A drug interaction prediction method based on bidirectional event semantic guidance
By employing a bidirectional event semantic guidance approach, deep molecular graph neural networks and biomedical language models are used to extract multimodal drug features, construct a bidirectional event semantic space, and optimize semantic supervision alignment. This approach solves the problems of semantic information neglect and unidirectional modeling in existing technologies, and achieves high accuracy and robustness in drug interaction prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF TECHNOLOGY
- Filing Date
- 2025-11-17
- Publication Date
- 2026-05-12
AI Technical Summary
Existing drug interaction prediction methods ignore semantic information in the text of interaction events. One-way modeling leads to performance degradation in imbalanced classes or scenarios with few samples, and it is difficult to form a symmetrical discrimination boundary between positive and negative samples.
A bidirectional event semantic guidance method is adopted, which extracts multimodal drug features through deep molecular graph neural network, multilayer perceptron and pre-trained biomedical language model, constructs bidirectional event semantic space, introduces uniform distribution constraint and orthogonal constraint, optimizes semantic supervision alignment, and adopts dynamic guidance coefficient adjustment strategy to improve model robustness.
It improves the accuracy and robustness of drug interaction prediction, especially in the identification of rare interaction types, reduces the risk of misjudgment, and enhances the model's generalization ability.
Smart Images
Figure CN121506310B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and bioinformatics, and in particular to a drug interaction prediction method based on bidirectional event semantic guidance. Background Technology
[0002] Drug interaction prediction is of great significance in drug safety monitoring and new drug development. Its core is identifying the types of interaction events that may occur when two or more drugs are used in combination. Current mainstream prediction methods can be mainly divided into three categories: one uses drug sequence information to extract features and perform classification; another represents the drug molecular structure as a graph and learns interaction features at the atomic or substructure level through graph neural networks; and the third constructs a heterogeneous network of biomedical entities such as drugs and targets, and obtains interaction features through graph embedding or relational reasoning.
[0003] However, existing methods have obvious limitations: most methods focus on drug structure or network features, but ignore the semantic information such as pharmacological effects, direction of action and indications contained in the interactive event text, which is crucial for distinguishing different interaction types; some methods that attempt to introduce event semantics use unidirectional modeling, only use positive semantics as supervision, and do not construct a negative semantic space, making it difficult to form a symmetrical discrimination boundary between positive and negative samples. In scenarios with class imbalance or few samples, the model is easily over-guided by high-frequency events, resulting in a decrease in prediction performance and robustness. Summary of the Invention
[0004] To address the problems mentioned in the background art, this invention provides a drug interaction prediction method based on bidirectional event semantic guidance, aiming to improve the accuracy of prediction.
[0005] This invention adopts the following technical solution: This invention provides a drug interaction prediction method based on bidirectional event semantic guidance, including: S1: Extracting multimodal drug features of the drug pair to be predicted; multimodal drug features refer to the comprehensive features obtained by extracting molecular graph structural features, sequence-level features, and textual semantic features from the drug's molecular graph, simplified molecular input linear canonical sequence, and drug text description, respectively, and then splicing and fusing the molecular graph structural features, sequence-level features, and textual semantic features; S2: Constructing a bidirectional event semantic space; the bidirectional event semantic space refers to constructing a corresponding negative drug interaction event description text for each positive drug interaction event description text to form a positive-negative semantic pair. After encoding the positive and negative semantic pairs, the semantic space is optimized by preset constraints; S3: The multimodal drug features obtained in step S1 are semantically supervised and aligned with the bidirectional event semantic space obtained in step S2; Semantic supervision alignment refers to optimizing the matching degree between the multimodal drug features and each positive and negative semantic vector in the bidirectional event semantic space through similarity calculation; S4: Based on the result of semantic supervision alignment completed in step S3, the semantic guidance intensity is scaled by combining a dynamic guidance coefficient adjustment strategy, and the drug interaction prediction result is output; The dynamic guidance coefficient adjustment strategy refers to the strategy of dynamically adjusting the semantic guidance intensity according to the occurrence frequency of drug interaction event categories and the model training rounds.
[0006] Step S1 extracts multimodal drug features of the drug pair to be predicted, including: using deep molecular graph neural network, multilayer perceptron and pre-trained biomedical language model respectively, extracting molecular graph structural features, sequence-level features and text semantic features from the molecular graph, simplified molecular input linear canonical sequence and drug text description of each drug in the drug pair to be predicted.
[0007] Step S2 involves optimization through preset constraints, including: calculating semantic centers for all positive and negative semantic vectors obtained after encoding, to obtain positive and negative semantic centers; applying uniform distribution constraints to the positive and negative semantic vectors based on the positive and negative semantic centers; and applying orthogonal constraints to the positive and negative semantic vectors to ensure that they are directionally independent in the semantic space.
[0008] In step S3, the matching degree is optimized by similarity calculation, including: calculating the positive similarity between multimodal drug features and each positive semantic vector in the bidirectional event semantic space, and the negative similarity between multimodal drug features and each negative semantic vector in the bidirectional event semantic space; and using a preset loss function to optimize the positive and negative similarities, thereby increasing the matching degree between positive samples and positive semantic vectors, and increasing the matching degree between negative samples and negative semantic vectors.
[0009] The dynamic guidance coefficient adjustment strategy in step S4 includes: dynamically determining the frequency adjustment coefficient using a preset function based on the difference between the occurrence frequency of each drug interaction event category and the average occurrence frequency of all drug interaction event categories; determining the round decay coefficient with a preset trend based on the relationship between the current model training rounds and the preset total training rounds; and calculating the dynamic guidance coefficient by weighting and combining the frequency adjustment coefficient and the round decay coefficient, which is used to scale the semantic guidance intensity.
[0010] The pre-trained biomedical language model is a pre-trained language model optimized for text in the biomedical field. It can semantically encode information related to pharmacological effects, indications, and contraindications in drug text descriptions.
[0011] In step S2, the mean of all positive semantic vectors is calculated to obtain the positive semantic center; the mean of all negative semantic vectors is calculated to obtain the negative semantic center.
[0012] The preset loss function used in step S3 is the information noise comparison estimation loss function. The information noise comparison estimation loss function adjusts the similarity weight by setting the temperature parameter, thereby optimizing the matching accuracy between multimodal drug features and semantic vectors.
[0013] The preset function used in step S4 is the hyperbolic tangent nonlinear activation function. The hyperbolic tangent nonlinear activation function processes the difference between the occurrence frequency of drug interaction event categories and the average occurrence frequency, amplifies the difference in the occurrence frequency of different drug interaction event categories, and obtains the frequency adjustment coefficient.
[0014] The round decay coefficient exhibits an exponential decay trend, which ensures that the model has strong semantic guidance in the early stages of training and maintains stable convergence in the later stages of training.
[0015] Compared to existing technologies, the beneficial effects are as follows:
[0016] 1. This application provides a drug interaction prediction method based on bidirectional event semantic guidance. By utilizing deep molecular graph neural networks, multilayer perceptrons, and pre-trained language models optimized for the biomedical field, it extracts and fuses the structural, sequence, and textual semantic features of drug molecules. This overcomes the limitations of existing methods that focus on single structures and network features while neglecting textual semantics, comprehensively covering drug molecular attributes and pharmacological descriptions, providing more robust feature support for interaction event discrimination. Furthermore, by constructing positive and negative semantic pairs and introducing uniform distribution constraints and orthogonal constraints, a bidirectional event semantic space is built, addressing the problem that existing unidirectional semantic modeling struggles to form symmetrical discrimination boundaries.
[0017] 2. Based on the information noise contrast estimation loss function, the matching degree between multimodal drug features and semantic vectors in the bidirectional semantic space is optimized, strengthening the correlation between features and target semantics, reducing non-corresponding semantic interference, and further improving the model's accuracy in recognizing different interaction types, thus reducing the risk of misjudgment. Combining the hyperbolic tangent function and exponential decay trend, a dynamic guidance coefficient adjustment strategy is constructed. The guidance intensity is dynamically scaled according to the frequency of interaction events and the training epochs, effectively solving the problem of over-guidance for high-frequency categories and insufficient guidance for low-frequency categories in traditional training. This ensures the prediction performance of rare interaction events and improves the overall robustness and generalization ability of the model. Attached Figure Description
[0018] Figure 1 This is a flowchart of a drug interaction prediction method based on bidirectional event semantic guidance provided in an embodiment of this application;
[0019] Figure 2 This is a comparison chart of experimental results provided in an embodiment of this application;
[0020] Figure 3 This is another comparative chart of experimental results provided in an embodiment of this application;
[0021] Figure 4 This is another comparative chart of experimental results provided in the embodiments of this application. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0023] This application provides a drug interaction prediction method based on bidirectional event semantic guidance, for example, such as... Figure 1 As shown, it includes:
[0024] S1: Extract multimodal drug features of the drug pair to be predicted; multimodal drug features refer to the comprehensive features obtained by extracting molecular graph structural features, sequence-level features and textual semantic features from the molecular graph of the drug, the simplified linear canonical sequence of the molecular input and the drug text description, respectively, and then splicing and fusing the molecular graph structural features, sequence-level features and textual semantic features.
[0025] As one possible implementation, step S1 extracts multimodal drug features of the drug pair to be predicted, including: using Deep Molecular Graph Neural Network (DeeperGCN), Multilayer Perceptron and pre-trained biomedical language model respectively, to extract molecular graph structural features, sequence-level features and textual semantic features from the molecular graphs, simplified molecular input linear canonical sequences and drug text descriptions of each drug in the drug pair to be predicted.
[0026] For example, after collecting drug molecular graphs, simplified molecular input linear canonical sequences (SMILES), and drug text descriptions for each drug, a deep molecular graph neural network is used to extract features from the drug molecular graphs obtained by transforming the simplified molecular input linear canonical sequences, thus obtaining drug molecular graph structural features. The simplified molecular input linear canonical sequences are encoded into high-dimensional vectors, and then processed by a multilayer perceptron to obtain drug sequence-level features. A pre-trained biomedical language model (BioBERT) is used to semantically encode the drug text descriptions, and then the encoding results are linearly transformed to obtain drug text semantic features. The drug molecular graph structural features, drug sequence-level features, and drug text semantic features are first fused, and then the fused features are concatenated with the interaction event label embedding features to obtain the comprehensive interaction features of the drug pairs, i.e., the comprehensive features.
[0027] Among them, the pre-trained biomedical language model is a pre-trained language model optimized for text in the biomedical field. The pre-trained biomedical language model can semantically encode information related to pharmacological effects, indications and contraindications in drug text descriptions.
[0028] As one possible implementation, the multimodal feature concatenation formula is as follows:
[0029]
[0030] in, These are features of the molecular diagram structure. For sequence-level features, For text semantic features, Embedded for event tags, This represents vector concatenation, where F is the composite feature.
[0031] S2: Construct a bidirectional event semantic space; the bidirectional event semantic space refers to constructing a corresponding negative drug interaction event description text for each positive drug interaction event description text to form a positive-negative semantic pair. After encoding the positive-negative semantic pairs, the semantic space is then optimized through preset constraints.
[0032] As one possible implementation, step S2 is optimized by pre-set constraints, including: calculating the semantic center for all positive semantic vectors and all negative semantic vectors obtained after encoding, to obtain the positive semantic center and the negative semantic center; based on the positive semantic center and the negative semantic center, performing uniform distribution constraint processing on the positive semantic vector and the negative semantic vector respectively; and performing orthogonal constraint processing on the positive semantic vector and the negative semantic vector to ensure that the two are directionally independent in the semantic space.
[0033] Before model training, for each positive drug interaction event description text, a negative drug interaction event description text semantically symmetrical to the positive text is constructed, forming a set of positive and negative semantic pairs. A pre-trained biomedical language model is used to semantically encode the positive and negative texts in each set of positive and negative semantic pairs, obtaining corresponding positive and negative semantic vectors. A linear mapping process is applied to the positive and negative semantic vectors, placing them in a feature space of the same dimension. To enhance the separation between the positive and negative semantic vectors, a bi-center uniform constraint loss and an orthogonal constraint loss are introduced: the bi-center uniform constraint loss ensures that the positive and negative semantic vectors are evenly distributed in the semantic space; the orthogonal constraint loss ensures that the directions of the positive and negative semantic vectors remain independent in the semantic space.
[0034] For example, in step S2, the mean of all positive semantic vectors is calculated to obtain the positive semantic center; the mean of all negative semantic vectors is calculated to obtain the negative semantic center.
[0035] As one possible implementation, the formula for the bicentric uniform constraint loss is as follows:
[0036]
[0037] in, Represents the set of positive events. Represents the set of negative events. This means removing j from the positive set. This indicates that the negative set is excluding k. and Let z represent the positive semantic center vector and the negative semantic center vector, respectively. + z - The feature representation of the corresponding event semantics.
[0038] For example, the formula for orthogonal constraint loss is shown below:
[0039]
[0040] in, and These are the positive and negative semantic vectors for the i-th type of event, respectively, and N is the negative semantic vector. l The loss represents the number of events, and it forces the two to be orthogonal in space, thereby maximizing the distinguishability between positive and negative classes.
[0041] This application constructs semantically symmetrical negative text for each positive event description text, forming positive-negative semantic pairs and building a bidirectional event semantic space. This allows the model to simultaneously learn the differences between positive and negative events, fundamentally solving the problem of ambiguous discrimination boundaries in unidirectional semantic modeling and significantly improving the ability to distinguish between positive and negative samples, especially improving the prediction performance of rare interaction types. Furthermore, addressing the issue of sample clustering and insufficient discriminative power between positive and negative semantics in the semantic space, this invention designs a bicentric uniform constraint loss and an orthogonal constraint loss: the bicentric uniform constraint loss ensures that positive and negative semantics are evenly distributed in the space, avoiding discrimination ambiguity caused by excessive sample clustering; the orthogonal constraint loss forces positive and negative semantic vectors to remain independent in spatial directions, maximizing class differences. The synergistic effect of these two types of constraints makes the feature distribution in the semantic space more reasonable and the class boundaries clearer, further improving the model's recognition accuracy for different interaction types and reducing the risk of misjudgment.
[0042] S3: Perform semantic supervision alignment between the multimodal drug features obtained in step S1 and the bidirectional event semantic space obtained in step S2; semantic supervision alignment refers to optimizing the matching degree between the multimodal drug features and each positive and negative semantic vector in the bidirectional event semantic space through similarity calculation.
[0043] As one possible implementation, step S3 optimizes the matching degree through similarity calculation, including: calculating the positive similarity between multimodal drug features and each positive semantic vector in the bidirectional event semantic space, and the negative similarity between multimodal drug features and each negative semantic vector in the bidirectional event semantic space; and using a preset loss function to optimize the positive and negative similarities, thereby increasing the matching degree between positive samples and positive semantic vectors, and increasing the matching degree between negative samples and negative semantic vectors.
[0044] Positive samples refer to drug pairs with specific drug interaction events and their corresponding data; negative samples refer to drug pairs without specific drug interaction events and their corresponding data.
[0045] For example, the preset loss function used in step S3 is a loss function based on information noise contrast estimation. The loss function based on information noise contrast estimation adjusts the weight of similarity by setting the temperature parameter, thereby optimizing the matching accuracy between multimodal drug features and semantic vectors.
[0046] As one possible implementation, the loss function formula based on information-noise contrast estimation is shown below:
[0047]
[0048] in, This represents the similarity between the comprehensive features of the drug pair to be predicted and the target positive semantics in the bidirectional event semantic space, where the target positive semantics corresponds to the positive semantic vector of the current drug interaction event type to be determined. The higher the similarity value, the stronger the match between the interaction features and the target positive semantics. This represents a temperature parameter used to adjust the weight scaling of similarity. This represents the similarity between the comprehensive features of the drug pair to be predicted and the non-target positive semantics in the bidirectional event semantic space. This represents the total number of interaction event types. exp represents an exponential function with the natural constant e as its base.
[0049] The core objective of step S3 is to optimize the matching accuracy between multimodal drug features and bidirectional event semantic space by establishing a connection between them, thereby providing a semantic basis for accurate output of drug interaction prediction results and solving the problem of the disconnect between drug features and the semantics of interaction events.
[0050] S4: Based on the results of semantic supervision alignment completed in step S3, the semantic guidance intensity is scaled by combining the dynamic guidance coefficient adjustment strategy, and the drug interaction prediction results are output. The dynamic guidance coefficient adjustment strategy refers to the strategy of dynamically adjusting the semantic guidance intensity according to the occurrence frequency of drug interaction event categories and the model training rounds.
[0051] As one possible implementation, the dynamic guidance coefficient adjustment strategy in step S4 includes: dynamically determining the frequency adjustment coefficient using a preset function based on the difference between the occurrence frequency of each drug interaction event category and the average occurrence frequency of all drug interaction event categories; determining the round decay coefficient with a preset trend based on the relationship between the current model training rounds and the preset total training rounds; and calculating the dynamic guidance coefficient by weighting the frequency adjustment coefficient and the round decay coefficient, which is used to scale the semantic guidance intensity.
[0052] For example, the preset function used in step S4 is the hyperbolic tangent nonlinear activation function. The hyperbolic tangent nonlinear activation function processes the difference between the occurrence frequency of drug interaction event categories and the average occurrence frequency, amplifies the difference in the occurrence frequency of different drug interaction event categories, and obtains the frequency adjustment coefficient.
[0053] To address the semantic guidance imbalance caused by class imbalance, this application proposes a dual-adjustment strategy for the guidance coefficient, also known as a pre-defined function. This strategy dynamically scales the guidance strength based on the frequency of event categories, reducing the weight of high-frequency events to avoid over-guidance and increasing the weight of low-frequency events to enhance the learning signal of rare events. The round decay coefficient exhibits an exponential decay trend, ensuring that the model has strong semantic guidance strength in the early stages of training and maintains stable convergence in the later stages.
[0054] For example, the dual adjustment strategy for the guidance coefficient is shown below:
[0055]
[0056] Where i represents the current training round; a is the initial coefficient; k is the control decay rate; and x is the number of times a certain event category occurs. represents the average frequency of occurrence of each category, and r is the amplification factor controlling the frequency difference. tanh represents the hyperbolic tangent function. The dual adjustment strategy for the guidance coefficient ensures that the guidance coefficient is influenced by frequency while also being dynamically adjusted over time.
[0057] This application also provides experimental verification, exemplary of which include D1 and D2:
[0058] D1: Use with R 2 The same dataset as the DDI dataset was used, with 2842 duplicate samples removed from the training, validation, and test sets, and misclassification issues corrected. The processed data contains 116,509 DDI triples, covering 1706 drugs and 62 interaction types. The ratio of negative to positive samples is 1:1. During data generation, rigorous checks were performed to ensure that negative sample data did not exist in the positive samples, thus ensuring correct training. In this embodiment, the DrugBank dataset is divided into training, validation, and test sets in a 6:2:2 ratio.
[0059] D2: This embodiment of the application also used all data in DrugBank as the experimental dataset. After processing, DrugBank contained 191,808 DDI triples, involving 1,706 drugs and 86 interactions. Each drug has only one interaction, which describes how one drug affects the metabolism of another drug. Negative samples were generated using the same method as in D1, and the training set, validation set, and test set were divided in a 6:2:2 ratio.
[0060] Reference Figure 2 , Figure 3 as well as Figure 4 , Figure 2This paper presents a comparison of the accuracy (ACC), area under the curve (AUC), mean precision (AP), and F1 score of this application (labeled Ours) with existing mainstream methods such as GAT-DDI, GMPNN, SSI-DDI, R2-DDI, DSN-DDI, and PEB-DDI on two public datasets, D1 and D2. Figure 2 It can be seen that this application performs best across all metrics in both datasets.
[0061] Figure 3 The radar charts show the accuracy of different methods across various drug interaction event categories. Each dimension corresponds to a type of interaction event, and the radius represents the prediction accuracy for that category. Existing methods generally exhibit high accuracy for high-frequency categories but low accuracy for rare categories, resulting in a noticeable dip in the radar chart along the rare category dimension. The method presented in this application, however, produces a more uniform and robust radar chart contour: it not only maintains high accuracy for high-frequency categories but also achieves a significant improvement in accuracy for rare interaction categories, achieving a breakthrough in robustness across all categories.
[0062] Figure 4 The performance trends of each method with training epochs (horizontal axis) and accuracy (vertical axis) are shown, intuitively reflecting the training convergence characteristics of the model. Compared with existing methods, the method provided in this application shows a rapid and sharp increase in performance in the early stages of training (epochs < 40), quickly reaching a high accuracy; while existing methods such as GAT-DDI, DeepDDI, and even the advanced PEB-DDI converge slowly, requiring more epochs to approach stable performance. This application maintains a stable high accuracy in the later stages of training (epochs > 60), without significant fluctuations; some existing methods (such as SRR-DDI) still have lower performance than this application even after convergence.
[0063] The above description, in conjunction with preferred technical solutions, provides a further detailed explanation of the present invention and should not be construed as limiting the specific implementation of the invention to these descriptions. For those skilled in the art, simple deductions and substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.
Claims
1. A drug interaction prediction method based on bidirectional event semantic guidance, characterized in that, include: S1: Extract multimodal drug features of the drug pair to be predicted; The multimodal drug features refer to the comprehensive features obtained by extracting molecular graph structural features, sequence-level features, and textual semantic features from the drug's molecular graph, simplified molecular input linear canonical sequence, and drug text description, respectively, and then concatenating and fusing the molecular graph structural features, sequence-level features, and textual semantic features; S2: Construct a bidirectional event semantic space; the bidirectional event semantic space refers to constructing a corresponding negative drug interaction event description text for each positive drug interaction event description text to form a positive-negative semantic pair, encoding the positive-negative semantic pair, and then optimizing it through preset constraints; wherein, the optimization through preset constraints in step S2 includes: calculating the semantic center for all positive semantic vectors and all negative semantic vectors obtained after encoding, obtaining the positive semantic center and the negative semantic center; based on the positive semantic center and the negative semantic center, performing uniform distribution constraint processing on the positive semantic vector and the negative semantic vector respectively; performing orthogonal constraint processing on the positive semantic vector and the negative semantic vector to ensure that the two are directionally independent in the semantic space; S3: Step S1: The multimodal drug features obtained in step S1 are semantically supervised and aligned with the bidirectional event semantic space obtained in step S2. This semantic supervision alignment refers to optimizing the matching degree between the multimodal drug features and each positive and negative semantic vector in the bidirectional event semantic space through similarity calculation. Step S4: Based on the results of the semantic supervision alignment completed in step S3, the semantic guidance intensity is scaled using a dynamic guidance coefficient adjustment strategy, and the drug interaction prediction results are output. This dynamic guidance coefficient adjustment strategy refers to a strategy that dynamically adjusts the semantic guidance intensity based on the frequency of occurrence of drug interaction event categories and the model training rounds. Specifically, the dynamic guidance coefficient adjustment strategy in step S4 includes: dynamically determining the frequency adjustment coefficient using a preset function based on the difference between the frequency of occurrence of each drug interaction event category and the average frequency of occurrence of all drug interaction event categories; determining a round decay coefficient with a preset trend based on the relationship between the current model training rounds and the preset total training rounds; and weighting and combining the frequency adjustment coefficient and the round decay coefficient to obtain a dynamic guidance coefficient, which is used to scale the semantic guidance intensity.
2. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 1, characterized in that, Step S1 extracts multimodal drug features of the drug pair to be predicted, including: using deep molecular graph neural network, multilayer perceptron and pre-trained biomedical language model respectively, extracting molecular graph structural features, sequence-level features and text semantic features from the molecular graph, simplified molecular input linear canonical sequence and drug text description of each drug in the drug pair to be predicted.
3. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 1, characterized in that, Step S3 optimizes the matching degree through similarity calculation, including: calculating the positive similarity between the multimodal drug features and each positive semantic vector in the bidirectional event semantic space, and the negative similarity between the multimodal drug features and each negative semantic vector in the bidirectional event semantic space; and using a preset loss function to optimize the positive and negative similarities, thereby increasing the matching degree between positive samples and positive semantic vectors, and increasing the matching degree between negative samples and negative semantic vectors.
4. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 2, characterized in that, The pre-trained biomedical language model is a pre-trained language model optimized for text in the biomedical field. The pre-trained biomedical language model can semantically encode information related to pharmacological effects, indications, and contraindications in drug text descriptions.
5. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 1, characterized in that, In step S2, the mean of all positive semantic vectors is calculated to obtain the positive semantic center; the mean of all negative semantic vectors is calculated to obtain the negative semantic center.
6. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 3, characterized in that, The preset loss function used in step S3 is a loss function based on information noise contrast estimation. The loss function based on information noise contrast estimation adjusts the weight of similarity by setting a temperature parameter, thereby optimizing the matching accuracy between multimodal drug features and semantic vectors.
7. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 1, characterized in that, The preset function used in step S4 is the hyperbolic tangent nonlinear activation function. The hyperbolic tangent nonlinear activation function amplifies the difference in the frequency of different drug interaction event categories by processing the difference between the occurrence frequency and the average occurrence frequency of drug interaction event categories, thereby obtaining the frequency adjustment coefficient.
8. The drug interaction prediction method based on bidirectional event semantic guidance according to claim 1, characterized in that, The round decay coefficient exhibits an exponential decay trend, which ensures that the model has strong semantic guidance strength in the early stage of training and maintains stable convergence in the later stage of training.