Drug target interaction prediction method based on large language model semantic enhancement

Abstract

This invention discloses a drug-target interaction prediction method based on semantic enhancement using a large language model, belonging to the fields of computer-aided drug discovery and deep learning. The steps are as follows: First, acquire biomedical data to construct a heterogeneous network. Then, use a domain-adapted large language model to generate semantic descriptions of drug and protein nodes and encode them as semantic features. Combine a semantic-attribute gating alignment mechanism and a semantic feature compensation strategy to construct a pluggable semantically enhanced heterogeneous graph neural network prediction model. Next, construct a joint loss function combining prediction loss and feature alignment constraints, and optimize the model training. Finally, acquire the data to be predicted, input it into the trained model, and output the predicted probability of drug-target interactions. This invention effectively alleviates the problem of missing node features or insufficient expressive power in biological networks, improves the accuracy and generalization ability of drug-target interaction prediction, and provides an efficient solution for reducing drug development costs and shortening the R&D cycle.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug discovery and deep learning technology, specifically involving a method for predicting drug target interactions based on semantic enhancement of a large language model. Background Technology

[0002] Drug-target interaction prediction is a core task in computer-aided drug development, aiming to identify the binding relationships between potential small drug molecules and protein receptors. Accurate drug-target interaction prediction is crucial for reducing drug development costs and shortening the development cycle. While traditional biological experimental methods offer high accuracy, they suffer from drawbacks such as high cost, lengthy development cycles, and poor scalability. Therefore, utilizing computational methods to efficiently predict potential drug-target relationships has become a key research focus in the fields of bioinformatics and cheminformatics.

[0003] With the development of graph representation learning techniques, an increasing number of studies are modeling drug-target interaction prediction as a link prediction problem in biomedical networks. Compared to early traditional machine learning methods that relied on manual feature engineering, graph neural networks can automatically learn low-dimensional representations of entities from topological structures, effectively capturing potential relationships in complex systems. In real biomedical systems, networks composed of entities such as drugs, proteins, diseases, and side effects exhibit high heterogeneity. To address this, heterogeneous graph neural networks, by uniformly modeling multiple types of nodes and various relationships and utilizing mechanisms such as meta-paths to aggregate multi-dimensional semantic information, demonstrate stronger expressive power than homogeneous graph methods in drug-target interaction prediction tasks.

[0004] However, despite significant progress made by heterogeneous graph neural networks in predicting drug-target interactions, they still face the following serious technical challenges when processing real-world biomedical big data: First, the lack of original node features and the "cold start" problem are prevalent. Existing heterogeneous graph neural network methods heavily rely on the initial attribute descriptions of nodes (such as the molecular fingerprint of a drug or the amino acid sequence characteristics of a protein) for information propagation. However, in practical applications, some biomedical datasets only provide network topology connection information; at the same time, for newly developed drugs or newly discovered targets (i.e., "cold start" nodes), there is often a lack of both complete original attribute descriptions and sufficient topological connections. This makes it difficult for the model to perform effective feature initialization, resulting in a precipitous drop in prediction performance.

[0005] Secondly, the expressive power of node features is insufficient. Existing attribute features are mainly based on the physical or chemical structure of entities, which is insufficient to fully represent complex biomedical semantic information. This limits the model's representation ability when processing entity pairs with deep functional relationships. Although large language models have shown great potential in biomedical text modeling in recent years, capable of extracting knowledge from massive amounts of literature, current research on drug-target interaction prediction still lacks an effective way to deeply integrate the semantic understanding capabilities of large models with the structural modeling capabilities of graph neural networks.

[0006] Finally, there is a lack of a general feature enhancement mechanism. Existing methods typically design feature processing strategies for specific graph models or data scenarios, often deeply coupled with the model architecture. This makes them difficult to adapt as general components to different types of heterogeneous graph neural network models, limiting the scalability and practical application value of these methods. In summary, how to construct a feature enhancement method that can deeply integrate biomedical semantic knowledge with complex heterogeneous graph topologies and possesses good generality, under conditions of incomplete or insufficient node features, has become a key scientific challenge for improving the prediction performance of drug-target interactions.

[0007] Therefore, this invention proposes a drug-target interaction prediction method based on semantic enhancement of a large language model, employing the semantically enhanced heterogeneous graph neural network prediction model LASE-HGNN, which boasts good pluggability. This model addresses existing problems by adapting the open-source large language model to a specific domain using LoRA, thereby generating semantic descriptions for drug and protein nodes, and then encoding them using Bio-ClinicalBERT to obtain a unified semantic representation. Furthermore, a unified multi-source feature modeling mechanism is designed for different data scenarios: for data with original attributes, a semantic-attribute gating alignment strategy is innovatively introduced to constrain the distribution of two types of features in a unified feature space to eliminate the modality gap and achieve consistent fusion; for data lacking original features, a semantic feature compensation strategy is proposed, utilizing the semantic features generated by LLM directly as supplementary information for collaborative learning with the topological structure, fundamentally alleviating the performance degradation caused by the dual lack of semantics and attributes. This invention possesses strong versatility and can be seamlessly integrated as an independent module into various heterogeneous graph neural networks, providing a general solution for drug-target interaction prediction tasks under low-resource and complex structures. Summary of the Invention

[0008] To address the issues of missing node features, insufficient semantic expression, and poor versatility in existing drug-target interaction prediction methods, this invention proposes a drug-target interaction prediction method based on semantic enhancement of a large language model. It employs the semantically enhanced heterogeneous graph neural network prediction model LASE-HGNN, which boasts good pluggability, and enhances semantic knowledge extraction and multi-scenario feature fusion strategies. This optimizes the node representation learning process of the heterogeneous graph neural network, improving the accuracy and robustness of the drug-target interaction prediction task and enhancing the model's generalization ability in feature-incomplete environments. Ultimately, this provides a general and pluggable technical solution for biomedical research and low-resource drug discovery tasks.

[0009] A drug target interaction prediction method based on semantic enhancement of a large language model includes the following steps: Step 1: Acquire biomedical data to construct a heterogeneous network. Use a domain-adapted large language model to generate structured semantic descriptions of drug and protein nodes and encode them as semantic features. Combine semantic-attribute gating alignment mechanism and semantic feature compensation strategy to construct a pluggable semantically enhanced heterogeneous graph neural network prediction model. Step 2: Construct a joint loss function by combining the prediction loss and feature alignment constraints, and optimize the training of the heterogeneous graph neural network prediction model; Step 3: Obtain the data of the drug and target nodes to be predicted, input them into the trained heterogeneous graph neural network prediction model, and output the predicted probability of the interaction between the drug and the target.

[0010] Furthermore, in step 1, the semantically enhanced heterogeneous graph neural network prediction model operates as follows: Step 1.1: Construct a heterogeneous biomedical network containing drug nodes, protein nodes and their various relationships, and use node type mapping functions and edge type mapping functions to uniformly model different types of entities and relationships; Step 1.2: Construct biomedical instruction data, use an efficient parameter fine-tuning method to adapt the open-source large language model to the domain, and generate semantic descriptions of drug and protein nodes; use Bio-ClinicalBERT to encode the semantic descriptions to obtain a unified semantic feature representation. Step 1.3: Design two types of feature enhancement strategies for different data conditions: For biomedical data with original node attributes, introduce a semantic-attribute gating alignment mechanism; for biomedical data lacking original node attributes, adopt a semantic feature compensation strategy. Step 1.4: Through a heterogeneous graph neural network model, learn the final embedding representation of protein and drug nodes that integrates semantic and structural information, which is used for prediction tasks; Step 1.5: Concatenate the final embedded representations of drug nodes and protein nodes to construct a joint representation of node pairs, and output the predicted probability of interaction between drug and target through a multilayer perceptron.

[0011] Furthermore, the specific process of step 1.2 is as follows: Step 1.2.1: Use the LoRA method to fine-tune the open-source large language model for domain adaptation. Let a certain linear transformation weight in the pre-trained large language model be... Incrementally update it Represented in low-rank matrix decomposition form: (1); Updated weights Represented as: (2); in, and For training, a low-rank matrix, Indicates the output feature dimension. Indicates the input feature dimension. Denotes the rank parameter of a low-rank decomposition, satisfying ; Step 1.2.2: After completing domain adaptation, utilize the fine-tuned pre-trained large language model. Generate semantic description text for drug nodes and protein nodes in heterogeneous biomedical networks: (3); (4); in, Represents the set of drug nodes. Represents a set of protein nodes. Represents the first element in the set. One drug node, Represents the first element in the set. One protein node; For drug nodes The corresponding semantic description text, For protein nodes The corresponding semantic description text; Step 1.2.3: After obtaining the semantic description text of the nodes, the semantic description text is semantically encoded using the biomedical pre-trained language model Bio-ClinicalBERT to obtain a unified semantic vector representation: (5); (6); in, This represents the encoding function of the pre-trained language model Bio-ClinicalBERT. Indicates drug node semantic feature vectors, Represents protein nodes The semantic feature vectors are then used to further construct the overall semantic feature matrix of drug and protein nodes. (7); (8); in, The semantic feature matrix for all drug nodes. This is the semantic feature matrix of all protein nodes. This represents the total number of drug nodes. This represents the total number of protein nodes. This represents the matrix transpose operation.

[0012] Furthermore, the specific process of step 1.3 is as follows: Step 1.3.1: For biomedical data containing original node attributes, feature enhancement is performed using the aforementioned semantic-attribute gating alignment mechanism. The original attribute feature matrices and semantic feature matrices of drugs and proteins are projected onto a unified feature space through a nonlinear mapping function, wherein protein nodes... The mapping result is represented as: (9); (10); in, Represents protein nodes The original attribute feature vector, This represents the original attribute features after mapping. This represents the semantic features after mapping. and The mapping function, composed of linear transformation, batch normalization, and nonlinear activation function, can be expressed in the following form: (11); in, This is the first layer of learnable parameter matrix. This is the learnable parameter matrix for the second layer. Indicates input features, This indicates a batch normalization operation. RelU is a non-linear activation function. Step 1.3.2: Before feature fusion, perform a normalization process on the two types of features: (12); (13); in, Representing vectors Norm, and These represent the normalized attribute features and semantic features, respectively; Step 1.3.3: Adaptively fuse the two types of features after normalization using a gating mechanism: (14); (15); in, This represents the vector concatenation operation. For the learnable parameter matrix in the gating mechanism, The learned gate vector takes values ​​between 0 and 1. Represents the element-wise multiplication operation of vectors; This is the initial representation of the fused protein nodes; Step 1.3.4: For biomedical data lacking original node attributes, a semantic feature compensation strategy is used for feature enhancement. The feature matrix obtained from semantic encoding is directly used as the initial feature of the nodes. The initial feature matrices for drug nodes and protein nodes are expressed as follows: (16); (17); in, This represents the initial feature matrix of the drug node. This represents the initial feature matrix of the protein node.

[0013] Furthermore, the specific process of step 1.4 is as follows: Step 1.4.1: In a heterogeneous graph attention network, compute the meta-path using a node-level attention mechanism. Next protein node Its neighboring nodes Attention score between : (18); in, Indicates the first Path of a single element and These represent protein nodes. and its neighboring nodes The input features are determined according to the feature enhancement strategy in step 1.3: for data with original attributes, the input features are the initial representations of the fused nodes; for data lacking original attributes, the input features are the initial feature matrices of the nodes. Let be the linear transformation matrix of the node features. For attention weight vectors, LeakyReLU is the transpose of this vector and is the rectified linear unit activation function with leakage. Step 1.4.2: Perform softmax normalization on the attention scores to obtain normalized attention weights. : (19); in, Represented by natural constant An exponential function with base 0. Indicates the metapath Next node The set of all neighboring nodes, The traversal index in the set; the node is obtained using the following formula. In metapath Embedded representation : (20); in, A learnable weight matrix for feature aggregation; Step 1.4.3: Introduce a semantic-level attention mechanism to fuse information from different meta-paths and calculate the first... Normalized importance coefficient of element path : (twenty one); in, Metapath The corresponding learnable weights, This represents the total number of metapaths in a heterogeneous network. For the first Learnable weights of element paths; Step 1.4.4: Integrate information from different metapaths to obtain protein nodes. The final embedding representation : (twenty two).

[0014] Further, step 1.5 specifically includes: For any drug node and protein nodes Finally embed it into the representation and Perform splicing to construct a joint representation of node pairs. : (twenty three); A multilayer perceptron is used to perform nonlinear modeling of the joint representation, and the predicted probability is output. : (twenty four); in, This represents a multilayer perceptron composed of multiple linear transformations and nonlinear activation functions. This represents the Sigmoid activation function. Indicates drug With protein There are predicted probability values ​​for interactions between them, ranging from 0 to 1.

[0015] Furthermore, in step 2, the specific process of constructing a joint loss function by combining the prediction loss and feature alignment constraints and optimizing the training is as follows: Step 2.1: Construct a training sample set using a negative sampling strategy. Calculate the weighted binary cross-entropy loss function used for optimization. : (25); in, This represents the total number of training samples. This represents a pair of positive and negative samples in the set. The weighting coefficient is used to adjust the proportion of positive samples; Step 2.2: Construct a joint optimization objective. For datasets containing original attribute features, it is necessary to calculate the semantic-attribute gating alignment loss. : (26); in, For protein nodes The corresponding gating weight vector; and By combining these, we obtain the joint optimization objective function. : (27); in, To control the alignment loss weighting coefficients of feature alignment strength, This represents the set of all learnable parameters in the model. The square of the parameter set Norm, To control The regularization coefficient for regularization strength; Step 2.3: For datasets lacking original attribute features, since feature alignment is not required, the following joint loss function, which includes prediction loss and regularization term, is directly used for optimization: (28).

[0016] Furthermore, the specific process of step 3 is as follows: Step 3.1: Determine the heterogeneous biomedical network and initial feature matrix of nodes; process the collected biomedical data to be predicted into a heterogeneous network form, where the nodes of the heterogeneous network represent different biomedical entities, including core drug nodes, protein nodes, and auxiliary disease, side effect or gene nodes; the edges of the heterogeneous network represent known interaction relationships between entities, including drug-protein relationship, drug-disease relationship, protein-disease relationship and drug-side effect relationship; use the big language model adapted to the biological domain through formula (1) and formula (2) to generate semantic description text of drug and protein nodes, and use the pre-trained model Bio-ClinicalBERT to generate semantic description text of drug and protein nodes through formula (3) to formula (8) and encode it into a semantic feature matrix; Step 3.2: Based on the feature conditions of the dataset to be predicted, perform forward fusion calculation of node representation using different feature enhancement strategies; if the dataset to be predicted contains original attribute features, then project the original attribute features and semantic features to a unified feature space according to formulas (9) to (11), normalize the two types of features according to formulas (12) and (13), and then use the gating parameter matrix obtained from training to perform semantic-attribute gating adaptive fusion according to formulas (14) and (15) to obtain the initial representation of the fused node; if the dataset to be predicted lacks original attribute features, then adopt the semantic feature compensation strategy according to formulas (16) and (17) and directly use the generated semantic feature matrix as the initial representation of drug and protein nodes; Step 3.3: Input the initial node representation obtained above into the heterogeneous graph attention network for representation learning; calculate the attention score between the node and its neighboring nodes under a specific meta-path according to formula (18), perform softmax normalization according to formula (19), and aggregate according to formula (20) to obtain the node embedding representation under the specific meta-path; then, calculate the semantic level attention weight of different meta-paths according to formula (21), and integrate the information of multiple meta-paths according to formula (22) to obtain the final embedding representation of the drug node and target protein node to be predicted; Step 3.4: Calculate the drug-target interaction probability and generate a prediction list. According to formula (23), the final embedding representations of the drug node to be predicted and the target protein node are concatenated to construct a joint representation of the node pair. The joint representation is input into the trained multilayer perceptron, and the prediction probability of the interaction between the drug and the target is calculated according to formula (24). Finally, the prediction probabilities between the drug to be predicted and all candidate protein targets are compared, and the prediction probability values ​​are sorted in descending order. The probability values ​​greater than the preset threshold or ranked first are selected. The target proteins are identified, and a list of predicted drug-protein target interactions is generated.

[0017] The beneficial technical effects of this invention are as follows: This invention addresses the problems of missing node features and difficulty in predicting new nodes. By directly using semantic features generated by a large language model as the initial features of nodes, the model maintains stable prediction performance and stronger adaptability even when node attributes are empty or network edges are few. Secondly, it improves the fusion effect of multi-source features. The semantic-attribute gating alignment mechanism designed in this invention can bring text semantics and graph structure features closer together, reducing errors and interference caused by directly splicing different types of data, and can automatically adjust the fusion ratio of the two features, making the expression of node features more accurate. Thirdly, this method has good versatility and flexibility. This feature enhancement framework does not depend on a specific graph network underlying structure and can be directly integrated into various existing heterogeneous graph neural networks like a plug-in, without the need to redevelop algorithms for different models. Finally, it significantly improves the accuracy of drug-target interaction prediction. Tests on multiple benchmark datasets show that this invention can comprehensively improve various prediction indicators, providing a more accurate and practical computational tool for new drug discovery and drug repurposing. Attached Figure Description

[0018] Figure 1 This is a flowchart of the drug target interaction prediction method based on semantic enhancement of a large language model according to the present invention.

[0019] Figure 2 This is a comparison chart of the prediction performance of different feature enhancement module combinations in the experiments of this invention; Among them, (a) is a comparison chart of prediction performance based on HAN; (b) is a comparison chart of prediction performance based on HPN; (c) is a comparison chart of prediction performance based on ie-HGCN; and (d) is a comparison chart of prediction performance based on DMHGNN.

[0020] Figure 3 This is a graph showing the performance improvement of the semantic feature compensation strategy on an attribute-free dataset in the experiments of this invention; Among them, (a) is the performance improvement diagram of the semantic feature compensation strategy based on HAN; (b) is the performance improvement diagram of the semantic feature compensation strategy based on HPN; (c) is the performance improvement diagram of the semantic feature compensation strategy based on ie-HGCN; and (d) is the performance improvement diagram of the semantic feature compensation strategy based on DMHGNN. Detailed Implementation

[0021] The specific embodiments of the present invention will be further described below with reference to specific examples: This embodiment provides a drug-target interaction prediction method based on semantic enhancement of a large language model, employing a semantically enhanced heterogeneous graph neural network prediction model (LASE-HGNN) with good pluggability. Heterogeneous graph neural networks, which mostly rely on the original attributes and topology of nodes, often face problems such as missing node features or insufficient feature representation capabilities when processing biomedical data, leading to significant performance degradation in real biological networks. Therefore, this invention combines the semantic modeling capabilities of a large language model with the structural mining capabilities of heterogeneous graph neural networks, proposing a general pluggable feature enhancement model. This method can effectively handle biomedical datasets containing drugs, proteins, and their complex relationships, aiming to solve the problems of incomplete features, lack of semantic representation, and poor model versatility in existing drug-target interaction prediction systems. By introducing a domain-adaptive large model semantic extraction mechanism and multi-scenario feature enhancement strategies, this invention significantly improves the accuracy and robustness of drug-target interaction prediction.

[0022] First, this invention leverages the powerful learning capabilities of large language models to address the issue of insufficient node feature representation through domain adaptation. Specifically, this embodiment employs the LoRa (Locally-Oriented Parameter Refinement) method to adapt the open-source large language model (Qwen-8B) to the biomedical domain. By constructing structured instruction templates, the fine-tuned model is guided to generate specialized semantic descriptions of drug and protein nodes. Subsequently, Bio-ClinicalBERT is used to encode the generated text, obtaining a unified and high-quality semantic vector representation. This method not only supplements the missing biomedical knowledge in the original attributes but also provides a solid semantic foundation for subsequent graph representation learning.

[0023] Secondly, this invention designs differentiated feature enhancement strategies for different data scenarios, improving the model's applicability in complex environments. For datasets with original attributes, this invention introduces a semantic-attribute gating alignment mechanism. This mechanism projects features from two different sources to a unified feature space through nonlinear mapping and introduces alignment constraints with dynamic weights. This effectively alleviates the semantic drift problem. Simultaneously, an adaptive gating mechanism is used to achieve deep fusion of multi-source features, and a stability-driven initialization strategy is adopted to ensure that the model is moderately biased towards original attributes in the early stages of training, thereby improving convergence stability. For data lacking original attributes, this invention proposes a semantic feature compensation strategy. Semantic embeddings generated by a large language model are directly used as initial features for nodes, replacing the traditional random initialization method. This effectively compensates for the lack of node attributes, giving node representations clear biomedical meaning.

[0024] Furthermore, this invention implements a highly versatile plug-in framework that can be seamlessly integrated into various heterogeneous graph neural network models. When predicting drug-target interactions, the LASE-HGNN model can be integrated with mainstream heterogeneous graph neural networks such as HAN, HPN, and ie-HGCN. Taking the HAN model as an example, the model first aggregates neighbor information at specific meta-paths through a node-level attention mechanism to capture subtle connections between nodes; then, it fuses information from different meta-paths through a semantic-level attention mechanism to generate a final node embedding that integrates semantic knowledge and heterogeneous structural information. This multi-level feature aggregation method enables the model to extract highly discriminative feature vectors from large-scale, sparse biomedical networks.

[0025] To ensure the accuracy and robustness of predictions, this invention employs a carefully designed joint loss function to guide model training. This joint loss function comprehensively considers the weighted binary cross-entropy prediction loss (…). Feature alignment constraints )as well as Regularization terms are added, along with a dynamic weight adjustment strategy. In practical drug-target interaction prediction tasks, to address the problem of imbalanced positive and negative samples, this invention employs a balanced negative sampling strategy to construct the training set and enhances the model's focus on positive samples through a class weighting mechanism, significantly improving prediction stability.

[0026] like Figure 1 As shown, a drug target interaction prediction method based on semantic enhancement of a large language model includes the following steps: Step 1: Acquire biomedical data to construct a heterogeneous network. Use a domain-adapted large language model to generate structured semantic descriptions of drug and protein nodes and encode them as semantic features. Combine semantic-attribute gating alignment mechanism and semantic feature compensation strategy to construct a pluggable semantically enhanced heterogeneous graph neural network prediction model LASE-HGNN. The working process of the semantically enhanced heterogeneous graph neural network prediction model LASE-HGNN is as follows: Step 1.1: Construct a heterogeneous biomedical network containing drug nodes, protein nodes and their various relationships, and use node type mapping functions and edge type mapping functions to uniformly model different types of entities and relationships in order to characterize complex biological interaction structures. Step 1.2: Construct biomedical instruction data, employ the LoRa (Locally-Oriented Parameter Refinement) method to adapt the open-source large language model to the domain, and generate semantic descriptions of drug and protein nodes; use Bio-ClinicalBERT to encode the semantic descriptions to obtain a unified semantic feature representation. The specific workflow is as follows: Step 1.2.1: Use the LoRA method to fine-tune the open-source large language model for domain adaptation. Let a certain linear transformation weight in the pre-trained large language model be... Incrementally update it Represented in low-rank matrix decomposition form: (1); Updated weights Represented as: (2); in, and For training, a low-rank matrix, Indicates the output feature dimension. Indicates the input feature dimension. Denotes the rank parameter of a low-rank decomposition, satisfying ; Step 1.2.2: After completing domain adaptation, utilize the fine-tuned pre-trained large language model. Generate semantic description text for drug nodes and protein nodes in heterogeneous biomedical networks: (3); (4); in, Represents the set of drug nodes. Represents a set of protein nodes. Represents the first element in the set. One drug node, Represents the first element in the set. One protein node; For drug nodes The corresponding semantic description text, For protein nodes The corresponding semantic description text; Step 1.2.3: After obtaining the semantic description text of the nodes, the semantic description text is semantically encoded using the biomedical pre-trained language model Bio-ClinicalBERT to obtain a unified semantic vector representation: (5); (6); in, This represents the encoding function of the pre-trained language model Bio-ClinicalBERT. Indicates drug node semantic feature vectors, Represents protein nodes The semantic feature vectors are then used to further construct the overall semantic feature matrix of drug and protein nodes. (7); (8); in, The semantic feature matrix for all drug nodes. This is the semantic feature matrix of all protein nodes. This represents the total number of drug nodes. This represents the total number of protein nodes. This represents the matrix transpose operation.

[0027] Step 1.3: Two types of feature enhancement strategies are designed for different data conditions: For biomedical data with original node attributes, a semantic-attribute gating alignment mechanism is introduced; for biomedical data lacking original node attributes, a semantic feature compensation strategy is adopted. The specific workflow is as follows: Step 1.3.1: For biomedical data containing original node attributes, feature enhancement is performed using the aforementioned semantic-attribute gating alignment mechanism. Since the original attribute features and semantic features have different sources and significantly different distributions, the original attribute feature matrices and semantic feature matrices of drugs and proteins are projected onto a unified feature space through a nonlinear mapping function, with protein nodes... For example, the mapping result is represented as: (9); (10); in, Represents protein nodes The original attribute feature vector, This represents the original attribute features after mapping. This represents the semantic features after mapping. and The mapping function, composed of linear transformation, batch normalization, and nonlinear activation function, can be expressed in the following form: (11); in, This is the first layer of learnable parameter matrix. This is the learnable parameter matrix for the second layer. Indicates input features, This indicates a batch normalization operation. RelU is a non-linear activation function. Step 1.3.2: Before feature fusion, perform a normalization process on the two types of features: (12); (13); in, Representing vectors Norm (i.e., vector length) and These represent the normalized attribute features and semantic features, respectively; Step 1.3.3: Adaptively fuse the two types of features after normalization using a gating mechanism: (14); (15); in, This represents the vector concatenation operation. For the learnable parameter matrix in the gating mechanism, The learned gate vector takes values ​​between 0 and 1. Represents the element-wise multiplication operation of vectors; This is the initial representation of the fused protein nodes; Drug nodes For example, using the same operation as above, the initial representation of the fused drug node is obtained. .

[0028] Step 1.3.4: For biomedical data lacking original node attributes, a semantic feature compensation strategy is used for feature enhancement. The feature matrix obtained from semantic encoding is directly used as the initial feature of the nodes. The initial feature matrices for drug nodes and protein nodes are expressed as follows: (16); (17); in, This represents the initial feature matrix of the drug node. This represents the initial feature matrix of the protein node.

[0029] Step 1.4: Using a heterogeneous graph neural network model, learn the final embedding representations of protein and drug nodes that integrate semantic and structural information, for use in prediction tasks. The specific workflow is as follows: Step 1.4.1: In the heterogeneous graph neural network model (taking a heterogeneous graph attention network as an example), the meta-path is calculated using a node-level attention mechanism. Next protein node Its neighboring nodes Attention score between : (18); in, Indicates the first Path of a single element and These represent protein nodes. and its neighboring nodes The input features are determined according to the feature enhancement strategy in step 1.3: for data with original attributes, the input features are the initial representations of the fused nodes; for data lacking original attributes, the input features are the initial feature matrices of the nodes. Let be the linear transformation matrix of the node features. For attention weight vectors, LeakyReLU is the transpose of this vector and is the rectified linear unit activation function with leakage. Step 1.4.2: Perform softmax normalization on the attention scores to obtain normalized attention weights. : (19); in, Represented by natural constant An exponential function with base 0. Indicates the metapath Next node The set of all neighboring nodes, The traversal index in the set; the node is obtained using the following formula. In metapath Embedded representation : (20); in, A learnable weight matrix for feature aggregation; Step 1.4.3: Introduce a semantic-level attention mechanism to fuse information from different meta-paths and calculate the first... Normalized importance coefficient of element path : (twenty one); in, Metapath The corresponding learnable weights, This represents the total number of metapaths in a heterogeneous network. For the first Learnable weights of element paths; Step 1.4.4: Integrate information from different metapaths to obtain protein nodes. The final embedding representation : (twenty two).

[0030] Using the same process as above, drug nodes are obtained. The final embedding representation .

[0031] Step 1.5: Concatenate the final embedding representations of the drug node and protein node to construct a joint representation of the node pair, and output the predicted probability of an interaction between the drug and target through a multilayer perceptron. The specific process is as follows: Step 2.1: For any drug node and protein nodes Finally embed it into the representation and Perform splicing to construct a joint representation of node pairs. : (twenty three); A multilayer perceptron is used to perform nonlinear modeling of the joint representation, and the predicted probability is output. : (twenty four); in, This represents a multilayer perceptron composed of multiple linear transformations and nonlinear activation functions. This represents the Sigmoid activation function. Indicates drug With protein The predicted probability values ​​between the two sides are 0 to 1. Step 2: Construct a joint loss function by combining the prediction loss and feature alignment constraints, and optimize and train the heterogeneous graph neural network prediction model LASE-HGNN; the specific process is as follows: Step 2.1: Construct a training sample set using a negative sampling strategy. Calculate the weighted binary cross-entropy loss function used for optimization. : (25); in, This represents the total number of training samples. This represents a pair of positive and negative samples in the set. The weighting coefficient is used to adjust the proportion of positive samples; Step 2.2: Construct a joint optimization objective. For datasets containing original attribute features, it is necessary to calculate the semantic-attribute gating alignment loss. : (26); in, For protein nodes The corresponding gating weight vector is used to dynamically adjust the alignment strength; and By combining these, we obtain the joint optimization objective function. : (27); in, To control the alignment loss weighting coefficients of feature alignment strength, This represents the set of all learnable parameters in the model. The square of the parameter set Norm, To control The regularization coefficient for regularization strength; Step 2.3: For datasets lacking original attribute features, since feature alignment is not required, the following joint loss function, which includes prediction loss and regularization term, is directly used for optimization: (28).

[0032] The model parameters are continuously updated and backpropagated according to formula (27) or formula (28).

[0033] Step 3: Obtain the drug and target node data to be predicted, input them into the trained heterogeneous graph neural network prediction model, and output the predicted probability of interaction between the drug and target. The specific process is as follows: Step 3.1: Determine the heterogeneous biomedical network and initial feature matrix of nodes; process the collected biomedical data to be predicted into a heterogeneous network, where the nodes of the heterogeneous network represent different biomedical entities, including core drug nodes, protein nodes, and auxiliary disease, side effect or gene nodes; the edges of the heterogeneous network represent known interaction relationships between entities, including drug-protein relationship, drug-disease relationship, protein-disease relationship and drug-side effect relationship and other biomedical relationships; use the big language model adapted to the biological domain through formula (1) and formula (2) to generate semantic description text of drug and protein nodes, and use the pre-trained model Bio-ClinicalBERT to generate semantic description text of drug and protein nodes through formula (3) to formula (8) and encode it into a semantic feature matrix; Step 3.2: Based on the feature conditions of the dataset to be predicted, perform forward fusion calculation of node representation using different feature enhancement strategies; if the dataset to be predicted contains original attribute features, then project the original attribute features and semantic features to a unified feature space according to formulas (9) to (11), normalize the two types of features according to formulas (12) and (13), and then use the gating parameter matrix obtained from training to perform semantic-attribute gating adaptive fusion according to formulas (14) and (15) to obtain the initial representation of the fused node; if the dataset to be predicted lacks original attribute features, then adopt the semantic feature compensation strategy according to formulas (16) and (17) and directly use the generated semantic feature matrix as the initial representation of drug and protein nodes; Step 3.3: Input the initial node representation obtained above into the heterogeneous graph neural network model (taking the heterogeneous graph attention network as an example) for representation learning; calculate the attention score between the node and its neighboring nodes under a specific meta-path according to formula (18), perform softmax normalization according to formula (19), and aggregate according to formula (20) to obtain the node embedding representation under the specific meta-path; then, calculate the semantic level attention weight of different meta-paths according to formula (21), and integrate the information of multiple meta-paths according to formula (22) to obtain the final embedding representation of the drug node and target protein node to be predicted; Step 3.4: Calculate the drug-target interaction probability and generate a prediction list. According to formula (23), the final embedding representations of the drug node to be predicted and the target protein node are concatenated to construct a joint representation of the node pair. The joint representation is input into the trained multilayer perceptron, and the prediction probability of the interaction between the drug and the target is calculated according to formula (24). Finally, the prediction probabilities between the drug to be predicted and all candidate protein targets are compared, and the prediction probability values ​​are sorted in descending order. The probability values ​​greater than the preset threshold or ranked first are selected. The target proteins are identified, and a list of predicted drug-protein target interactions is generated.

[0034] To demonstrate the feasibility and superiority of this invention, the following comparative experiments were conducted. The proposed LASE-HGNN model was compared and analyzed with six baseline models: HAN, HPN, ie-HGCN, SGCL-DTI, DMHGNN, and MVCL-DTI. Among them, HAN is a network that combines node-level and semantic-level attention mechanisms, aiming to capture subtle connections between different types of nodes and edges in heterogeneous graphs and improve the performance of representation learning; HPN is a path-based heterogeneous graph neural network that focuses on learning node representations by utilizing different types of paths in the graph, enhancing the understanding of complex relationships; ie-HGCN is a model that adopts a hierarchical aggregation strategy, automatically identifying and strengthening useful meta-path patterns through iterative object-level and type-level information aggregation; SGCL-DT builds an end-to-end model through supervised contrastive learning, mining drug and target representations from meta-path-based neighbors, constructing dual views with drug-target pairs as nodes, comparing global topological and semantic features, and improving model prediction performance by constructing contrastive loss; DMHGNN models high-order semantic information and topological dependencies through a dual-channel structure, enhancing the ability to model complex interaction patterns; MVCL-DTI models high-order semantic information and topological dependencies through a dual-channel structure, enhancing the ability to model complex interaction patterns.

[0035] To ensure fair comparison, the experimental parameters of all baseline models and the model proposed in this invention were uniformly set: the hidden layer dimension was set to 64, the learning rate to 0.001, and the weight decay to 1.0 × 10⁻⁶. -4 The feature dropout rate was 0.5, the maximum number of iterations was set to 1200, and the Adam optimizer was used to update the model parameters. The training set and test set were set to 10% and 80% of the total dataset, respectively.

[0036] This invention selected three real-world biomedical heterogeneous network datasets—Ning, Luo, and Peng—for comparative experiments. The Ning and Luo datasets contain the raw physical / chemical properties of drug molecules (e.g., molecular fingerprints) and protein targets (e.g., amino acid sequences); while the Peng dataset is a typical attribute-free dataset, containing only network topology connections between entities. For the different feature conditions of the datasets, the LASE-HGNN model of this invention sets differentiated feature enhancement strategies and parameters. In the Ning and Luo datasets, which contain raw features, the model activates a semantic-attribute gating alignment mechanism and sets the alignment loss weight coefficients. To constrain the distribution consistency of multimodal features, the model adaptively learns gating bias parameters through the network to dynamically adjust the fusion ratio of semantics and attributes. In the Peng dataset, which lacks original features, the model activates a semantic feature compensation strategy, directly projecting the 768-dimensional semantic feature matrix generated by the large model onto the hidden layer as the initial representation of the nodes. In all three datasets, the ratio of positive to negative samples is strictly controlled to 1:1 using a random negative sampling strategy to avoid evaluation bias caused by class imbalance.

[0037] The model was evaluated on a drug-target interaction (binary classification) task using a multilayer perceptron (MLP) combined with a sigmoid activation function as the final predictive classifier. Accuracy (ACC), area under the receiver operating characteristic (AUROC) curve, area under the precision-recall (AUPR) curve, and F1 score were used as the core evaluation metrics. Table 1 shows the performance comparison results of six baseline methods with the method of this invention on different datasets, with the best results shown in bold.

[0038] Table 1. Performance comparison (%) of six baseline methods and feature enhancement strategies on different datasets. In all three datasets (Ning, Luo, Peng) and all six baseline models (from traditional HAN and HPN to the newer contrastive learning model MVCL-DTI), the proposed +LASE-HGNN feature enhancement strategy achieved comprehensive and stable top performance across all evaluation metrics (ACC, AUROC, AUPR, F1). This fully demonstrates that the framework is not tied to any specific underlying architecture and possesses strong pluggability and versatility. For the Ning and Luo datasets, although directly using the original attribute features can bring some improvement compared to using only topological structures, the performance after feature enhancement with LASE-HGNN is still significantly better than directly using the original attribute features. For example, on the HAN model in the Luo dataset, the ACC jumped from 92.60% to 96.62% after feature enhancement with LASE-HGNN. This indicates that the deep semantic features extracted by the present invention through a large model, combined with a semantic-attribute gating alignment mechanism, have stronger discriminative power than simply using the original physicochemical attributes, successfully bridging the modality gap and eliminating semantic drift.

[0039] The Peng dataset belongs to networks without original attributes (therefore, "+original attribute features" is N / A). In this dataset, traditional baselines can only rely on extremely limited topology for learning, leading to a precipitous drop in the performance of some models (e.g., HAN's ACC is only 58.44%, and ie-HGCN's ACC is 70.15%). However, after introducing the semantic feature compensation strategy of LASE-HGNN, the model performance experienced an explosive increase (e.g., HAN's ACC soared to 96.59%). This intuitively demonstrates that the present invention, by utilizing the semantic features of large models as supplementary information, can fundamentally solve the performance bottleneck caused by feature loss and endow the model with extremely high robustness in incomplete environments.

[0040] To further verify the stability, robustness, and generalization ability of the semantically enhanced heterogeneous graph neural network prediction model (LASE-HGNN) under different data partitioning conditions, this embodiment conducted a five-fold cross-validation experiment on three representative heterogeneous graph neural network baseline models (HAN, HPN, and ie-HGCN) on the Ning, Luo, and Peng datasets. The experiment used the mean and standard deviation (mean ± standard deviation) of five independent tests as the final evaluation index. The specific results are shown in Table 2. Table 2 Performance stability evaluation results of five-fold cross-validation on different datasets (mean ± standard deviation, %) The five-fold cross-validation experiment further confirms that the proposed framework (LASE-HGNN) possesses extremely high stability, robustness, and broad architectural versatility. Table 2 shows that regardless of the heterogeneous graph neural network (such as HAN, HPN, or ie-HGCN) used, after introducing the feature enhancement strategy of this invention, the model achieves comprehensive leadership in core evaluation metrics across all datasets, and performance fluctuations are strictly controlled within a very small range (e.g., the AUROC standard deviation of HAN on the Peng dataset is as low as ±0.0002). Particularly in the Peng dataset, which lacks original attributes, the semantic feature compensation strategy of this invention effectively overcomes the performance precipitate and high volatility caused by traditional models relying solely on sparse topological features (e.g., significantly increasing the AUROC of ie-HGCN from 72.05% to 96.89% and significantly reducing variance). In summary, the high-quality semantic knowledge extracted by this invention can not only seamlessly adapt to various network architectures as a general-purpose plugin, but also fundamentally solve the problem of missing features, endowing the model with excellent generalization capabilities under different data partitions and low-resource scenarios.

[0041] To provide a more intuitive evaluation of the internal mechanism of the framework proposed in this invention and its performance under extreme data, this invention conducted ablation experiments and feature missing scenario tests on the model, and performed visualization analysis on the results.

[0042] Figure 2 This is a comparison of the prediction performance of different feature enhancement module combinations in the experiments of this invention. The figure shows the changes in model performance after removing different key components on a dataset containing the original attributes (Ning dataset). Figure 2 As shown in the figure, the performance differences of the proposed method on the baseline model are presented after removing large language model fine-tuning, removing alignment constraints, removing adaptive gating mechanisms, and so on. It is evident that when large language model fine-tuning is removed, the model's prediction performance significantly decreases because it can only acquire shallow textual information and lacks deep biomedical semantics. When alignment constraints or gating mechanisms are removed, the model experiences significant information interference when directly splicing or fusing original physical attributes with textual semantics, resulting in damage to multiple evaluation metrics. In contrast, the performance histogram of the complete framework of this invention consistently ranks highest, with all metrics reaching optimal levels. This intuitively verifies that each micro-module of the multimodal feature fusion mechanism in this invention (domain adaptation, alignment constraints, and gating fusion) is indispensable; they work together to achieve consistent and optimal modeling of multi-source features.

[0043] Figure 3 This image shows the performance improvement effect of the semantic feature compensation strategy on attribute-free datasets in the experiments of this invention. In real biomedical networks, many nodes often lack physicochemical property features, posing a significant challenge to traditional graph neural networks when processing such data. Figure 3 As shown in the Peng dataset, traditional models (baseline methods relying solely on topology) generally exhibit low prediction accuracy and low AUROC metrics due to their reliance on extremely limited and sparse network topology edges for node embedding learning. However, after introducing the semantic feature compensation strategy proposed in this invention, the performance histograms of each baseline model show a dramatic increase, with the histogram heights showing a significant difference. This visualization strongly demonstrates that, under extreme conditions lacking original attributes, this invention directly utilizes the structured semantic features generated by a large language model as the initial representation of nodes, effectively compensating for the inherent deficiencies caused by attribute loss and fundamentally solving the performance bottleneck of the model in low-resource and cold-start scenarios.

[0044] The experimental results demonstrate that the framework of this invention excels in drug-target interaction prediction and multi-source feature fusion, accurately capturing and aligning deep biomedical semantics with complex graph topologies. These superior semantic compensation and gating alignment effects directly enhance the robustness of heterogeneous graph neural networks, making potential target predictions more accurate even in extreme feature-deficient (cold-start) environments. Therefore, the method of this invention is not only highly effective in computer-aided drug discovery but also significantly superior to existing baseline methods, revealing and predicting the complex interaction mechanisms between drug molecules and protein receptors at a deeper level.

[0045] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A drug target interaction prediction method based on semantic enhancement of a large language model, characterized in that, Includes the following steps: Step 1: Acquire biomedical data to construct a heterogeneous network. Use a domain-adapted large language model to generate structured semantic descriptions of drug and protein nodes and encode them as semantic features. Combine semantic-attribute gating alignment mechanism and semantic feature compensation strategy to construct a pluggable semantically enhanced heterogeneous graph neural network prediction model. Step 2: Construct a joint loss function by combining the prediction loss and feature alignment constraints, and optimize the training of the heterogeneous graph neural network prediction model; Step 3: Obtain the data of the drug and target nodes to be predicted, input them into the trained heterogeneous graph neural network prediction model, and output the predicted probability of the interaction between the drug and the target.

2. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 1, characterized in that, In step 1, the semantically enhanced heterogeneous graph neural network prediction model works as follows: Step 1.1: Construct a heterogeneous biomedical network containing drug nodes, protein nodes and their various relationships, and use node type mapping functions and edge type mapping functions to uniformly model different types of entities and relationships; Step 1.2: Construct biomedical instruction data, use an efficient parameter fine-tuning method to adapt the open-source large language model to the domain, and generate semantic descriptions of drug and protein nodes; use Bio-ClinicalBERT to encode the semantic descriptions to obtain a unified semantic feature representation. Step 1.3: Design two types of feature enhancement strategies for different data conditions: For biomedical data with original node attributes, introduce a semantic-attribute gating alignment mechanism; for biomedical data lacking original node attributes, adopt a semantic feature compensation strategy. Step 1.4: Through a heterogeneous graph neural network model, learn the final embedding representation of protein and drug nodes that integrates semantic and structural information, which is used for prediction tasks; Step 1.5: Concatenate the final embedded representations of drug nodes and protein nodes to construct a joint representation of node pairs, and output the predicted probability of interaction between drug and target through a multilayer perceptron.

3. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 2, characterized in that, The specific process of step 1.2 is as follows: Step 1.2.1: Use the LoRA method to fine-tune the open-source large language model for domain adaptation. Let a certain linear transformation weight in the pre-trained large language model be... Incrementally update it Represented in low-rank matrix decomposition form: (1)； Updated weights Represented as: (2)； in, and For training, a low-rank matrix, Indicates the output feature dimension. Indicates the input feature dimension. Denotes the rank parameter of a low-rank decomposition, satisfying ; Step 1.2.2: After completing domain adaptation, utilize the fine-tuned pre-trained large language model. Generate semantic description text for drug nodes and protein nodes in heterogeneous biomedical networks: (3)； (4)； in, Represents the set of drug nodes. Represents a set of protein nodes. Represents the first element in the set. One drug node, Represents the first element in the set. One protein node; For drug nodes The corresponding semantic description text, For protein nodes The corresponding semantic description text; Step 1.2.3: After obtaining the semantic description text of the nodes, the semantic description text is semantically encoded using the biomedical pre-trained language model Bio-ClinicalBERT to obtain a unified semantic vector representation: (5)； (6)； in, This represents the encoding function of the pre-trained language model Bio-ClinicalBERT. Indicates drug node semantic feature vectors, Represents protein nodes The semantic feature vectors are then used to further construct the overall semantic feature matrix of drug and protein nodes. (7)； (8)； in, The semantic feature matrix for all drug nodes. This is the semantic feature matrix of all protein nodes. This represents the total number of drug nodes. This represents the total number of protein nodes. This represents the matrix transpose operation.

4. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 3, characterized in that, The specific process of step 1.3 is as follows: Step 1.3.1: For biomedical data containing original node attributes, feature enhancement is performed using the aforementioned semantic-attribute gating alignment mechanism. The original attribute feature matrices and semantic feature matrices of drugs and proteins are projected onto a unified feature space through a nonlinear mapping function, wherein protein nodes... The mapping result is represented as: (9)； (10)； in, Represents protein nodes The original attribute feature vector, This represents the original attribute features after mapping. This represents the semantic features after mapping. and The mapping function, composed of linear transformation, batch normalization, and nonlinear activation function, is expressed in the following form: (11)； in, This is the first layer of learnable parameter matrix. This is the learnable parameter matrix for the second layer. Indicates input features, This indicates a batch normalization operation. RelU is a non-linear activation function. Step 1.3.2: Before feature fusion, perform a normalization process on the two types of features: (12)； (13)； in, Representing vectors Norm, and These represent the normalized attribute features and semantic features, respectively; Step 1.3.3: Adaptively fuse the two types of features after normalization using a gating mechanism: (14)； (15)； in, This represents the vector concatenation operation. For the learnable parameter matrix in the gating mechanism, The learned gate vector takes values ​​between 0 and 1. Represents the element-wise multiplication operation of vectors; This is the initial representation of the fused protein nodes; Step 1.3.4: For biomedical data lacking original node attributes, a semantic feature compensation strategy is used for feature enhancement. The feature matrix obtained from semantic encoding is directly used as the initial feature of the nodes. The initial feature matrices for drug nodes and protein nodes are expressed as follows: (16)； (17)； in, This represents the initial feature matrix of the drug node. This represents the initial feature matrix of the protein node.

5. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 4, characterized in that, The specific process of step 1.4 is as follows: Step 1.4.1: In a heterogeneous graph attention network, compute the meta-path using a node-level attention mechanism. Next protein node Its neighboring nodes Attention score between : (18)； in, Indicates the first Path of a single element and These represent protein nodes. and its neighboring nodes The input features are determined according to the feature enhancement strategy in step 1.3: for data with original attributes, the input features are the initial representations of the fused nodes; for data lacking original attributes, the input features are the initial feature matrices of the nodes. Let be the linear transformation matrix of the node features. For attention weight vectors, LeakyReLU is the transpose of this vector and is the rectified linear unit activation function with leakage. Step 1.4.2: Perform softmax normalization on the attention scores to obtain normalized attention weights. : (19)； in, Represented by natural constant An exponential function with base 0. Indicates the metapath Next node The set of all neighboring nodes, The traversal index in the set; the node is obtained using the following formula. In metapath Embedded representation : (20)； in, A learnable weight matrix for feature aggregation; Step 1.4.3: Introduce a semantic-level attention mechanism to fuse information from different meta-paths and calculate the first... Normalized importance coefficient of element path : (21)； in, Metapath The corresponding learnable weights, This represents the total number of metapaths in a heterogeneous network. For the first Learnable weights of element paths; Step 1.4.4: Integrate information from different metapaths to obtain protein nodes. The final embedding representation : (22)。 6. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 5, characterized in that, Step 1.5 specifically involves: For any drug node and protein nodes Finally embed it into the representation and Perform splicing to construct a joint representation of node pairs. : (23)； A multilayer perceptron is used to perform nonlinear modeling of the joint representation, and the predicted probability is output. : (24)； in, This represents a multilayer perceptron composed of multiple linear transformations and nonlinear activation functions. This represents the Sigmoid activation function. Indicates drug With protein There are predicted probability values ​​for interactions between them, ranging from 0 to 1.

7. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 6, characterized in that, In step 2, the specific process of constructing a joint loss function by combining the prediction loss and feature alignment constraints and optimizing the training is as follows: Step 2.1: Construct a training sample set using a negative sampling strategy. Calculate the weighted binary cross-entropy loss function used for optimization. : (25)； in, This represents the total number of training samples. This represents a pair of positive and negative samples in the set. The weighting coefficient is used to adjust the proportion of positive samples; Step 2.2: Construct a joint optimization objective. For datasets containing original attribute features, it is necessary to calculate the semantic-attribute gating alignment loss. : (26)； in, For protein nodes The corresponding gating weight vector; and By combining these, we obtain the joint optimization objective function. : (27)； in, To control the alignment loss weighting coefficients of feature alignment strength, This represents the set of all learnable parameters in the model. The square of the parameter set Norm, To control The regularization coefficient for regularization strength; Step 2.3: For datasets lacking original attribute features, since feature alignment is not required, the following joint loss function, which includes prediction loss and regularization term, is directly used for optimization: (28)。 8. The drug target interaction prediction method based on semantic enhancement of a large language model according to claim 7, characterized in that, The specific process of step 3 is as follows: Step 3.1: Determine the heterogeneous biomedical network and initial feature matrix of nodes; process the collected biomedical data to be predicted into a heterogeneous network form, where the nodes of the heterogeneous network represent different biomedical entities, including core drug nodes, protein nodes, and auxiliary disease, side effect or gene nodes; the edges of the heterogeneous network represent known interaction relationships between entities, including drug-protein relationship, drug-disease relationship, protein-disease relationship and drug-side effect relationship; Using a large language model adapted to the biological domain through formulas (1) and (2), semantic description texts of drug and protein nodes are generated. The pre-trained model Bio-ClinicalBERT is used to generate semantic description texts of drug and protein nodes through formulas (3) to (8) and encode them into semantic feature matrices. Step 3.2: Based on the feature conditions of the dataset to be predicted, perform forward fusion calculation of node representation using different feature enhancement strategies; if the data to be predicted contains original attribute features, then project the original attribute features and semantic features to a unified feature space according to formulas (9) to (11), normalize the two types of features according to formulas (12) and (13), and then use the gating parameter matrix obtained from training to perform semantic-attribute gating adaptive fusion according to formulas (14) and (15) to obtain the initial representation of the fused node; If the data to be predicted lacks original attribute features, then according to formulas (16) and (17), a semantic feature compensation strategy is adopted, and the generated semantic feature matrix is ​​directly used as the initial representation of drug and protein nodes. Step 3.3: Input the initial node representations obtained above into the heterogeneous graph attention network for representation learning; The attention scores of nodes and their neighbors under a specific metapath are calculated according to formula (18), softmax normalization is performed according to formula (19), and aggregation is performed according to formula (20) to obtain the node embedding representation under the specific metapath. Subsequently, the semantic attention weights of different meta-paths are calculated according to formula (21), and the information of multiple meta-paths is integrated according to formula (22) to obtain the final embedding representation of the drug node to be predicted and the target protein node. Step 3.4: Calculate the drug-target interaction probability and generate a prediction list. According to formula (23), the final embedding representations of the drug node to be predicted and the target protein node are concatenated to construct a joint representation of the node pair. The joint representation is input into the trained multilayer perceptron, and the prediction probability of the interaction between the drug and the target is calculated according to formula (24). Finally, the prediction probabilities between the drug to be predicted and all candidate protein targets are compared, and the prediction probability values ​​are sorted in descending order. The probability values ​​greater than the preset threshold or ranked first are selected. The target proteins are identified, and a list of predicted drug-protein target interactions is generated.

Images