A drug-target interaction prediction method based on phenotype guide subgraph correction

Abstract

The application discloses a drug-target interaction prediction method based on phenotype guide subgraph correction, and belongs to the technical field of biological information. In order to solve the problem that high-frequency disease, side effect and other phenotype nodes easily interfere with the specific relationship expression of a candidate pair, the application takes a candidate drug-target pair as the center, discovers diseases, side effects and related entities along multiple meta-paths, screens according to the structural specificity of the phenotype nodes and the semantic diversity of the meta-paths, inhibits highly-numbered general nodes, and retains selected phenotype-related paths to construct a candidate pair-specific heterogeneous subgraph. A graph neural network and a pre-trained protein language model are used to extract drug and target features respectively to form entity pair representation; a heterogeneous graph attention network and a query-aware attention mechanism are used to encode the heterogeneous subgraph to obtain relationship evidence representation. A relationship evidence is converted into a controlled correction amount through a gating residual correction module, the entity pair representation is corrected, and a classification predictor is used to output a drug-target interaction prediction result.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a drug-target interaction prediction method based on phenotypic guide graph correction. Background Technology

[0002] Drug-target interactions refer to the specific binding or functional regulation relationships between drug molecules and protein targets, serving as a crucial foundation for drug discovery, mechanism of action studies, drug retargeting, and adverse reaction analysis. Accurately identifying potential drug-target interactions helps narrow down the pool of candidate compounds, reduces the cost of new drug development, and provides a basis for developing new indications for existing drugs. Traditional methods for verifying drug-target interactions primarily involve biochemical experiments, affinity assays, and cell experiments. However, these experiments are typically time-consuming, costly, and involve a large number of candidate combinations, making it difficult to comprehensively screen large-scale drug-target combinations. Therefore, using computational methods to predict unknown drug-target interactions has become an important research direction in bioinformatics and computational drug development.

[0003] Early computational methods primarily utilized drug chemical structural similarity, target sequence similarity, and known interaction matrices for prediction, including ligand-based methods, molecular docking methods, matrix factorization methods, and traditional machine learning methods. These methods typically rely on artificially constructed molecular descriptors, protein sequence features, or similarity matrices, and their predictive performance is easily affected by feature representation capabilities and the number of known samples. Furthermore, the interaction between drugs and targets involves multiple factors such as molecular structure, sequence composition, and biological function; simple similarity propagation or linear mapping cannot fully describe the complex relationship between the two.

[0004] With the development of deep learning technology, convolutional neural networks, recurrent neural networks, Transformers, and graph neural networks are increasingly being used for drug-target interaction prediction. Existing methods typically represent drugs as smils and targets as amino acid sequences, using deep neural networks to learn higher-order representations of drugs and targets. These methods improve the ability to model the molecular structure of drugs and the semantics of target sequences, providing a relatively stable foundation of entity information for drug-target interaction prediction. Building upon this foundation, further incorporating phenotypic node association information such as diseases and side effects can supplement the entity representation from the perspectives of biological function and network relationships, thus forming a more comprehensive predictive basis.

[0005] To leverage multi-source biomedical relationships, some existing technologies construct heterogeneous biomedical networks from entities such as drugs, targets, diseases, and side effects, along with their relationships, and utilize graph neural networks to learn the association information between different types of entities. Since different nodes and relationships possess different biological semantics, existing methods typically employ meta-paths to describe the complex associations between drugs and targets, obtaining relationship representations through path statistics, path instance encoding, or attention aggregation. For example, drugs can indirectly connect with targets through relationships such as diseases, side effects, similar drugs, and drug-to-drug interactions.

[0006] However, existing methods primarily focus on the importance of metapath types or path instances, neglecting the connection frequency of intermediate phenotypic nodes. In heterogeneous biomedical networks, some common diseases and side effects are linked to a large number of drugs or targets, frequently appearing in multiple metapath types, leading to similar relationship evidence for different candidate drug-target pairs. Conversely, some phenotypic nodes with fewer connections but strong candidate pair specificity may be difficult to fully express due to the limited number of paths. Furthermore, relying solely on the number of paths makes it difficult to distinguish between the recurrence of similar paths and the joint support of multiple relational semantics, thus affecting the specificity of metapath relation representations.

[0007] Existing methods typically fuse metapath relationship representations with drug / target entity representations using concatenation, summation, or fixed weighting, making it difficult to adjust the contributions of the two types of information based on the specific relationship structure of candidate pairs. Therefore, current technologies still need to address the impact of high-frequency phenotypic nodes on the distribution of metapath evidence, and the lack of adaptive coordination between entity attributes and relationship information. How to balance the structural specificity of phenotypic nodes with the semantic diversity of metapaths, construct candidate pair-specific relationship structures, and utilize these structures to effectively modify drug-target entity representations is a key technical problem that needs to be solved in drug-target interaction prediction.

[0008] In summary, existing drug-target interaction prediction techniques can jointly model drug molecular structures, target protein sequences, and heterogeneous biomedical networks. However, during the construction of metapath subgraphs, there remains an imbalance between the connection frequency of phenotypic nodes and their local semantic contributions, with high-frequency phenotypic nodes occupying a large proportion, affecting the expression of candidate pair-specific relationships. Furthermore, a more flexible collaborative mechanism is needed between entity representation and subgraph relationship representation. Therefore, it is necessary to investigate a prediction method that can highlight specific phenotypic relationships, maintain the semantic diversity of metapaths, and adaptively correct drug-target entity representations using heterogeneous subgraph relationship evidence. Summary of the Invention

[0009] To address the problems in existing drug-target interaction prediction methods when utilizing multi-source biomedical relationships, such as the tendency of high-frequency disease and side-effect phenotypic nodes to interfere with the expression of candidate pair-specific relationships, and the lack of adaptive synergy between entity attribute representation and heterogeneous relationship representation, this invention proposes a drug-target interaction prediction method based on phenotypic guide graph correction.

[0010] A drug-target interaction prediction method based on phenotypic guide graph correction includes:

[0011] For candidate drug-target pairs, local meta-path instances are extracted to obtain a set of meta-path instances. Phenotypic nodes in the meta-path instances are specifically screened to obtain a set of retained disease nodes and side effect nodes. This leads to the construction of candidate pair-specific isomer subgraphs. For each node in the candidate pair-specific heterogeneous subgraph Construct node input features Then, a heterogeneous graph attention network is used to process the heterogeneous subgraphs. The middle node represents the update; utilizing candidate drug nodes and candidate target nodes The final representation constructs the query vector. Then based on and nodes In heterogeneous subgraphs The final representation in the graph is calculated for each subgraph node. Attention score relative to candidate drug-target pair ,Will Normalization is performed on all nodes of the subgraph to obtain the normalized attention score. Ultimately, based on node In heterogeneous subgraphs The final representation in the graph is then attention-weighted aggregated to obtain the phenotypic guided isomer subgraph representation of the drug-target pair. ;

[0012] For candidate drugs Drug molecular maps were constructed based on their molecular structures. For drug molecule graphs, a drug molecule graph encoder is used for encoding, and each atom node in the drug molecule graph is encoded. Construct atomic nodes initial atomic features Based on atoms and atoms Determining the chemical bond characteristics between them ;Will As input to the graph neural network, the atomic node representations are updated using the graph neural network. During the update process, chemical bond features are incorporated, ultimately yielding the final atomic representation layer of the graph neural network, which in turn provides the molecular structure representation of the candidate drug. ;

[0013] Constructing sequence encoding representations of candidate targets Representing the molecular structure of candidate drugs Representation of candidate target sequences Perform element-by-element interactions to obtain candidate entity interaction features. Candidate entity interaction features Mapped as entity interaction representation of drug-target pairs ;based on and Obtain fusion features The fused features are input into a feedforward neural network to generate a subgraph relationship correction vector. ;according to and Get corrected gating ;based on , combined and The corrected candidate drug-target pair representation is obtained. ;Will Inputting the data into a classification and prediction network yields candidate drugs. With candidate targets The classification results.

[0014] Furthermore, the process of specifically filtering phenotypic nodes in metapath instances includes:

[0015] From the metapath instance collection Extract disease nodes and side effect nodes to form candidate disease node sets and candidate side effect node sets; calculate phenotypic nodes. structural specificity ,in Represents a phenotypic node The degree of global connectivity in the original interconnected network. Represents nodes A set of nodes of the same type, where u is a node in the set;

[0016] Statistical phenotype nodes The semantic diversity is obtained by counting the number of different metapath types covered under the current candidate drug-target pair. , Represents a path instance The associated metapath template, Represents a path instance The corresponding phenotypic anchor points, For indicator functions; This refers to the set of meta-path instances obtained based on local path instances in a multi-source biomedical heterogeneous network.

[0017] Calculate the overall score ,in The weight parameter is used to determine the semantic diversity of the metapath. Candidate disease nodes and candidate side effect nodes are sorted from highest to lowest according to their comprehensive scores and then filtered to obtain the retained sets of disease nodes and side effect nodes. .

[0018] Furthermore, the process of constructing a candidate pair-specific isomorphic subgraph includes:

[0019] reserve The metapath instance of a node is a local node set composed of candidate drug nodes, candidate target nodes, and all entity nodes in the retained path. , This represents the set of path instances retained after filtering by phenotypic nodes;

[0020] Deduplication and merging are performed based on the global entity identifier, ensuring that when the same entity appears through multiple meta-path instances, only one node is retained in the local node set, thereby based on the local node set. Local relation edges are reconstructed from multi-source biomedical heterogeneous networks to obtain candidate pair heterogeneous subgraphs. ,in Represents the set of local relation edges.

[0021] Furthermore, for each node in the candidate pair-specific heterogeneous subgraph , construct node input features ,in Encoding for node type; This refers to the corresponding query anchor point of a node relative to the current candidate drug-target pair, indicating the candidate drug node and candidate target node in the current candidate pair to be predicted. For nodes Relative to the structural position characteristics of the current candidate drug-target pair, by nodes The shortest path distance to the current candidate drug node and the current candidate target node is determined in the candidate pair specific isomer subgraph; The feature indicates whether a node participates in a specific meta-path relationship. Its dimension corresponds to the number of meta-path types, and each dimension corresponds to one meta-path type. If the node participates in the corresponding meta-path, the value is 1; otherwise, it is 0. This represents the number of adjacent nodes that a node connects to under different relationships.

[0022] Furthermore, a heterogeneous graph attention network is used to focus on heterogeneous subgraphs. The process of updating a node includes:

[0023] Will The initial features of the nodes are mapped to the latent space, and the node features are updated using a multi-layer heterogeneous graph attention network; for the ... Layered heterogeneous graph attention network, based on nodes ,node In the Layer feature representation of computation nodes and neighboring nodes Attention score for relation types And normalize them within the same relation type. , Indicates the relationship With nodes A set of connected neighbors;

[0024] Relationship Weighted aggregation of neighbor information , For nodes In the The layer's feature representation is then used; the information obtained from different relationships is then fused by summation, and combined with residual links to obtain nodes. The Layer update node representation ,go through The final representation of the node is obtained after the layer is updated. .

[0025] Furthermore, As input to a graph neural network, the process of updating the representation of atomic nodes using a graph neural network includes:

[0026] For candidate drugs Drug molecular maps were constructed based on their molecular structures. For drug molecule graphs, a drug molecule graph encoder is used for encoding, and each atom node in the drug molecule graph is encoded. Construct atomic nodes initial atomic features Based on atoms and atoms Determining the chemical bond characteristics between them ;Will As input to the graph neural network, the representation of the atomic nodes is updated using the graph neural network. Layered graph neural networks first combine atomic nodes With neighboring atomic nodes The chemical bond features between atoms are mapped to the same feature space as the atom hidden representation, and the mapped chemical bond features are compared with those of neighboring atom nodes. The upper layer represents addition, and after nonlinear activation, neighboring atomic nodes are obtained. To the central atomic node The message being delivered and for the central atomic node The messages generated by all directly connected neighboring atomic nodes are summed and then compared with the central atomic node. By combining the feature representations from the previous layer, we obtain the neighborhood aggregation representation of the central atomic node. Then, atomic nodes are added through residual connections. In the feature representation of the previous layer, atomic nodes are obtained. In the Layer update representation .

[0027] Furthermore, construct sequence encoding representations of candidate targets. The process includes:

[0028] For candidate targets The amino acid sequence of the target protein is obtained, and the target sequence embedding is obtained using a protein language model. Based on the target sequence embedding, a global sequence representation of the protein sequence is obtained, and based on the global sequence representation of the protein sequence, a sequence encoding representation of the candidate target is obtained. .

[0029] Furthermore, the candidate pair entity interaction features ,in Represents element-wise product. Represents the absolute difference element by element; This indicates feature splicing.

[0030] Furthermore, the fusion feature ,in Represents element-wise product. Represents the absolute difference element by element; This indicates feature splicing.

[0031] Furthermore, the revised candidate drug-target pair representation ,in This is the overall residual correction factor.

[0032] The beneficial effects of this invention are:

[0033] This invention effectively integrates multi-source biomedical relationships, including drug molecular structure, target protein sequence, and disease, side effects, similar drugs, similar targets, drug-to-target interactions, and protein-to-protein interactions, by introducing a phenotype-guided candidate pair heterogeneous subgraph correction framework. On one hand, a phenotype node screening mechanism based on structure specificity and metapath semantic diversity reduces the interference of common phenotype nodes such as high-frequency diseases and side effects on local relationship evidence of candidate drug-to-target pairs, enhancing the expressive power of candidate pair-specific relationship structures. On the other hand, a heterogeneous graph attention encoder models multiple types of nodes and relationships in the phenotype-guided heterogeneous subgraph, fully capturing the complex indirect associations between drugs and targets formed through phenotype nodes. This invention can effectively fuse drug-to-target entity attribute information with heterogeneous subgraph relationship evidence, and adaptively correct entity representations based on the relationship structures of different candidate pairs through a gated residual mechanism, thereby improving the accuracy and robustness of drug-to-target interaction prediction. Experimental results show that this invention improves the rationality and generalizability of the model prediction process in multiple datasets and scenarios. Therefore, this invention can more accurately identify potential drug-target interactions, providing effective support for candidate compound screening, drug action mechanism research, and drug repositioning. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0035] Figure 2 This is a comparison chart of the AUC and AUPR performance of the method of the present invention in a randomly partitioned li dataset;

[0036] Figure 3 This is a comparison chart of the AUC and AUPR performance of the method of the present invention in a random partition of the luo dataset;

[0037] Figure 4 This is a comparison chart of the AUC and AUPR performance of the method of the present invention in the drug cold start setting;

[0038] Figure 5 This is a comparison chart of the AUC and AUPR performance of the method of the present invention in the target cold start setting;

[0039] Figure 6 This is a comparison chart of the AUC and AUPR performance of the method of the present invention and its ablation variants under cold start settings. Detailed Implementation

[0040] To address the problems of existing drug-target interaction prediction methods when utilizing heterogeneous biomedical networks, such as the tendency of high-frequency disease and side-effect phenotypic nodes to weaken the specific relationship expression of candidate drug-target pairs, and the lack of adaptive synergy between entity attribute representation and subgraph relationship representation, this invention proposes a drug-target interaction prediction method based on phenotypic-guided subgraph correction. This method uses candidate drug-target pairs as query centers to construct phenotypic-guided heterogeneous subgraphs of candidate pairs. It then utilizes a heterogeneous graph attention network to obtain subgraph relationship evidence, and uses this evidence as correction information to adaptively apply to the drug-target entity representation, thereby outputting the drug-target interaction prediction result. The invention will be described in detail below with specific implementation details.

[0041] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0042] This embodiment is a drug-target interaction prediction method based on phenotypic guided subgraph correction, including a data and relation network construction module, a phenotypic guided heterogeneous subgraph generation and encoding module, a candidate drug-target pair entity representation construction module, an entity representation adaptive correction module based on subgraph relation evidence, and a drug-target interaction classification prediction module.

[0043] The data and relationship network construction module specifically performs the following steps:

[0044] Step a1: Obtain data on drugs, targets, and multi-source biomedical relationships.

[0045] Specifically, acquiring drug sets and target set and a known set of drug-target interaction samples. ; wherein, each drug record in the drug set This includes a unique drug identifier and a corresponding SDF-format molecular structure file, the SDF-format molecular structure file being used to parse and obtain a drug molecule diagram; each target in the target set records... This includes a unique target identifier and a corresponding FASTA-formatted protein amino acid sequence file, the FASTA-formatted protein amino acid sequence file being used to generate a target protein sequence representation; and a set of known drug-target interaction samples. This includes samples consisting of drug unique identifiers, target unique identifiers, and interaction tags.

[0046] Simultaneously, it acquires multi-source biomedical relationship data, including drug-disease associations, target-disease associations, drug-side effect associations, drug-interactions, protein-protein interactions, drug similarity, and target similarity.

[0047] Step a2: Construct a multi-source biomedical heterogeneous network.

[0048] Specifically, drugs, targets, diseases, and side effects are treated as different types of nodes, and drug-disease associations, target-disease associations, drug-side effect associations, drug similarity, target similarity, drug-to-disease interactions, and protein-to-protein interactions are treated as different types of edges, resulting in a heterogeneous network:

[0049]

[0050] in, Represents a set of nodes. Denotes the set of edges. Represents node type mapping, Represents a relation type mapping.

[0051] Step a3: Divide the data for the candidate prediction task.

[0052] Specifically, a five-fold cross-validation dataset is constructed based on known drug-target interaction samples. Random partitioning prediction scenarios, drug cold-start prediction scenarios, and target cold-start prediction scenarios are set according to the partitioning method of the prediction objects. In the random partitioning prediction scenario, the drug-target interaction samples are randomly divided into five subsets. Each time, one subset is selected as the validation set, and the remaining four subsets are combined as the training set. In the drug cold-start prediction scenario, the drug set is divided into five subsets. Each time, the drug-target interaction samples corresponding to one drug subset are selected as the validation set, ensuring that candidate drugs in the validation set do not appear in the training set. In the target cold-start prediction scenario, the target set is divided into five subsets. Each time, the drug-target interaction samples corresponding to one target subset are selected as the validation set, ensuring that candidate targets in the validation set do not appear in the training set.

[0053] The phenotypic-guided heterogeneous subgraph generation and encoding module specifically performs the following steps:

[0054] Step b1: Extract local metapath instances for candidate drug-target pairs.

[0055] Specifically, for the candidate drug-target pair to be predicted drug nodes and target nodes To find the anchor point, the heterogeneous network constructed in step a2 searches for local path instances connecting the two according to a preset meta-path pattern, resulting in a set of meta-path instances:

[0056]

[0057] in, Indicates the first Metapath pattern Indicates connection to candidate drugs With candidate targets Path instance , Represents a path instance The metapath template to which it belongs.

[0058] In one implementation, the meta-path pattern includes non-drug-target interaction meta-paths such as drug-disease-target, drug-similar drug-disease-target, drug-disease-similar target-target, drug-side effect-drug-disease-target, drug-drug interaction drug-disease-target, and drug-disease-protein interaction target-target, as well as drug-target interaction meta-paths within the current training iteration; when adding drug-target interaction meta-paths, the current candidate edges are excluded to avoid using the label to be predicted as path evidence.

[0059] Step b2: Perform specific filtering on phenotypic nodes in the metapath instance.

[0060] Specifically, disease nodes and side effect nodes in heterogeneous biomedical networks are collectively referred to as phenotypic nodes. Disease nodes represent disease phenotypes associated with drugs or targets, while side effect nodes represent side effect phenotypes associated with drugs. These phenotypic nodes, within the metapath instance, carry the semantics of disease or side effect associations and do not include drug or target nodes. (From the metapath instance set...) Disease nodes and side effect nodes are extracted to form a set of candidate disease nodes and a set of candidate side effect nodes.

[0061] For a disease node, its global connectivity is determined by the drug-disease association matrix and the target-disease association matrix, that is, by counting the sum of the number of drug nodes and the number of target nodes directly associated with the disease node.

[0062] For a side effect node, its global connectivity is determined by the drug-side effect association matrix, which is the number of drug nodes directly associated with the side effect node.

[0063] The global connectivity is pre-statistically calculated on the original interconnected network.

[0064] For phenotypic nodes Calculate its structural specificity:

[0065]

[0066] in, Represents a phenotypic node In the original interconnected network (the heterogeneous network), the degree of global connectivity is the number of other entities directly associated with that node. Represents nodes A set of nodes of the same type, where u is a node in the set. Disease or side-effect nodes with higher connectivity tend to have lower structure specificity. By applying a logarithmic transformation and inverse mapping to the connectivity, the structure specificity is made to decrease monotonically with increasing connectivity. Furthermore, by normalizing within nodes of the same type, disease and side-effect nodes can be compared at their respective connectivity scales. This reduces the likelihood that highly connected disease or side-effect nodes will receive a higher retention priority simply because they have a larger number of associated entities.

[0067] Statistical phenotype nodes The semantic diversity is obtained by considering the number of different meta-path types covered under the current candidate drug-target pair:

[0068]

[0069] And calculate the overall score:

[0070]

[0071] in, The weighting parameter for the semantic diversity of metapaths; Represents a path instance The associated metapath template, Represents a path instance The corresponding set of phenotypic nodes; This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. Essentially, it counts how many meta-path templates node v contains, rather than the total number of path instances passing through that node. Even if the same meta-path template appears multiple times, it only contributes 1 point. A larger value indicates that node v can connect candidate drugs and targets from more diverse relational patterns, resulting in more diverse semantic evidence.

[0072] Subsequently, the candidate disease node set and the candidate side effect node set are sorted from high to low according to their comprehensive scores, respectively, to obtain the sets of retained disease nodes and side effect nodes. To avoid the two types of phenotypic nodes crowding out each other's retention slots due to differences in quantity and connectivity distribution, in a preferred embodiment, the top 20 disease nodes and the top 10 side effect nodes are retained; the number of nodes retained can be adjusted according to the data scale and computing resources.

[0073] Step b3: Construct and encode candidate pair-specific isomorphic subgraphs.

[0074] Specifically, retain the set of nodes containing the filtered phenotype. The metapath instance of the middle node, and the candidate drug node, candidate target node, and all entity nodes in the retained path are combined into a local node set:

[0075]

[0076] in, This represents the set of path instances retained after filtering by phenotypic nodes.

[0077] All nodes in the retained path are deduplicated and merged according to the global entity identifier, so that when the same entity appears through multiple meta-path instances, only one node is retained in the local node set. In this way, mutually independent and potentially duplicated path instances are transformed into candidate pairs of local topologies with shared nodes, enabling relational evidence from different meta-path templates to be jointly message-passed in the same heterogeneous subgraph.

[0078] Based on the local node set Reconstructing local relation edges from the original heterogeneous network yields candidate heterogeneous subgraphs:

[0079]

[0080] in, Represents the set of local relation edges.

[0081] For each node in the candidate pair-specific isomorphic subgraph Construct node input features:

[0082]

[0083] in, One-hot encoding for node types, representing drugs, diseases, side effects, and target proteins, respectively; This is the corresponding query anchor point for a node relative to the current candidate drug-target pair. It refers to the candidate drug node and candidate target node in the current candidate pair to be predicted. Both originate from the candidate drug-target pair input to the model and are always retained when constructing the candidate pair-specific heterogeneous subgraph. This feature identifies the nodes respectively. Is it a current candidate drug node and a current candidate target node? If the node... If a node is a candidate drug node, it is marked as a drug query anchor point; if the node If a node is a candidate target node, it will be marked as a target query anchor point; disease nodes, side effect nodes, and other drug nodes and target nodes other than the current candidate drug and candidate target in the heterogeneous subgraph will not be marked as query anchor points. For nodes Relative to the structural position characteristics of the current candidate drug-target pair, by nodes The shortest path distance to the current candidate drug node and the current candidate target node is determined in the candidate pair specific heterogeneous subgraph; where the current candidate drug node and the current candidate target node are the two query anchor points of the current candidate pair to be predicted. The feature indicates whether a node participates in a specific meta-path relationship. Its dimension corresponds to the number of meta-path types, and each dimension corresponds to one meta-path type. If the node participates in the corresponding meta-path, the value is 1; otherwise, it is 0. This represents the number of adjacent nodes that a node connects to under different relationships.

[0084] Subsequently, a heterogeneous graph attention network was used to analyze the heterogeneous subgraphs. The node representation is updated. Specifically, first, the initial features of the nodes are mapped to the latent space:

[0085]

[0086] in, Represents a node Input features, This represents the initial implicit representation of the node.

[0087] Then, a multi-layer heterogeneous graph attention network is used to update the node features. For the first... Layered heterogeneous graph attention network, assuming For relation types, Indicates the relationship With nodes The set of connected neighbors. For a node and neighboring nodes Calculate the attention score for the relation type:

[0088]

[0089] in, , For nodes ,node In the Layer feature representation; , For the corresponding weights, Learnable attention parameters specific to relation types;

[0090] And normalize within the same relation type:

[0091]

[0092] Subsequently, regarding the relationship The neighbor information is weighted and aggregated:

[0093]

[0094] Then, the information obtained from different relations is summed and merged, and the node representation is updated by combining the residual links:

[0095]

[0096] in, This indicates a relation type association pointing to node v; GELU is the GELU activation function.

[0097] go through After the layer is updated, the final representation of the node is obtained. .

[0098] Then, using candidate drug nodes and candidate target nodes The final representation constructs the query vector:

[0099]

[0100] in, This indicates vector concatenation. The query vector contains node representations of both candidate drugs and candidate targets.

[0101] Based on this, calculate each subgraph node. Attention score relative to the candidate drug-target pair:

[0102]

[0103] in, and represents the corresponding weight, and tanh is the hyperbolic tangent function.

[0104] Normalization is performed on all nodes of the subgraph:

[0105]

[0106] Finally, attention-weighted aggregation of the node representations is performed to obtain the subgraph context representation of the candidate drug-target pair:

[0107]

[0108] in, Represents a node Relative to the attention weights of candidate drug-target pairs, This represents the phenotypic-guided isomorphic subgraph representation of the candidate drug-target pair. This subgraph representation is used for subsequent adaptive correction of the entity representation of the candidate drug-target pair.

[0109] Through the above processing, this invention first utilizes global structural specificity and the semantic diversity of metapaths under candidate pair conditions to control the retention range of phenotypic nodes, then merges shared entities in the filtered paths, and reconstructs the multi-relational local topology. Finally, it uses query vectors composed of candidate drugs and candidate targets to conditionally aggregate local nodes, thereby enabling the obtained subgraph relation representation to simultaneously possess phenotypic node screening constraints, candidate pair relevance, and multi-relational topology information.

[0110] The candidate drug-target pair entity representation construction module specifically performs the following steps:

[0111] Step c1: Construct a molecular diagram representation of the candidate drug.

[0112] Specifically, for candidate drugs A drug molecule map was constructed based on its molecular structure:

[0113]

[0114] in, This represents the set of atomic nodes in a drug molecule. It represents the set of chemical bonds between atoms.

[0115] For drug molecule graphs, a drug molecule graph encoder is used for encoding, which encodes each atomic node in the drug molecule graph. Construct its initial atomic features This characteristic is mainly composed of atomic type, atomicity, formal charge, hybridization mode, number of hydrogen atoms, chirality, aromaticity, ring structure, and atomic mass.

[0116] For atoms and atoms Chemical bonds between them, determining the characteristics of chemical bonds. In this embodiment, the chemical bond characteristics are mainly composed of bond type, conjugation, whether it is located in a ring, stereochemistry, bond orientation, and aromaticity. .

[0117] Will As input to the graph neural network, that is:

[0118]

[0119] Subsequently, a two-layer GINE network is used to update the representation of the atomic nodes. For the first... Layered GINE networks first combine atomic nodes With neighboring atomic nodes The chemical bond features between them are mapped to the same feature space as the atom hidden representation, and the mapped chemical bond features are compared with those of neighboring atom nodes. The upper layer represents addition, and after nonlinear activation, neighboring atomic nodes are obtained. To the central atomic node The message being conveyed:

[0120]

[0121] in, Indicates the first Neighboring atomic nodes in the layer To the central atomic node The message being delivered; Representing neighboring atomic nodes In the Layer feature representation; Represents atomic nodes With neighboring atomic nodes Characteristics of chemical bonds between them; Indicates the first The learnable weight matrix of the layer.

[0122] For atomic nodes The messages generated by all directly connected neighboring atomic nodes are summed, and the sum is then compared with the messages generated by the atomic nodes. By combining the feature representations from the previous layer, atomic nodes are obtained. Neighborhood aggregation representation:

[0123]

[0124] in, Indicates atomic nodes A set of neighboring atomic nodes directly connected by chemical bonds; Indicates the first The feedforward mapping network in the layered GINE network consists of linear transformation, layer normalization, nonlinear activation, and random deactivation layers.

[0125] The neighborhood aggregation representation is added to the atomic node via residual connections. In the feature representation of the previous layer, atomic nodes are obtained. In the Layer update representation:

[0126]

[0127] Through the above processing, each atomic node not only retains the feature information of its previous layer, but also aggregates the features of its neighboring atoms and the chemical bond features connecting them. Therefore, even if neighboring atoms are of the same type, different neighborhood messages can be generated when they are connected to the central atom through different types of chemical bonds.

[0128] A two-layer GINE network is used to sequentially update the atomic node representations, resulting in the second layer of atomic representations. Finally, the atomic node representations are pooled and passed through a feature mapping layer to obtain the molecular structure representation of the candidate drug. .

[0129] Step c2: Construct protein sequence representations of candidate targets.

[0130] Specifically, for candidate targets The amino acid sequence is obtained, and the context embedding representation of each amino acid position in the sequence is obtained using a pre-trained protein language model (ESM). Let the target sequence embedding be... ,in Indicates the first An embedding vector for each amino acid position.

[0131] To reduce the impact of padding and sequence end positions on target representation, masked average pooling is performed on effective amino acid positions to obtain the global sequence representation of the pooled protein sequence corresponding to the target:

[0132]

[0133] in, This represents the set of valid amino acid positions in the candidate target sequence.

[0134] By using masked average pooling, only the context embedding vectors corresponding to the actual amino acid positions are aggregated, avoiding interference from sequence start markers, sequence end markers, and padding positions on the target sequence representation. Simultaneously, average pooling can convert amino acid sequences of different lengths into global sequence representations with the same dimension, facilitating subsequent joint modeling with the molecular structure representation of candidate drugs.

[0135] The global sequence representation is sequentially input into a linear mapping layer, a layer normalization layer, a nonlinear activation layer, and a random deactivation layer to obtain the sequence encoding representation of the candidate target:

[0136]

[0137] in, This represents the learnable weight matrix during the target sequence feature mapping process; The representation layer normalization operation is used to adjust the numerical distribution of different feature dimensions; Represents a non-linear activation function; This indicates a random deactivation operation, used to reduce overfitting during model training; Indicates candidate targets The protein sequence is represented by the protein sequence encoding.

[0138] The linear mapping layer transforms the sequence representation output by the ESM model into the same hidden feature space as the candidate drug molecule structure representation; layer normalization and nonlinear activation are used to improve the feature transformation capability; and the random deactivation layer is used to enhance the training stability and generalization ability of the model. Thus, a candidate target sequence representation that can interact and fuse with the candidate drug molecule structure representation is obtained.

[0139] Step c3: Construct entity interaction representations of candidate drug-target pairs.

[0140] Specifically, the molecular structure of the candidate drug is represented. Representation of candidate target sequences Perform element-by-element interactions to obtain candidate entity interaction features:

[0141]

[0142] in, Represents element-wise product. The two terms represent element-wise absolute differences, and are used to describe the similarity and difference relationships between drug and target representations, respectively.

[0143] Subsequently, a feedforward mapping layer consisting of linear transformation, layer normalization, and the nonlinear activation function Dropout is used to map the interaction features into entity interaction representations of drug-target pairs:

[0144]

[0145] in, Indicates learnable weights, Representation layer normalization.

[0146] The entity representation adaptive correction module based on subgraph relationship evidence and the drug-target interaction classification and prediction module specifically perform the following steps:

[0147] Step d1: Generate relation correction information based on entity representation and subgraph relation representation.

[0148] Specifically, the candidate drug-target pair obtained in step c3 is represented as an entity. Relationship with the phenotypic guided isomer subgraph obtained in step b3 By combining the features, we obtain the fusion features:

[0149]

[0150] This fusion feature simultaneously includes entity attribute information, subgraph relationship information, and information on the interaction differences between the two.

[0151] The fused features are input into a feedforward neural network to generate a subgraph relationship correction vector:

[0152]

[0153] in, This indicates the entity representation correction information provided by the phenotypic-guided heterogeneous subgraph relation evidence.

[0154] Step d2: Adaptively correct the entity representation based on the gating residual mechanism. Specifically, based on the candidate drug-target pair, the entity representation is... Subgraph Relation Representation Computational dimensionality correction gating:

[0155]

[0156] in, express and splicing Indicates a gating network. This represents the Sigmoid activation function. Used to control the correction strength of subgraph relationship evidence on different feature dimensions.

[0157] Furthermore, the overall residual correction factor is calculated using learnable parameters. Used to control the overall magnitude of the correction applied to entity representations by subgraph relation evidence.

[0158] Based on the aforementioned dimensionality-level corrected gating and overall residual correction coefficients, the corrected candidate drug-target pair representation is obtained:

[0159]

[0160] Thus, the model can dynamically adjust the contribution of subgraph relation evidence to entity representation based on the local relation structure of different candidate drug-target pairs, avoiding insufficient information fusion caused by using fixed splicing or fixed weighting methods.

[0161] Step d3: Perform drug-target interaction classification and prediction. Specifically, represent the corrected candidate drug-target pairs. Inputting the data into a classification and prediction network yields candidate drugs. With candidate targets Classification output:

[0162]

[0163] in, This represents a classification prediction network. This is the output vector for binary classification.

[0164] Subsequently, candidate drugs were obtained through Softmax normalization. With candidate targets Predicted probabilities of interaction between the parties:

[0165]

[0166] When the predicted probability is greater than a preset threshold, it is determined that there is a potential interaction between the candidate drug and the candidate target; otherwise, it is determined that there is no interaction between the candidate drug and the candidate target.

[0167] The model training module specifically performs the following steps:

[0168] Step e1: Construct training samples and labels. Specifically, construct a positive sample set based on the known drug-target interaction matrix, and construct a negative sample set according to a preset negative sampling ratio; for candidate drug-target pairs... If there is a known interaction between the two, then the label Otherwise, the label The samples were then divided into training and validation sets, and batches were constructed during training according to the ratio of positive to negative samples to reduce the impact of class imbalance on model training.

[0169] Step e2: Construct auxiliary supervision branches. Specifically, the model includes a final prediction branch, an entity auxiliary branch, and a subgraph auxiliary branch. The final prediction branch takes the candidate drug-target pair representation, which has been adaptively corrected by the subgraph relationship evidence, as input and outputs the final interaction prediction result between the candidate drug and the candidate target. This final prediction result serves as the drug-target interaction prediction result used by the model during the inference phase.

[0170] To enable the candidate drug-target entity pair representation and the phenotype-guided heterogeneous subgraph relation representation before adaptive correction to learn discriminative information related to drug-target interaction labels, an entity-aided classifier and a subgraph-aided classifier are further set up during the training phase.

[0171] The entity-aided classifier represents the candidate drug-target entity pairs obtained in step c3 before adaptive correction. As input, entity-assisted prediction results are obtained through independent linear mapping. The subgraph-assisted classifier uses the phenotypic guided heterogeneous subgraph relationships obtained in step b3. As input, the subgraph-assisted prediction result is obtained through another independent linear mapping. .

[0172] Neither the entity-aided classifier nor the subgraph-aided classifier performs adaptive correction or fusion between subgraph relation evidence and entity pair representations, nor do they input their auxiliary prediction results into the final prediction branch. They are only used to apply auxiliary supervision to the entity representation branch and the subgraph relation representation branch, respectively, during model training; during model inference, the prediction result output by the final prediction branch is used as the drug-target interaction prediction result.

[0173] Step e3: Calculate the training loss and update the model parameters. Specifically, calculate the cross-entropy loss for the final prediction branch, the entity-assisted branch, and the subgraph-assisted branch respectively:

[0174]

[0175] And obtain the total training loss:

[0176]

[0177] in, and These represent the weight parameters of the entity-assisted loss and the subgraph-assisted loss, respectively. In a preferred embodiment, the entity-assisted loss weight is set to 0, and the subgraph-assisted loss weight is set to 0.2, so that the training process mainly optimizes the final corrected candidate pair representation, while retaining the auxiliary constraints on the subgraph relation representation.

[0178] Step e4: Model optimization and validation. Specifically, the AdamW optimizer is used to jointly train the drug molecule graph encoder, target sequence encoder, heterogeneous subgraph encoder, gated residual correction module, and classification prediction network. In the normal prediction scenario, the training process allows the reconstruction of local subgraph edges using drug-target interactions within the current training iteration, but excludes drug-target edges from the current candidate. In both drug and target cold-start prediction scenarios, training drug-target interaction edges are not introduced during subgraph generation and edge reconstruction, thus ensuring that the relational evidence in the cold-start scenario does not depend on the known interaction labels of the entities to be predicted.

[0179] This invention can reduce the meta-path bias caused by high-frequency phenotypic nodes, and improve the accuracy and robustness of predictions in both normal and cold-start scenarios.

[0180] To test and evaluate the performance of this application, it was evaluated on a drug-target interaction prediction dataset.

[0181] In one implementation, experimental validation is performed using publicly available drug-target interaction datasets, the luo and li datasets, which contain drugs, targets, and known interactions. Experimental scenarios include a standard prediction scenario, a drug cold-start prediction scenario, and a target cold-start prediction scenario. The standard prediction scenario evaluates the model's predictive ability for common candidate drug-target combinations; the drug cold-start prediction scenario evaluates the model's generalization ability for unseen drugs; and the target cold-start prediction scenario evaluates the model's generalization ability for unseen targets.

[0182] During the experiment, known drug-target interaction samples were used as positive samples, while drug-target combinations with no observed interactions were sampled as negative samples according to a preset ratio. A five-fold cross-validation method was used for model training and validation. For each fold, positive samples from the training fold were used to construct training relationship information, and candidate drug-target edges were excluded when constructing candidate pair graphs or reconstructing training DTI edges to avoid leakage of label information.

[0183] To verify and evaluate the performance of this application, accuracy, precision, recall, F1 score, area under the ROC curve (AUC), and area under the precision-recall curve (AUPR) are used as evaluation metrics. The specific calculation formulas are as follows:

[0184]

[0185]

[0186]

[0187]

[0188] in, This indicates samples where drug-target interaction actually exists and is correctly predicted by the model to have an interaction. This indicates samples where no drug-target interaction actually exists and which are correctly predicted by the model to have no interaction. This refers to samples where there is no actual drug-target interaction, but the model incorrectly predicted that there is an interaction. This indicates samples where drug-target interaction actually exists but was incorrectly predicted by the model as not having an interaction.

[0189] To verify the effectiveness of the method of the present invention under different datasets and prediction scenarios, the PGSC-DTI method of the present invention was compared with existing drug-target interaction prediction methods CCL-DTI, DCGAT-DTI, HGCML-DTI, MIDTI, and SHGCL-DTI. The experimental setup included three scenarios with a positive-to-negative sample ratio of 1:1, 1:3, and 1:5 under random partitioning, as well as two generalization prediction scenarios: drug cold start and target cold start. Detailed experimental results are shown in Tables 1 to 6. Figures 2 to 5 .

[0190] Table 1. Performance comparison results of the li dataset in the random partitioning scenario.

[0191]

[0192] Table 2 Performance comparison results of drug cold start scenario on the li dataset

[0193]

[0194] Table 3 Performance comparison results of target cold start scenario on the li dataset

[0195]

[0196] Table 4. Performance comparison results of the luo dataset in the random partitioning scenario.

[0197]

[0198] Table 5 Performance comparison results of drug cold start scenario on the luo dataset.

[0199]

[0200] Table 6. Performance comparison results of target cold start scenario on the luo dataset.

[0201]

[0202] The experimental results show that the method of this invention achieves superior prediction performance in multiple experimental scenarios on both the LI and Luo datasets. Compared with the best-performing existing comparative methods in each scenario, the method of this invention improves on F1, AUC, and AUPR metrics, indicating that this invention can more accurately identify potential drug-target interactions and maintain good generalization ability in scenarios involving the prediction of new drugs or new targets.

[0203] To further verify the contribution of each component module to the overall prediction performance, ablation experiments were conducted. The ablation settings included: removing heterogeneous subgraph relationship branches, retaining only the drug-target entity representation; removing drug and target entity representation branches, retaining only the subgraph relationship representation; replacing the pre-trained protein sequence representation with ordinary sequence encoding; replacing the drug molecular graph encoding with molecular fingerprint representation; and simplifying the heterogeneous subgraph encoding module into an attention aggregation module based on metapath instances. Detailed experimental results are shown in Tables 7 and 8. Figure 6 .

[0204] Table 7 Ablation Experiment Results under the li Dataset

[0205]

[0206] Table 8 Ablation Experiment Results on the luo Dataset

[0207]

[0208] Specific ablation experiments were conducted in drug cold-start and target cold-start scenarios. The results showed that the complete model was more stable on most core evaluation metrics, especially AUC and AUPR. When only subgraph representation branches were retained, the model's performance significantly decreased in the cold-start scenario, indicating that relying solely on heterogeneous network path evidence is insufficient to fully characterize the entity attributes of new drugs or targets. When only entity representation branches were retained, the model lacked candidate pair-specific subgraph relationship supplementation, leading to an overall decrease in metrics such as AUC and AUPR, indicating that phenotypic guided subgraphs can provide effective relationship evidence for cold-start prediction. Replacing the pre-trained ESM representation or molecular graph encoding also resulted in varying degrees of performance degradation, demonstrating the significant contribution of drug molecule structure representation and target pre-trained semantic representation to the prediction results. Furthermore, even when meta-path instance attention aggregation was used instead of complete subgraph encoding, its overall performance was still lower than the complete model, indicating the effectiveness of the candidate pair heterogeneous subgraph encoding and adaptive correction mechanism of this invention. The above results demonstrate that the phenotype-guided candidate pair heterogeneous subgraph, drug-target entity representation, and adaptive correction mechanism based on subgraph relationship evidence all contribute positively to the predictive performance of this invention.

[0209] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A drug-target interaction prediction method based on phenotypic guide graph correction, characterized in that, include: For candidate drug-target pairs, local meta-path instances are extracted to obtain a set of meta-path instances. Phenotypic nodes in the meta-path instances are specifically screened to obtain a set of retained disease nodes and side effect nodes. This leads to the construction of candidate pair-specific isomer subgraphs. For each node in the candidate pair-specific heterogeneous subgraph Construct node input features Then, a heterogeneous graph attention network is used to process the heterogeneous subgraphs. The middle node represents the update; utilizing candidate drug nodes and candidate target nodes The final representation constructs the query vector. Then based on and nodes In heterogeneous subgraphs The final representation in the graph is calculated for each subgraph node. Attention score relative to candidate drug-target pair ,Will Normalization is performed on all nodes of the subgraph to obtain the normalized attention score. Ultimately, based on node In heterogeneous subgraphs The final representation in the graph is then attention-weighted aggregated to obtain the phenotypic guided isomer subgraph representation of the drug-target pair. ; For candidate drugs Drug molecular maps were constructed based on their molecular structures. For drug molecule graphs, a drug molecule graph encoder is used for encoding, and each atom node in the drug molecule graph is encoded. Construct atomic nodes initial atomic features ; Based on atoms and atoms Determining the chemical bond characteristics between them ;Will As input to the graph neural network, the atomic node representations are updated using the graph neural network. During the update process, chemical bond features are incorporated, ultimately yielding the final atomic representation layer of the graph neural network, which in turn provides the molecular structure representation of the candidate drug. ; Constructing sequence encoding representations of candidate targets Representing the molecular structure of candidate drugs Representation of candidate target sequences Perform element-by-element interactions to obtain candidate entity interaction features. Candidate entity interaction features Mapped as entity interaction representation of drug-target pairs ;based on and Obtain fusion features The fused features are input into a feedforward neural network to generate a subgraph relationship correction vector. ;according to and Get corrected gating ;based on , combined and The corrected candidate drug-target pair representation is obtained. ;Will Inputting the data into a classification and prediction network yields candidate drugs. With candidate targets The classification results.

2. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, The process of specifically filtering phenotypic nodes in metapath instances includes: From the metapath instance collection Extract disease nodes and side effect nodes to form candidate disease node sets and candidate side effect node sets; calculate phenotypic nodes. structural specificity ,in Represents a phenotypic node The degree of global connectivity in the original interconnected network. Represents nodes A set of nodes of the same type, where u is a node in the set; Statistical phenotype nodes The semantic diversity is obtained by counting the number of different metapath types covered under the current candidate drug-target pair. , Represents a path instance The associated metapath template, Represents a path instance The corresponding phenotypic anchor points, For indicator functions; This refers to the set of meta-path instances obtained based on local path instances in a multi-source biomedical heterogeneous network. Calculate the overall score ,in The weight parameter is used to determine the semantic diversity of the metapath. Candidate disease nodes and candidate side effect nodes are sorted from highest to lowest according to their comprehensive scores and then filtered to obtain the retained sets of disease nodes and side effect nodes. .

3. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 2, characterized in that, The process of constructing a candidate pair-specific isomorphic subgraph includes: reserve The metapath instance of a node is a local node set composed of candidate drug nodes, candidate target nodes, and all entity nodes in the retained path. , This represents the set of path instances retained after filtering by phenotypic nodes; Deduplication and merging are performed based on the global entity identifier, ensuring that when the same entity appears through multiple meta-path instances, only one node is retained in the local node set, thereby based on the local node set. Local relation edges are reconstructed from multi-source biomedical heterogeneous networks to obtain candidate pair heterogeneous subgraphs. ,in Represents the set of local relation edges.

4. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, For each node in the candidate pair-specific isomorphic subgraph , construct node input features ,in Encoding for node type; This refers to the corresponding query anchor point of a node relative to the current candidate drug-target pair, indicating the candidate drug node and candidate target node in the current candidate pair to be predicted. For nodes Relative to the structural position characteristics of the current candidate drug-target pair, by nodes The shortest path distance to the current candidate drug node and the current candidate target node is determined in the candidate pair specific isomer subgraph; The feature indicates whether a node participates in a specific meta-path relationship. Its dimension corresponds to the number of meta-path types, and each dimension corresponds to one meta-path type. If the node participates in the corresponding meta-path, the value is 1; otherwise, it is 0. This represents the number of adjacent nodes that a node connects to under different relationships.

5. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, Using heterogeneous graph attention networks to study heterogeneous subgraphs The process of updating a node includes: Will The initial features of the nodes are mapped to the latent space, and the node features are updated using a multi-layer heterogeneous graph attention network; for the ... Layered heterogeneous graph attention network, based on nodes ,node In the Layer feature representation of computation nodes and neighboring nodes Attention score for relation types and normalize them within the same relation type. , Indicates the relationship With nodes A set of connected neighbors; Relationship Weighted aggregation of neighbor information , For nodes In the The layer's feature representation is then used; the information obtained from different relationships is then fused through summation, and combined with residual links to obtain nodes. The Layer update node representation ,go through The final representation of the node is obtained after the layer is updated. .

6. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, Will As input to a graph neural network, the process of updating the representation of atomic nodes using a graph neural network includes: For candidate drugs Drug molecular maps were constructed based on their molecular structures. For drug molecule graphs, a drug molecule graph encoder is used for encoding, and each atom node in the drug molecule graph is encoded. Construct atomic nodes initial atomic features Based on atoms and atoms Determining the chemical bond characteristics between them ;Will As input to the graph neural network, the representation of the atomic nodes is updated using the graph neural network. Layered graph neural networks first combine atomic nodes With neighboring atomic nodes The chemical bond features between them are mapped to the same feature space as the atom hidden representation, and the mapped chemical bond features are compared with those of neighboring atom nodes. The upper layer represents addition, and after nonlinear activation, neighboring atomic nodes are obtained. To the central atomic node The message being delivered and for the central atomic node The messages generated by all directly connected neighboring atomic nodes are summed and then compared with the central atomic node. By combining the feature representations from the previous layer, we obtain the neighborhood aggregation representation of the central atomic node. Then, atomic nodes are added through residual connections. In the feature representation of the previous layer, atomic nodes are obtained. In the Layer update representation .

7. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, Constructing sequence encoding representations of candidate targets The process includes: For candidate targets The amino acid sequence of the target protein is obtained, and the target sequence embedding is obtained using a protein language model. Based on the target sequence embedding, a global sequence representation of the protein sequence is obtained, and based on the global sequence representation of the protein sequence, a sequence encoding representation of the candidate target is obtained. .

8. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, The candidate entity interaction features ,in Represents element-wise product. Represents the absolute difference element by element; This indicates feature splicing.

9. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, The fusion feature ,in Represents element-wise product. Represents the absolute difference element by element; This indicates feature splicing.

10. The drug-target interaction prediction method based on phenotypic guide graph correction according to claim 1, characterized in that, Corrected candidate drug-target pair representation ,in This is the overall residual correction factor.

Images