graph through the heterogeneous graph neural network includes:

[0022] Calculate the message vector from the source node to the target node of the knowledge graph:

[0023]

[0024] in, Represents the parameter matrix related to edge type, e s represents the source node, r represents the relationship, et Indicates the target node; ATT(e s ,r,e t ) indicates that each relationship r calculates the scaled dot product attention score corresponding to the h-th attention head;

[0025]

[0026] in, represents the biomedical knowledge graph, ε and Represents entity and relationship sets respectively; each triple <e s ,r,e t >Describe a piece of factual knowledge, namely the head entity e h and tail entity e t connected by relation r; Indicated in Upper node e t The vector representation of the lth layer, d represents the vector dimension, given a triple <e s ,r,e t >, the corresponding node types and edge types are φ(s), φ(t) and The target node e t With the source node e s Projected into multi-head query vector, key vector and value vector, that is, as well as Where h represents the hth attention head, and the total number of attention heads is H; as well as is the parameter matrix related to the node type;

[0027] Update the vector representation of the target node using the message vector:

[0028]

[0029] in, Trainable parameters related to node types;

[0030] After L layers of calculation, the entity representation output by the last layer is recorded as |ε| is the total number of entities.

[0031] Furthermore, the method of using the disease vector representation and the drug vector representation in the entity to predict the association probability between the two through an association predictor includes:

[0032] Drug expression h u With the disease vector h v They are connected in series and input into a multi-layer perceptron to calculate the probability of drug u being associated with disease v:

[0033] y u,v=sigmoid(MLP(h u ||h v ))

[0034] Drug-disease association prediction is optimized using the cross-entropy loss function:

[0035]

[0036] Among them, S is the set of positive and negative training samples, is the true label.

[0037] Furthermore, the disease node binary classification task includes:

[0038] The sigmoid function is used to calculate the probability score corresponding to any disease node:

[0039]

[0040] Among them, W b is a trainable parameter;

[0041] The disease node binary classification task is optimized by the cross entropy loss function:

[0042]

[0043] Among them, |ε DI | is the total number of disease type entities, is the true label value of entity i.

[0044] Furthermore, the drug type multi-classification task includes:

[0045] The drug node vector representation is input into the multi-layer perceptron, and the category probability distribution is calculated using the softmax function:

[0046]

[0047] The drug type multi-classification task is optimized using a cross entropy loss function.

[0048] Furthermore, the node similarity comparison task loss function is:

[0049]

[0050] Among them, topK(·) is the top-K nearest neighbors corresponding to the entity, and τ is the temperature coefficient.

[0051] Furthermore, the loss function of the combined main task and auxiliary task is used as the loss function of the model:

[0052]

[0053] Furthermore, the correction of the auxiliary task optimization gradient in each round of training includes:

[0054] Compute the gradient of the shared parameters θ:

[0055]

[0056] The shared parameters θ are updated via gradient descent: Where α is the learning rate;

[0057] According to the main task gradient Dynamically adjust the gradient of each auxiliary task The direction and magnitude of

[0058] like Correct the gradient direction corresponding to the auxiliary task, Project to Normal plane:

[0059]

[0060] Among them, · represents the dot product operation;

[0061] In order to avoid the excessive gradient of the auxiliary task dominating the overall optimization trend, a smoothing factor ω is introduced to make and The gradient magnitudes are close:

[0062]

[0063] Here, ω is a hyperparameter between 0 and 1.

[0064] The self-supervised drug-disease association prediction method system based on knowledge graph includes:

[0065] A node representation learning module that learns the vector representation of each entity from the biomedical knowledge graph through a heterogeneous graph neural network;

[0066] A self-supervised learning module for providing additional training labels, including disease node binary classification tasks, drug type multi-classification tasks, and node similarity comparison tasks;

[0067] The association predictor module includes: concatenating drug and disease vector representations, obtaining drug-disease association probability through the association predictor, and optimizing it through the cross entropy loss function;

[0068] An adaptive gradient balancing module that calculates whether the gradient directions between the auxiliary task and the main task conflict and whether the gradient magnitude difference is too large, and corrects the optimization gradient of the self-supervised task in each training round.

[0069] Advantages and positive effects of the present invention:

[0070] The present invention adopts a heterogeneous graph neural network (Heterogeneous Graph Transformer) model to encode the biomedical knowledge graph, so as to better capture the heterogeneous properties of the biomedical knowledge graph.

[0071] In addition to introducing contrastive learning tasks, the present invention further considers auxiliary tasks related to drugs and diseases, namely, binary classification of disease nodes and multi-classification of drug types. The self-supervision signal is mined from unlabeled data, and the model is jointly trained to learn the representation vectors that are valuable for the main task. The existing graph neural network-based model is improved. Only a single contrastive learning auxiliary task cannot fully obtain the intrinsic implicit attributes of various types of biological entities, as well as the complex semantic associations between drugs and diseases.

[0072] The present invention corrects the gradient direction and magnitude of the self-supervised task through an adaptive gradient balancing strategy, alleviating the problem of inconsistent gradient optimization direction and magnitude between the self-supervised task and the main task, thereby better transferring the implicit knowledge learned in the auxiliary task to the main task of drug-disease association prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0074] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention.

[0075] Figure 2 Schematic diagram of the biomedical knowledge graph in an embodiment of the present invention.

[0076] Figure 3 This is a diagram of the overall architecture of the model of an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0078] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0079] The present invention provides a self-supervised drug-disease association prediction method and system based on a knowledge graph, and designs an auxiliary self-supervised learning task for the main task of drug-disease association prediction; in order to solve the problem that the intrinsic implicit attributes of various types of biological entities and the complex semantic associations between drugs and diseases cannot be fully obtained; the self-supervised learning tasks of the present invention include: disease node binary classification, drug type multi-classification and node similarity comparison tasks; at the same time, in order to solve the problem that the gradient optimization direction and magnitude between the self-supervised task and the main task are seriously inconsistent in the existing self-supervised auxiliary tasks, resulting in a decrease in model prediction accuracy, an adaptive gradient balance is introduced, and in each training round, the direction and magnitude of the gradient of each auxiliary task are dynamically adjusted according to the main task gradient.

[0080] Combined with Figure 1 The embodiment flow chart shown further illustrates the method of the present invention:

[0081] S1. Define the biomedical knowledge graph: where ε and Represents entity and relationship sets respectively. Each triple <e s ,r,e t >Describe a piece of factual knowledge, namely the head entity e h and tail entity e t They are connected by the relationship r. In addition, both entities and relationships include multiple types, such as entity types include drugs, diseases, genes, phenotypes, pathways, and side effects, and relationship types include drug-disease, gene-pathway, and disease-phenotype.

[0082] Based on the above biomedical knowledge graph The Heterogeneous Graph Transformer model is used to learn the vector representations corresponding to each entity in the knowledge graph:

[0083] make Indicated in Upper node e t The vector representation of the lth layer, d represents the vector dimension. Given a triple <e s ,r,e t >, the corresponding node types and edge types are φ(s), φ(t) and ; Set the target node e t With the source node e s Projected into multi-head query vector, key vector and value vector, that is, as well as Where h represents the hth attention head, and the total number of attention heads is H; as well as It is the parameter matrix related to the node type.

[0084] For each relation r, calculate the scaled dot product attention score corresponding to the h-th attention head:

[0085]

[0086] in The target node e t exist The neighbor nodes on Parameters related to edge types. From the source node e s Follow relationship r to target node e t The message vector is:

[0087]

[0088] in, It is also a parameter matrix related to edge type. t The vector representation update formula at the (l+1)th layer is:

[0089]

[0090] in is a trainable parameter related to the node type. After L layers of calculation, the entity representation output by the last layer is recorded as |ε| is the total number of entities. To simplify the representation, remove the superscript L, that is,

[0091] S2. In order to fully learn more robust entity representations under limited training sample scenarios, three types of self-supervised learning auxiliary tasks are designed, namely disease node binary classification, drug type multi-classification, and node similarity comparison tasks.

[0092] Disease node binary classification task: Design an auxiliary binary classification task based on the available label information of the disease, that is, to determine whether the disease is a common disease or a rare disease.

[0093] Given any disease type node e1∈ε DI , ε DI for All disease entities in the node vector obtained in step S1 are expressed as Use the sigmoid function to calculate the probability score corresponding to the node:

[0094]

[0095] Among them, W b is a trainable parameter. This auxiliary task is optimized by the cross entropy loss function:

[0096]

[0097] where |ε DI | is the total number of disease type entities, is the true label value of entity i.

[0098] Drug type multi-classification task: Identify the drug category according to the anatomical therapeutic chemical classification system (ATC) and define the drug type multi-label classification auxiliary task. In this embodiment, for any drug type entity e2∈ε DR , ε DR for All drug entities in the same node vector obtained in step S1 are expressed as And input it into the Multilayer Perceptron (MLP), and use the softmax function to calculate the category probability distribution:

[0099]

[0100] This auxiliary task is also optimized by the cross entropy loss function, which is denoted as

[0101] For the node similarity comparison task, nodes with similar attributes should be distributed as close as possible in the representation space. Therefore, to measure the similarity between entities of the same type in the biomedical knowledge graph: first define the similarity metric for drug and disease entities. In this embodiment, the similarity between any drugs is calculated based on the drug fingerprint similarity and the Gaussian interaction spectral kernel similarity. For each drug, the top-K nearest neighbor drug entities are screened out. Similarly, the similarity between any diseases is calculated based on the disease phenotype similarity and the Gaussian interaction spectral kernel similarity, and for each disease, the top-K nearest neighbor disease entities are screened out. Based on the above-mentioned node similarity measurement method, the node similarity comparison task is designed using the contrastive learning paradigm, so that the representation vectors corresponding to semantically similar instances (positive samples) are closer in the representation space, and the representation distances corresponding to semantically irrelevant instances (negative samples) are farther. Given any node e, the top-K nearest neighbor entities corresponding to the node are regarded as positive examples, and other entities with low similarity are regarded as negative examples. The contrastive learning loss function is defined as:

[0102]

[0103] where topK(·) is the top-K nearest neighbors corresponding to the entity, and τ is the temperature coefficient.

[0104] S3. Given a drug-disease entity pair (u, v) to be predicted, represent the drug obtained in step S1 as h u With the disease vector h v They are concatenated and input into a multilayer perceptron to calculate the probability of drug u being associated with disease v:

[0105] y u,v =sigmoid(MLP(h u ||h v ))

[0106] Drug-disease association prediction is optimized using the cross-entropy loss function:

[0107]

[0108] Where S is a set of positive and negative training samples, is the true label.

[0109] S4, self-supervised auxiliary tasks (i.e. as well as ) and the main task of drug-disease association prediction (i.e. ) In this embodiment, multi-task learning is used for joint optimization, and the overall loss function is: During the tth training iteration, The gradient calculation formula for the shared parameter θ is:

[0110]

[0111] The shared parameters θ are updated via gradient descent: Where α is the learning rate. Therefore, if (||·|| is the L2 norm), the optimizer tends to update parameters towards the i-th self-supervised auxiliary task rather than the main task, which causes a serious optimization imbalance problem and leads to reduced prediction accuracy.

[0112] Since the gradient direction and magnitude corresponding to the main task and the auxiliary task change dynamically throughout the optimization process, in each training round, the gradient of the main task Dynamically adjust the gradient of each auxiliary task direction and magnitude.

[0113] In this embodiment, if Correct the gradient direction corresponding to the auxiliary task, that is, Project to Normal plane:

[0114]

[0115] where · represents the dot product operation.

[0116] In order to avoid the excessive gradient of the auxiliary task dominating the overall optimization trend, a smoothing factor ω is introduced to make and The gradient magnitudes are close:

[0117]

[0118] Where ω is a hyperparameter between 0 and 1. After the above corrections to the gradient direction and magnitude of the auxiliary task, the implicit knowledge learned in the auxiliary task can be better transferred to the main task of drug-disease association prediction. Finally, the main task and the self-supervised auxiliary task are still jointly optimized through multi-task learning.

[0119] The present invention also provides a self-supervised drug-disease association prediction system based on a knowledge graph, comprising: a node representation learning module, a self-supervised learning module, an association predictor module, and an adaptive gradient balancing module.

[0120] Since the gradient optimization direction and magnitude of the self-supervised auxiliary task and the main task of drug-disease association prediction change dynamically in each training round, if the gradient direction of the auxiliary task is opposite to that of the main task, or the gradient magnitude of the auxiliary task is much greater than that of the main task, it will cause parameter optimization to be carried out towards the auxiliary task rather than the main task, thereby increasing the training time and reducing the model prediction accuracy. To solve this problem, the present invention proposes an adaptive gradient balancing strategy to correct the gradient direction and magnitude of the self-supervised task, so as to better transfer the implicit knowledge learned in the auxiliary task to the main task of drug-disease association prediction.

[0121] Combination Figure 3 The structure diagram further explains:

[0122] The node representation learning module is responsible for extracting the vector representation of any entity from the biomedical knowledge graph;

[0123] The self-supervised learning module is responsible for formulating self-supervised auxiliary tasks to fully learn more robust entity representations under limited training sample scenarios; self-supervised auxiliary tasks include: disease node binary classification, drug type multi-classification, and node similarity comparison tasks.

[0124] The association predictor module calculates the drug-disease association probability through the association predictor and optimizes it through the cross-entropy loss function;

[0125] The adaptive gradient balancing module is used to dynamically correct the gradient optimization direction and magnitude of the self-supervised task in each training round.

[0126] The present invention solves the problem of being unable to fully acquire the intrinsic implicit attributes of various types of biological entities and the complex semantic associations between drugs and diseases by designing multiple self-supervised learning tasks; and introduces adaptive gradient balance for self-supervised auxiliary tasks, dynamically adjusting the direction and magnitude of each auxiliary task gradient according to the main task gradient in each training round. This solves the problem of serious inconsistency in the direction and magnitude of gradient optimization between the self-supervised task and the main task in existing self-supervised auxiliary tasks, which leads to reduced model prediction accuracy.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-supervised drug-disease association prediction method based on knowledge graph, characterized in that: include: Construct and train a drug-disease association prediction model; the model includes: The main task is drug-disease association prediction and the auxiliary task is self-supervised learning; The main tasks include: Learn the vector representation of each entity from the biomedical knowledge graph through a heterogeneous graph neural network; Using the disease vector representation and the drug vector representation, the association probability between the two is predicted through an association predictor; The model training includes: The loss function of the combined main task and auxiliary task is used as the loss function of the model; Provide additional training labels for the main task through auxiliary tasks; The direction and magnitude of gradient optimization for the self-supervised auxiliary task are modified in each training epoch.

2. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 1, characterized in that: The auxiliary tasks include: The disease node binary classification task is used to determine whether the disease is a common disease or a rare disease based on the disease label information; Drug type multi-classification task, used to identify the class to which a drug belongs based on anatomical, therapeutic, and chemical classification systems; The node similarity comparison task makes the representation vectors corresponding to semantically similar entities closer in the representation space, and makes the representation vectors corresponding to semantically unrelated entities farther in the representation space.

3. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 1, characterized in that: The method of learning the vector representation corresponding to each entity from the biomedical knowledge graph through the heterogeneous graph neural network includes: Calculate the message vector from the source node to the target node of the knowledge graph: in, Represents the parameter matrix related to edge type, e s represents the source node, r represents the relationship, e t Indicates the target node; ATT(e s ,r,e t ) indicates that each relationship r calculates the scaled dot product attention score corresponding to the h-th attention head; in, represents the biomedical knowledge graph, ε and Represents entity and relationship sets respectively; each triple <e s ,r,e t >Describe a piece of factual knowledge, namely the head entity e h and tail entity e t connected by relation r; Indicated in Upper node e t The vector representation of the lth layer, d represents the vector dimension, given a triple <e s ,r,e t >, the corresponding node types and edge types are φ(s), φ(t) and The target node e t With the source node e s Projected into multi-head query vector, key vector and value vector, that is, as well as Where h represents the hth attention head, and the total number of attention heads is H; as well as is the parameter matrix related to the node type; Update the vector representation of the target node using the message vector: in, Trainable parameters related to node types; After L layers of calculation, the entity representation output by the last layer is recorded as |ε| is the total number of entities.

4. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 1, characterized in that: The method of using the disease vector representation and the drug vector representation in the entity to predict the association probability between the two through an association predictor includes: Drug expression h u With the disease vector h v They are connected in series and input into a multi-layer perceptron to calculate the probability of drug u being associated with disease v: y u,v =sigmoid(MLP(h u ||h v )) Drug-disease association prediction is optimized using the cross-entropy loss function:

5. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 2, characterized in that: The disease node binary classification task includes: The sigmoid function is used to calculate the probability score corresponding to any disease node: Among them, W b is a trainable parameter; The disease node binary classification task is optimized by the cross entropy loss function: Among them, |ε DI | is the total number of disease type entities, is the true label value of entity i.

6. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 2, characterized in that: The drug type multi-classification task includes: The drug node vector representation is input into the multi-layer perceptron, and the category probability distribution is calculated using the softmax function: The drug type multi-classification task is optimized using a cross entropy loss function.

7. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 2, characterized in that: The node similarity comparison task loss function is: Among them, topK(·) is the top-K nearest neighbors corresponding to the entity, and τ is the temperature coefficient.

8. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 1, characterized in that: The loss function of the combined main task and auxiliary task is used as the loss function of the model:

9. The self-supervised drug-disease association prediction method based on knowledge graph according to claim 1, characterized in that: Each round of training performs correction of auxiliary task optimization gradient, including: Compute the gradient of the shared parameters θ: The shared parameters θ are updated via gradient descent: Where α is the learning rate; According to the main task gradient Dynamically adjust the gradient of each auxiliary task The direction and magnitude of like Correct the gradient direction corresponding to the auxiliary task, Project to Normal plane: Among them, · represents the dot product operation; In order to avoid the excessive gradient of the auxiliary task dominating the overall optimization trend, a smoothing factor ω is introduced to make and The gradient magnitudes are close: Here, ω is a hyperparameter between 0 and 1.

10. A self-supervised drug-disease association prediction method system based on knowledge graph, characterized in that: include: A node representation learning module that learns the vector representation of each entity from the biomedical knowledge graph through a heterogeneous graph neural network; A self-supervised learning module for providing additional training labels, including disease node binary classification tasks, drug type multi-classification tasks, and node similarity comparison tasks; The association predictor module includes: concatenating drug and disease vector representations, obtaining drug-disease association probability through the association predictor, and optimizing it through the cross entropy loss function; An adaptive gradient balancing module that calculates whether the gradient directions between the auxiliary task and the main task conflict and whether the gradient magnitude difference is too large, and corrects the optimization gradient of the self-supervised task in each training round.

Images